Beyond the Spectrum: Decoding Autism Heterogeneity Through Structural and Functional MRI Subtypes

Abstract

Autism Spectrum Disorder (ASD) is characterized by significant clinical and biological heterogeneity, which has been a major obstacle in understanding its neurobiology and developing effective, targeted treatments. This article synthesizes recent advances in neuroimaging that leverage both structural (sMRI) and functional MRI (fMRI) to identify biologically distinct subtypes of ASD. We explore the foundational evidence for brain-based subtypes, detail methodological approaches for their identification—including multimodal data fusion and machine learning—and address the historical challenges in drug development linked to patient heterogeneity. Furthermore, we examine how these neurosubtypes are validated through their correlation with unique genetic profiles, clinical symptom presentations, and behavioral outcomes. For researchers, scientists, and drug development professionals, this review provides a comprehensive framework for moving towards a precision medicine approach in autism, paving the way for biologically informed diagnostics and personalized interventions.

The Neurobiological Imperative: Why ASD Subtyping is Essential for Progress

Autism Spectrum Disorder (ASD) represents one of the most complex challenges in neurodevelopmental psychiatry due to its profound heterogeneity in clinical presentation, underlying biology, and developmental trajectories. This diversity has long hampered efforts to establish reliable biomarkers, develop targeted interventions, and understand the fundamental mechanisms driving the condition. Historically, the diagnosis of ASD has relied exclusively on behavioral observation and clinical assessment, with the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) categorizing ASD as a single broad spectrum despite recognizing significant variations in symptom profiles, cognitive abilities, and co-occurring conditions [1]. The clinical heterogeneity of autism manifests across multiple dimensions, including differences in social communication challenges, restricted and repetitive behaviors, sensory processing profiles, cognitive functioning, and the presence of comorbid conditions such as anxiety, depression, and attention-deficit/hyperactivity disorder (ADHD).

The emerging paradigm in autism research recognizes that this clinical diversity reflects distinct biological subtypes with different genetic architectures, neural circuitry profiles, and developmental pathways. Advances in neuroimaging and genetics have begun to deconstruct this heterogeneity by identifying biologically meaningful subgroups that transcend behavioral observations alone. The integration of structural magnetic resonance imaging (sMRI), functional MRI (fMRI), diffusion tensor imaging (DTI), and genomic data has enabled researchers to move beyond descriptive phenomenology toward a mechanistic understanding of autism's varied manifestations [2] [1] [3]. This review systematically compares how structural and functional neuroimaging approaches have contributed to identifying ASD subtypes, examining their respective methodologies, findings, limitations, and implications for personalized diagnostic and therapeutic strategies.

Neuroimaging Approaches to Deconstructing Heterogeneity

Structural MRI: Mapping Brain Anatomy in ASD

Structural MRI provides detailed information about brain anatomy, including cortical thickness, gray matter volume, and overall brain structure. This approach has identified several structural biomarkers associated with ASD, though findings have often been inconsistent due to the disorder's heterogeneity [3]. Early structural studies focused on single anatomical features such as enlarged amygdalae, increased cerebellar size with decreased vermal size, larger caudate nuclei, atypical gyrification, and changes in hippocampal volume and shape [3]. However, these individual biomarkers demonstrated limited classificatory power on their own, prompting a shift toward multivariate approaches that integrate multiple structural parameters.

A novel multivariate classification method using structural MRI data developed a summed total index (TI) that indicates whether an individual's gross morphological pattern aligns more closely with ASD or neurotypical controls [3]. This approach achieved 78.9% classification accuracy in a pilot study, performing comparably to more complex machine learning methods while offering greater transparency and simplicity [3]. Significant structural differences were particularly noted in subcortical gray matter structures and limbic areas, with no significant difference in total brain volume between ASD and control groups [3]. The TI correlated well with the autism quotient (AQ) across groups (R = 0.51), suggesting a relationship between brain structure and behavioral manifestations of autism [3].

Functional MRI: Examining Brain Network Dynamics

Functional MRI, particularly resting-state fMRI (rsfMRI), measures brain activity by detecting changes in blood flow and oxygenation, providing insights into functional connectivity between different brain regions. This approach has revealed that individuals with ASD often exhibit atypical connectivity patterns, characterized by both local hyperconnectivity and global hypoconnectivity [4]. These functional alterations appear more pronounced in networks critical for social cognition and information integration, including the default mode network (DMN), salience-executive network, and fronto-parietal network [5] [6].

Recent large-scale studies using semi-supervised clustering methods have identified two primary functional connectivity subtypes in ASD: a hyper-connectivity subtype and a hypo-connectivity subtype [4]. These subtypes exhibit distinct connectivity patterns both within and between major brain networks, with the hyper-connectivity subtype showing increased connectivity within major large networks and mixed (both hyper and hypo) connectivity between networks, while the hypo-connectivity subtype displays the opposite pattern [4]. Furthermore, these functional subtypes demonstrate different correlations between connectivity patterns and core ASD symptoms, suggesting they may represent distinct pathophysiological mechanisms with implications for personalized treatment approaches [4].

Multimodal Integration: Combining Structural and Functional Insights

The most recent advances in ASD subtyping have come from approaches that integrate multiple neuroimaging modalities to capture the complex interplay between brain structure and function. One innovative study combined structural and functional MRI data through a skeleton-based white matter functional analysis, enabling voxel-wise function-structure coupling by projecting fMRI signals onto a white matter skeleton [2]. Using white matter low-frequency oscillations (LFOs) as input features for clustering algorithms, this approach identified two distinct neurosubtypes of ASD with unique biomarkers [2].

Subtype 1 displayed significantly lower fractional anisotropy (FA) in the posterior cingulate cortex (PCC) compared to neurotypical controls, while Subtype 2 exhibited reduced FA in the anterior cingulate cortex, middle temporal gyrus, parahippocampus, and thalamus [2]. Additionally, Subtype 2 had markedly higher mean diffusivity in the middle temporal gyrus, parahippocampus, and thalamus than controls, a pattern not seen in Subtype 1 [2]. The full-scale intelligence quotient (FIQ) and performance IQ (PIQ) scores were also lower for Subtype 2 compared to Subtype 1, demonstrating how neuroimaging-based subtypes can capture meaningful clinical differences [2].

Table 1: Comparison of Structural and Functional MRI Approaches in ASD Subtyping

Feature	Structural MRI	Functional MRI	Multimodal Integration
Primary Measures	Cortical thickness, gray matter volume, brain structure	Functional connectivity, network dynamics, brain activity	Both structural and functional coupling
Identified Subtypes	Limited consensus on specific subtypes	Hyper-connectivity vs. hypo-connectivity subtypes [4]	Two neurosubtypes with distinct white matter profiles [2]
Key Biomarkers	Subcortical gray matter, limbic areas [3]	DMN, frontoparietal, sensory networks [5] [6]	White matter integrity, structure-function coupling [2]
Classification Accuracy	78.9% with multivariate approach [3]	Varies; enhanced with semi-supervised methods [4]	Improved diagnostic prediction compared to general ASD classification [2]
Clinical Correlations	Correlates with AQ (R=0.51) [3]	Distinct brain-behavior relationships across subtypes [4]	IQ differences between subtypes [2]
Main Advantages	Clear anatomical references, established analysis methods	Direct assessment of brain network function	Comprehensive view of brain organization
Limitations	Limited explanatory power for symptoms	Variable reliability, influenced by state factors	Computational complexity, data requirements

Methodological Approaches in ASD Subtyping Research

Data Acquisition and Preprocessing

The quality and consistency of neuroimaging data are fundamental to reliable subtype identification. Most recent studies utilize data from large, multi-site datasets such as the Autism Brain Imaging Data Exchange (ABIDE I and II), which collectively provide structural and functional MRI data from thousands of participants across multiple international sites [6] [4]. Standardized preprocessing pipelines are critical for minimizing site-specific variations and ensuring data comparability. Common preprocessing steps typically include motion correction, spatial normalization to standard templates (e.g., MNI152), band-pass filtering (0.01-0.1 Hz for fMRI), and regression of confounding signals [5].

For functional connectivity analyses, regions of interest (ROIs) are typically defined using established atlases. For instance, one study utilized the Dosenbach 160 ROIs, which are derived from meta-analyses covering multiple cognitive domains including error processing, default mode, memory, language, and sensorimotor functions [6]. Both static and dynamic functional connectivity features are often extracted, with static functional connectivity strength (SFCS) calculated using Pearson correlation and dynamic functional connectivity assessed through measures such as dynamic conditional correlation (DCC) for capturing instant dynamic FC strength (DFCS) and variance (DFCV) [6].

Feature Extraction and Dimension Reduction

The high dimensionality of neuroimaging data necessitates effective feature reduction strategies to avoid overfitting and enhance interpretability. Different studies have employed various approaches for this purpose. Tensor decomposition methods have been used to extract compressed feature sets from resting-state fMRI data, capturing different brain communities across ASD subtypes [5]. Other common functional features include the amplitude of low-frequency fluctuation (ALFF) and fractional ALFF (fALFF), which measure the intensity of spontaneous brain activity [5]. For structural analyses, gray matter volume (GMV) derived from MRI serves as a key feature for evaluating structural variations among subtypes [5].

More advanced dimension reduction techniques include orthonormal projective non-negative matrix factorization (OPNNMF), which has been applied to high-dimensional functional connectivity data to create lower-dimensional representations suitable for clustering [4]. This approach has demonstrated superiority over traditional feature reduction methods when combined with semi-supervised clustering algorithms, yielding more robust and reproducible subtypes [4].

Clustering Algorithms and Validation

The choice of clustering algorithm significantly impacts subtype identification. While traditional unsupervised methods like K-means clustering have been widely used, recent studies have demonstrated the advantages of semi-supervised approaches. The HeterogeneitY through Discriminative Analysis (HYDRA) method incorporates diagnostic labels (ASD vs. controls) to guide the clustering process, resulting in more clinically meaningful and neurobiologically distinct subtypes [4].

Cluster validity is typically assessed through multiple measures, including reliability indices, silhouette scores, and clinical correlation analyses. The optimal number of clusters is determined through systematic evaluation rather than a priori assumptions. For instance, one comprehensive analysis of 1046 participants identified two distinct neural ASD subtypes based on normative modeling of functional connectivity [6], while another study of 5,392 individuals revealed four clinically and genetically distinct subtypes using a person-centered approach [7]. These differences in optimal cluster number highlight how methodological choices and sample characteristics influence subtype identification.

Table 2: Key Clustering Methods in ASD Neuroimaging Research

Method	Approach	Key Features	Identified Subtypes	Advantages
K-means Clustering	Unsupervised	Partitions data into K clusters based on distance metrics	Varies by study; typically 2-4 subtypes	Simple implementation, computationally efficient
HYDRA [4]	Semi-supervised	Incorporates diagnostic labels to guide clustering	Two subtypes: hyper-connectivity and hypo-connectivity	Enhanced clinical relevance, improved separation
Normative Modeling [6]	Individual-level deviation	Quantifies deviations from typical neurodevelopmental trajectories	Two subtypes with opposite deviation patterns	Accounts for developmental effects, personalized
Growth Mixture Modeling [8]	Latent trajectory	Identifies subgroups based on longitudinal patterns	Early childhood emergent vs. late childhood emergent trajectories	Captures developmental heterogeneity, temporal dynamics
Tensor Decomposition [5]	Multivariate pattern analysis	Extracts compressed feature sets from high-dimensional data	Three subtypes (autism, Asperger's, PDD-NOS)	Captures complex interactions, multimodal integration

Significantly Identified ASD Subtypes and Their Characteristics

Clinically-Defined Subtypes from Large Cohort Studies

A landmark study analyzing data from over 5,000 children in the SPARK cohort identified four clinically and biologically distinct subtypes of autism using a "person-centered" approach that considered over 230 traits [7]. These subtypes exhibit distinct developmental trajectories, medical, behavioral, and psychiatric traits, and different patterns of genetic variation:

• Social and Behavioral Challenges Subtype (37% of participants): Individuals in this group show core autism traits, including social challenges and repetitive behaviors, but generally reach developmental milestones at a pace similar to children without autism. They frequently experience co-occurring conditions like ADHD, anxiety, depression, or obsessive-compulsive disorder alongside autism [7].

• Mixed ASD with Developmental Delay Subtype (19% of participants): This group tends to reach developmental milestones, such as walking and talking, later than children without autism, but usually does not show signs of anxiety, depression, or disruptive behaviors. "Mixed" refers to differences within this group regarding repetitive behaviors and social challenges [7].

• Moderate Challenges Subtype (34% of participants): Individuals in this category show core autism-related behaviors but less strongly than those in other groups, and typically reach developmental milestones on a similar track to those without autism. They generally do not experience co-occurring psychiatric conditions [7].

• Broadly Affected Subtype (10% of participants): This group faces more extreme and wide-ranging challenges, including developmental delays, social and communication difficulties, repetitive behaviors, and co-occurring psychiatric conditions like anxiety, depression, and mood dysregulation [7].

Neuroimaging-Defined Subtypes

Neuroimaging studies have revealed subtypes that cut across clinical classifications, suggesting distinct neural substrates underlying ASD heterogeneity. One comprehensive analysis of 1046 participants identified two neural ASD subtypes with unique functional brain network profiles despite comparable clinical presentations [6]. One subtype was characterized by positive deviations in the occipital network and cerebellar network, coupled with negative deviations in the frontoparietal network, default mode network, and cingulo-opercular network. The other subtype exhibited an inverse pattern of functional deviations across these networks [6]. These neural subtypes were also associated with distinct gaze patterns assessed by autism-sensitive eye-tracking tasks, demonstrating their behavioral relevance [6].

Another study utilizing semi-supervised clustering on functional connectivity data from approximately 1800 individuals similarly identified two robust subtypes: a hyper-connectivity subtype showing hyper-connectivity within major large networks and mixed connectivity between networks, and a hypo-connectivity subtype displaying the opposite pattern [4]. These subtypes demonstrated distinct neuro-behavioral correlations, suggesting they may require different intervention approaches despite similar clinical presentations [4].

Genetic Subtypes and Developmental Trajectories

Recent genetic studies have provided crucial insights into the biological underpinnings of ASD heterogeneity. Analysis of polygenic architectures revealed that autism's genetic architecture can be decomposed into two modestly genetically correlated (rg = 0.38) polygenic factors [8]. One factor is associated with earlier autism diagnosis and lower social and communication abilities in early childhood, with only moderate genetic correlations with ADHD and mental-health conditions. The second factor is associated with later autism diagnosis and increased socioemotional and behavioral difficulties in adolescence, with moderate to high positive genetic correlations with ADHD and mental-health conditions [8].

Longitudinal data from birth cohorts demonstrate that these genetic profiles align with different developmental trajectories. The "early childhood emergent latent trajectory" is characterized by difficulties in early childhood that remain stable or modestly attenuate in adolescence, while the "late childhood emergent latent trajectory" shows fewer difficulties in early childhood that increase in late childhood and adolescence [8]. Autistic individuals in the early childhood emergent trajectory are more likely to be diagnosed in childhood than those in the late childhood emergent trajectory [8].

Experimental Protocols and Methodological Details

Multimodal Neuroimaging Subtyping Protocol

The following experimental workflow outlines the comprehensive approach used in recent studies integrating structural and functional neuroimaging data for ASD subtyping:

Diagram 1: Multimodal neuroimaging subtyping protocol. This workflow illustrates the comprehensive pipeline for identifying ASD subtypes through integrated analysis of structural and functional neuroimaging data, from initial data collection through final validation [2] [6] [4].

Genetic Analysis Protocol

Genetic studies of ASD subtypes employ sophisticated polygenic analysis methods to identify distinct genetic architectures underlying different phenotypic presentations:

Diagram 2: Genetic analysis protocol for ASD subtypes. This workflow outlines the process for identifying genetically distinct ASD subtypes through integrated analysis of deep phenotypic data and genetic information from large cohorts [7] [8] [9].

Table 3: Research Reagent Solutions for ASD Subtyping Studies

Resource Category	Specific Tools/Measures	Function/Application	Key Features
Neuroimaging Databases	ABIDE I & II [6] [4]	Large-scale, multi-site neuroimaging data	Standardized preprocessing, phenotypic data, >2000 participants
Genetic Cohorts	SPARK [7] [9]	Genetic analysis of ASD subtypes	50,000+ families, deep phenotyping, genomic data
Clinical Assessments	ADOS, ADI-R, SRS [6]	Standardized behavioral assessment	Gold-standard diagnostic measures, quantitative traits
Eye-Tracking Paradigms	Face emotion processing, Joint attention tasks [6]	Social attention measurement	Objective behavioral metrics, autism-sensitive tasks
Processing Software	fMRIPrep, FreeSurfer [6] [3]	Automated image processing	Standardized pipelines, reproducibility, quality control
Analysis Frameworks	HYDRA [4], Normative modeling [6]	Advanced clustering algorithms	Semi-supervised approach, individual deviation quantification
Genetic Analysis Tools	Growth mixture models [8], Polygenic risk scoring	Genetic architecture decomposition	Latent trajectory identification, genetic correlation analysis

The decomposition of autism heterogeneity into biologically meaningful subtypes represents a paradigm shift in how we conceptualize, diagnose, and treat this complex condition. The integration of structural and functional neuroimaging with genetic and deep phenotypic data has revealed distinct subtypes with different developmental trajectories, neural circuitry profiles, and genetic architectures. While structural MRI provides valuable information about brain anatomy and has demonstrated respectable classification accuracy in multivariate approaches, functional MRI offers unique insights into brain network dynamics that may more directly relate to clinical symptoms. The most promising approaches, however, combine multiple modalities to capture the complex interplay between brain structure and function.

These advances have important implications for both research and clinical practice. From a research perspective, they provide a framework for deconstructing autism heterogeneity, enabling more homogeneous grouping for mechanistic studies and clinical trials. For clinical practice, they pave the way for precision medicine approaches that move beyond one-size-fits-all interventions toward personalized strategies tailored to an individual's specific neurobiological subtype. As these methods continue to refine and validate ASD subtypes, we anticipate a transformation in how autism is diagnosed and treated, ultimately leading to more effective, individualized interventions that improve quality of life for autistic individuals across the lifespan.

The understanding and classification of Autism Spectrum Disorder (ASD) have undergone a profound transformation, shifting from behaviorally defined subgroups to categories grounded in measurable neurobiological variation. Historically, diagnostic manuals categorized autism into distinct behavioral subtypes such as autistic disorder, Asperger's syndrome, and Pervasive Developmental Disorder-Not Otherwise Specified (PDD-NOS) [5]. While clinically useful, these categories often masked underlying biological heterogeneity. The advent of advanced neuroimaging techniques, particularly structural (sMRI) and functional Magnetic Resonance Imaging (fMRI), has revolutionized this paradigm. This guide objectively compares how sMRI and fMRI approaches are used to identify autism subtypes, detailing their experimental protocols, findings, and applications for researchers and drug development professionals.

Structural MRI (sMRI) Approaches to Subtyping

Structural MRI investigates the neuroanatomy of the brain, providing static measures of volume, thickness, and shape of various brain structures. Its application has been pivotal in demonstrating that autism is not a single, monolithic entity but comprises subgroups with distinct neuroanatomical profiles.

Core Experimental Protocols

A common modern protocol for sMRI subtyping involves population modeling and machine learning:

• Data Acquisition and Processing: T1-weighted MRI scans are processed using software like FreeSurfer to extract morphometric features such as cortical thickness, surface area, and grey matter volume from parcellated brain regions (e.g., using the Desikan-Killiany atlas) [10].
• Population Modeling (Normative Modeling): Generalized Additive Models of Location Scale and Shape (GAMLSS) are used to generate population-level "growth charts" for brain features based on thousands of typically developing individuals. Each autistic individual is then assigned a centile score representing their deviation from this normative trajectory, accounting for age and sex [10].
• Clustering Analysis: Semi-supervised machine learning algorithms, such as HYDRA (HeterogeneitY through DiscRiminative Analysis), cluster individuals based on their patterns of neuroanatomical deviation relative to controls. This method is designed to identify subgroups that may exhibit opposite neuroanatomical patterns (e.g., one subgroup with increased volume and another with decreased volume) [10].

Key Subtyping Findings from sMRI

Research using these protocols has successfully parsed heterogeneity, though the number and nature of subgroups can vary based on the methods and features used [10]. A pivotal finding has been the identification of subgroups with opposing neuroanatomical alterations. For instance, some autistic individuals show patterns of increased total brain volume and cortical surface area, while others show decreased values, patterns that would cancel each other out in traditional case-control analyses [10]. These subgroups, however, have not yet shown consistent correlations with distinct clinical profiles, highlighting a challenge in linking sMRI-based subtypes to behavior [10].

Table 1: Key sMRI Studies on Autism Subtyping

Study Focus	Dataset	Key Method	Identified Subtypes	Key Findings
Transdiagnostic Subgrouping [10]	4,115 participants (1,305 ASD, 987 ADHD, 1,823 controls)	HYDRA clustering on cortical centile scores	Subgroups with opposing neuroanatomical patterns (number varies by method)	Confirms neuroanatomical heterogeneity within ASD/ADHD; subgroups often lack clear clinical differentiation.
Traditional Subtype Comparison [5]	234 participants (152 Autism, 54 Asperger's, 28 PDD-NOS)	Tensor decomposition of fMRI; Gray Matter Volume (GMV) analysis	Autism, Asperger's, PDD-NOS (DSM-IV)	Found significant GMV differences in subcortical and default mode networks between the autism subtype and the other two.

Functional MRI (fMRI) Approaches to Subtyping

Functional MRI measures brain activity by detecting changes in blood flow, providing insights into the dynamics of neural circuitry. Subtyping efforts using fMRI focus on how different brain regions communicate, both at rest and during tasks.

Core Experimental Protocols

fMRI subtyping often employs a multi-level analysis of functional connectivity:

• Data Acquisition and Preprocessing: Resting-state fMRI (rsfMRI) data is preprocessed using standardized pipelines like fMRIPrep, which includes steps for motion correction, normalization to a standard brain template (e.g., MNI152), and band-pass filtering [6].
• Multilevel Functional Connectivity Features:
  • Static Functional Connectivity (SFCS): Calculated using Pearson correlation between the average blood-oxygen-level-dependent (BOLD) signals from different brain regions (e.g., the Dosenbach 160 ROIs) to create a correlation matrix representing the strength of connections at rest [6].
  • Dynamic Functional Connectivity (DFCS/DFCV): Assessed using Dynamic Conditional Correlation (DCC) to measure the strength and variability of connections between regions over time, capturing the instant dynamics of brain networks [6].
• Normative Modeling and Clustering: Similar to the sMRI approach, normative models of multilevel FC are built from a typically developing (TD) group. Individual deviations from this norm are calculated for the ASD group, which are then used in clustering analyses (e.g., k-means) to identify subtypes [6].

Key Subtyping Findings from fMRI

This approach has successfully identified clinically relevant neural subtypes. A 2025 study of 1,046 participants identified two distinct ASD subtypes with unique functional brain network profiles despite comparable clinical symptom scores [6]. One subtype showed positive deviations (hyperconnectivity) in the occipital and cerebellar networks, coupled with negative deviations (hypoconnectivity) in frontoparietal, default mode, and cingulo-opercular networks. The other subtype exhibited the inverse pattern [6]. Crucially, these neural subtypes were validated with an independent measure, showing different gaze patterns in social eye-tracking tasks, confirming a link between neural subtypes and behavioral function [6].

Table 2: Key fMRI Studies on Autism Subtyping

Study Focus	Dataset	Key Method	Identified Subtypes	Key Findings
Functional Subtypes [6]	1,046 participants (479 ASD, 567 TD)	Normative modeling of static/dynamic functional connectivity	2 neural subtypes with opposite connectivity patterns	Subtypes showed distinct eye-gaze patterns, confirming a neurobehavioral link beyond clinical symptoms.
Underconnectivity & Default Mode [11]	Literature Review	Task-based and resting-state fMRI	Not Applicable (Review Article)	Provided early evidence of underconnectivity in distributed cortical networks and abnormalities in the default-mode network in ASD.

Integrated Analysis: Bridging Genetics, Biology, and Subtypes

Beyond imaging, large-scale datasets are now integrating genotypic and deep phenotypic data to define subtypes holistically. A landmark 2025 study used generative mixture modeling on over 230 traits in 5,392 individuals from the SPARK cohort, identifying four clinically and biologically distinct subtypes [7] [12].

• Social and Behavioral Challenges (37%): Core ASD traits with co-occurring conditions (ADHD, anxiety) but no developmental delays; linked to genetic mutations in genes active after birth [7] [13].
• Mixed ASD with Developmental Delay (19%): Prominent developmental delays but fewer co-occurring psychiatric conditions; linked to rare inherited variants and genes active prenatally [7] [13].
• Moderate Challenges (34%): Milder core ASD traits and fewer co-occurring conditions [12].
• Broadly Affected (10%): Severe, wide-ranging challenges including core symptoms, delays, and psychiatric conditions; showed the highest rate of damaging de novo mutations [7] [13].

This work demonstrates that biologically defined subtypes have distinct genetic underpinnings and developmental trajectories, offering a powerful framework for future research and targeted therapeutic development.

Table 3: Key Resources for Autism Subtyping Research

Resource Name	Type	Primary Function in Research
ABIDE I & II [6]	Data Repository	Publicly shared repository of brain imaging and phenotypic data from ASD individuals and controls, enabling large-scale analyses.
SPARK Cohort [7] [12]	Research Cohort	Largest US study of autism, providing genetic and deep phenotypic data for identifying biologically distinct subgroups.
FreeSurfer [10]	Software Toolbox	Processes structural MRI data to extract morphometric features like cortical thickness and surface area.
fMRIPrep [6]	Software Toolbox	Standardizes and automates the preprocessing of fMRI data, ensuring reproducibility and reducing pipeline variability.
HYDRA [10]	Algorithm	A semi-supervised clustering algorithm that groups individuals based on their differences from a control sample.
Normative Modeling [6] [10]	Statistical Framework	Quantifies individual-level deviation from a normative population benchmark, embracing heterogeneity.
Dosenbach 160 ROIs [6]	Brain Atlas	A predefined set of brain regions used to extract signals for functional connectivity analysis.

Visualizing Research Workflows

The following diagrams illustrate the core experimental protocols for sMRI and fMRI subtyping, highlighting the logical flow from data acquisition to subtype identification.

Figure 1: fMRI Subtyping Workflow. This diagram outlines the process for identifying autism subtypes using functional MRI data, from acquisition to validation.

Figure 2: sMRI Subtyping Workflow. This diagram illustrates the structural MRI subtyping process, which uses population modeling and machine learning to find subgroups based on anatomical differences.

Autism spectrum disorder (ASD) is characterized by substantial neurobiological heterogeneity, driving research efforts to identify meaningful subtypes that can inform diagnosis and treatment. Central to this investigation is the consistent observation of early brain overgrowth followed by atypical developmental trajectories in a significant subset of autistic individuals. This article synthesizes key neuroanatomical findings within the context of structural versus functional MRI research, providing a comparative analysis of methodological approaches and empirical data shaping current understanding. The examination of brain development patterns not only offers insights into potential biological mechanisms but also facilitates the stratification of ASD into more homogeneous subgroups based on neuroanatomical profiles rather than solely behavioral manifestations.

Quantitative Synthesis of Key Neuroanatomical Findings

Table 1: Comparative Analysis of Early Brain Overgrowth Findings

Study Reference	Sample Characteristics	Key Metric	Developmental Pattern	Magnitude of Effect
Courchesne (2004) [14]	Retrospective analysis	Brain volume	Early overgrowth followed by premature arrest	Most deviant overgrowth in cerebral, cerebellar, and limbic structures by 2-4 years
Nature (2017) [15]	106 high-risk infants, 42 low-risk	Cortical surface area	Hyperexpansion 6-12 months, precedes volume overgrowth (12-24 months)	Predicted ASD diagnosis with 81% PPV, 88% sensitivity
Yale PET Study (2025) [16]	12 autistic adults, 20 neurotypical	Synaptic density	Reduced synaptic density in adulthood	17% lower synaptic density across whole brain
Feng et al. [17]	46 FXS, 90 idiopathic ASD, 54 TD	Gray matter volume	Faster GMV growth rates in ASD vs FXS and TD	Distinct spatial patterns: FXS increased subcortical GMV, ASD varied

Table 2: Neuroanatomical Profiles Across ASD Subtypes and Related Conditions

Condition/Subtype	Structural Findings	Functional Connectivity Patterns	Developmental Trajectory
ASD Functional Subtype 1 [6]	N/A	Positive deviations: occipital and cerebellar networks; Negative deviations: frontoparietal, DMN, cingulo-opercular networks	Associated with distinct gaze patterns in social cue tasks
ASD Functional Subtype 2 [6]	N/A	Inverse pattern of Subtype 1	Different functional developmental pattern despite similar clinical presentation
Fragile X Syndrome [17]	Increased GMV in caudate, Crus I cerebellum; Decreased GMV in frontal insular regions, cerebellar vermis	N/A	Divergent trajectory from idiopathic ASD, despite behavioral overlap
High-Functioning Autism [18]	N/A	Generalized visuoperceptual processing deficit	Atypical configural processing across development

Experimental Protocols and Methodological Approaches

Prospective Infant Neuroimaging

The landmark 2017 Nature study established a protocol for identifying early biomarkers in infants at high familial risk for ASD [15]. This longitudinal approach involved:

• Participant Recruitment: 106 infants at high familial risk and 42 low-risk infants
• Imaging Timepoints: 6, 12, and 24 months of age
• Key Analytical Method: Deep learning algorithm applied to cortical surface area measurements from 6-12 months to predict ASD diagnosis at 24 months
• Primary Findings: Cortical surface area hyperexpansion between 6-12 months preceded brain volume overgrowth observed between 12-24 months
• Validation: Algorithm achieved positive predictive value of 81% and sensitivity of 88% in individual predictions

This protocol demonstrated that early brain changes occur during the emergence of autistic behaviors, providing a potential window for very early intervention.

Functional Subtyping in Heterogeneous ASD

The 2025 Molecular Psychiatry study addressed ASD heterogeneity through functional subtyping [6]:

• Sample: 1,046 participants (479 ASD, 567 typical development) from ABIDE I and II datasets
• Analytical Framework: Normative modeling of multilevel functional connectivity features
• Connectivity Measures: Static functional connectivity strength (SFCS), dynamic functional connectivity strength (DFCS), and dynamic functional connectivity variance (DFCV)
• Validation Cohort: Independent cohort of 21 ASD individuals with resting-state fMRI and eye-tracking data
• Clustering Approach: Identification of subtypes based on deviations from normative functional connectivity trajectories

This approach revealed two distinct neural subtypes with unique functional network profiles despite comparable clinical presentations, underscoring the value of data-driven subtyping approaches.

Synaptic Density Measurement in Living Brains

The Yale study pioneered direct measurement of synaptic density in living autistic individuals [16]:

• Participants: 12 autistic adults and 20 neurotypical controls
• Imaging Technique: PET scanning with novel radiotracer 11C-UCB-J
• Complementary Imaging: MRI for anatomical reference
• Clinical Correlation: ADOS assessment and self-report measures of autistic features
• Key Finding: 17% lower synaptic density across whole brain correlated with social-communication differences

This protocol provided the first direct evidence of reduced synaptic density in living autistic people, with potential implications for understanding connectivity abnormalities.

Diagram 1: Developmental Trajectory of Brain Changes in Autism

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Critical Research Resources for Autism Neurodevelopment Studies

Resource Category	Specific Examples	Research Application
Neuroimaging Databases	ABIDE I & II [19] [6]	Large-scale, multi-site datasets for normative modeling and subtyping
Radiotracers	11C-UCB-J [16]	In vivo synaptic density measurement via PET imaging
Eye-Tracking Paradigms	Face emotion processing, Joint attention tasks [6]	Quantifying social attention patterns linked to neural subtypes
Analytical Pipelines	Normative modeling [6], Tensor decomposition [19], Deep learning algorithms [15]	Identifying deviations from typical development and predictive patterns
Behavioral Assessment Tools	ADOS, ADI-R, SRS [6] [18]	Standardized clinical phenotyping and correlation with neural measures

Structural versus Functional Subtyping: Convergent and Divergent Patterns

The distinction between structural and functional approaches to ASD subtyping reveals complementary insights into neurobiological mechanisms. Structural approaches have identified early overgrowth trajectories [14] [15] and volumetric differences across subtypes [17], while functional approaches have revealed distinct connectivity profiles that cross-cut traditional diagnostic boundaries [6].

Diagram 2: Structural versus Functional Subtyping Approaches in Autism Research

Implications for Targeted Interventions

The identification of neuroanatomical and functional subtypes holds significant promise for developing targeted interventions. The finding that cortical overgrowth begins between 6-12 months [15] suggests a critical window for early intervention. Similarly, the discovery of reduced synaptic density in autistic adults [16] points to potential targets for pharmacological interventions. Furthermore, the correlation between functional connectivity subtypes and differential response to oxytocin treatment [6] underscores the clinical utility of neurobiological stratification.

The integration of structural and functional neuroimaging has substantially advanced our understanding of atypical development in autism. The consistent observation of early brain overgrowth in a subset of autistic children, followed by divergent developmental trajectories, provides a neurobiological framework for parsing heterogeneity. Contemporary research approaches that leverage large-scale datasets, normative modeling, and multimodal imaging are increasingly capable of identifying biologically meaningful subtypes that transcend behavioral phenomenology. These advances promise more personalized approaches to intervention aligned with an individual's specific neurodevelopmental profile, ultimately improving outcomes across the autism spectrum.

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition characterized by challenges in social communication, interaction, and the presence of restricted, repetitive behaviors. The quest to identify its neurobiological underpinnings has increasingly focused on the concept of large-scale brain network dysconnectivity. Among the most studied networks are the default mode network (DMN), crucial for self-referential thought and social cognition, and the salience network (SN), responsible for detecting and orienting attention toward relevant stimuli. This guide objectively compares the roles of these two networks in ASD, synthesizing current experimental data to provide a clear resource for researchers and drug development professionals working within the context of structural versus functional MRI research on autism subtypes.

Converging evidence from multimodal neuroimaging indicates that altered functional and structural organization of both the DMN and SN are prominent neurobiological features of ASD [20]. The DMN, comprising key nodes such as the posterior cingulate cortex (PCC), medial prefrontal cortex (mPFC), and temporoparietal junction (TPJ), is fundamentally involved in processes like theory of mind and self-referential thinking—functions known to be impaired in ASD [20]. Conversely, the SN, anchored in the anterior insula (AI) and anterior cingulate cortex (ACC), acts as a switch between the DMN and task-positive networks like the central executive network, guiding attention to the most salient internal and external events [21] [22]. Dysfunction in this switching mechanism is theorized to underlie the atypical sensory processing and social attention observed in ASD.

Default Mode Network (DMN) in ASD

Functional Neuroanatomy and Social Cognition

The DMN is a strongly interconnected system that is most active during rest and engages during social cognitive processes. Its core nodes support distinct aspects of social cognition:

• Posterior Cingulate Cortex (PCC): Acts as a central functional hub with a high baseline metabolic rate. It is implicated in autobiographical memory, imagining one's future, and evaluating the mental states of others [20].
• Medial Prefrontal Cortex (mPFC): This region is involved in monitoring both one's own mental states and the mental states of others. The ventral mPFC is more associated with self-referential processing, while the dorsal mPFC is engaged during mentalizing about others [20].
• Temporoparietal Junction (TPJ): Preferentially encodes 'other-relevant' information, including the beliefs and intentions of others. It is critical for distinguishing self-relevant from other-relevant information and predicting others' behavior during social interaction [20].

Evidence of DMN Dysconnectivity in ASD

Research across multiple modalities consistently reveals DMN alterations in ASD, which are linked to core social cognitive deficits.

Table 1: Task-Based fMRI Findings of DMN Dysfunction in ASD

Social Cognitive Domain	Task Examples	Key DMN Findings in ASD	Associated Brain Regions
Self-Referential Processing	Judgments about self vs. other	Reduced activation; Atypical self-representation; Reduced functional connectivity	Ventral mPFC, PCC [20]
Theory of Mind/Mentalizing	Viewing images or stories to infer mental states	Decreased recruitment; Hypoactivation; Mixed findings in developmental studies	Dorsal mPFC, TPJ, PCC [20]

Table 2: Intrinsic Functional Connectivity and Structural Findings of the DMN in ASD

Modality	Experimental Measure	Key Findings in ASD	Interpretation
Resting-State fMRI	Functional Integration & Segregation	Reduced long-range within-network connectivity; Increased between-network connectivity [23]	Functional underconnectivity within the DMN and reduced network segregation
Graph Theory Analysis	Modularity, Local Efficiency	Reduced modularity and local efficiency (clustering) [23]	Less optimized, less segregated functional network organization
Diffusion Tensor Imaging	White Matter Integrity	Lower white matter integrity despite higher numbers of streamlines [23]	Structural underconnectivity supporting functional findings

A recent meta-analysis of adults with ASD found consistent hypo-activation in the left amygdala, a region with strong functional connections to DMN nodes. This cluster was found to co-activate with the cerebellum and fusiform gyrus, regions implicated in social cognition, suggesting a broader disrupted network [24]. Furthermore, studies investigating ASD subtypes have found that impairments within the DMN are a major factor that differentiates the autism subtype from Asperger's and PDD-NOS [5].

Salience Network (SN) in ASD

Functional Neuroanatomy and Its Role in Attention

The SN is an early-emerging network critical for identifying the most subjectively relevant stimuli from among multiple internal and external inputs. Its core function is to guide behavior by dynamically switching attention between the internally-focused DMN and the externally-focused, goal-oriented central executive network (CEN) [21] [25]. The key nodes of the SN include:

• Anterior Insula (AI): Considered the hub of the SN, it is involved in interoceptive awareness and perception of emotionally salient information.
• Anterior Cingulate Cortex (ACC): Works with the AI to integrate sensory, emotional, and cognitive information to guide attentional resources.

In typical development, robust SN connectivity with prefrontal regions supports attention to socially relevant information, such as faces [22].

Evidence of SN Dysconnectivity in ASD

Altered SN connectivity, particularly hyperconnectivity with primary sensory regions, is a robust finding in ASD and is strongly linked to sensory symptoms.

Table 3: Key Findings on Salience Network Dysconnectivity in ASD

Study Demographic	Experimental Method	Key SN Findings in ASD	Correlation with Behavior
Children & Adolescents	Resting-state fMRI	SN hyperconnectivity with sensorimotor and attention regions [21]	Associated with sensory over-responsivity (SOR) symptoms
6-Week-Old Infants (High-Likelihood for ASD)	Resting-state fMRI	Stronger SN connectivity with sensorimotor regions; Weaker SN connectivity with prefrontal regions [22]	Predicted subsequent sensory hypersensitivity and attenuated social attention
Adolescents	Resting-state fMRI (ICA)	Atypical within-network and between-network connectivity as part of a triple-network model [25]	Associated with core social impairments

The hyperconnectivity between the SN and sensorimotor/limbic regions (e.g., amygdala) creates a neural substrate where basic sensory information is assigned excessive salience, leading to sensory over-responsivity (SOR) [21]. This is further supported by findings that this hyperconnectivity is directly correlated with the extent of brain activation in response to mildly aversive auditory and tactile stimuli [21]. A trade-off has been observed in infancy, whereby stronger SN-sensorimotor connectivity is inversely correlated with weaker SN-prefrontal connectivity, providing a mechanistic account for the co-emergence of sensory and social symptoms in ASD [22].

Direct Comparison of DMN and SN Alterations

The DMN and SN exhibit distinct, yet potentially interrelated, patterns of dysconnectivity in ASD. The table below provides a structured, direct comparison based on the synthesized research findings.

Table 4: Comparative Overview of DMN and SN Dysconnectivity in ASD

Aspect	Default Mode Network (DMN)	Salience Network (SN)
Primary Functional Role	Self-referential thought, mentalizing, autobiographical memory [20]	Detecting salient stimuli, switching between internal (DMN) and external (CEN) attention [21] [25]
Key Altered Nodes	Medial Prefrontal Cortex (mPFC), Posterior Cingulate Cortex (PCC), Temporoparietal Junction (TPJ) [20]	Anterior Insula (AI), Anterior Cingulate Cortex (ACC) [21]
Predominant Connectivity Pattern in ASD	Functional Underconnectivity within the network and with other social brain regions [20] [23]	Hyperconnectivity with sensorimotor and limbic regions (e.g., amygdala) [21] [22]
Associated Clinical Symptoms	Deficits in theory of mind, self-other processing, social communication [20] [26]	Sensory over-responsivity, atypical social attention, anxiety, restrictive/repetitive behaviors [21] [22]
Developmental Trajectory	Atypical developmental trajectory; mixed patterns of hypo-/hyper-connectivity in children vs. adolescents [20]	Alterations present as early as 6 weeks of age in high-likelihood infants; predicts later symptoms [22]
Relationship with Other Networks	Reduced segregation from other functional systems [23]	Failed deactivation of the DMN during tasks due to impaired SN switching [25]

The Triple-Network Model and Inter-Network Dynamics

The DMN and SN do not operate in isolation. They are core components of the triple-network model, which also includes the central executive network (CEN) [25]. In this model, the SN is hypothesized to regulate the dynamic interplay between the DMN and CEN. Dysfunction of the SN in ASD can therefore lead to a failure to suppress the DMN during externally demanding tasks, and/or a failure to engage the CEN appropriately. This provides a parsimonious framework for understanding how distinct alterations in both the DMN (social cognition) and SN (sensory salience) might arise from a common source of dysregulation in network switching [25].

The following diagram illustrates the typical and hypothesized ASD-specific interactions within this triple-network model:

Experimental Protocols and Methodologies

This section details the standard experimental protocols used to generate the key findings cited in this guide, providing a reference for researchers seeking to replicate or build upon this work.

Resting-State Functional MRI (rs-fMRI) Protocol

Resting-state fMRI is the primary method for investigating intrinsic functional connectivity within and between large-scale networks like the DMN and SN.

Table 5: Standard rs-fMRI Acquisition and Analysis Protocol

Protocol Phase	Key Parameters & Procedures	Considerations for ASD Research
Data Acquisition	• Scanner: 3T MRI scanner• Sequence: Gradient-echo EPI• Parameters: TR/TE = 2000/30 ms, voxel size = 3-4 mm³• Duration: 5-10 minutes of rest (eyes open/closed) [23] [21]	• Participant comfort and acclimatization are critical.• Use of mock scanners.• For infants, data is acquired during natural sleep [22].
Preprocessing	• Software: DPARSF, FSL, SPM• Steps: Discard initial volumes, slice-time correction, realignment, normalization to MNI space, spatial smoothing (FWHM 5-8 mm) [27] [25]	• Rigorous motion correction is essential (e.g., scrubbing, regression).• Global signal regression is debated but often used.
Functional Connectivity Analysis	• Seed-Based Correlation: Placing a seed region (e.g., rAI for SN, PCC for DMN) and correlating its time-course with all other brain voxels [21] [22].• Independent Component Analysis (ICA): Data-driven approach to identify networks without a priori seeds [25].• Graph Theory: Models the brain as a network of nodes and edges to compute metrics like efficiency and modularity [23].	• Seed selection must be justified.• ICA requires estimation of component numbers.• Graph theory necessitates defining network nodes (e.g., using a brain atlas).

Task-Based fMRI Protocol for Social Cognition

Task-based fMRI is used to probe network function during specific cognitive processes.

Table 6: Key Elements of Social Cognitive Task-fMRI Protocols

Component	Description	Examples for DMN/SN Engagement
Task Paradigm	Block or event-related design presenting stimuli.	• Theory of Mind: Animations or stories requiring mental state inference [20].• Self-Referential Processing: Judging whether adjectives describe oneself or a familiar other [20].• Sensory Processing: Exposure to mildly aversive tactile or auditory stimuli to probe SN [21].
Analysis	General Linear Model (GLM) to identify brain regions with BOLD signal changes correlated with task conditions.	Contrasts are created (e.g., "Self" > "Other" or "Social" > "Non-social") to identify regions differentially active in ASD vs. control groups [20].

The Scientist's Toolkit: Essential Research Reagents & Materials

For researchers investigating DMN and SN connectivity in ASD, the following table details key resources and analytical tools.

Table 7: Essential Reagents and Resources for Network Dysconnectivity Research

Item / Resource	Function / Purpose	Specific Examples / Notes
ABIDE Database	Publicly shared repository of preprocessed neuroimaging data from ASD individuals and controls.	Primary data source for large-scale analyses; includes anatomical and resting-state fMRI data [5] [28].
Preprocessing Pipelines	Software for standardizing the initial steps of MRI data analysis.	• DPARSF: Integrated pipeline for rs-fMRI data preprocessing [25].• CCS (Connectome Computation System): Provides a standardized preprocessing protocol for ABIDE data [5].
Analysis Toolboxes	Specialized software for functional connectivity and network analysis.	• GIFT: Toolbox for performing Independent Component Analysis (ICA) [25].• FNC & BrainGraph: Tools for functional network connectivity and graph theory analysis [23] [25].
Statistical & Meta-Analytic Tools	For synthesizing findings across studies and performing robust statistical inference.	• ES-SDM: Software for voxel-based meta-analysis of neuroimaging studies [27].• GingerALE: Tool for Activation Likelihood Estimation (ALE) meta-analysis [24].
Contrast Subgraph Algorithms	Advanced network comparison technique to identify maximally different connectivity patterns between groups.	Used to identify mesoscopic-scale structures that are hyper- or hypo-connected in ASD vs. controls, reconciling conflicting connectivity reports [28].

The quest to identify robust neurobiological subtypes of Autism Spectrum Disorder (ASD) has been significantly hampered by the condition's profound heterogeneity. For decades, neuroimaging research has vacillated between structural (sMRI) and functional (fMRI) magnetic resonance imaging approaches, often with conflicting results. This guide objectively compares the performance of these unimodal methodologies against emerging multimodal fusion techniques, which integrate sMRI and fMRI data to capture complementary aspects of brain organization. By synthesizing experimental data from recent studies, we demonstrate that multimodal integration consistently outperforms single-modality analyses in classifying ASD subtypes, predicting symptom severity, and identifying biologically coherent subgroups. The comparative data presented herein provide researchers and drug development professionals with a evidence-based framework for selecting neuroimaging approaches that maximize sensitivity to the complex, system-level brain alterations characteristic of ASD.

Autism Spectrum Disorder (ASD) encompasses a range of neurodevelopmental conditions characterized by deficits in social communication and restricted, repetitive behaviors [29]. Its remarkable heterogeneity across biological etiologies, neural systems, and clinical presentations presents a substantial challenge for diagnosis and precision treatment [29] [30]. Historically, neuroimaging studies have attempted to parse this heterogeneity using either structural or functional modalities in isolation, yielding inconsistent and often non-reproducible findings [31].

The theoretical rationale for multimodal integration stems from the understanding that pathophysiological processes in ASD manifest across multiple aspects of neurobiology [32]. Structural modalities (e.g., sMRI, DTI) reveal static anatomical properties like gray matter volume (GMV) and white matter integrity, while functional modalities (e.g., resting-state fMRI) capture dynamic patterns of neural activity and connectivity [33] [31]. Neither perspective alone sufficiently characterizes the complex brain-behavior relationships in ASD, necessitating combined approaches that leverage their complementary strengths [32].

Comparative Performance: Unimodal vs. Multimodal Approaches

Diagnostic Classification Accuracy

Table 1: Classification Performance of Neuroimaging Modalities in ASD

Modality Approach	Features Used	Sample Size (ASD/Controls)	Classification Accuracy	Study/Model
Multimodal (sMRI + rs-fMRI)	ALFF + sMRI maps	702 (351/351)	76.9% ± 2.34	3D-DenseNet (Two-channel) [34]
Functional (rs-fMRI)	ALFF maps only	702 (351/351)	72.0% ± 3.1	3D-DenseNet (One-channel) [34]
Multimodal (rs-fMRI + sMRI)	Functional connectivity + Gray matter	Not specified	65.6%	Traditional ML Fusion [34]
Structural (sMRI)	Gray matter only	Not specified	63.9%	Traditional ML [34]
Functional (rs-fMRI)	Functional connectivity only	Not specified	60.6%	Traditional ML [34]
Structural (sMRI)	White matter only	Not specified	59.7%	Traditional ML [34]

Deep learning models leveraging combined ALFF-sMRI inputs achieve superior classification accuracy compared to unimodal approaches, demonstrating the value of integrating functional and structural information [34]. The two-channel model's performance highlights how structural and functional data provide non-redundant information that collectively enhances ASD identification.

Subtype Identification and Behavioral Correlation

Table 2: Subtype Identification Across Methodological Approaches

Imaging Approach	Subtypes Identified	Key Neural Correlates	Behavioral Correlations	Study
Multimodal Fusion (SNF)	2 male ASD subtypes	Opposite GMV changes; distinct ALFF patterns	ALFF alterations predicted social communication severity in Subtype 1	Gao et al. 2025 [35]
DSM-IV Subtype Comparison	Asperger's, PDD-NOS, Autistic	Common: DLPFC, temporal cortex; Unique: subcortical fALFF patterns	Each pattern correlated with different ADOS subdomains	[29]
rTMS Intervention	Pre-post changes	Increased GMV: Cerebellar Vermis, Caudate; Enhanced FC: Frontal-Temporal	Neuroimaging changes correlated with behavioral improvements	[36]

Multimodal approaches successfully parse heterogeneity by revealing distinct neurobiological subtypes with differential clinical profiles. Unlike unimodal classifications, multimodal fusion can identify subgroups with opposite structural patterns (e.g., increased vs. decreased GMV) coupled with unique functional alterations that directly predict specific symptom domains [35] [29].

Experimental Protocols in Multimodal Research

Multimodal Fusion Using Similarity Network Fusion (SNF)

Protocol Overview: This unsupervised learning method identifies data-driven subtypes by fusing structural and functional distance networks [35].

• Sample Characteristics: 207 male children (105 ASD, 102 healthy controls) from ABIDE database
• Feature Extraction:
  • Structural: Gray matter volume (GMV) from sMRI
  • Functional: Amplitude of low-frequency fluctuation (ALFF) from resting-state fMRI
• Fusion Process: Structural and functional distance networks are constructed separately then integrated using SNF algorithm
• Clustering: Spectral clustering applied to the fused network to identify subtypes
• Validation: Multivariate support vector regression analyzes relationship between multimodal alterations and symptom severity

Linked Independent Component Analysis (LICA)

Protocol Overview: This data-driven technique decomposes multimodal data to identify coherent patterns of variation across modalities [32].

• Sample Characteristics: 206 autistic and 196 non-autistic participants from EU-AIMS LEAP project
• Feature Extraction:
  • Structural: Grey matter density maps from T1-weighted MRI; probabilistic tractography from DWI
  • Functional: Connectopic maps from resting-state fMRI
• Integration: LICA decomposes multimodal data into independent components representing shared variance across modalities
• Analysis: Linear mixed-effects models evaluate component relationships with diagnosis and behavioral measures

Deep Learning with Twinned Neuroimaging Inputs

Protocol Overview: Two-channel 3D-DenseNet architecture processes structural and functional inputs simultaneously [34].

• Sample Characteristics: 702 participants (351 ASD, 351 controls) from ABIDE I
• Input Preparation:
  • Channel 1: sMRI maps (T1-weighted, downsampled to 3mm isotropic)
  • Channel 2: ALFF/fALFF maps from resting-state fMRI
• Preprocessing: Minimal preprocessing with intentional data diversity retention
• Data Augmentation: Random rotation (±30° z-axis) and zoom (0.7-1.3×) during training
• Architecture: Two-channel 3D-DenseNet with adaptive average pooling to handle varying matrix sizes

Multimodal Integration Workflow

The Scientist's Toolkit: Essential Research Reagents

Table 3: Key Reagents and Analytical Tools for Multimodal ASD Research

Tool/Reagent	Type	Function in Research	Example Application
ABIDE Database	Data Resource	Preprocessed neuroimaging data from multiple sites	Provides standardized datasets for method development [35] [29]
11C-UCB-J Radiotracer	PET Tracer	Quantifies synaptic density in living humans	First direct measurement of reduced synapses in ASD [16]
Yiruide YRDCCY-1 rTMS	Intervention Device	Non-invasive neuromodulation	Investigates causal structure-function relationships [36]
Similarity Network Fusion (SNF)	Algorithm	Unsupervised multimodal fusion	Identifies data-driven ASD subtypes [35]
Linked ICA (LICA)	Algorithm	Multimodal data decomposition	Identifies cross-modal biomarkers [32]
3D-DenseNet	Deep Learning Architecture	Classification from neuroimaging inputs	Twinned networks for sMRI-fMRI integration [34]

The cumulative evidence from comparative studies strongly supports multimodal integration as a superior approach for delineating the neurobiological architecture of ASD. By simultaneously capturing structural and functional aspects of brain organization, multimodal protocols achieve enhanced classification accuracy, reveal clinically meaningful subtypes, and identify robust brain-behavior relationships that elude unimodal methods.

For drug development professionals, these advances offer promising pathways toward target validation and patient stratification. The identification of biologically coherent subgroups through multimodal imaging may enable more targeted clinical trials and personalized intervention approaches [30] [37]. Furthermore, the demonstration that intervention-induced structural changes correlate with functional and behavioral improvements provides a crucial framework for evaluating treatment efficacy [36].

As the field progresses, future research should prioritize the standardization of multimodal protocols, development of normative references, and integration of genetic and molecular data with neuroimaging phenotypes. Such efforts will ultimately realize the promise of precision medicine for individuals with ASD, moving beyond behavioral symptomatology to target underlying neurobiological mechanisms.

Multimodal Neuroimaging in Action: Techniques for Identifying ASD Subtypes

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by deficits in social communication and repetitive behaviors, affecting approximately 1% of the population worldwide [38]. The profound heterogeneity in its presentation has posed significant challenges for identifying consistent neural biomarkers, with studies often reporting contradictory findings regarding brain structure and function [39]. This variability stems from multiple factors, including biological heterogeneity, differences in imaging protocols, and small sample sizes typical of single-site studies [38]. To address these challenges, the field has increasingly turned to large-scale, multi-site datasets that provide the statistical power needed to detect robust signals amidst noise and variability. Two pivotal resources have emerged in this landscape: the Autism Brain Imaging Data Exchange (ABIDE I and II) and the SPARK (SParsity-based Analysis of Reliable k-hubness) analytical framework. ABIDE provides the foundational data infrastructure by aggregating neuroimaging data across international sites, while SPARK offers a novel methodological approach for extracting meaningful patterns from this complex data. Together, these resources represent complementary forces advancing the search for reliable neurosubtypes in ASD, potentially differentiating between structural and functional manifestations of the condition.

Table 1: Key Characteristics of Major Autism Neuroimaging Resources

Resource Name	Type	Primary Data Modalities	Sample Size	Key Features
ABIDE I	Data Repository	rs-fMRI, sMRI	1,112 participants (17 sites)	First large-scale autism neuroimaging data sharing initiative
ABIDE II	Data Repository	rs-fMRI, sMRI	Over 1,000 participants (19 sites)	Extension with more detailed phenotypic data
SPARK	Analytical Method	rs-fMRI	Variable application across datasets	Identifies connector hubs and overlapping networks

Dataset Deep Dive: ABIDE I and II

Architecture and Composition

The Autism Brain Imaging Data Exchange (ABIDE) represents a landmark achievement in open neuroscience, created through the aggregation of datasets independently collected across more than 24 international brain imaging laboratories [40]. The initiative was founded to accelerate the understanding of neural bases of autism by providing large-scale samples essential for revealing brain mechanisms underlying ASD heterogeneity [40]. ABIDE I, the initial collection, included functional and structural brain imaging data from 1,112 participants across 17 international sites, comprising 539 individuals with ASD and 573 age-matched typical controls [41]. ABIDE II expanded this resource further with additional datasets and more detailed phenotypic information. This unprecedented data-sharing initiative has consistently adhered to open science principles, making data freely available to researchers worldwide and fundamentally transforming the scale at which autism neuroimaging research can be conducted.

Impact on Research Practices

The availability of ABIDE data has democratized autism neuroimaging research, enabling investigations into brain connectivity patterns across larger and more demographically heterogeneous samples that better reflect clinical reality [41]. Prior to ABIDE, most classification studies achieved high accuracy (often above 90%) but were constrained to dozens of participants from single sites [41]. The scale of ABIDE has revealed a crucial inverse relationship between sample size and classification accuracy—studies using the full dataset typically achieve more modest accuracy (60-70%) but likely produce more generalizable and robust findings [38] [39]. This resource has facilitated the application of diverse machine learning approaches, from support vector machines to deep learning networks, while highlighting the critical importance of standardized evaluation frameworks and the challenges of multi-site data harmonization.

Analytical Innovation: The SPARK Framework

Methodological Foundations

SPARK (SParsity-based Analysis of Reliable k-hubness) represents a significant methodological advancement specifically designed to identify connector hubs and overlapping network structures from resting-state fMRI data [42]. Traditional approaches to hub identification in functional connectivity often rely on thresholded correlation matrices, which are particularly vulnerable to the multicollinearity between temporal dynamics within functional networks [42]. SPARK addresses this fundamental limitation through a unified framework that combines a data-driven sparse General Linear Model (GLM) with bootstrap resampling strategies. The method's novelty lies in its ability to not only count the number of networks involved in each voxel but also identify which specific networks are actually involved, providing a more nuanced view of functional brain organization in autism.

Technical Implementation and Advantages

The SPARK framework introduces the concept of "k-hubness," which denotes the number of networks overlapping in each voxel [42]. This approach bypasses multicollinearity issues by avoiding dependence on simple connection counting from thresholded matrices. The integration of bootstrap resampling provides statistical assessment of reproducibility at the single-subject level, addressing the critical need for reliability in neuroimaging biomarkers [42]. Validation studies using both dimensional box simulations and realistic simulations with artificial hubs generated on real data have demonstrated SPARK's accuracy and robustness [42]. Furthermore, test-retest reliability assessments using the 1000 Functional Connectome Project database, which includes data from 25 healthy subjects at three different occasions, have shown that SPARK provides accurate and reliable estimation of k-hubness, suggesting its utility for understanding hub organization in resting-state fMRI studies of autism [42].

Comparative Performance: Experimental Data and Protocols

Machine Learning Classification Performance

The availability of ABIDE data has enabled systematic comparisons of different machine learning approaches applied to autism classification. Under standardized evaluation frameworks, multiple models demonstrate remarkably similar performance, suggesting that variations in reported accuracy across literature may stem more from differences in inclusion criteria, data modalities, and evaluation pipelines rather than fundamental algorithmic advantages [38] [39].

Table 2: Performance Comparison of Machine Learning Models on ABIDE Data

Model	Accuracy	AUC	Modality	Key Features
Support Vector Machine (SVM)	70.1%	0.77	fMRI	Classic approach, strong performance
Graph Convolutional Network (GCN)	72.2%	0.78	fMRI + sMRI	Ensemble method, highest accuracy
Autoencoder + Fully Connected Network	70.0%	-	fMRI	Feature learning + classification
Edge-Variational GCN (EV-GCN)	81.0%*	-	fMRI	*Reported in literature with different framework
Deep Learning Model (University of Plymouth)	98.0%	-	fMRI	Explainable AI, probability scores

Recent research leveraging ABIDE data has pushed accuracy boundaries while emphasizing explainability. A deep learning model developed by the University of Plymouth achieved 98% cross-validated accuracy for ASD classification using resting-state fMRI data, coupled with explainable AI techniques that produce maps of brain regions most influential to its decisions [43]. This approach not only provides classification but also model-estimated probability scores and interpretable visualizations, potentially offering clinicians both accurate results and clear, explainable insights to inform assessment decisions [43].

Experimental Protocols and Workflows

Standardized experimental protocols have emerged as critical factors for robust classification performance. The most reliable frameworks typically involve nested cross-validation, where models are fitted to training sets, parameters tuned on validation sets, and performance evaluated on completely held-out test sets [39]. This process is performed multiple times with rotating test sets to ensure robust performance independent of specific data partitions. For functional connectivity analysis, common preprocessing pipelines include motion correction, spatial normalization, and band-pass filtering, followed by construction of connectivity matrices using Pearson's correlation between region-wise time series [38]. Structural analyses typically involve cortical thickness measurements, subcortical volumetry, and voxel-based morphometry [39].

Diagram 1: Experimental Workflow for Autism Subtype Identification

Table 3: Essential Research Reagents and Resources

Resource	Type	Function	Application in ASD Research
ABIDE I/II	Data Repository	Provides standardized neuroimaging data	Large-scale classification studies; method validation
SPARK Code	Analytical Tool	Identifies connector hubs and overlapping networks	Functional hub analysis; network connectivity studies
SmoothGrad	Interpretation Method	Improves stability of feature identification	Model interpretability; biomarker discovery
MSDL Atlas	Parcellation Atlas	Defines regions of interest for connectivity	Functional connectivity matrix construction
FreeSurfer	Software Tool	Extracts structural features (cortical thickness, volume)	Structural MRI analysis; volumetric studies
Bootstrap Resampling	Statistical Method	Assesses reproducibility of findings	Reliability assessment; validation of results

Integration with Autism Subtyping Research

Structural vs. Functional Insights

The combined application of ABIDE data and advanced analytical methods like SPARK has accelerated progress in differentiating structural and functional autism subtypes. Structural MRI studies have consistently reported differences in total brain volume, brain asymmetry, cortical thickness, and subcortical volume in participants with autism [39]. Functional investigations, particularly using resting-state fMRI, have revealed widespread reductions in connectivity spanning unimodal, heteromodal, primary somatosensory, and limbic regions, with increased connectivity in some subcortical nodes [39]. More recently, a reproducible pattern of hyperconnectivity in prefrontal and parietal cortices with hypoconnectivity in sensory-motor regions has emerged across cohorts [39]. The anterior-posterior disruption in brain connectivity, particularly the anticorrelation between anterior and posterior areas, has been consistently identified as a key functional signature of ASD [41].

Converging Evidence for Subtypes

Machine learning approaches applied to ABIDE data have begun to identify neurobiological subtypes that may correspond to clinically meaningful subgroups. Structural features from ventricles and temporal cortex regions have demonstrated consistent predictive value for autism identification across multiple models [38]. Functional analyses have highlighted the importance of regions including the Paracingulate Gyrus, Supramarginal Gyrus, and Middle Temporal Gyrus in differentiating ASD from controls [41]. The combination of structural and functional modalities has consistently exhibited higher predictive ability compared to single-modality approaches, with ensemble methods further improving performance [39]. These converging lines of evidence suggest distinct neural subtypes that may eventually inform targeted interventions and personalized treatment approaches.

Diagram 2: Structural and Functional Autism Subtypes Hypothesis

The synergy between large-scale datasets like ABIDE and advanced analytical frameworks like SPARK has fundamentally transformed autism neuroimaging research. ABIDE has provided the essential infrastructure for robust, generalizable findings through its multi-site, large-sample design, while SPARK has addressed specific methodological challenges in identifying reliable functional connectivity patterns. Together, they have enabled significant progress in delineating structural and functional subtypes of autism, moving the field closer to biologically-based diagnostic and stratification approaches. Future research directions include further validation of proposed subtypes in independent cohorts, integration of genetic data to explore biological mechanisms, and translation of these findings into clinical applications that can reduce diagnostic delays and personalize interventions. As these resources continue to evolve and be supplemented with new data types, they hold the promise of unraveling the complexity of autism and improving outcomes for autistic individuals worldwide.

Structural MRI (sMRI) provides critical insights into the neuroanatomical underpinnings of Autism Spectrum Disorder (ASD), primarily through measures of gray matter volume (GMV) and cortical morphometry. In the broader research landscape comparing structural and functional MRI for identifying autism subtypes, sMRI offers distinct advantages in delineating stable, quantifiable brain structural features, while functional MRI (fMRI) reveals dynamic brain activity patterns. This guide objectively compares the performance of sMRI biomarkers against other modalities, detailing experimental protocols and key findings that inform their application in research and drug development.

The table below summarizes the core characteristics of sMRI biomarkers in comparison to other common neuroimaging approaches in autism research.

Table 1: Comparison of Primary Neuroimaging Modalities in Autism Research

Modality	Primary Biomarker Type	Key Strengths	Inherent Limitations	Exemplary Classification Accuracy
sMRI (Structural MRI)	Gray Matter Volume (GMV), Cortical Thickness, Surface Area	High spatial resolution; Quantifies stable neuroanatomical traits; Excellent reproducibility [1]	Does not directly measure brain function; Findings can be heterogeneous [44]	75-100% (sMRI alone) [45] [1]
fMRI (Functional MRI)	Functional Connectivity (FC), ALFF/fALFF	Measures dynamic brain network activity; Can identify state-dependent biomarkers [19]	Lower spatial resolution; Sensitive to motion and physiological noise [1]	79-100% (fMRI alone) [45] [1]
Multimodal MRI	Combined structural, functional, and sometimes diffusion features	Holistic view of brain structure and function; Mitigates limitations of single modalities [45]	High computational complexity; Data integration challenges [46]	Up to 85% (Multimodal fusion) [46]

Detailed Performance Metrics of sMRI Biomarkers

Research consistently shows that individuals with ASD exhibit atypical neurodevelopmental trajectories, which are captured by sMRI. A key meta-analysis of 25 years of voxel-based morphometry research highlights that GMV and gray matter concentration (GMC) represent distinct yet synergistic indices, with non-overlapping patterns of alteration in ASD [44]. The following table synthesizes quantitative findings on specific sMRI biomarkers from recent literature.

Table 2: Key sMRI Biomarkers in Autism Spectrum Disorder

Brain Region	Biomarker Type	Alteration in ASD	Associated Clinical Function	Supporting Evidence
Isthmus Cingulate	Cortical Surface Area	Individual differences correlate with symptom severity [47]; Subgroup differentiation [48]	Default Mode Network; Memory & social function [47]	SHAP value: ~0.08 (High predictive feature) [48]
Middle & Inferior Temporal Gyri	GMV / Cortical Thickness	Reduced in subgroup "L"; Associated with symptom severity [48] [47]	Language, semantic processing, visual perception [47]	High abnormality prevalence in subgroup "L" [48]
Anterior Cingulate Cortex	GMV / Cortical Thickness	Reduced volume; Associated with social impairment [49] [47]	Social-cognitive functions, error monitoring [49]	Feature in predictive models of symptom severity [47]
Frontal Lobe (Various)	GMV & Surface Area	Increased in prefrontal cortex; Region-specific alterations [46]	Executive function, social behavior [47]	Increased GMV reported in multiple studies [46]
Insula	GMV & Surface Area	Increased volume in males with ASD; Subgroup differentiation [48] [46]	Salience network; Interoception & social emotion [47]	High abnormality prevalence in subgroup "H" [48]
Cerebellum	GMV	Decreased volume [44]	Motor control, coordination	Coordinate-based meta-analysis finding [44]

Experimental Protocols for Key sMRI Studies

Normative Modeling and Subgroup Identification

A 2025 study profiled brain morphology in ASD using large-scale, cross-cultural consortia (ABIDE and CABIC) to address neurodevelopmental heterogeneity [48].

• Data Acquisition: T1-weighted structural MRI scans were obtained from 2 large-scale consortia: ABIDE (international) and CABIC (China).
• Preprocessing: Brain morphometry was quantified by extracting regional cortical volumes based on the 34 bilateral Desikan-Killiany atlas regions.
• Normative Modeling: Individual-level Out-of-Sample (OoS) centile scores were computed for each brain region by leveraging the Lifespan Brain Chart Consortium (LBCC) models, which provide population-based references for neurodevelopment. This quantified how much each participant's regional volumes deviated from the normative benchmark.
• Subgrouping: Spectral clustering was applied to the OoS centile scores to identify data-driven subgroups.
• Validation: A Support Vector Machine (SVM) classifier with Recursive Feature Elimination and Cross-Validation (RFECV) was trained on one dataset and applied to the other to test the reproducibility of the identified subgroups.

Figure 1: Workflow for Normative Modeling and Subgroup Identification in ASD

Subject-Level Distance-Based Prediction

A 2019 study introduced a novel method to quantify how individual differences in brain morphometry underlie symptom severity in ASD [47].

• Participant Matching: Each individual with ASD (n=100) was strictly matched to a typically developing control based on age, sex, IQ, and data acquisition site.
• Intrapair Distance Calculation: Euclidean distance was computed within each matched pair for both MRI features (cortical thickness and surface area) and symptom severity scores (Social Responsiveness Scale - SRS).
• Machine Learning Pipeline: A regularized linear regression model with elastic net penalty was implemented on the intrapair distance data for feature selection and prediction.
• Validation: Rigorous out-of-sample validation and permutation testing (n=5000) were conducted to ensure the robustness of findings. The model was also tested on independent validation samples.

Deep Learning with Gray Matter Maps

A 2025 study developed a deep learning network (GM-VGG-Net) for ASD classification using solely sMRI-based gray matter maps [46].

• Data: T1-weighted images from 272 subjects (132 controls, 140 ASD) from the ABIDE dataset.
• Preprocessing: Images were processed using SPM12 and the DARTEL toolbox. Gray matter probability maps were segmented, normalized to MNI space, and smoothed.
• Model Architecture: A VGG-based convolutional neural network (CNN) was modified and trained. The model comprised multiple convolutional layers with ReLU activation, followed by pooling layers and fully connected layers.
• Training & Validation: The model was trained for 50 epochs, with a portion of the data reserved for validation to monitor for overfitting. An independent t-test confirmed no significant age difference between groups (p=0.23).

Successfully conducting sMRI biomarker research requires a suite of specialized data, software, and methodological tools. The following table catalogs key resources referenced in the featured studies.

Table 3: Essential Research Reagents and Resources for sMRI Biomarker Studies

Resource Name	Type	Primary Function in Research	Exemplary Use Case
ABIDE (Autism Brain Imaging Data Exchange)	Data Consortium	Provides aggregated, shared sMRI and fMRI datasets from individuals with ASD and controls [48] [46].	Serves as a primary data source for developing and validating classification models [45] [46].
CABIC (China Autism Brain Imaging Consortium)	Data Consortium	Offers cross-cultural sMRI data, enabling validation of findings across diverse populations [48].	Used for cross-cultural replication of ASD subgroups identified in ABIDE [48].
Lifespan Brain Chart Consortium (LBCC)	Normative Model	Provides population-based reference charts for brain morphometry across the lifespan [48].	Generates normative benchmarks for calculating individual deviation scores in ASD [48].
Voxel-Based Morphometry (VBM)	Analytical Method	A whole-brain, user-independent technique for quantifying regional GMV and GMC differences [44] [49].	Used in meta-analyses to identify consistent patterns of GM alteration in ASD [44].
Desikan-Killiany Atlas	Parcellation Atlas	A standardized atlas for partitioning the cerebral cortex into distinct regions for volumetric analysis [48].	Used to extract regional cortical volumes for normative modeling and subgrouping [48].
Support Vector Machine (SVM)	Classifier	A machine learning model effective for high-dimensional data, used for classification and feature selection [48].	Identifying key brain regions that differentiate ASD subgroups via Recursive Feature Elimination [48].
GM-VGG-Net	Deep Learning Model	A customized convolutional neural network designed to classify ASD from control using gray matter maps [46].	Achieving high classification accuracy (96% validation) using sMRI data alone [46].

Figure 2: Logical Flow of Resources and Methods in sMRI Biomarker Research

Integrated Discussion: Clinical Translation and Future Directions

The convergence of evidence from traditional morphometric analyses, normative modeling, and advanced deep learning confirms the substantial promise of sMRI biomarkers for delineating neuroanatomical subtypes in ASD. The identification of biologically distinct subgroups (e.g., subgroups "L" and "H" with divergent volumetric patterns) [48] moves the field beyond a unitary "ASD vs. control" paradigm and towards a precision medicine framework. This is critical for drug development, as it enables the targeting of candidate therapeutics to populations most likely to respond based on their underlying neurobiology.

The performance of sMRI-only models is robust, with deep learning approaches like GM-VGG-Net achieving high validation accuracy [46]. However, multimodal integration that combines sMRI with fMRI may ultimately provide the most comprehensive pathophysiological understanding [45]. Future research should prioritize longitudinal designs to track neurodevelopmental trajectories, increase sample sizes to capture the full spectrum of heterogeneity, and rigorously validate biomarkers for specific clinical purposes, such as predicting treatment response or stratifying patients for clinical trials [50].

Functional Magnetic Resonance Imaging (fMRI) has revolutionized our ability to non-invasively study brain organization and dysfunction in neurodevelopmental conditions such as Autism Spectrum Disorder (ASD). Within this field, the analysis of functional connectivity (FC)—the temporal correlation of neural signals between different brain regions—has emerged as a primary tool for understanding the neurobiological underpinnings of ASD's heterogeneous presentation. FC analysis has evolved into two complementary approaches: Static Functional Connectivity (sFC), which calculates the average connectivity strength over an entire scanning session, and Dynamic Functional Connectivity (dFC), which captures time-varying fluctuations in connectivity patterns throughout the scan [51]. The application of these methods within the context of a broader thesis on structural versus functional MRI for identifying autism subtypes reveals that while sFC provides a stable overview of brain network organization, dFC offers a window into the flexible reconfiguration of brain networks that supports complex cognitive functions—a process potentially disrupted in ASD [51] [52]. This comparison guide objectively examines the methodological protocols, empirical findings, and practical applications of both sFC and dFC for researchers and drug development professionals working to advance ASD biomarker discovery and therapeutic development.

Theoretical Foundations and Comparative Framework

Fundamental Definitions and Physiological Basis

Static Functional Connectivity (sFC) operates under the principle "what is wired together, fires together," representing the average correlation of blood-oxygen-level-dependent (BOLD) time series between brain regions across typically 5-10 minutes of resting-state scanning [51]. It provides a stable, summary measure of brain network integration, reflecting the underlying structural connectivity architecture while being influenced by moment-to-moment physiological fluctuations. In contrast, Dynamic Functional Connectivity (dFC) adheres to the more nuanced principle "what is wired together, fires together... unless at that time it's firing somewhere else," capturing the temporal variability in connectivity patterns that occurs even at rest [51]. This temporal variability is thought to reflect the brain's inherent capacity for network reorganization in response to internal and external demands, with alterations in these dynamics observed across neurodevelopmental disorders including ASD [51] [52].

Comparative Analytical Properties

Table 1: Core Analytical Properties of Static vs. Dynamic Functional Connectivity

Property	Static FC (sFC)	Dynamic FC (dFC)
Temporal Assumption	Stationarity (connectivity stable across scan)	Non-stationarity (connectivity fluctuates over time)
Primary Metric	Pearson correlation coefficient across entire time series	Time-varying correlation measured through sliding windows
Typical Output	Single connectivity matrix per subject	Multiple connectivity matrices representing different states
Biological Interpretation	Average strength of functional communication between regions	Flexibility and temporal characteristics of functional communication
Sensitivity to Cognitive States	Lower (averaged across states)	Higher (can identify distinct brain states)
Data Requirements	Standard resting-state fMRI (5-10 mins)	Longer scans preferred for better state estimation (10+ mins)
Computational Complexity	Lower	Higher (requires multiple computations and state classification)

Methodological Protocols and Experimental Designs

Standardized sFC Analysis Pipeline

The established protocol for sFC analysis begins with extensive fMRI preprocessing using tools like fMRIPrep or Connectome Computation System (CCS), which includes slice-time correction, motion realignment, normalization to standard space (e.g., MNI152), and band-pass filtering (typically 0.01-0.1 Hz) [5] [6]. Following preprocessing, region of interest (ROI) definition is performed using standardized atlases such as the Dosenbach 160 ROI set, which covers regions derived from meta-analyses across multiple cognitive domains including error processing, default mode, memory, language, and sensorimotor functions [6]. The core connectivity calculation involves extracting average BOLD time series from each ROI and computing Pearson correlation coefficients between all region pairs, creating an N×N connectivity matrix for each participant where N represents the number of regions [6]. These matrices are then subjected to statistical analysis, typically involving group-level comparisons (ASD vs. controls) using general linear models or machine learning classifiers, with rigorous multiple comparison correction [5].

Advanced dFC Analysis Methodologies

Dynamic FC analysis incorporates more complex temporal analyses, with the sliding window approach being most widely used [51] [52] [53]. This method involves creating multiple overlapping windows across the BOLD time series, with typical window lengths of 30-60 seconds (often optimized at 50 seconds) [51] [52]. For each window, a separate connectivity matrix is calculated, creating a time series of connectivity matrices that capture the evolving functional architecture [53]. Subsequent clustering analysis using methods like k-means clustering is then applied to these time-varying matrices to identify recurring connectivity "states" that the brain transitions through during the scan [52] [6]. Key dynamic metrics calculated include dwell time (percentage of time spent in each state), transition probability (likelihood of moving between states), and connectivity variability (standard deviation of connectivity strength over time) [52]. For more sophisticated analysis, instant dynamic FC assessment using Dynamic Conditional Correlation (DCC) can be implemented to capture moment-to-moment connectivity fluctuations without predefined windows [6].

Integrated Multi-Modal Subtyping Approaches

Cutting-edge protocols for ASD subtyping now integrate multiple connectivity measures with normative modeling. For instance, a comprehensive analysis of 1,046 participants established a normative model of typical functional development using both sFC and dFC measures from the Dosenbach 160 atlas, then quantified individual ASD participants' deviations from this normative trajectory [6]. Clustering analyses applied to these deviation profiles have successfully identified distinct ASD subtypes with unique functional network signatures despite comparable clinical presentations [6]. Another innovative approach combines structural and functional coupling through skeleton-based white matter functional analysis, projecting fMRI signals onto a white matter skeleton to enable voxel-wise function-structure coupling quantification [2]. These multi-modal approaches demonstrate enhanced sensitivity for identifying biologically distinct ASD subgroups compared to single-modality designs.

Empirical Findings and Data Synthesis in Autism Spectrum Disorder

Characteristic Connectivity Alterations in ASD

Table 2: Characteristic Functional Connectivity Alterations in Autism Spectrum Disorder

Brain Network/Region	Static FC Findings in ASD	Dynamic FC Findings in ASD	Clinical Correlations
Default Mode Network (DMN)	Mixed hyper-/hypoconnectivity patterns [51]	Longer dwell time in disconnected states [52]	Social communication deficits [52]
Frontoparietal Network	Decreased connectivity in one subtype [6]	Altered dynamic control network functions [52]	Executive function impairments [6]
Amygdala Subregions	Decreased amygdala-thalamus connectivity [54]	Increased dFC between BLA and temporal lobe [54]	Social symptom severity [54]
Thalamocortical Pathways	Not typically reported	Significant hyperconnectivity (82.58% difference in degree) [55]	Potential sensory processing alterations
Cerebellar Network	Positive deviations in one subtype [6]	Not specifically reported	Repetitive behaviors, motor coordination
Subcortical Networks	Major differences between ASD subtypes [5]	Greater dynamic connectivity variability [56]	Heterogeneous clinical presentations

Subtype-Specific Connectivity Signatures

Research has consistently identified two primary ASD subtypes based on functional connectivity profiles. Subtype 1 demonstrates positive deviations in the occipital and cerebellar networks coupled with negative deviations in the frontoparietal network, default mode network, and cingulo-opercular network [6]. In contrast, Subtype 2 exhibits the inverse pattern, with negative deviations in occipital and cerebellar networks and positive deviations in frontoparietal, default mode, and cingulo-opercular networks [6]. These neural subtypes manifest in distinct behavioral profiles, with Subtype 2 showing lower full-scale and performance IQ scores compared to Subtype 1 [2]. Importantly, these subtypes, while neurally distinct, may present with comparable clinical symptoms, highlighting the limitation of relying solely on behavioral assessments for ASD stratification [6].

Another subtyping approach based on structural-functional coupling identified two neurosubtypes with distinct white matter profiles: Subtype 1 showed significantly lower fractional anisotropy (FA) only in the posterior cingulate cortex compared to controls, while Subtype 2 exhibited reduced FA in the anterior cingulate cortex, middle temporal gyrus, parahippocampus, and thalamus [2]. This pattern suggests different underlying microstructural abnormalities contributing to the heterogeneous ASD presentation.

Visualization of Analytical Workflows and Neural Signatures

Comprehensive fMRI Connectivity Analysis Pipeline

ASD Neural Subtype Classification Model

Essential Research Reagents and Computational Tools

Table 3: Essential Research Resources for fMRI Connectivity Studies in ASD

Resource Category	Specific Tools/Platforms	Primary Function	Application Notes
Neuroimaging Databases	ABIDE I/II (N>2000) [5] [6]	Multi-site fMRI data repository	Essential for large-scale analyses; includes phenotypic data
Preprocessing Tools	fMRIPrep [6], CCS Pipeline [5]	Automated fMRI preprocessing	Standardizes processing across sites; reduces manual errors
Connectivity Analysis	Dosenbach 160 Atlas [6]	ROI-based connectivity definition	Covers multiple cognitive domains; enables cross-study comparisons
Dynamic FC Toolboxes	Sliding Window Algorithms [53] [54]	Time-varying connectivity estimation	Window length critical (30-60s typical); requires optimization
Statistical Frameworks	Normative Modeling [6]	Individual deviation quantification	Maps ASD heterogeneity relative to typical development
Cluster Analysis	k-means, Hierarchical Clustering [6] [2]	ASD subtype identification	Identifies data-driven subgroups beyond clinical categories
Validation Tools	Eye-tracking (Tobii TX300) [6]	Behavioral correlation	Links neural subtypes to objective behavioral measures

Discussion and Research Applications

Clinical Translation and Biomarker Development

The integration of sFC and dFC measures shows significant promise for developing clinically viable biomarkers for ASD. Research demonstrates that combining these modalities enhances diagnostic prediction accuracy compared to using either approach alone [2]. Furthermore, the identification of distinct neurosubtypes with unique functional connectivity signatures provides a biological framework for understanding ASD heterogeneity and developing personalized intervention strategies [6] [2]. Notably, a recent study found that different ASD connectivity subtypes showed dramatically different response rates to intranasal oxytocin treatment (61.5% vs. 13.3%), highlighting the potential treatment implications of connectivity-based subtyping [6].

Methodological Considerations and Limitations

While promising, several methodological challenges require consideration in connectivity studies of ASD. Motion artifacts remain a significant confound, particularly in pediatric ASD populations, necessitating rigorous motion correction and exclusion criteria (typically mean framewise displacement <0.3mm) [6]. Multi-site heterogeneity in scanning parameters introduces variability, though methods like image-based meta-analysis (IBMA) can identify consistent effects across sites [54]. Temporal resolution limitations of fMRI constrain the interpretation of dFC findings, with the optimal balance between window length and state detection stability requiring careful consideration [51]. Finally, sample size requirements for robust subtyping are substantial, with recent studies leveraging samples exceeding 1,000 participants to ensure generalizability [6].

The complementary application of static and dynamic functional connectivity analyses has substantially advanced our understanding of autism spectrum disorder's neurobiological foundations. While sFC provides a stable mapping of brain network organization alterations in ASD, dFC captures the disrupted temporal dynamics and network flexibility underlying the condition's cognitive and behavioral features. The integration of these approaches within a broader structural-functional framework has revealed biologically distinct ASD subtypes that cross traditional diagnostic boundaries, offering a path toward personalized interventions targeting specific neural mechanisms. For drug development professionals, these connectivity signatures provide potential biomarkers for patient stratification and treatment response monitoring in clinical trials. Future research directions should include longitudinal studies to track connectivity development, integration with genetic and molecular data to elucidate biological mechanisms, and translation of these findings into clinically accessible biomarkers that can guide therapeutic decisions for individuals with ASD.

Autism Spectrum Disorder (ASD) is characterized by substantial phenotypic, biological, and etiologic heterogeneity that has complicated neurobiological investigation and treatment development. The conventional case-control paradigm, which assumes homogeneous diagnostic entities, has proven inadequate for capturing the complex and continuous nature of cognitive and neural variations in autism [57] [58]. This analytical limitation has resulted in highly inconsistent neuroimaging findings across studies, with some reporting increases in cortical thickness, others decreases, and still others finding no significant differences, even within the same brain regions [57]. The emergence of big data resources in neuroscience, characterized by both breadth (large sample sizes) and depth (multiple data levels per individual), has created unprecedented opportunities to decompose this heterogeneity through advanced computational approaches [58].

Two particularly promising frameworks—tensor decomposition and normative modeling—have recently demonstrated utility in parsing neurobiological heterogeneity in ASD. These approaches represent complementary paradigms for moving beyond group-level comparisons to individualized characterization of brain organization. Tensor decomposition provides a multivariate framework for analyzing high-dimensional neuroimaging data, while normative modeling establishes subject-specific references for quantifying individual deviations from typical brain development patterns [6] [5]. When applied to the ongoing debate regarding structural versus functional MRI biomarkers in autism, these advanced analytical approaches offer new pathways for identifying biologically meaningful subtypes with potential implications for personalized intervention strategies [59].

This review systematically compares the methodological foundations, experimental implementations, and empirical findings of tensor decomposition and normative modeling approaches for identifying autism subtypes using structural and functional neuroimaging data. We provide detailed experimental protocols, quantitative comparisons, and practical resources to guide researchers and drug development professionals in selecting and implementing these advanced analytical frameworks.

Methodological Foundations and Comparative Frameworks

Tensor Decomposition for Neuroimaging Data Analysis

Tensor decomposition represents a multidimensional extension of matrix factorization that is particularly well-suited for analyzing complex neuroimaging data structures. In the context of fMRI analysis, tensors naturally represent the multidimensional nature of brain data, which typically encompasses spatial dimensions (brain regions or voxels), temporal dimensions (time points), and participant dimensions (individual subjects or groups) [5]. The core principle involves decomposing the original tensor into a set of factor matrices and a core tensor that capture the latent patterns underlying the observed data.

When applied to autism subtyping, tensor decomposition operates on a three-mode tensor constructed from resting-state fMRI data across multiple participants. The methodology enables simultaneous decomposition of functional connectivity patterns across brain networks, temporal dynamics, and individual subjects, revealing distinctive subpopulations with common functional brain organization [5]. A key advantage of this approach is its ability to capture complex multi-way interactions in the data that would be lost in conventional matrix factorization approaches applied to flattened connectivity matrices.

The experimental workflow for tensor decomposition analysis typically involves several stages: (1) data preprocessing and tensor construction, (2) dimensionality estimation, (3) tensor decomposition using algorithms such as Canonical Polyadic Decomposition or Tucker decomposition, (4) factor interpretation and clustering, and (5) subtype validation and characterization. This approach has demonstrated particular utility for distinguishing between historically defined ASD subtypes (autism, Asperger's, and PDD-NOS) based on their distinctive functional network profiles [5].

Normative Modeling for Individualized Deviation Mapping

Normative modeling represents a fundamentally different approach that quantifies neurobiological heterogeneity at the level of the individual participant rather than seeking clusters within the population. This framework establishes a reference model of typical brain development based on a large sample of neurotypical individuals, then quantifies how each autistic individual deviates from this normative trajectory [57] [6]. The core output is a personalized map of anatomical or functional abnormalities that captures the unique neurobiological signature of each person with autism.

The mathematical foundation of normative modeling typically involves Gaussian process regression, which provides a flexible, probabilistic framework for estimating non-linear developmental trajectories while accounting for covariates such as age, sex, and intelligence quotient [57]. For each brain feature (e.g., cortical thickness at each vertex, or functional connectivity for each edge), the algorithm learns a distribution of typical variation across the neurotypical reference sample. Individuals with ASD are then projected onto this normative range, generating Z-scores that quantify their deviation from the expected pattern at each brain location.

The experimental workflow encompasses: (1) reference cohort acquisition and quality control, (2) normative model estimation using Gaussian process regression, (3) projection of ASD participants onto the normative range, (4) computation of individualized deviation maps, and (5) optional clustering of deviation patterns to identify subtypes [57]. This approach has revealed that different autistic individuals show highly individualized patterns of atypicality across nearly the entire cortex, with some showing relatively widespread decreased cortical thickness and others showing relatively increased cortical thickness [57].

Table 1: Core Methodological Differences Between Tensor Decomposition and Normative Modeling

Analytical Feature	Tensor Decomposition	Normative Modeling
Primary objective	Identify subgroups with similar brain patterns	Quantify individual deviations from typical development
Reference framework	Data-driven clustering within ASD population	Comparison to neurotypical reference cohort
Mathematical foundation	Multilinear algebra and tensor factorization	Gaussian process regression and outlier detection
Heterogeneity conceptualization	Categorical subtypes	Continuous dimensional deviations
Primary data structure	Three+ dimensional tensors	Vectorized features with covariate modeling
Output for each participant	Cluster assignment	Personalized deviation map (Z-scores)
Temporal dynamics handling	Directly incorporated into tensor structure	Requires separate models for each time point

Comparative Workflow Visualization

Diagram 1: Comparative analytical workflows for tensor decomposition (red) and normative modeling (blue) approaches to autism subtyping.

Experimental Protocols and Implementation

Tensor Decomposition Protocol for ASD Subtyping

The implementation of tensor decomposition for autism subtyping follows a structured pipeline with specific parameters and data requirements. A representative protocol from a recent study using the ABIDE I dataset illustrates the key methodological steps [5]:

Data Acquisition and Preprocessing: The study utilized resting-state fMRI and anatomical data from 152 patients with autism, 54 with Asperger's, and 28 with PDD-NOS from the ABIDE I database. Data preprocessing followed the Connectome Computation System pipeline, including slice timing correction, head motion correction, spatial normalization to MNI152 standard space, and band-pass filtering (0.01-0.1 Hz). Global signal regression was applied to reduce nonspecific physiological effects [5].

Feature Extraction and Tensor Construction: For each participant, functional connectivity matrices were computed by extracting average BOLD signals from predefined brain regions (e.g., Dosenbach 160 ROIs) and calculating Pearson correlation coefficients between regional time series. These matrices were then used to construct a three-mode tensor with dimensions: participants × functional connections × temporal windows (for dynamic FC) or a participants × regions × regions tensor for static FC analysis [5].

Decomposition and Clustering: The Tucker decomposition algorithm was applied to the constructed tensor, factorizing it into a core tensor and factor matrices corresponding to each mode. The number of components was determined using cross-validation and core consistency diagnostics. Subsequently, spectral clustering with a cosine similarity affinity matrix was applied to the participant factor matrix to identify discrete ASD subtypes [5].

Validation and Characterization: The identified subtypes were validated through: (1) split-half reproducibility analysis, (2) comparison of clinical and demographic profiles across subtypes, and (3) neurobiological characterization using additional imaging modalities (e.g., gray matter volume, ALFF/fALFF). Statistical tests including ANOVAs and post-hoc comparisons were used to evaluate subtype differences [5].

Normative Modeling Protocol for ASD Heterogeneity

The implementation of normative modeling for quantifying neurobiological heterogeneity in autism involves distinct methodological stages, as demonstrated in large-scale studies such as the Longitudinal European Autism Project (LEAP) [57]:

Reference Cohort Establishment: The normative model was established using structural MRI data from 206 neurotypical participants (79 female, aged 17.5 ± 6.1 years) across 6 sites. Cortical thickness was estimated at each vertex across the cortical surface using FreeSurfer v5.3. The sample was carefully selected to represent a broad age range (6-31 years) and balanced sex distribution [57].

Normative Model Estimation: Gaussian process regression was used to predict vertex-wise cortical thickness as a function of age and sex, while accounting for nuisance covariates including full-scale IQ, site effects, and data quality metrics (Euler number). This approach generates both a predicted cortical thickness and an estimate of uncertainty at each vertex for any given age and sex combination [57].

Deviation Quantification: For each of the 316 participants with ASD (88 female, aged 17.2 ± 5.7 years), normative probability maps were computed by subtracting the predicted from the actual cortical thickness, divided by the estimated standard deviation at each vertex. This generates a subject- and vertex-specific Z-score representing the deviation from the normative range [57].

Subtype Derivation: Spectral clustering with a cosine similarity affinity matrix was applied to the unthresholded normative probability maps from ASD participants. The optimal number of clusters (K=5) was determined using multi-class linear support vector machine separability and stability analysis through leave-one-out procedures [57].

Clinical and Genetic Validation: The derived subtypes were validated by examining their associations with clinical measures (ADI-R, ADOS-2, SRS-2, RBS-R, SSP) and polygenic scores for seven neuropsychiatric traits. This validation tested whether the neuroanatomical subtypes showed differential loading onto symptoms and genetic risk factors [57].

Table 2: Representative Experimental Parameters from Recent Studies

Parameter	Tensor Decomposition Study [5]	Normative Modeling Study [57]
Sample size	234 ASD (152 Autism, 54 Asperger's, 28 PDD-NOS)	316 ASD, 206 Neurotypical
Age range	Children and adolescents (exact range not specified)	6-31 years
Data sources	ABIDE I	EU-AIMS LEAP (6 sites)
Primary features	Functional connectivity, ALFF/fALFF, GMV	Cortical thickness (vertex-wise)
Clustering method	Tensor decomposition + spectral clustering	Normative modeling + spectral clustering
Number of subtypes	3 (based on DSM-IV categories)	5 data-driven clusters
Validation approach	Cross-site reproducibility, clinical differentiation	Clinical correlates, polygenic risk scores
Software tools	Custom MATLAB implementations, CCS pipeline	Gaussian process regression, FreeSurfer

Empirical Findings and Subtype Characterization

Neurobiological Subtypes Identified Through Tensor Decomposition

Tensor decomposition approaches applied to functional and structural neuroimaging data have revealed subtype differences that align with historically defined diagnostic categories while providing new insights into their neurobiological foundations. A recent study examining three ASD subtypes (autism, Asperger's, and PDD-NOS) found distinctive functional and structural patterns that differentiated these groups [5].

Specifically, the autism subtype showed prominent impairments in the subcortical network and default mode network compared to the other subtypes. These functional differences were complemented by structural alterations in gray matter volume, particularly in regions associated with social cognition and sensory processing. The Asperger's subtype demonstrated a distinct pattern characterized by relative preservation of these networks but alterations in frontoparietal connectivity. The PDD-NOS subtype showed an intermediate pattern with less pronounced alterations across networks [5].

The functional basis for these distinctions was further elaborated through analysis of amplitude of low-frequency fluctuations (ALFF) and fractional ALFF (fALFF), which revealed subtype-specific patterns of intrinsic brain activity. When combined with gray matter volume measures, these multimodal features provided a comprehensive characterization of the neurobiological differences between subtypes, suggesting that historically defined diagnostic categories may reflect distinct patterns of brain organization [5].

Data-Driven Subtypes Revealed Through Normative Modeling

Normative modeling approaches have identified neuroanatomical subtypes that cut across conventional diagnostic boundaries and demonstrate distinctive genetic and clinical profiles. Application of this method to cortical thickness data from 316 autistic individuals revealed five biologically based subtypes with differential patterns of neuroanatomical deviation [57].

Three clusters showed relatively widespread decreased cortical thickness and two showed relatively increased cortical thickness compared to the neurotypical reference. These subtypes demonstrated significant morphometric differences from one another, providing a potential explanation for inconsistent case-control findings in autism. Crucially, the subtypes loaded differentially onto clinical symptoms and polygenic risk, with one cluster (Cluster 2) containing subjects with higher impairment, having lower IQ and more severe core autism symptoms across diagnostic instruments [57].

The clinical utility of this approach was further demonstrated by the finding that the neuroanatomical subtypes showed stronger associations with symptoms and genetic risk than the undifferentiated ASD group, indicating that conventional case-control approaches dilute clinical effects across heterogeneous cohorts. This provides empirical support for the hypothesis that decomposing heterogeneity can enhance biomarker discovery and facilitate precision medicine approaches [57].

Functional Connectivity Subtypes and Cross-Species Validation

Recent advances in functional connectivity analysis have identified reproducible subtypes characterized by distinct patterns of functional network organization. A groundbreaking cross-species study examining fMRI connectivity in 20 distinct mouse models of autism (n=549 individual mice) and extending these findings to humans (n=940 autistic and n=1036 neurotypical individuals) identified two prominent hypo- and hyperconnectivity subtypes that were replicable across independent cohorts [60].

Remarkably, these functionally defined subtypes accounted for 25.1% of all autism data and exhibited distinct functional network architecture, behavioral profiles, and underlying biological mechanisms. The hypoconnectivity subtype was associated with synaptic dysfunction, while the hyperconnectivity subtype reflected transcriptional and immune-related alterations, demonstrating that connectivity-based subtypes capture distinct pathway-specific etiologies [60].

Complementing these findings, a comprehensive analysis of 1,046 participants (479 with ASD, 567 with typical development) identified two distinct neural ASD subtypes with unique functional brain network profiles despite comparable clinical presentations [6]. One subtype was characterized by positive deviations in the occipital and cerebellar networks coupled with negative deviations in the frontoparietal network, default mode network, and cingulo-opercular network. The other subtype exhibited the inverse pattern. These neural subtypes were also associated with distinct gaze patterns assessed by autism-sensitive eye-tracking tasks, providing a crucial link between brain network organization and behavioral presentation [6].

Table 3: Comparative Findings from Tensor Decomposition and Normative Modeling Studies

Research Dimension	Tensor Decomposition Findings	Normative Modeling Findings
Primary subtype distinctions	Differences in subcortical network, default mode network, and frontoparietal connectivity	Widespread decreased vs. increased cortical thickness patterns
Clinical correlations	Alignment with historical diagnostic categories (autism, Asperger's, PDD-NOS)	Distinct symptom profiles and cognitive functioning levels
Genetic associations	Not explicitly examined in studies reviewed	Differential loading onto polygenic risk scores for neuropsychiatric traits
Reproducibility	Cross-site validation using ABIDE I dataset	Replication across multiple sites in EU-AIMS LEAP
Behavioral validation	Not extensively reported	Association with eye-tracking patterns and sensory profiles
Potential therapeutic implications	Subtype-specific network targets for neuromodulation	Personalized interventions based on individual deviation profiles

Integration with Complementary Analytical Frameworks

Synergies with Cross-Species Validation Approaches

The combination of advanced analytical approaches with cross-species validation represents a particularly powerful paradigm for deconstructing autism heterogeneity. The identification of analogous hypo- and hyperconnectivity subtypes in both mouse models and humans provides compelling evidence that these subtypes reflect fundamental biological divisions with distinct underlying mechanisms [60].

This cross-species approach enables researchers to move beyond descriptive subtyping to mechanistic investigations of the biological pathways driving different autism presentations. The hypoconnectivity subtype's association with synaptic dysfunction suggests potential responsiveness to interventions targeting synaptic function, while the hyperconnectivity subtype's link to immune mechanisms suggests different therapeutic targets [60]. This exemplifies how advanced analytics can inform targeted therapeutic development.

Relationship to Precision Neurodiversity Frameworks

Both tensor decomposition and normative modeling align with the emerging paradigm of precision neurodiversity, which marks a shift from pathological models to personalized frameworks that view neurological differences as adaptive variations [59]. This perspective emphasizes individual-specific neural fingerprints rather than categorical diagnostic assignments.

Recent advances in personalized brain network analysis demonstrate that individual network profiles reliably predict cognitive, behavioral, and sensory phenomena [59]. The integration of tensor decomposition and normative modeling with this framework offers promising avenues for developing individually tailored support strategies that accommodate neurobiological variation rather than attempting to normalize it.

Complementary Neuroimaging Modalities

Diffusion Tensor Imaging (DTI) studies provide important complementary evidence for structural connectivity alterations in autism that parallel findings from functional connectivity analyses. A systematic review of DTI studies in adults with ASD found regionally diverse white matter alterations, particularly in frontal and interhemispheric tracts, association fibers, and subcortical projection pathways [61].

These structural connectivity patterns show both convergent and divergent relationships with functional connectivity findings, suggesting that a complete understanding of autism subtypes will require integration across multiple imaging modalities. The combination of functional and structural connectivity measures within tensor decomposition or normative modeling frameworks represents a promising direction for future research.

Research Reagents and Computational Tools

Table 4: Essential Research Resources for Implementing Advanced Analytical Approaches

Resource Category	Specific Tools/Databases	Primary Function	Access Information
Neuroimaging Datasets	ABIDE I & II (Autism Brain Imaging Data Exchange)	Provide large-scale, multi-site fMRI data for discovery and validation	http://fcon_1000.projects.nitrc.org/indi/abide/
Neuroimaging Datasets	EU-AIMS LEAP (Longitudinal European Autism Project)	Offer deeply phenotyped sample with multiple imaging modalities	Controlled access through EU-AIMS consortium
Data Processing Pipelines	Connectome Computation System (CCS)	Comprehensive fMRI preprocessing and feature extraction	https://github.com/zuoxinian/CCS
Data Processing Pipelines	fMRIPrep	Robust preprocessing of functional MRI data	https://fmriprep.org/en/stable/
Software Libraries	TensorLy	Python library for tensor decomposition and analysis	http://tensorly.org/stable/
Software Libraries	PCNToolkit	MATLAB/Python tools for normative modeling	https://github.com/amarquand/PCNtoolkit
Analysis Platforms	FreeSurfer	Cortical reconstruction and anatomical analysis	https://surfer.nmr.mgh.harvard.edu/
Analysis Platforms	SPM, FSL, AFNI	General neuroimaging analysis and statistical processing	Various academic distributions

Visualization of Analytical Framework Integration

Diagram 2: Integration framework showing how multiple data sources (green) feed into advanced analytical approaches (yellow) to generate clinically relevant outputs (blue) for autism research.

Tensor decomposition and normative modeling represent complementary advanced analytical approaches for deconstructing the profound heterogeneity in autism spectrum disorder. While tensor decomposition excels at identifying discrete subgroups based on patterns of functional network organization, normative modeling provides a framework for quantifying individualized deviations from typical neurodevelopmental trajectories. Both approaches demonstrate superior sensitivity to biologically meaningful distinctions compared to conventional case-control paradigms.

The integration of these computational approaches with large-scale neuroimaging datasets, cross-species validation, and deep phenotyping represents a promising pathway toward precision medicine in autism research. Future directions should focus on: (1) longitudinal applications to characterize neurodevelopmental trajectories, (2) multi-modal integration of structural, functional, and genetic data, (3) clinical translation for treatment targeting and prognosis prediction, and (4) open-source tool development to facilitate widespread implementation. As these advanced analytical frameworks mature, they offer the potential to transform autism from a behaviorally defined diagnostic category into a neurobiologically informed spectrum with personalized support strategies.

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by heterogeneous manifestations across social communication, behavior, and cognitive functioning. The current diagnostic paradigm relies primarily on behavioral assessments and clinical observation, a process that can be subjective, time-consuming, and challenging to standardize. This creates a pressing need for objective, biologically-based diagnostic tools. Neuroimaging techniques, particularly structural and functional Magnetic Resonance Imaging (sMRI and fMRI), offer promising avenues for identifying robust biomarkers by revealing the distinct neural underpinnings of ASD. While sMRI provides detailed information on brain anatomy, including cortical thickness and gray matter volume, fMRI captures dynamic brain activity and functional connectivity patterns. Individually, each modality offers valuable but incomplete insights. However, their integration, powered by advanced machine learning techniques like semi-supervised learning, creates a powerful framework for achieving a more comprehensive and accurate classification of ASD, potentially transforming both diagnosis and the understanding of its neurobiological subtypes.

Performance Comparison of Multimodal Classification Approaches

Research consistently demonstrates that integrating sMRI and fMRI data outperforms single-modality approaches in classifying individuals with ASD versus typical controls. The table below summarizes key performance metrics from recent studies utilizing different data fusion strategies.

Table 1: Performance Comparison of Multimodal vs. Single-Modality ASD Classification Models

Study & Model Description	Primary Modality/Fusion Type	Accuracy	F1-Score	Key Fusion Strategy
ALFF-sMRI Two-Channel 3D-DenseNet [34]	fMRI (ALFF maps) + sMRI	76.9%	-	Two-channel deep network with twinned neuroimaging inputs
fMRI 3D-CNN (50-middle slices) [62]	fMRI (Multi-slice)	87.1%	0.826	3D-CNN on mean of 4D images across axial, coronal, sagittal planes
Attention-Enhanced sMRI-fMRI Fusion (ASFF) [63]	sMRI + fMRI (Multimodal)	-	-	Mutual cross-attention fusion module with semantic constraint
sMRI Transfer Learning (ConvNeXt) [62]	sMRI (Multi-slice)	-	-	Vision transformer fine-tuned on 50-middle sMRI slices
LSTM on rs-fMRI Time-Series [64]	fMRI (Time-series)	68.5%	-	Recurrent neural network analyzing full temporal dynamics

The quantitative data reveals a clear efficacy hierarchy. The highest performance is achieved by models that effectively leverage complementary information: 3D-CNNs on processed fMRI data can achieve exceptional accuracy (87.1%), while two-channel architectures fusing sMRI with fMRI-derived maps like ALFF also show significant promise (76.9% accuracy) [62] [34]. These results substantiate the core thesis that the fusion of structural and functional data provides a more complete phenotypic signature of ASD than either modality alone.

Detailed Experimental Protocols for Multimodal Integration

Two-Channel Deep Learning with sMRI and ALFF/fALFF

A. Objective: To classify young individuals with ASD using a balanced dataset by leveraging the complementary information from sMRI and resting-state fMRI-derived amplitude maps (ALFF/fALFF) in a unified deep learning architecture [34].

B. Dataset:

• Source: Autism Brain Imaging Data Exchange (ABIDE I) [34].
• Cohort: 702 participants (351 ASD, 351 typically developing controls), aged 2-30 years.
• Exclusion Criteria: Major comorbidities, and data with motion artifacts, ghosting, or incomplete brain coverage.

C. Preprocessing & Feature Extraction:

• sMRI Processing: T1-weighted images were downsampled to 3mm isotropic resolution. Skull-stripping was performed using SynthStrip in FreeSurfer [34].
• fMRI Processing & ALFF/fALFF Calculation:
  • The first four fMRI volumes were discarded for signal stabilization.
  • Motion correction and skull-stripping were applied.
  • Steps included despiking, linear detrending, spatial smoothing (6mm FWHM), and band-pass filtering (0.01–0.08 Hz).
  • The power spectrum of the time series was computed. ALFF was defined as the total power within the 0.01–0.08 Hz band. fALFF was the ratio of this power to the total power across the entire frequency range [34].

D. Model Architecture & Training:

• Base Model: A 3D-DenseNet architecture was employed.
• Input Configurations:
  • One-channel: Used only sMRI, ALFF, or fALFF maps.
  • Two-channel: Combined sMRI with ALFF or sMRI with fALFF maps as dual inputs.
• Training Regime:
  • Data Augmentation: Random rotation (±30° around z-axis) and zoom (0.7–1.3x) were applied during training.
  • Validation: Tenfold cross-validation was used for performance evaluation [34].

Attention-Enhanced Structural and Functional MRI Fusion (ASFF)

A. Objective: To develop an interpretable fusion framework that models the underlying mutual relationships between sMRI and fMRI features for improved classification of neurocognitive disorders [63].

B. Multimodal Feature Extraction:

• sMRI Features:
  • Data-driven: A 3D CNN with four blocks processed raw T1-weighted images.
  • Handcrafted: FreeSurfer was used to extract 4,858 features (e.g., cortical thickness, gray matter volume).
  • Both feature sets were processed by MLPs and combined into a final sMRI feature vector S [63].
• fMRI Features:
  • The brain was parcellated using the AAL atlas.
  • The fMRI time series was partitioned using a sliding window.
  • For each window, a graph was constructed where nodes represented ROIs and edges represented Pearson's correlation coefficients.
  • A Graph Isomorphism Network (GIN) updated node representations, and a Transformer modeled temporal dynamics across windows, outputting a spatiotemporal feature vector F [63].

C. Fusion Module:

• Mutual Cross-Attention: The core of the ASFF framework. It takes the sMRI feature S and fMRI feature F and generates three vectors for each (Query, Key, Value). The cross-attention mechanism is computed as [63]:
  • S_A = softmax((Q_F · K_S^T) / √M) · V_S (fMRI query attends to sMRI key)
  • F_A = softmax((Q_S · K_F^T) / √M) · V_F (sMRI query attends to fMRI key)
• Semantic Inter-Modality Constraint: A loss function (L_S) was designed to maximize the semantic consistency between the attended features (S_A and F_A) from the same subject while pushing apart features from different subjects [63].

D. Objective Function: The total loss for training combines the classification loss (L_C), the modality-specific loss (L_M), and the semantic inter-modality constraint (L_S) [63].

Diagram 1: Workflow of the Attention-Enhanced sMRI-fMRI Fusion (ASFF) Framework.

Successful execution of sMRI-fMRI fusion studies requires a curated suite of data, software, and computational tools. The following table details the key components of the modern computational neuroscientist's toolkit.

Table 2: Essential Research Reagents and Resources for sMRI-fMRI Fusion Analysis

Tool/Resource Name	Type	Primary Function in Research	Key Utility
ABIDE (I & II) [65] [34]	Data Repository	Aggregates sMRI, fMRI, and phenotypic data from individuals with ASD and controls.	Provides large-scale, publicly available datasets for training and testing models.
FreeSurfer [66] [63]	Software Toolbox	Automated processing and analysis of sMRI data.	Extracts cortical and subcortical morphological features (thickness, volume).
FSL [34]	Software Library	FMRIB Software Library for MRI data analysis.	Used for brain extraction, tissue segmentation, and functional connectivity analysis.
AFNI [34]	Software Library	Analysis of functional neuroimages.	Aids in preprocessing and statistical analysis of fMRI data.
3D-DenseNet [34]	Deep Learning Model	3D convolutional neural network for volumetric data.	Classifies subjects using sMRI, ALFF/fALFF maps, or their combination.
Graph Isomorphism Network (GIN) [63]	Deep Learning Model	Graph neural network for non-Euclidean data.	Models functional connectivity patterns from fMRI data represented as graphs.
Transformer Networks [63]	Deep Learning Model	Architecture for sequence-to-sequence learning.	Captures temporal dynamics and long-range dependencies in fMRI time-series.
U-Net [66]	Deep Learning Model	Convolutional network for biomedical image segmentation.	Used in semi-supervised learning for tasks like brain tissue segmentation.

The integration of sMRI and fMRI, particularly when enhanced by sophisticated fusion algorithms and semi-supervised learning paradigms, represents a paradigm shift in autism research. The experimental data and protocols detailed in this guide underscore a fundamental conclusion: the combined power of structural and functional biomarkers significantly surpasses the capabilities of either modality in isolation. Frameworks like the ASFF and two-channel 3D-DenseNet move beyond simple feature concatenation, actively modeling the rich, mutual relationships between brain structure and function. This approach not only boosts classification accuracy but also enhances the interpretability of models, allowing researchers to identify modality-shared and modality-specific brain regions that may serve as core biomarkers for distinct autism subtypes. As these tools mature and are applied to larger, more diverse datasets, they hold the promise of delivering the objective, biologically grounded diagnostic aids necessary to personalize interventions and advance drug development for the heterogenous population of individuals with ASD.

Navigating the Valley of Death: Overcoming Barriers in ASD Biomarker Discovery

The development of effective pharmacological treatments for autism spectrum disorder (ASD) represents one of the most significant challenges in contemporary neuropsychiatry. Despite substantial investments in basic neuroscience research and clinical trials, translational failure rates remain exceptionally high, with approximately 95% of experimental drugs entering human trials failing to gain regulatory approval [67]. This crisis exists within a paradoxical landscape: while our fundamental knowledge of ASD genetics and neurobiology has expanded dramatically, this progress has not yielded commensurate therapeutic advances [68] [69]. The translational gap—often termed the "valley of death"—between promising preclinical findings and clinical application has proven particularly difficult to bridge for ASD, a condition characterized by exceptional heterogeneity in its clinical presentation and underlying biology [67]. This review examines the causes of these translational failures and explores how emerging approaches to ASD subtyping using structural and functional neuroimaging may provide a pathway toward more successful therapeutic development.

The Scale and Scope of Translational Challenges

Quantifying the Problem

The magnitude of the translational challenge in ASD drug development is evidenced by several critical metrics:

Table 1: Quantifying Translational Challenges in ASD Drug Development

Challenge Area	Metric	Impact
Overall Success Rates	0.1% of candidates proceed from preclinical to approved drug [67]	Extremely low return on research investment
Clinical Trial Failure	~95% failure rate for drugs entering human trials [67]	High attrition despite promising preclinical data
Development Timeline	>13 years from discovery to approval [67]	Extended timeframe increases costs and delays treatments
Development Costs	~$2.6 billion per approved drug [67]	Substantial financial barriers to ASD therapeutic development
Treatment Coverage	<600 of ~8000 human diseases have approved treatments [68]	Majority of conditions lack regulatory-approved interventions

The complexity of ASD presents unique additional challenges. Of the numerous controlled clinical trials conducted for ASD core features, none have demonstrated conclusive efficacy, with only a few medications receiving regulatory approval for associated symptoms such as irritability [69]. This failure occurs despite compelling preclinical theories, most notably the excitatory/inhibitory (E/I) imbalance hypothesis, which successfully guided drug development in animal models but failed to translate to human populations [69].

Root Causes of Translational Failure

Multiple interrelated factors contribute to the high failure rate in ASD therapeutic development:

• Biological Heterogeneity: ASD encompasses a spectrum of conditions with diverse genetic underpinnings and neurobiological mechanisms, making single-treatment approaches unlikely to benefit the entire population [69].

• Preclinical Model Limitations: Animal models used in preclinical testing often fail to recapitulate the complexity of human ASD, particularly for social and cognitive endpoints [70]. Studies in rodent models frequently suffer from inadequate statistical power, insufficient blinding, and failure to account for genetic and environmental variables [70].

• Diagnostic and Assessment Challenges: Traditional ASD diagnosis relies on behavioral observations rather than objective biomarkers, introducing subjectivity and heterogeneity into clinical trial populations [1]. Current endpoints may lack sensitivity to detect meaningful clinical changes, particularly for core social and communication deficits [69].

• Target Identification Problems: The transition from genetic discoveries to viable drug targets has proven difficult, with poor understanding of how to modulate relevant neural circuits in humans [69].

Neuroimaging Subtypes as a Potential Solution

The Case for Biological Subtyping in ASD

The exceptional heterogeneity of ASD suggests that dividing the spectrum into more biologically homogeneous subgroups may be essential for therapeutic progress. Emerging evidence from neuroimaging studies supports this approach, revealing distinct neurobiological subtypes that cut across traditional behaviorally defined categories [2] [5] [71]. These subtypes demonstrate different patterns of brain structure and function, potentially reflecting distinct underlying pathophysiological mechanisms that may respond differently to targeted interventions.

Table 2: Comparison of MRI Modalities for ASD Subtyping

Modality	Key Metrics	Strengths	Subtyping Evidence
Structural MRI (sMRI)	Gray matter volume, cortical thickness, gyrification	High spatial resolution, widely available	Distinct gray-white matter boundary contrast patterns in toddlers later diagnosed with ASD [72]
Functional MRI (fMRI)	Functional connectivity, ALFF/fALFF, network properties	Measures brain activity and connectivity	Tensor decomposition reveals different brain community patterns in ASD subtypes [5]
Diffusion MRI (dMRI)	Fractional anisotropy, mean diffusivity, fiber density	Maps white matter microstructure	Two neurosubtypes identified based on structural-functional coupling [2]
Multimodal Integration	Combined structural, functional, and diffusion metrics	Comprehensive characterization of brain organization	Enhanced differentiation of autism, Asperger's, and PDD-NOS subtypes [5]

Evidence for Neurobiological Subtypes

Recent research has identified several promising neurobiological subtypes within ASD:

Structural-Functional Coupling Subtypes: A 2025 study by Qiao et al. identified two distinct neurosubtypes of ASD using combined DTI and fMRI data [2]. Subtype 1 displayed significantly lower fractional anisotropy (FA) in the posterior cingulate cortex compared to neurotypical controls, while Subtype 2 exhibited reduced FA in the anterior cingulate cortex, middle temporal gyrus, parahippocampus, and thalamus [2]. These subtypes also differed clinically, with Subtype 2 showing lower full-scale and performance IQ scores [2].

Gray-White Matter Contrast Trajectories: Studies of gray-white matter contrast (GWC) have revealed distinct developmental trajectories associated with ASD outcomes. In toddlers at high familial risk for ASD, diagnostic outcome at age 3 was associated with widespread increases in GWC between 12 and 24 months of age across cortical regions implicated in social processing and language acquisition [72]. Another study found that a faster rate of GWC decline in ASD, particularly within the motor system, predicted diagnostic outcome with 86% accuracy [71].

Functional and Structural Subtypes: Research comparing traditional ASD subtypes (autism, Asperger's, and PDD-NOS) using multiple neuroimaging features found that alterations in the subcortical network and default mode network primarily distinguished the autism subtype from Asperger's and PDD-NOS [5]. These findings suggest that previously behaviorally defined categories may have distinct neurobiological correlates.

Experimental Approaches and Methodologies

Neuroimaging Protocols for ASD Subtyping

The identification of biologically meaningful ASD subtypes requires standardized, rigorous methodological approaches across multiple imaging modalities:

Multimodal Data Acquisition: Comprehensive subtyping protocols typically include structural T1-weighted imaging, resting-state functional MRI, and diffusion-weighted imaging [2] [73] [5]. For diffusion imaging, advanced techniques such as diffusion kurtosis imaging (DKI) and neurite orientation dispersion and density imaging (NODDI) provide more specific information about microstructural properties than conventional diffusion tensor imaging [73].

Processing and Analysis Pipelines: Surface-based analysis using tools such as CIVET enables precise measurement of gray-white matter contrast by placing surfaces at the gray-white matter boundary and sampling intensities at specific distances from this boundary [71]. Functional connectivity analysis typically involves parcellating the brain into regions of interest and computing correlation matrices between time series [5] [1]. Machine learning approaches, including support vector machines and deep learning models, are increasingly employed to identify multivariate patterns distinguishing subtypes [1].

Longitudinal Designs: Tracking neurodevelopmental trajectories through longitudinal designs provides particularly powerful insights into ASD subtypes, as rate of change often reveals differences not apparent in cross-sectional analyses [72] [71]. These designs require careful attention to interscan intervals, consistency of acquisition parameters, and appropriate modeling of nonlinear developmental patterns.

Integration with Clinical Trials

The application of neuroimaging subtypes to clinical trials involves several methodological considerations:

Stratification Designs: Clinical trials can incorporate neuroimaging biomarkers as stratification factors, randomizing participants within subtypes to ensure balanced allocation of biological subgroups across treatment arms [69].

Target Engagement Metrics: Imaging measures can serve as biomarkers of target engagement, providing early indication of whether an intervention is affecting the intended neural system [69].

Endpoint Development: Imaging measures sensitive to incremental change may serve as intermediate endpoints, potentially providing more sensitive measures of treatment response than behavioral measures alone [73].

Diagram 1: Neuroimaging-Guided Precision Medicine Framework for ASD

Table 3: Research Reagent Solutions for ASD Subtyping Studies

Resource Category	Specific Tools	Application in ASD Research
Neuroimaging Databases	ABIDE I/II [2] [5] [71]	Large-scale aggregated datasets enabling discovery of robust subtypes across sites
Processing Pipelines	CIVET [71], CCS [5], FSL, FreeSurfer	Standardized processing of structural and functional MRI data
Analysis Tools	FSL, DPARSF, Connectome Computation System	Functional connectivity analysis, tensor decomposition, network analysis
Biophysical Models	NODDI [73], WMTI [73]	Advanced microstructure characterization beyond conventional DTI
Classification Algorithms	Support Vector Machines [2] [1], Deep Learning [1]	Multivariate pattern classification for subtyping and prediction
Clinical Instruments	ADOS [71], ADI-R	Standardized behavioral phenotyping for correlation with neurobiological measures

Pathway Forward: Recommendations for Future Trials

Implementing Biomarker-Guided Approaches

To overcome historical translational failures, future ASD clinical trials should incorporate several key strategies:

Precision Medicine Frameworks: Clinical trials should adopt biomarker-stratified designs that assign participants to targeted interventions based on their neurobiological subtype [69]. This approach requires prior identification of robust subtypes with differential predicted response to specific mechanisms of action.

Endpoint Development: There is a critical need for validated biomarkers that can serve as intermediate endpoints in clinical trials, providing earlier readouts of potential efficacy than long-term behavioral measures [69]. Candidates include EEG measures, eye-tracking paradigms, and functional MRI tasks probing specific neural circuits.

Standardization and Reproducibility: The field requires improved standardization of acquisition parameters, processing pipelines, and analytical approaches across sites to enhance reproducibility [73] [1]. This is particularly important for multisite trials, where scanner differences can introduce significant variance.

Integration of Multiple Data Types: Combining neuroimaging with genetic, electrophysiological, and behavioral measures will enable more comprehensive subtyping approaches that capture the multidimensional nature of ASD [69].

Promising Directions

Several emerging approaches offer particular promise for advancing ASD therapeutics:

Circuit-Based Targeting: Rather than focusing exclusively on molecular targets, interventions can be developed to modulate specific neural circuits identified as aberrant in particular ASD subtypes [69].

Developmental Timing Considerations: Interventions may have differential effects depending on developmental stage, suggesting that clinical trials should carefully consider the age range most likely to benefit from a particular mechanism [72] [71].

Multimodal Biomarker Integration: Combining multiple biomarker modalities (e.g., structural and functional MRI with genetic data) may enhance predictive power for treatment response beyond single modalities [5].

Diagram 2: Translational Failure Causes and Neuroimaging Solutions

The historical failure rate in ASD therapeutic development reflects fundamental challenges in bridging the gap between basic neurobiology and clinical application. The exceptional heterogeneity of ASD necessitates a move beyond one-size-fits-all approaches toward precision medicine strategies that account for neurobiological diversity within the spectrum. Emerging research on neuroimaging-based subtypes provides a promising foundation for this transition, offering biologically meaningful stratification approaches that may enhance clinical trial success. Future progress will depend on standardized multimodal assessment, replication of subtype findings across diverse samples, and the implementation of biomarker-stratified trial designs that target interventions to those most likely to benefit. Through these approaches, the field may finally navigate the "valley of death" that has long separated basic ASD research from meaningful clinical breakthroughs.

Autism Spectrum Disorder (ASD) is characterized by significant phenotypic and biological heterogeneity, presenting major challenges for research and therapeutic development. The traditional diagnostic approach, which treats ASD as a single entity, is increasingly recognized as insufficient for capturing the diverse clinical presentations and underlying neurobiological variations. This limitation has stimulated the emergence of data-driven stratification methods that decompose heterogeneity into meaningful subtypes using neuroimaging and behavioral data. Within this field, a central investigation involves comparing stratification approaches based on structural magnetic resonance imaging (sMRI) versus functional MRI (fMRI) to determine their respective utilities in identifying biologically coherent subgroups.

This guide provides a comparative analysis of data-driven stratification methodologies, focusing on their application to ASD subtyping. We evaluate experimental protocols, analytical frameworks, and outcome measures associated with structural and functional neuroimaging approaches, providing researchers with objective performance data to inform methodological selection for precision medicine initiatives.

Experimental Protocols in Data-Driven Subtyping

Structural MRI-Based Subtyping Protocol

Structural approaches typically utilize features such as diffusion tensor imaging (DTI) parameters and gray matter volume (GMV) to identify subtypes based on neuroanatomical differences.

• Data Acquisition: Structural protocols acquire high-resolution T1-weighted images for GMV analysis and DTI sequences for white matter characterization. DTI parameters include fractional anisotropy (FA) and mean diffusivity (MD), which reflect white matter integrity and organization [2].
• Preprocessing: Structural data undergoes spatial normalization, tissue segmentation, and feature extraction. For DTI, tract-based spatial statistics skeletonize white matter tracts for voxel-wise analysis [2].
• Feature Extraction: Key features include voxel-based morphometry maps for GMV, and skeletonized FA/MD maps for white matter microstructure [5].
• Stratification Analysis: Clustering algorithms applied to structural features identify subgroups. Validation involves comparing clinical profiles and cognitive measures between subtypes [2].

Functional MRI-Based Subtyping Protocol

Functional approaches leverage resting-state fMRI to characterize subtypes based on brain network dynamics and connectivity patterns.

• Data Acquisition: Resting-state fMRI data collected during wakeful rest, typically 6-10 minutes. Preprocessing includes head motion correction, band-pass filtering, and global signal regression [6].
• Feature Extraction: Multiple connectivity metrics extracted:
  • Static Functional Connectivity Strength: Pearson correlation between brain region time-series [6].
  • Dynamic Functional Connectivity: Time-varying connectivity assessed using dynamic conditional correlation for strength and variability [6].
• Stratification Analysis: Normative modeling establishes typical developmental trajectories, then individual deviations are clustered. Validation includes eye-tracking and clinical measures [6].

Cognitive Performance-Based Stratification

Some approaches stratify based on behavioral performance, such as the Reading the Mind in the Eyes Test, using item-level response patterns rather than neuroimaging data [74].

Table 1: Comparative Overview of Data-Driven Stratification Methodologies

Methodological Component	Structural MRI Approach	Functional MRI Approach	Cognitive Performance Approach
Primary Data Source	DTI & T1-weighted MRI	Resting-state fMRI	Behavioral task performance
Key Features	FA, MD, GMV	Static & dynamic functional connectivity	Item-level response patterns
Analytical Technique	Voxel-wise statistical analysis, clustering	Normative modeling, deviation mapping, clustering	Hierarchical clustering
Sample Sizes	92 ASD, 65 controls [2]	479 ASD, 567 TD (discovery); 21 ASD (validation) [6]	694 (discovery), 249 (replication) [74]
Identified Subtypes	2 neurosubtypes [2]	2 neural subtypes [6]	5 ASC subgroups [74]

Comparative Performance of Stratification Approaches

Subtype Differentiation Capabilities

Different methodological approaches yield varying subtype solutions with distinct neurobiological and clinical profiles.

Table 2: Subtype Characteristics Identified by Different Methodological Approaches

Stratification Method	Identified Subtypes	Distinguishing Features	Clinical Correlations
Structure-Function Coupling	Subtype 1	Lower FA in posterior cingulate cortex [2]	No significant FIQ/PIQ differences from controls [2]
	Subtype 2	Reduced FA in anterior cingulate, temporal regions; higher MD [2]	Lower FIQ and PIQ scores [2]
Multilevel Functional Connectivity	Subtype 1	Positive deviations: occipital, cerebellar networks; Negative deviations: frontoparietal, DMN [6]	Similar clinical symptoms but distinct gaze patterns [6]
	Subtype 2	Inverse pattern of Subtype 1 [6]	Different social attention profiles [6]
Mentalizing Heterogeneity	3 impaired subgroups (45-62%)	Large performance impairments (d = -1.03 to -11.21) [74]	Varying social communication abilities
	2 unimpaired subgroups	Performance within typical range [74]	Compensatory mechanisms likely

Predictive Accuracy and Clinical Utility

• Diagnostic Prediction: Structural-functional coupling approaches demonstrated enhanced diagnostic prediction accuracy when distinguishing between subtypes compared to general ASD classification [2].
• Treatment Response Prediction: Functional connectivity subtypes show promise for predicting intervention response, with one study reporting 61.5% response rate to intranasal oxytocin in one subtype versus 13.3% in another [6].
• Cognitive Heterogeneity: Cognitive stratification revealed that 45-62% of ASC individuals show significant mentalizing impairments, while others perform within normal ranges [74].

Research Reagent Solutions

Table 3: Essential Research Materials and Analytical Tools for ASD Stratification Studies

Research Tool	Specifications	Primary Function	Example Applications
ABIDE Database	Multi-site dataset; 1,112 participants (ABIDE I) [6]	Provides standardized neuroimaging and phenotypic data	Data source for discovery and validation cohorts [2] [6]
fMRIPrep	Version 20.2.1 [6]	Standardized fMRI preprocessing pipeline	Image normalization, motion correction, quality control [6]
DTI Analysis	Tract-based spatial statistics	Voxel-wise white matter analysis	Skeleton projection of FA/MD metrics [2]
Normative Modeling	RsfMRI from TD participants (N=567) [6]	Establishes expected developmental trajectories	Quantifies individual deviations in ASD patients [6]
Dynamic Conditional Correlation	Applied to BOLD time-series [6]	Measures instant dynamic functional connectivity	Calculates connectivity strength and variability [6]

Integration Pathways for Multimodal Subtyping

The following diagram illustrates the conceptual relationship between different data modalities and analytical approaches in ASD stratification research:

Figure 1: Integrated Workflow for Multimodal ASD Subtyping

The comparative analysis of structural and functional MRI approaches to ASD stratification reveals complementary strengths. Structural methods provide stable neuroanatomical biomarkers with direct pathophysiological interpretations, while functional approaches capture dynamic network properties with strong links to cognitive processes and treatment response. Cognitive performance-based stratification offers practical assessment without specialized equipment.

The emerging consensus indicates that multimodal integration of structural, functional, and behavioral data holds the greatest promise for delineating clinically meaningful subtypes. Future research directions should prioritize longitudinal designs to establish subtype stability, and validate stratification schemes against treatment outcomes to fulfill the promise of precision medicine in autism.

The pursuit of biological markers for autism spectrum disorder (ASD) represents a fundamental shift from behavior-based diagnostics toward precision medicine. ASD is a prevalent neurodevelopmental condition, currently affecting an estimated 1 in 36 children according to recent data from the Centers for Disease Control and Prevention [75]. This complex disorder is characterized by significant heterogeneity in clinical presentation, underlying biology, and developmental trajectories, creating a substantial "biomarker gap" between behavioral observations and objective biological measures. The traditional reliance on behavioral assessments—while crucial for diagnosis—introduces subjectivity and variability that complicate early intervention, prognosis prediction, and treatment development [75] [19].

Current biomarker research spans multiple biological domains, investigating genetic vulnerabilities, neuroimaging signatures, metabolomic profiles, and immune system alterations [75] [76]. Among these, neuroimaging approaches—particularly structural and functional MRI—have emerged as powerful tools for delineating the neurobiological underpinnings of ASD. Functional magnetic resonance imaging (fMRI) enables the investigation of functional connectivity (FC) patterns across distributed brain networks, while structural MRI (sMRI) characterizes alterations in brain anatomy, such as gray matter volume (GMV) and cortical thickness [19] [77]. These complementary approaches offer unprecedented insights into the brain organization of autistic individuals, potentially revealing subtypes with distinct clinical profiles and treatment needs [6] [7].

The integration of large-scale datasets and advanced computational methods is accelerating progress in this field. Initiatives such as the Autism Brain Imaging Data Exchange (ABIDE) have consolidated neuroimaging data from thousands of participants across international sites, enabling robust analyses of brain patterns in ASD [19] [6]. Concurrently, advances in data analysis—including normative modeling, tensor decomposition, and clustering algorithms—are refining our ability to identify biologically meaningful subgroups within the autism spectrum [19] [6] [78]. As research transitions from case-control comparisons to individualized approaches, the potential emerges for biomarkers that not only aid diagnosis but also predict outcomes and guide targeted interventions.

Neuroimaging Approaches: Structural versus Functional MRI

Structural and functional MRI provide complementary yet distinct windows into the autistic brain. Structural MRI (sMRI) captures the brain's anatomical architecture, including gray matter volume (GMV), cortical thickness, and white matter integrity. These measures reflect the brain's physical infrastructure, which undergoes atypical development in ASD. Studies have revealed atypical development in GMV and gray matter density in autistic individuals, with age-associated aberrant changes in auditory, visual, and fronto-parietal networks [19]. These structural differences potentially underlie the cognitive and behavioral characteristics associated with ASD.

In contrast, functional MRI (fMRI) measures brain activity, typically by detecting changes in blood oxygenation (BOLD signal). Resting-state fMRI (rsfMRI) examines spontaneous neural activity while the brain is at rest, revealing patterns of functional connectivity (FC) between different regions. Two primary analytical approaches dominate fMRI research: static functional connectivity, which measures the average temporal correlation between brain regions over an entire scan session, and dynamic functional connectivity, which captures how these correlations fluctuate over time [6]. These functional measures provide insights into the brain's operational state, revealing both stable and time-varying properties of large-scale networks.

The comparative strengths of sMRI and fMRI for identifying ASD subtypes are outlined in the table below:

Table 1: Comparison of Structural and Functional MRI for ASD Subtyping

Feature	Structural MRI (sMRI)	Functional MRI (fMRI)
Primary Measure	Gray matter volume, cortical thickness, brain structure	Blood oxygenation level-dependent (BOLD) signal, functional connectivity
Temporal Resolution	Static snapshot	High (captures dynamics over seconds)
Key Findings in ASD	Atypical GMV development; age-associated aberrant changes in specific networks	Altered within- and between-network connectivity; compressed primary gradient of functional organization
Subtyping Utility	Identifies structural subtypes with potential clinical relevance [19]	Reveals functional subtypes with distinct clinical profiles and treatment responses [6]
Analytical Methods	Voxel-based morphometry, surface-based analysis	Seed-based correlation, independent component analysis (ICA), graph theory
Relationship to Behavior	Moderate; structural differences may reflect long-term neurodevelopmental patterns	Potentially stronger; functional states may more directly relate to current symptoms

Research increasingly suggests that functional connectivity measures may offer superior sensitivity for capturing the neural correlates of ASD heterogeneity. A foundational study examining functional connectivity subtypes found they associate robustly with ASD diagnosis and generalize across independent datasets [78]. Interestingly, these FC subtypes converged on a common topography across different networks, consistent with a compression of the primary gradient of functional brain organization previously reported in ASD literature [78]. This suggests that despite the heterogeneous clinical presentation, there may be consistent alterations in the fundamental organization of functional brain networks in autism.

Methodological Frameworks for ASD Subtyping

Data-Driven Clustering Approaches

Unsupervised machine learning techniques, particularly clustering algorithms, have become instrumental for identifying neurobiological subtypes in ASD without a priori diagnostic categories. Hierarchical agglomerative clustering has been widely applied to functional connectivity data, successfully decomposing ASD heterogeneity into reproducible subtypes [78]. This approach typically begins with computing spatial dissimilarity matrices between individual functional connectivity maps, followed by clustering individuals based on similarity patterns. Studies applying this method have demonstrated that 97% of individuals can be reliably assigned to functional connectivity subtypes across multiple brain networks, with these subtypes showing moderate associations with clinical ASD diagnosis [78].

The robustness of subtype assignments varies between discrete (categorical) and continuous approaches. Research indicates that continuous assignments—computed as spatial correlation between FC subtype maps and individual connectivity patterns—demonstrate superior reliability compared to discrete categorical assignments [78]. This finding has important implications for biomarker development, suggesting that dimensional approaches may better capture the complex neurobiological spectra underlying autism than rigid categorical boundaries. Furthermore, these data-driven FC subtypes have shown generalizability across independent datasets, strengthening their potential as valid biomarkers [78].

Normative Modeling and Heterogeneity Mapping

Normative modeling has emerged as a powerful statistical framework for quantifying individual differences in brain organization. This approach characterizes typical neurodevelopmental trajectories across a reference population (typically neurotypical individuals), then measures how each autistic individual deviates from these expected patterns [6]. Unlike case-control comparisons that assume group homogeneity, normative models explicitly account for heterogeneity by identifying both hyper-connected and hypo-connected subtypes within ASD populations.

A recent comprehensive study applied this approach to 1,046 participants (479 with ASD, 567 typical development) from the ABIDE I and II datasets, incorporating both static and dynamic functional connectivity features within normative models [6]. The research identified two distinct neural ASD subtypes with unique functional brain network profiles despite comparable clinical presentations: one subtype characterized by positive deviations in the occipital and cerebellar networks coupled with negative deviations in frontoparietal, default mode, and cingulo-opercular networks, while the other subtype exhibited the inverse pattern [6]. These neural subtypes were additionally associated with different gaze patterns in eye-tracking tasks, confirming their behavioral relevance.

Hybrid and Multimodal Integration Methods

Hybrid approaches that integrate prior biological knowledge with data-driven refinement offer a promising middle ground between fully theory-driven and purely exploratory methods. The NeuroMark pipeline, for example, employs spatially constrained independent component analysis (ICA) that incorporates templates derived from large datasets to identify replicable components, which are then used as spatial priors for single-subject analysis [79]. This hybrid approach maintains correspondence across individuals while capturing individual-specific variability, addressing a key limitation of both predefined atlases and fully data-driven approaches [79].

Functional decompositions can be systematically categorized along three key attributes: source (anatomic, functional, multimodal), mode (categorical, dimensional), and fit (predefined, data-driven, hybrid) [79]. Hybrid models like NeuroMark are functionally defined, dimensional, and data-driven, offering advantages for capturing individual variability and spatial dynamics of brain networks [79]. The integration of multiple data modalities—including structural MRI, functional MRI, eye-tracking, and genetic measures—further enhances our ability to identify clinically meaningful subtypes with distinct biological foundations [6] [7].

Key Experimental Protocols and Findings

Functional Connectivity Subtyping Protocol

A robust protocol for identifying functional connectivity subtypes involves multiple standardized steps, as demonstrated in a study that systematically explored 18 different brain networks [78]:

• Data Acquisition and Preprocessing: Resting-state fMRI data are acquired following standardized protocols (e.g., ABIDE consortium parameters). Preprocessing typically includes motion correction, slice-timing correction, normalization to standard space (e.g., MNI152), and band-pass filtering (0.01-0.1 Hz).

• Functional Connectivity Calculation: Seed-based functional connectivity is computed between predefined regions of interest (e.g., using the MIST_20 parcellation with 18 non-cerebellar seed networks) and all other brain voxels.

• Confound Regression: Covariates of non-interest (recording site, age, head motion) are regressed out from individual seed-FC maps to minimize non-neural influences.

• Dissimilarity Matrix Construction: Spatial dissimilarity between individual seed-FC maps is computed as 1 minus their spatial correlation.

• Hierarchical Clustering: Agglomerative hierarchical clustering is performed on the dissimilarity matrix with criteria for subtype inclusion (average spatial dissimilarity <1 and minimum of 20 individuals per subtype).

This protocol has demonstrated exceptional robustness, with FC subtypes not driven by confounds and stable across hyper-parameter choices [78].

Large-Scale Genetic Subtyping Framework

A groundbreaking study published in 2025 analyzed data from over 5,000 children in the SPARK autism cohort, employing a "person-centered" computational model that considered over 230 traits per individual [7]. The methodological workflow included:

• Multidimensional Phenotyping: Comprehensive assessment of social interactions, repetitive behaviors, developmental milestones, and co-occurring psychiatric conditions.

• Computational Clustering: Application of a data-driven clustering algorithm to identify naturally occurring groupings across the phenotypic landscape.

• Genetic Association Analysis: Linking identified subtypes to distinct genetic profiles, including damaging de novo mutations and rare inherited variants.

• Biological Pathway Mapping: Identification of divergent biological processes affected in each subtype and the developmental timing of genetic disruptions.

This approach revealed four clinically and biologically distinct subtypes of autism, each with distinct developmental trajectories, genetic architectures, and clinical presentations [7].

Table 2: Characteristics of Autism Subtypes Identified in Large-Scale Genetic Study

Subtype	Prevalence	Clinical Features	Genetic Profile
Social and Behavioral Challenges	37%	Core autism traits with typical developmental milestones; frequent co-occurring conditions (ADHD, anxiety, depression)	Mutations in genes active later in childhood
Mixed ASD with Developmental Delay	19%	Developmental delays (walking, talking) without anxiety/depression; variable repetitive behaviors and social challenges	Elevated rare inherited genetic variants
Moderate Challenges	34%	Milder core autism behaviors; typical developmental milestones; minimal co-occurring psychiatric conditions	Not specified
Broadly Affected	10%	Severe, wide-ranging challenges including developmental delays, social-communication difficulties, and co-occurring psychiatric conditions	Highest proportion of damaging de novo mutations

Signaling Pathways and Experimental Workflows

The integration of multiple data modalities and analytical techniques requires sophisticated workflows that can capture complex relationships between biological mechanisms and clinical manifestations. The following diagram illustrates a comprehensive framework for identifying autism subtypes through multi-modal data integration:

Diagram 1: Multi-Modal Framework for ASD Subtype Identification (Title: ASD Subtyping Framework)

The functional connectivity analytical process involves specific steps for quantifying both static and dynamic properties of brain networks, as illustrated in the following workflow:

Diagram 2: Functional Connectivity Analysis Workflow (Title: FC Analysis Pipeline)

Table 3: Research Reagent Solutions for Autism Biomarker Studies

Resource	Type	Function	Example Use Cases
ABIDE I & II Datasets	Data Repository	Consolidated resting-state fMRI, anatomical, and phenotypic data from multiple international sites	Large-scale analyses of functional connectivity; normative model development; subtype validation [19] [6]
fMRIPrep	Software Pipeline	Standardized fMRI data preprocessing	Ensuring reproducible processing across studies; minimizing analytical variability [6]
COMBAT-GAM	Harmonization Tool	Removing site effects in multi-center studies	Enabling combined analysis of datasets from different scanners and protocols [77]
NeuroMark Pipeline	Analytical Tool	Hybrid ICA-based functional network identification	Extracting individual-specific functional networks while maintaining cross-subject correspondence [79]
MIST Parcellation	Brain Atlas	Defining regions of interest for connectivity analysis	Standardized seed-based connectivity analysis; network definition [78]
Generalized Additive Models (GAMs)	Statistical Tool	Flexible nonlinear modeling of developmental trajectories	Capturing complex age-related changes in functional connectivity without predefined shapes [77]
Tobii Eye-Tracking Systems	Behavioral Instrument	Quantifying gaze patterns and social attention	Linking neural subtypes to behavioral measures of social cognition [6]

The quest for objective diagnostic and predictive markers in autism is transitioning from a theoretical possibility to a tangible reality. Research over the past decade has demonstrated that neurobiologically distinct subtypes of autism exist, with distinct functional brain network profiles, genetic architectures, and developmental trajectories [6] [7] [78]. The convergence of large-scale datasets, advanced analytical methods, and multimodal integration is rapidly bridging the biomarker gap that has long separated behavioral observations from biological mechanisms.

Functional MRI has emerged as a particularly powerful tool for delineating ASD heterogeneity, with functional connectivity measures demonstrating robust associations with diagnosis and potential for predicting treatment response [6] [78]. The identification of subtypes such as those characterized by divergent connectivity in the default mode, frontoparietal, and somatomotor networks provides a neurobiological foundation for understanding the diverse clinical presentations of autism [6]. Furthermore, the connection between these neural subtypes and differences in eye-gaze patterns during social tasks creates an important bridge between biology and behavior [6].

The translation of these findings into clinical practice faces several challenges, including the reliability of biomarker measurements, the heterogeneity of ASD, and the need for large-scale validation studies [76]. However, the accelerating pace of discovery—exemplified by recent studies identifying genetically distinct subtypes with different clinical trajectories—suggests that precision medicine approaches for autism are within reach [7]. As these biomarkers mature, they hold the potential to revolutionize autism care by enabling earlier detection, prognostication of outcomes, and personalized interventions tailored to an individual's specific neurobiological profile.

The future of autism biomarker research lies in continued methodological refinement, cross-disciplinary collaboration, and the development of integrative frameworks that incorporate genetic, neuroimaging, behavioral, and clinical data. Initiatives such as Princeton Precision Health, which combines artificial intelligence and computational modeling to integrate across biological and clinical domains, represent the vanguard of this approach [7]. By embracing these sophisticated methodologies and maintaining focus on clinically meaningful endpoints, the field moves closer to delivering on the promise of precision medicine for autistic individuals.

Autism Spectrum Disorder (ASD) is characterized by significant neurobiological heterogeneity, complicating the identification of consistent biomarkers and effective interventions. Neuroimaging research has increasingly shifted from traditional case-control designs toward data-driven subtyping approaches to parse this heterogeneity. This comparison guide evaluates computational pipelines for identifying autism subtypes using structural MRI (sMRI) and functional MRI (fMRI), with particular emphasis on their integration with Explainable AI (XAI) frameworks to enhance clinical interpretability. The fundamental challenge in this domain lies in reconciling the high predictive performance of deep learning models with the clinical necessity for transparent, interpretable decision-making processes that researchers and clinicians can trust and utilize effectively.

Advanced machine learning techniques now enable the identification of neurobiological subtypes that may transcend conventional diagnostic boundaries, offering potential pathways toward personalized intervention strategies. However, the selection of optimal computational pipelines depends on multiple factors including data modality, sample characteristics, and clinical objectives. This guide provides an objective comparison of current methodologies, their performance metrics, and implementation requirements to inform researchers' pipeline selection decisions.

Comparative Performance Analysis of Computational Approaches

Table 1: Performance Metrics of Featured Computational Pipelines

Pipeline Approach	Data Modality	Sample Size	Key Algorithm(s)	Reported Accuracy/Performance	Primary Application
XAI Diagnostic Framework [80]	Behavioral Questionnaires	Not specified	TabPFNMix + SHAP	91.5% accuracy, 92.7% recall, 94.3% AUC-ROC	ASD screening and diagnosis
Stacked Ensemble Model [81]	Behavioral Questionnaires	Multiple datasets	RF + ET + CB + ANN	96.96%-99.89% accuracy across age groups	ASD screening across developmental stages
HYDRA Clustering [82]	Structural MRI	4,115 participants	HYDRA + population modeling	Distinct subgroups identified	Neuroanatomical subtyping in ASD/ADHD
Functional Subtyping [6]	Resting-state fMRI	1,046 participants	Normative modeling + clustering	Two neural subtypes identified	Functional connectivity subtyping in ASD
Multimodal Fusion [35]	sMRI + fMRI	207 participants	Similarity Network Fusion (SNF)	Two subtypes with opposite GMV changes	Multimodal subtyping in male children with ASD

Table 2: Clinical and Technical Implementation Considerations

Pipeline Approach	Clinical Correlations	Technical Complexity	Computational Demand	Interpretability Strength
XAI Diagnostic Framework [80]	Social responsiveness, repetitive behaviors, parental age	Moderate	Moderate	High (explicit feature importance via SHAP)
Stacked Ensemble Model [81]	Cross-developmental stage applicability	High	High	High (SHAP + uncertainty quantification)
HYDRA Clustering [82]	Limited clinical differentiation observed	High	High	Moderate (subgroup identification without clinical correlation)
Functional Subtyping [6]	Distinct gaze patterns in social tasks	Moderate-High	Moderate	Moderate (neural subtypes linked to behavior)
Multimodal Fusion [35]	ALFF predicted social impairment in one subtype	High	High	Moderate (biological heterogeneity captured)

Experimental Protocols and Methodologies

Structural MRI Subtyping with Population Modeling

The HYDRA (HeterogeneitY through DiscRiminative Analysis) pipeline represents a semi-supervised machine learning approach that clusters individuals based on neuroanatomical differences relative to a control sample [82]. The methodology employs population modeling to generate centile scores for cortical thickness, surface area, and grey matter volume, analogous to pediatric growth charts. These centile scores capture an individual's structural variation relative to same-sex and same-age controls in a common reference space, facilitating multi-site comparisons.

Key Experimental Protocol:

• Data Processing: T1-weighted MRI scans processed through FreeSurfer v6.0.1 with Desikan-Killiany atlas parcellation
• Population Modeling: Centile scores derived using Generalised Additive Models of Location Scale and Shape (GAMLSS) based on Brain Chart models
• Clustering Algorithm: HYDRA implementation compared against traditional k-medoids with UMAP dimensionality reduction
• Validation Approach: Analysis of both within-diagnosis (ASD alone) and transdiagnostic (ASD+ADHD) subgroups

This approach identified distinct subgroups with often opposite neuroanatomical alterations relative to controls, characterized by different combinations of increased or decreased morphometric patterns [82]. However, the number of subgroups and their membership varied considerably based on feature selection and algorithm choice, highlighting the methodological sensitivity of subtyping approaches.

Functional MRI Subtyping through Normative Modeling

The functional subtyping pipeline employs normative modeling of both static and dynamic functional connectivity to identify individuals with extreme deviations from typical developmental trajectories [6]. This approach characterizes multilevel functional connectivity using Pearson correlation for static functional connectivity strength (SFCS) and dynamic conditional correlation (DCC) for instant dynamic functional connectivity assessment (DFCS and DFCV).

Key Experimental Protocol:

• Data Acquisition: Resting-state fMRI from ABIDE-I and ABIDE-II datasets (1,046 participants)
• Feature Extraction: Multilevel functional connectivity features from Dosenbach 160 ROIs
• Normative Modeling: Individual-level deviation quantification from typical development trajectories
• Validation: Independent cohort with eye-tracking tasks for social attention assessment
• Clinical Correlation: Analysis of gaze patterns in face emotion processing and joint attention tasks

This methodology identified two distinct neural ASD subtypes with unique functional brain network profiles despite comparable clinical presentations [6]. One subtype showed positive deviations in the occipital and cerebellar networks with negative deviations in frontoparietal, default mode, and cingulo-opercular networks, while the other exhibited the inverse pattern. These neural subtypes were associated with distinct gaze patterns during social tasks, providing behavioral validation.

Multimodal Fusion with Similarity Network Fusion

The Similarity Network Fusion (SNF) approach integrates structural and functional MRI features to capture complementary aspects of neurobiological heterogeneity [35]. This method constructs separate structural and functional distance networks then fuses them to create a comprehensive representation of neurobiological similarity between participants.

Key Experimental Protocol:

• Feature Extraction: Gray matter volume (GMV) for structure, amplitude of low-frequency fluctuation (ALFF) for function
• Network Construction: Separate structural and functional distance networks constructed then fused via SNF
• Clustering: Spectral clustering applied to the fused network
• Clinical Correlation: Multivariate support vector regression to investigate subtype-symptom relationships

This multimodal approach identified two ASD subtypes with opposite GMV changes and distinct ALFF alterations [35]. Furthermore, ALFF alterations predicted the severity of social communication impairments in one subtype but not the other, demonstrating the clinical relevance of the subtyping approach and highlighting differential structure-function relationships across subtypes.

XAI-Enhanced Diagnostic Frameworks

Explainable AI frameworks integrate high-performance classifiers with interpretation modules to make transparent predictions. The TabPFNMix model exemplifies this approach, combining advanced classification with SHAP (SHapley Additive exPlanations) for interpretability [80].

Key Experimental Protocol:

• Data Preparation: Addresses class imbalance using Safe-Level SMOTE
• Dimensionality Reduction: Employs Principal Component Analysis (PCA)
• Feature Selection: Uses Mutual Information and Pearson correlation
• Model Architecture: TabPFNMix regressor optimized for structured medical data
• Interpretation Module: SHAP analysis for feature importance visualization
• Uncertainty Quantification: Monte Carlo Dropout for confidence estimation

This approach achieved 91.5% accuracy while identifying social responsiveness scores, repetitive behavior scales, and parental age at birth as the most influential factors in ASD diagnosis [80]. The integration of XAI provides transparent reasoning behind model decisions, building trust with clinicians and caregivers.

Visualization of Computational Workflows

Diagram 1: Comprehensive Computational Pipeline for Autism Subtyping

Research Reagent Solutions: Essential Computational Tools

Table 3: Key Computational Tools and Frameworks

Tool/Framework	Primary Function	Application Context	Implementation Considerations
FreeSurfer [82]	Automated cortical reconstruction & subcortical segmentation	Structural feature extraction	Standardized processing pipeline for multi-site data
ICA-AROMA [83]	Automatic removal of motion artifacts from fMRI data	Functional connectivity analysis	Superior motion correction for ASD populations
SHAP [80] [81]	Model interpretation and feature importance	Explainable AI implementations	Compatible with various ML models; provides intuitive visualizations
HYDRA [82]	Semi-supervised clustering based on control differences	Neuroanatomical subtyping	Requires representative control group for reference
Similarity Network Fusion [35]	Integrates multiple data types into fused network	Multimodal data integration	Effectively combines structural and functional features
TabPFNMix [80]	Tabular data classification	Questionnaire-based screening	Specifically optimized for structured medical data
Safe-Level SMOTE [81]	Handles class imbalance in medical datasets	Data preprocessing	Critical for ASD detection with imbalanced datasets

Discussion: Integration Pathways and Clinical Translation

The comparative analysis reveals distinctive strengths across computational approaches. Structural MRI pipelines with population modeling effectively capture neuroanatomical heterogeneity but show limited clinical correlations [82]. Functional approaches identify subtypes with meaningful behavioral correlates, particularly in social domains [6], while multimodal fusion leverages complementary information for enhanced subtyping [35]. XAI frameworks bridge the critical gap between predictive accuracy and clinical interpretability [80] [81].

For research and drug development applications, pipeline selection should align with specific objectives. Structural approaches may better inform neurobiological mechanisms, while functional and multimodal subtyping offers stronger behavioral correlations. XAI-enhanced screening tools provide scalable assessment with transparent decision-making, particularly valuable for early detection and clinical trials stratification.

Future development should prioritize: (1) standardized preprocessing protocols to enhance reproducibility, (2) unified frameworks for multimodal data integration, and (3) longitudinal validation of subtype stability and treatment response predictions. The integration of neuroimaging with genetic and behavioral data through explainable computational pipelines represents the most promising pathway toward personalized interventions in autism.

The identification of autism spectrum disorder (ASD) subtypes through neuroimaging is a cornerstone of modern psychiatry and neurology, promising to unravel the condition's profound biological heterogeneity. Research increasingly leverages both structural magnetic resonance imaging (sMRI) and functional MRI (fMRI) to delineate neurologically distinct subgroups, aiming to pave the way for personalized diagnostics and interventions [2] [5]. However, this endeavor is fraught with methodological challenges that threaten the validity and generalizability of findings. Key among these are site effects originating from multi-site data acquisition, motion artifacts that corrupt data integrity, and the overarching crisis of reproducibility. This guide objectively compares the performance of different methodological approaches designed to mitigate these pitfalls, providing researchers with a critical framework for evaluating and selecting the most robust strategies for ASD subtyping studies.

Comparative Analysis of Structural and Functional Neuroimaging for ASD Subtyping

The pursuit of ASD subtypes has employed diverse analytical approaches applied to structural and functional neuroimaging data. The table below summarizes the core methodologies, their findings, and a direct comparison of their performance in identifying biologically distinct subgroups.

Table 1: Comparison of ASD Subtyping Studies Based on Structural and Functional Neuroimaging

Study Focus	Data Modality & Source	Core Analytical Method	Identified Subtypes	Key Neurobiological Findings	Performance & Limitations
Function-Structure Coupling [2]	DTI & fMRI (ABIDE, 4 sites)	Semi-supervised clustering on white matter low-frequency oscillations	Subtype 1 & Subtype 2	Subtype 1: Lower FA in Posterior Cingulate Cortex. Subtype 2: Lower FA in Anterior Cingulate Cortex, thalamus; higher MD in temporal areas; lower FIQ/PIQ. [2]	Enhanced diagnostic prediction accuracy over general ASD classification. Limited by a focus on white matter coupling.
Traditional Subtype Comparison [5]	fMRI & sMRI (ABIDE I)	Tensor decomposition, ALFF/fALFF, and Gray Matter Volume analysis	Autism, Asperger's, PDD-NOS	Major differences in subcortical network and default mode network for autism vs. Asperger's/PDD-NOS. [5]	Provides systematic evidence for subtype discrimination. Relies on clinically defined labels, which may not reflect neurobiology.

The synthesis of these studies indicates that data-driven, biologically anchored approaches (e.g., clustering on neuroimaging features) can reveal subtypes that enhance diagnostic models [2]. In contrast, methods that rely on pre-defined clinical diagnoses, while valuable, may be constrained by the subjectivity of those original labels [5]. The most promising findings consistently implicate specific large-scale networks and structural pathways, such as the default mode network and white matter integrity in cingulate and temporal regions, as key differentiators between ASD subgroups [2] [5].

Experimental Protocols for Key Methodologies

Protocol for Function-Structure Coupling and Clustering

This protocol outlines the methodology for identifying ASD subtypes through the integration of diffusion tensor imaging (DTI) and functional MRI data [2].

• Data Acquisition and Sources:

  • Acquire neuroimaging data from a multi-site repository such as the Autism Brain Imaging Data Exchange (ABIDE). The study by Qiao et al. utilized data from 92 individuals with ASD and 65 neurotypical controls across four sites [2].
  • Collect both Diffusion Tensor Imaging (DTI), which measures white matter integrity through metrics like Fractional Anisotropy (FA) and Mean Diffusivity (MD), and resting-state functional MRI (fMRI) data [2].

• Data Preprocessing:

  • Process structural and functional data through a standardized pipeline (e.g., the Connectome Computation System) [5].
  • For fMRI, steps include slice-time correction, motion realignment, band-pass filtering (e.g., 0.01–0.1 Hz), and global signal regression [5].
  • For DTI, perform eddy current correction and tensor fitting to derive FA and MD maps.

• Feature Integration:

  • Integrate structural and functional data through a skeleton-based white matter functional analysis.
  • Project the fMRI signals onto a white matter skeleton to achieve voxel-wise function-structure coupling.
  • Use white matter low-frequency oscillations (LFOs) as the primary input features for the clustering algorithm [2].

• Subtype Identification:

  • Employ a semi-supervised clustering algorithm on the integrated LFO features to categorize individuals with ASD into distinct neurological subgroups [2].
  • Ensure the number of clusters is validated using appropriate stability measures.

• Validation and Statistical Analysis:

  • Perform statistical analyses (e.g., ANCOVA) to identify significant differences in FA, MD, and clinical measures (e.g., FIQ, PIQ) between the identified subtypes and the neurotypical control group [2].
  • Evaluate the clinical relevance of the subtypes by testing if a support vector machine (SVM) classifier using these subgroups can enhance diagnostic prediction accuracy for ASD compared to a generic ASD model [2].

Protocol for Motion Artifact Simulation and Correction

This protocol details a method for evaluating the efficacy of deep learning models in correcting for motion artifacts in MRI, a critical step for ensuring data quality [84].

• Data Acquisition:

  • Acquire T2-weighted axial head MRI images from healthy volunteers using a fast spin-echo sequence. A sample study used a 1.5T scanner with parameters: TR = 3000 ms, TE = 90 ms, matrix size = 256 × 256, and 24 slices per patient [84].

• Generation of Simulated Motion Artifacts:

  • Create a library of motion-corrupted images from the original, clean images.
  • Apply a series of rigid-body transformations (translations and rotations) to the original images. For example, shift images by ±10 pixels in horizontal, vertical, and diagonal directions, and rotate by ±5° [84].
  • Use the Fourier transform to convert these transformed images into k-space data.
  • To simulate motion during a scan, randomly select sequential phase-encoding lines from the k-space of the transformed images and composite them into a new, motion-corrupted k-space dataset.
  • Apply the inverse Fourier transform to generate the simulated motion artifact image. Generate a large dataset (e.g., 5,500 images per artifact direction) for model training [84].

• Deep Learning Model Training:

  • Use a conditional generative adversarial network (CGAN) architecture, which includes a generator and a discriminator network [84].
  • Train the model using paired datasets: the simulated motion-corrupted images as input and the corresponding original, clean images as the ground truth.
  • Optionally, train separate models for artifacts in specific phase-encoding directions (horizontal vs. vertical) or a combined model for robustness [84].
  • For comparison, also train other models like U-net and Autoencoder (AE) models on the same dataset [84].

• Model Evaluation:

  • Evaluate the quality of the motion-corrected images produced by the models using quantitative metrics by comparing them to the original, clean images.
  • Use Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) as primary metrics [84].
  • The model with the highest SSIM and PSNR scores, and which produces the most visually authentic corrections, is deemed most effective.

Visualization of Experimental Workflows

Workflow for ASD Subtype Identification

The following diagram illustrates the integrated analytical pipeline for identifying autism spectrum disorder subtypes from multi-modal neuroimaging data.

ASD Subtyping Analytical Pipeline

Workflow for Motion Artifact Correction

The diagram below outlines the process of using deep learning, specifically a Conditional Generative Adversarial Network (CGAN), to reduce motion artifacts in MRI images.

MRI Motion Artifact Correction with CGAN

The Scientist's Toolkit: Essential Research Reagents and Materials

For researchers designing studies on ASD subtyping, especially those addressing methodological pitfalls, the following tools and resources are essential.

Table 2: Key Research Reagents and Materials for ASD Neuroimaging Studies

Item Name	Function/Application	Relevance to Pitfalls
ABIDE Database	A publicly available repository of brain imaging data from individuals with ASD and typical controls.	Mitigates site effects by providing a large, multi-site dataset that allows for testing and correcting for scanner-related variance. [2] [5]
Prospective Motion Correction (PMC)	An external tracking system (e.g., optical camera) that dynamically updates the MRI field of view in real-time to match head movement. [85]	Directly reduces motion artifacts at the source during data acquisition, improving data integrity. [85]
Conditional GAN (CGAN)	A deep learning model architecture for image-to-image translation, trained to map motion-corrupted images to clean versions. [84]	Provides a powerful post-processing tool for motion artifact reduction, often outperforming traditional methods and other network models like U-Net. [84]
Tensor Decomposition	A computational method for extracting compressed features and brain community patterns from high-dimensional fMRI data. [5]	Aids in managing data complexity and heterogeneity, enhancing the reproducibility of findings by revealing robust, data-driven patterns.
Function-Structure Coupling	An analytical approach that projects fMRI signals onto a white matter skeleton to integrate multimodal information. [2]	Addresses biological heterogeneity by creating more informative biomarkers for subtyping, potentially leading to more reproducible subgroup classifications. [2]

From Clusters to Clinic: Validating and Comparing ASD Subtypes

Autism Spectrum Disorder (ASD) is characterized by significant clinical and biological heterogeneity, which has long hampered the development of targeted diagnostics and treatments. Historically, research often treated autism as a single disorder, a approach that masked the distinct biological pathways underlying its varied manifestations. Groundbreaking research is now fundamentally reshaping this understanding by integrating large-scale phenotypic and genetic data to identify biologically distinct ASD subtypes [7] [12]. This paradigm shift from a trait-centered to a person-centered approach—considering the full spectrum of an individual's traits rather than searching for genetic links to single characteristics—has enabled the discovery of clinically relevant subtypes with distinct genetic profiles and developmental trajectories [7] [13]. These advances, powered by computational analyses of large cohorts like the SPARK study, are revealing the "multiple distinct narratives" of autism's biology, moving the field beyond the former "jigsaw puzzle" of mixed genetic signals [7]. This review synthesizes the latest findings on ASD subtypes defined through neuroimaging and genetic approaches, providing a comparative analysis of their underlying biological narratives.

Phenotype-Driven Subtype Discovery: A Data-Driven Framework

The Person-Centered Analytical Approach

The identification of robust autism subtypes has been enabled by innovative computational methods applied to large, deeply phenotyped cohorts. The seminal study by Troyanskaya and colleagues analyzed data from over 5,000 children in the SPARK autism cohort, employing a generative finite mixture model (GFMM) to parse heterogeneity [7] [12]. This model analyzed 239 phenotypic features—including item-level responses from standard diagnostic questionnaires (SCQ, RBS-R, CBCL), developmental milestones, and medical history—accommodating mixed data types (continuous, binary, categorical) without fragmenting individuals into separate phenotypic categories [12]. The person-centered framework maintained representation of the whole individual, allowing researchers to define groups of individuals with shared phenotypic profiles that translated to clinically similar presentations [13]. Model selection considered six standard statistical fit measures, with the four-class solution providing the optimal balance of statistical fit and clinical interpretability [12].

Four Clinically Distinct Phenotypic Subtypes

The analysis revealed four clinically distinct ASD subtypes with characteristic phenotypic profiles:

Table 1: Phenotypic Profiles of Autism Subtypes

Subtype Name	Prevalence	Core Clinical Features	Developmental Trajectory	Co-occurring Conditions
Social & Behavioral Challenges	~37%	Significant social challenges, repetitive behaviors, disruptive behaviors	Typical milestone achievement; later diagnosis	High rates of ADHD, anxiety, depression, OCD
Mixed ASD with Developmental Delay	~19%	Variable social/repetitive behaviors, developmental delays	Later walking/talking; earlier diagnosis	Language delays, intellectual disability; low anxiety/depression
Moderate Challenges	~34%	Milder core autism symptoms	Typical milestone achievement	Few co-occurring psychiatric conditions
Broadly Affected	~10%	Severe across all core domains	Significant developmental delays	Multiple: anxiety, depression, mood dysregulation

These subtypes were externally validated through medical history data not included in the original model, with patterns of co-occurring conditions (ADHD, language delays, intellectual disability) aligning with the subtype classifications [12]. The model demonstrated strong replication in the independent Simons Simplex Collection cohort, confirming the robustness of these phenotypic classes [12].

Neurobiological Validation Through Brain Imaging Profiles

Neuroimaging Approaches to Subtyping

Complementing the phenotypically defined subgroups, multiple research groups have used neuroimaging data to identify ASD subtypes based on brain structure and function. These approaches typically utilize clustering algorithms applied to neuroimaging-derived features:

• Semi-supervised clustering methods like HYDRA (HeterogeneitY through DiscRiminative Analysis) incorporate diagnostic labels (ASD vs. controls) during the clustering process [4]
• Tensor decomposition of functional MRI data to capture different brain community patterns across subtypes [5]
• Normative modeling of functional connectivity to quantify individual deviations from typical neurodevelopmental trajectories [6]
• Multimodal integration combining structural and functional MRI data through methods like skeleton-based white matter functional analysis [2]

These neuroimaging approaches consistently identify 2-4 biological subtypes within ASD populations, each with distinct brain network profiles [5] [6] [2].

Distinct Neural Subtypes and Their Characteristics

Research leveraging the ABIDE dataset (Aggregated Brain Imaging Data Exchange) has revealed reproducible neural subtypes:

Table 2: Neuroimaging-Defined ASD Subtypes

Study	Subtypes Identified	Key Neural Characteristics	Behavioral Correlates
Wang et al. [4]	Hyper-connectivity (n=847)	Increased within-network connectivity (DMN, attention); altered between-network connectivity	Distinct neuro-behavioral relationships critical for individualized treatment
	Hypo-connectivity	Opposite connectivity patterns to hyper-connectivity subtype
Qiao et al. [2]	Subtype 1 (n=92)	Lower fractional anisotropy (FA) in posterior cingulate cortex	No significant FIQ/PIQ differences from controls
	Subtype 2	Reduced FA in anterior cingulate, middle temporal gyrus, parahippocampus, thalamus; higher mean diffusivity	Lower FIQ and PIQ compared to Subtype 1
Normative Modeling Study [6]	Subtype A (n=479)	Positive deviations: occipital, cerebellar networks; Negative deviations: frontoparietal, DMN, cingulo-opercular networks	Distinct gaze patterns in eye-tracking tasks despite similar clinical scores
	Subtype B	Inverse functional deviation pattern to Subtype A

These neuroimaging subtypes demonstrate the biological validity of ASD stratification, with studies showing that the identified subgroups respond differently to interventions [6]. For instance, one study noted a 61.5% response rate to chronic intranasal oxytocin in one ASD subtype compared to only 13.3% in another [6].

Figure 1: Integrated Workflow for ASD Subtype Discovery. This framework illustrates how multimodal data integration through computational modeling yields distinct subtype classifications that inform precision medicine approaches.

Genetic Architecture of Autism Subtypes

Distinct Genetic Profiles Across Subtypes

The phenotypically defined subtypes exhibit striking differences in their genetic architecture, explaining previous challenges in identifying consistent genetic markers for autism:

Table 3: Genetic Characteristics of ASD Subtypes

Subtype	Genetic Variation Type	Key Biological Pathways	Developmental Timing
Social & Behavioral Challenges	Common variation; genes active during childhood	Neuronal action potentials, synaptic function	Predominantly postnatal gene expression
Mixed ASD with Developmental Delay	Rare inherited variants	Chromatin organization, transcriptional regulation	Predominantly prenatal gene expression
Broadly Affected	High burden of damaging de novo mutations	Multiple affected pathways	Widespread developmental disruption
Moderate Challenges	Mixed genetic profile	Less severe pathway disruptions	Varied developmental timing

The Social/Behavioral subtype showed the strongest common variant signal through polygenic score analysis, while the Broadly Affected subtype carried the highest burden of damaging de novo mutations [7] [12]. Notably, the Mixed ASD with Developmental Delay subtype was uniquely enriched for rare inherited variants [12]. Importantly, there was "little to no overlap in the impacted pathways between the classes," with each subtype affecting largely distinct biological processes despite all being previously implicated in autism [13].

Developmental Timing of Genetic Effects

A remarkable finding concerns the developmental timing of genetic disruptions across subtypes. Researchers discovered that genes affected in the Social and Behavioral Challenges subtype—characterized by later diagnosis and minimal developmental delays—were predominantly active after birth [7] [13]. Conversely, the Mixed ASD with Developmental Delay subtype, which presents with early developmental delays, involved genes predominantly active during prenatal development [13]. This alignment between the timing of genetic disruptions and clinical presentation represents a significant advance in understanding autism's heterogeneous developmental trajectories.

Experimental Protocols and Methodologies

Key Analytical Workflows

The breakthrough findings in autism subtyping rely on sophisticated analytical workflows:

Phenotypic Subtyping Protocol (SPARK Cohort)

• Data Collection: Comprehensive phenotypic data from 5,392 individuals with ASD, including 239 item-level and composite features from SCQ, RBS-R, CBCL, and developmental history [12]
• Model Training: General Finite Mixture Model (GFMM) application accommodating heterogeneous data types (continuous, binary, categorical) [12]
• Class Determination: Evaluation of models with 2-10 latent classes using Bayesian Information Criterion and clinical interpretability [12]
• Validation: External validation using medical history data not included in modeling and replication in independent SSC cohort [12]

Neuroimaging Subtyping Protocol (ABIDE Cohort)

• Data Acquisition: Resting-state fMRI and structural MRI from 1,046 participants across ABIDE I and II [6] [4]
• Feature Extraction: Calculation of static and dynamic functional connectivity measures across brain networks [6]
• Dimension Reduction: Orthogonal Projective Non-Negative Matrix Factorization to reduce high-dimensional connectivity data [4]
• Clustering: Application of HYDRA semi-supervised clustering informed by ASD-control diagnostic labels [4]

Figure 2: Genetic Discovery Workflow in SPARK Cohort. This diagram outlines the process from data collection through phenotypic and genetic analysis to identify subtype-specific biological mechanisms.

Table 4: Key Research Resources for ASD Subtyping Studies

Resource	Type	Function	Example Use
ABIDE Database	Data Repository	Multi-site aggregation of brain imaging & phenotypic data	Neuroimaging studies of functional connectivity subtypes [5] [6] [4]
SPARK Cohort	Patient Cohort	Largest US study of autism with genetic & phenotypic data	Phenotypic subtyping and genetic correlation [7] [12] [13]
General Finite Mixture Models	Statistical Model	Identify latent classes in mixed data types	Phenotypic subclassification of 5,392 individuals [12]
HYDRA	Algorithm	Semi-supervised clustering using diagnostic labels	Neurosubtyping based on functional connectivity [4]
Normative Modeling	Analytical Framework	Quantifies individual deviations from typical development	Identification of neural subtypes with distinct trajectories [6]
fMRIPrep	Software Tool	Robust preprocessing of functional MRI data	Standardized processing across multiple study sites [6]

The delineation of autism subtypes based on integrated phenotypic, neuroimaging, and genetic data represents a transformative advance in the field. The consistent identification of biologically meaningful subgroups across multiple independent studies and methodologies provides strong evidence for distinct "biological narratives" underlying autism's heterogeneity. These findings explain previous challenges in identifying unified biological explanations for autism—researchers were essentially "trying to solve a jigsaw puzzle without realizing we were actually looking at multiple different puzzles mixed together" [7].

The implications for research and clinical practice are substantial. For neuroscientists and drug development professionals, these findings enable more targeted investigations of biological mechanisms and potential therapeutic approaches. The distinct genetic pathways and developmental timelines associated with each subtype suggest they may respond differently to interventions, explaining the limited success of one-size-fits-all treatment approaches. For clinical practice, validated subtyping frameworks could eventually enable earlier identification of challenges and more personalized intervention strategies, helping individuals access appropriate supports based on their specific subtype profile [7] [13].

Future research directions should focus on further refining these subtypes, investigating non-coding genetic variation, developing clinical tools for subtype identification, and testing subtype-specific interventions. As this field progresses, the vision of precision medicine for neurodevelopmental conditions like autism is becoming increasingly attainable, promising more effective and individualized approaches to support for the diverse autism community.

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in both its clinical presentation and underlying neurobiology, presenting a substantial challenge for diagnosis and the development of targeted interventions [2] [6]. This variability has motivated a paradigm shift from treating ASD as a single disorder toward identifying discrete neurosubtypes with distinct biological signatures and clinical correlates [29]. The integration of neuroimaging technologies, particularly structural and functional magnetic resonance imaging (sMRI and fMRI), has been pivotal in advancing this approach, allowing researchers to parse the heterogeneity of ASD based on differences in brain structure and function [35] [5]. Identifying how these neurobiological subtypes map onto specific symptom profiles and developmental trajectories is crucial for advancing precision medicine in autism, potentially enabling clinicians to predict individual outcomes and select the most effective, personalized treatment strategies [6].

Comparative Analysis of ASD Neurosubtypes

Research utilizing multimodal neuroimaging has consistently identified several reproducible neurosubtypes of autism, each defined by a unique pattern of brain alterations and associated with a distinct clinical profile. The tables below synthesize the key characteristics of these subtypes from recent studies.

Table 1: Neural and Clinical Profiles of Data-Driven Neurosubtypes

Neurosubtype	Neural Profile (Structural & Functional)	Clinical & Cognitive Correlates
Subtype 1 (Function-Structure Coupling) [2]	- Lower fractional anisotropy (FA) in the posterior cingulate cortex (PCC) [2]- No significant differences in anterior cingulate or temporal regions [2]	- Higher full-scale IQ (FIQ) and performance IQ (PIQ) compared to Subtype 2 [2]
Subtype 2 (Function-Structure Coupling) [2]	- Reduced FA in anterior cingulate cortex, middle temporal gyrus, parahippocampus, and thalamus [2]- Higher mean diffusivity in middle temporal gyrus, parahippocampus, and thalamus [2]	- Lower full-scale IQ (FIQ) and performance IQ (PIQ) [2]
Subtype A (Functional Connectivity) [6]	- Positive deviations in the occipital and cerebellar networks [6]- Negative deviations in the frontoparietal network, default mode network, and cingulo-opercular network [6]	- Comparable clinical symptoms to Subtype B [6]- Distinct gaze patterns in eye-tracking tasks [6]
Subtype B (Functional Connectivity) [6]	- Inverse pattern of Subtype A across the mentioned networks [6]	- Comparable clinical symptoms to Subtype A [6]- Distinct gaze patterns in eye-tracking tasks [6]

Table 2: Profiles of Traditional DSM-IV Subtypes Based on Multimodal Neuroimaging

DSM-IV Subtype	Key Neural Features	Symptom Correlation
Asperger's Disorder [29]	- Negative functional features in subcortical areas (putamen-parahippocampus fALFF) [29]	- Correlated with specific ADOS subdomains, with social interaction as a common deficit [29]
PDD-NOS [29]	- Negative fALFF in the anterior cingulate cortex [29]	- Correlated with specific ADOS subdomains, with social interaction as a common deficit [29]
Autistic Disorder [29]	- Negative fALFF in thalamus-amygdala-caudate circuit [29]- Major functional impairments in the subcortical network and default mode network [5]	- Correlated with specific ADOS subdomains, with social interaction as a common deficit [29]

Experimental Protocols for Subtype Identification

The identification of neurosubtypes relies on sophisticated analytical protocols applied to neuroimaging data. Below are the detailed methodologies for two key approaches cited in this review.

Multimodal Fusion of Structural and Functional Coupling

This protocol is designed to identify subtypes by integrating white matter structure and brain function [2].

• Data Acquisition and Preprocessing: Collect Diffusion Tensor Imaging (DTI) and resting-state functional MRI (fMRI) data. Preprocess the DTI data to create a white matter skeleton and extract fMRI signals.
• Feature Extraction: Project the fMRI signals onto the white matter skeleton to perform voxel-wise function-structure coupling. Use White Matter Low-Frequency Oscillations (WM LFOs) as the primary input feature for clustering.
• Clustering Analysis: Employ an unsupervised clustering algorithm (e.g., similarity network fusion) on the fused function-structure features to identify distinct subgroups without prior diagnostic labels.
• Statistical Validation: Compare the identified clusters using statistical tests to find significant differences in microstructural properties (Fractional Anisotropy (FA) and Mean Diffusivity (MD)) and clinical measures (e.g., IQ scores) between the subtypes and against a neurotypical control group.
• Predictive Validation: Use a support vector machine (SVM) to evaluate whether the identified subtypes can enhance the diagnostic prediction accuracy for ASD compared to a general ASD classification model.

Normative Modeling of Functional Connectivity and Eye-Tracking

This protocol identifies subtypes based on deviations from typical brain development and links them to behavioral measures [6].

• Cohort Establishment: Assemble a large, multi-site discovery cohort (e.g., from ABIDE I and II) with resting-state fMRI data from individuals with ASD and typically developing (TD) controls. Include an independent validation cohort with paired fMRI and eye-tracking data.
• Multilevel Functional Connectivity Analysis: For each participant, calculate multiple types of functional connectivity: Static Functional Connectivity Strength (SFCS) using Pearson correlation, and instant Dynamic Functional Connectivity (DFC) assessing both strength (DFCS) and variance (DFCV) using Dynamic Conditional Correlation.
• Normative Model Construction: Build a model defining the typical developmental trajectory of multilevel FC features using data from the TD control group.
• Deviation Mapping: Calculate individual-level deviation scores for each participant with ASD from the established normative model.
• Clustering and Behavioral Correlation: Apply clustering analysis to the deviation scores to identify potential ASD subtypes. Correlate these neural subtypes with clinical symptom scores and, in the validation cohort, with performance on eye-tracking tasks (e.g., face emotion processing and joint attention).

Visualizing Workflows and Relationships

The following diagrams illustrate the logical workflows for the key experimental protocols and the relationship between neural features and clinical symptoms.

Neurosubtyping Identification Workflow

Neural-to-Clinical Symptom Mapping

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Reagents and Solutions for ASD Neurosubtyping Research

Tool Category	Specific Tool/Technique	Primary Function in Research
Imaging Modalities	Diffusion Tensor Imaging (DTI)	Quantifies white matter microstructure through fractional anisotropy (FA) and mean diffusivity (MD) [2].
	Resting-state fMRI (rs-fMRI)	Measures spontaneous brain activity to map functional connectivity (FC) between neural networks [6].
	Structural MRI (sMRI)	Provides high-resolution images of brain anatomy for measuring gray matter volume (GMV) [35] [5].
Analytical Frameworks	Normative Modeling	A statistical framework that quantifies individual deviations from typical neurodevelopmental trajectories [6].
	Similarity Network Fusion (SNF)	An unsupervised clustering algorithm that integrates multiple data types (e.g., sMRI and fMRI) to identify subtypes [35].
	Tensor Decomposition	A method for extracting compressed feature sets from high-dimensional fMRI data to identify brain community patterns [5].
Biomarkers & Features	Fractional Amplitude of Low-Frequency Fluctuations (fALFF/ALFF)	Measures the intensity of spontaneous brain activity from fMRI data, reflecting regional neural activity [29] [5].
	Gray Matter Volume (GMV)	A quantitative measure from sMRI indicating the volume of neural tissue in brain regions [35] [5].
	Functional Connectivity (FC)	Quantifies the temporal correlation of neural activity between different brain regions, indicating network integration [6].
Behavioral Assessments	Autism Diagnostic Observation Schedule (ADOS)	A gold-standard, semi-structured assessment used to diagnose and measure ASD symptom severity [6] [29].
	Eye-Tracking (e.g., Tobii TX300)	Precisely measures gaze patterns during tasks (e.g., face emotion), quantifying social attention deficits [6].

Discussion and Future Directions

The consistent identification of neurosubtypes across multiple studies, despite variations in methodological approaches, strongly validates the existence of biologically distinct subgroups within the autism spectrum. The convergence of findings is particularly notable, with several studies isolating two primary subtypes characterized by divergent patterns of neural alterations—often showing opposite changes in GMV or deviations in functional networks—despite comparable clinical presentations at a gross level [2] [6] [35]. This dissociation between neural pathology and overt symptomatology underscores the limitation of relying solely on behavioral checklists for diagnosis and highlights the potential of biomarkers to provide a more granular understanding of the disorder.

A critical insight from this research is that unique neural pathways can lead to similar behavioral symptoms in ASD [29]. This has profound implications for therapeutic development, as it suggests that a "one-size-fits-all" treatment is unlikely to be effective for all individuals with a diagnosis of ASD [6]. Future work must focus on longitudinal studies to determine the stability of these subtypes over time and their predictive value for individual developmental trajectories and treatment outcomes. Furthermore, integrating genetic and molecular data with neuroimaging findings will be essential for uncovering the etiological mechanisms driving these distinct neurosubtypes, ultimately paving the way for truly personalized interventions in autism.

The classification of autism has undergone a profound transformation, shifting from clinically defined categories in the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) to modern, data-driven approaches that integrate neuroimaging and genetics. This evolution reflects the growing recognition of autism's heterogeneity and the limitations of behavior-based diagnostic frameworks [86]. The DSM-IV, published in 1994, established autism as a spectrum disorder but divided it into several distinct diagnoses, including Autistic Disorder, Asperger's Syndrome (AS), and Pervasive Developmental Disorder-Not Otherwise Specified (PDD-NOS) [87] [86]. This framework aimed to capture varying presentations but relied exclusively on behavioral observations and clinical checklists without biological validation [19] [5].

Contemporary research has challenged this categorical approach, leveraging advanced computational methods and large-scale datasets to identify subtypes grounded in neurobiological and genetic evidence. The integration of structural magnetic resonance imaging (sMRI), functional MRI (fMRI), and whole-genome sequencing has enabled a more nuanced understanding of autism's complex architecture [19] [7] [45]. This paradigm shift has substantial implications for research, diagnosis, and therapeutic development, moving the field toward precision medicine approaches that account for the underlying biological mechanisms driving different autism presentations [7] [13]. This analysis systematically contrasts these two subtyping frameworks, examining their methodologies, empirical foundations, and clinical implications within the context of neuroimaging research.

Historical DSM-IV Subtyping Framework

Diagnostic Categories and Clinical Criteria

The DSM-IV classification system recognized five pervasive developmental disorders, with three constituting the core autism spectrum: Autistic Disorder, Asperger's Disorder, and Pervasive Developmental Disorder-Not Otherwise Specified (PDD-NOS) [86]. Each category featured specific diagnostic criteria centered on behavioral manifestations. Autistic Disorder required impairments in social interaction, communication, and the presence of restricted, repetitive behaviors, often with associated language delays and intellectual disability [87]. Asperger's Syndrome was characterized by social interaction impairments and restricted interests/behaviors but with no significant general delay in language or cognitive development [87] [88]. PDD-NOS served as a subthreshold diagnosis for individuals who exhibited some autistic behaviors but did not meet the full criteria for other disorders [87].

The differentiation between Asperger's Syndrome and High-Functioning Autism (HFA) was particularly contentious. Both groups exhibited average or above-average cognitive functioning, leading to substantial debate about whether they represented distinct conditions or variants of a single disorder [87]. Comparative studies revealed subtle differences in language, cognitive profiles, school functioning, and comorbidity patterns, though these distinctions were often quantitative rather than qualitative [87]. This diagnostic confusion ultimately contributed to the consolidation of these categories into the unified autism spectrum disorder (ASD) diagnosis in DSM-5 [87] [86].

Table 1: DSM-IV Autism Subtype Diagnostic Criteria

Subtype	Social Interaction	Communication	Restricted/Repetitive Behaviors	Cognitive/Language Development
Autistic Disorder	Impairment required	Impairment required	Required	Often delayed
Asperger's Syndrome	Impairment required	No significant delay	Required	No significant delay
PDD-NOS	Atypical or subthreshold symptoms	Variable presentation	May be present	Variable

Limitations of Behavioral Classification

The DSM-IV framework presented several significant limitations. Diagnosis relied heavily on clinician observation and behavioral checklists, introducing substantial subjectivity [19] [5]. The system demonstrated poor reliability in differentiating between subtypes, particularly between AS and HFA, as evidenced by inconsistent findings across 125 comparative studies [87]. Additionally, the categorical approach failed to capture the continuous nature of autistic traits and the substantial heterogeneity within each diagnostic category [87] [19]. Perhaps most critically, the subtypes lacked validation through biological markers, creating a classification system detached from underlying neurodevelopmental mechanisms [19] [7].

Modern Data-Driven Subtyping Approaches

Neuroimaging-Based Subtypes

Advanced neuroimaging techniques have enabled the identification of subtypes based on differences in brain structure and function. Using data from the Autism Brain Imaging Data Exchange (ABIDE I), researchers have extracted features from fMRI and sMRI to distinguish between traditional DSM-IV subtypes, finding significant differences in brain network organization [19] [5]. Tensor decomposition of resting-state fMRI data has revealed distinct brain communities and functional connectivity patterns that differentiate autism, Asperger's, and PDD-NOS subtypes [19] [5]. Specifically, impairments in the subcortical network and default mode network were found to primarily distinguish the autism subtype from Asperger's and PDD-NOS [19] [5].

Additional features such as the amplitude of low-frequency fluctuation (ALFF), fractional ALFF (fALFF), and gray matter volume (GMV) have provided further discrimination between subtypes, indicating both functional and structural neurological differences [19] [5]. These neuroimaging findings demonstrate that historically defined categories do indeed have biological correlates, though the relationship is more complex than initially conceptualized in the DSM-IV framework.

Table 2: Key Neuroimaging Features Differentiating ASD Subtypes

Feature	Description	Significant Findings
Functional Connectivity (FC)	Temporal correlation between brain regions	Different brain community patterns across subtypes [19]
ALFF/fALFF	Spontaneous brain activity intensity	Significant differences in specific neural circuits [19] [5]
Gray Matter Volume (GMV)	Volume of gray matter brain regions	Structural variations between subtypes [19] [5]
Default Mode Network	Brain network active at rest	Impaired function distinguishes autism subtype [19]
Subcortical Network	Deep brain structures for emotion/memory	Impaired function distinguishes autism subtype [19]

Genetically Informed Phenotypic Classes

A landmark 2025 study analyzing data from over 5,000 individuals in the SPARK cohort identified four biologically distinct autism subtypes using a "person-centered" computational approach that considered 239 phenotypic features [7] [13] [12]. This method employed general finite mixture modeling (GFMM) to handle diverse data types (continuous, binary, categorical) and identify latent classes based on shared trait profiles [13] [12]. The resulting subtypes exhibit distinct developmental trajectories, medical profiles, and patterns of co-occurring conditions:

• Social and Behavioral Challenges (37%): Characterized by core autism traits with typical developmental milestone attainment but high rates of co-occurring ADHD, anxiety, and depression [7] [13].
• Mixed ASD with Developmental Delay (19%): Features developmental delays but lower rates of co-occurring psychiatric conditions; associated with rare inherited genetic variants [7] [13].
• Moderate Challenges (34%): Milder manifestation of core autism behaviors without developmental delays or significant co-occurring psychiatric conditions [7] [13].
• Broadly Affected (10%): Presents with widespread challenges including developmental delays, social-communication difficulties, and multiple co-occurring conditions; shows the highest burden of damaging de novo mutations [7] [13].

This approach represents a paradigm shift from symptom-focused classification to a holistic framework that captures the complex interplay of traits within individuals and links them to specific biological mechanisms.

Methodological Comparison

Experimental Protocols and Workflows

The methodologies underlying historical versus contemporary subtyping approaches differ fundamentally in their design, data requirements, and analytical techniques.

Table 3: Comparison of Methodological Approaches

Aspect	DSM-IV Subtyping	Data-Driven Subtyping
Primary Data	Clinical behavioral observations	Multimodal: neuroimaging, genetics, developmental history
Assessment Tools	ADOS, ADI-R, clinical checklists	fMRI, sMRI, whole-genome sequencing, standardized questionnaires
Analytical Approach	Categorical classification	Computational modeling (e.g., finite mixture models, tensor decomposition)
Sample Sizes	Small to moderate cohorts	Large cohorts (n>5,000) [7] [12]
Validation	Clinical consensus	Biological replication (genetic pathways, neuroimaging profiles)

Key Research Reagents and Materials

Contemporary autism subtyping research relies on sophisticated methodological tools and resources:

• ABIDE I Dataset: Publicly shared neuroimaging dataset containing resting-state fMRI, anatomical data, and phenotypic information from 539 individuals with ASD across 17 international sites [19] [5]. Essential for neuroimaging studies of ASD subtypes.

• SPARK Cohort: Largest autism cohort study with over 150,000 participants, integrating genetic data with extensive phenotypic measures [7] [13]. Enabled identification of genetically informed subtypes.

• General Finite Mixture Models (GFMM): Computational approach that handles heterogeneous data types (continuous, binary, categorical) to identify latent classes without fragmenting individual phenotypic profiles [13] [12].

• Tensor Decomposition Methods: Feature extraction technique for analyzing high-dimensional fMRI data (combining brain regions, time, and patients) to identify distinct brain communities across subtypes [19] [5].

• Connectome Computation System (CCS): Standardized pipeline for preprocessing fMRI data, including band-pass filtering (0.01-0.1 Hz) and global signal regression, ensuring consistency across analyses [19] [5].

Comparative Analysis and Implications

Biological Plausibility and Validation

The fundamental distinction between these approaches lies in their relationship to biological mechanisms. DSM-IV subtypes were defined purely based on behavioral manifestations, with subsequent research attempting to identify corresponding biological differences [87] [19]. In contrast, modern data-driven subtypes emerge from patterns in biological data itself, creating categories with inherent neurobiological validity [7] [13].

Genetic analyses reveal that the four data-driven subtypes exhibit distinct molecular signatures with minimal pathway overlap [13] [12]. Crucially, the timing of genetic expression differs between subtypes, aligning with their clinical presentations. For the Social and Behavioral Challenges group, affected genes primarily activate after birth, consistent with their typical early development and later diagnosis [7] [13]. Conversely, the Mixed ASD with Developmental Delay subtype involves genes active prenatally, corresponding to earlier apparent developmental differences [7] [13].

Clinical and Research Applications

The implications of these differing approaches extend to diagnosis, treatment development, and research design. The DSM-IV framework offered clinicians familiar categories with established, though imperfect, intervention approaches [87] [86]. Data-driven subtypes provide more precise prognostic information and potential for personalized interventions based on underlying biological mechanisms [7] [13].

For pharmaceutical development, the new subtypes offer opportunities for targeted clinical trials based on shared biological pathways rather than heterogeneous behavioral presentations [7]. This could significantly improve treatment efficacy by addressing the specific mechanisms underlying different autism presentations.

Visualizing Research Workflows

Data-Driven Subtype Identification Pipeline

Neuroimaging Feature Extraction Protocol

The transition from DSM-IV to data-driven autism subtyping represents a fundamental shift in nosological philosophy—from categorical classification based on surface-level behaviors to dimensional characterization grounded in neurobiological mechanisms. While the DSM-IV framework established autism as a spectrum and recognized its heterogeneity, its subtypes lacked biological validation and reliable boundaries [87] [86]. Contemporary approaches leveraging large-scale multimodal data have identified biologically distinct subgroups with distinct genetic profiles, developmental trajectories, and neurobiological signatures [7] [13] [12].

This paradigm shift has profound implications for autism research and clinical practice. The identification of biologically meaningful subtypes enables more precise prognosis, targeted interventions, and stratified clinical trials [7]. Furthermore, recognizing that different genetic programs and developmental timelines underlie various autism presentations challenges the notion of a unified "autism biology" and suggests the need for diverse therapeutic approaches [13]. As these data-driven frameworks continue to evolve, they promise to advance both our understanding of autism's complex etiology and our ability to provide personalized support for autistic individuals across the lifespan.

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in clinical symptoms, neurobiology, and developmental trajectories, presenting substantial challenges for biomarker identification and intervention development [6]. This heterogeneity necessitates research approaches that can identify reproducible neurobiological patterns across diverse populations. The integration of structural (sMRI) and functional MRI (fMRI) has emerged as a promising approach for delineating ASD subtypes, though the replicability and generalizability of findings across independent cohorts remain critical hurdles [89] [6]. Cross-cohort replication studies provide an essential framework for validating potential biomarkers and establishing their utility for diagnostic classification, prognostic prediction, and treatment personalization in ASD research.

Recent advances in multivariate analytical methods and machine learning have enabled more sophisticated approaches to ASD subtyping, yet these must be validated through rigorous independent testing to avoid overestimation of predictive capabilities [90]. This comparison guide examines the current landscape of cross-cohort replication in ASD neuroimaging research, objectively evaluating methodological approaches, performance metrics, and clinical applicability of findings from key studies in the field.

Comparative Analysis of Structural vs. Functional MRI Approaches for ASD Subtyping

Table 1: Performance comparison of structural and functional MRI approaches for ASD subtyping across independent cohorts

Study	Modality	Sample Size (ASD/TD)	Analytical Approach	Cross-Cohort Validation	Key Findings	Limitations
Chen et al. (2025) [6]	Resting-state fMRI	479/567 (Discovery) 21/15 (Validation)	Normative modeling, functional connectivity clustering	Internal validation across ABIDE I/II; external validation with independent cohort	Identified two neural ASD subtypes with distinct functional connectivity profiles despite comparable clinical symptoms	Limited clinical behavioral correlation; sample size differences between cohorts
ABIDE I Subtype Analysis (2024) [19]	fMRI & sMRI	152/0 (Autism) 54/0 (Asperger's) 28/0 (PDD-NOS)	Tensor decomposition, ALFF/fALFF, GMV	Multi-site sample but no explicit external validation	Found significant functional differences in subcortical and default mode networks between ASD subtypes	No healthy control comparison; unequal subgroup sizes
HCP Brain-Behavior Study (2022) [89]	Resting-state fMRI	1,200+ (across multiple cohorts)	Region-wise and whole-brain predictive modeling	Cross-dataset replication across HCP-YA, HCP-A, eNKI-RS	Moderate generalizability for fluid cognition prediction; low replicability for openness prediction	Modest prediction accuracy; limited behavioral measures
Brain Structure-Behavior Study (2022) [91]	sMRI (GMV, CT, SA)	1,200+ (HCP-YA & HCP-A)	Regularized canonical correlation analysis	Replication across HCP-YA and HCP-A cohorts	Identified one heritable, replicable latent dimension linking brain structure to executive function and positive affect	Focus on healthy adults only; limited ASD-specific findings

Table 2: Quantitative performance metrics of cross-cohort validation studies in neuropsychiatric disorders

Study	Disorder	Biomarker Type	Internal Validation Accuracy	External Validation Accuracy	Clinical Translation Potential
Smooth Pursuit Eye Movements (2024) [90]	Psychosis	Sensorimotor (eye tracking)	64% (B-SNIP1 sample)	65% (B-SNIP2), 66% (PARDIP), 58% (FOR2107)	High - simple, quantifiable, low-cost assessment
Chen et al. (2025) [6]	ASD	Functional connectivity	Not explicitly reported	Subtypes replicated in independent cohort with distinct gaze patterns	Moderate - subtypes showed different treatment response patterns
Brain Metastases Monitoring (2022) [92]	Oncology (brain metastases)	Structural MRI (automated color-coding)	74% (conventional reading)	91% (with automated color-coding)	High - immediately clinically applicable for treatment monitoring
HCP Fluid Cognition Prediction (2022) [89]	Healthy adults	Functional connectivity	Variable within cohorts	Moderate cross-cohort generalizability	Low - modest effect sizes and limited clinical application

Experimental Protocols for Cross-Cohort Validation

Normative Modeling and Functional Subtyping Protocol

The most robust protocol for identifying ASD subtypes involves a comprehensive analytical pipeline that integrates multiple data modalities and validation steps [6]:

• Data Acquisition and Preprocessing: Collect resting-state fMRI data from multiple sites using standardized protocols. The ABIDE consortium dataset provides an exemplary model, with data from over 1,000 participants across 23 international sites. Preprocessing should include motion correction, normalization to standard space (e.g., MNI152), and registration using combined linear and non-linear transforms [19] [6].

• Multilevel Functional Connectivity Assessment: Calculate both static and dynamic functional connectivity features. Static functional connectivity strength (SFCS) can be derived using Pearson correlation, while dynamic functional connectivity strength (DFCS) and variance (DFCV) can be assessed using dynamic conditional correlation (DCC) approaches [6].

• Normative Modeling: Develop normative models based on multilevel FC features observed in typically developing (TD) control groups. These models quantify individual-level deviations in ASD participants from normative neurodevelopmental trajectories [6].

• Clustering Analysis: Apply unsupervised clustering algorithms to the individual-level deviation maps to identify potential ASD subtypes with distinct functional connectivity profiles.

• Behavioral Correlation: Associate identified neural subtypes with behavioral measures, including clinical symptoms (ADOS, SRS) and performance on eye-tracking tasks assessing social attention and emotion processing [6].

• Cross-Cohort Validation: Validate identified subtypes in independent cohorts with both neuroimaging and behavioral data to assess generalizability and stability of the subtypes.

Machine Learning Validation Framework for Biomarker Generalizability

A rigorous machine learning framework for cross-cohort validation should incorporate the following steps [90]:

• Initial Model Training: Train predictive models on a large, heterogeneous sample (e.g., B-SNIP1 sample with N=674 psychosis probands and 305 controls) using multivariate pattern analysis.

• Internal Validation: Assess model performance using robust cross-validation techniques within the discovery sample to avoid overfitting.

• External Validation: Apply the trained model to multiple independent samples with varying characteristics:

  • Similar clinical population from different recruitment sites (B-SNIP2)
  • Related disorders with and without the target phenotype (bipolar disorder with and without psychosis)
  • Distinct clinical populations (affective disorders vs. psychosis)
  • Early-stage and high-risk populations (clinical high-risk for psychosis)

• Performance Metrics: Report balanced accuracy, sensitivity, specificity, and likelihood ratios for each validation step to provide comprehensive assessment of model performance [90].

Visualization of Research Workflows and Methodological Approaches

ASD Subtyping and Validation Workflow: This diagram illustrates the comprehensive pipeline for identifying and validating autism subtypes across multiple cohorts, integrating neuroimaging, eye-tracking, and clinical data.

Cross-Cohort Validation Framework: This diagram outlines the multi-cohort approach for independent validation of neuroimaging findings, demonstrating how models trained on discovery cohorts are tested across independent samples with varying characteristics.

The Scientist's Toolkit: Essential Research Reagent Solutions

Table 3: Essential research tools and resources for cross-cohort ASD neuroimaging studies

Tool/Resource	Type	Function	Example Implementation
ABIDE Database	Data Repository	Provides aggregated neuroimaging and phenotypic data from multiple international sites for discovery and validation	Primary data source for ASD subtype identification studies; includes over 1,000 participants across 23 sites [19] [6]
Normative Modeling	Analytical Framework	Quantifies individual deviations from typical neurodevelopmental trajectories	Identification of extreme deviations in functional connectivity patterns that define ASD subtypes [6]
fMRIPrep	Processing Tool	Standardized preprocessing pipeline for fMRI data	Ensures consistent data quality and preprocessing across multi-site studies [6]
Multilevel Functional Connectivity	Analytical Metric	Captures both static and dynamic properties of brain network organization	Differentiation of ASD subtypes based on SFCS, DFCS, and DFCV profiles [6]
Automated Color-Coding Software	Visualization Tool	Highlights longitudinal changes in MRI scans through color-coded overlays	Improved diagnostic accuracy in monitoring brain metastasis progression (91% vs 74% with conventional reading) [92]
Eye-Tracking Paradigms	Behavioral Assessment	Quantifies social attention and emotion processing through gaze patterns	Validation of neural subtypes through distinct behavioral profiles in social attention tasks [6]
Machine Learning Validation Framework	Analytical Protocol	Tests stability and generalizability of multivariate pattern analysis	Comprehensive validation of smooth pursuit eye movement biomarkers across 4 independent cohorts [90]

The rigorous application of cross-cohort validation methodologies represents a critical pathway toward identifying clinically useful biomarkers for ASD. Current evidence suggests that functional connectivity approaches show particular promise for identifying neurobiologically distinct ASD subtypes that transcend conventional behavioral diagnoses [6]. The replication of these subtypes across independent cohorts and their association with distinct gaze patterns provides compelling evidence for their validity.

Nevertheless, important challenges remain. Modest effect sizes, limited generalizability across diverse populations, and the complexity of integrating multimodal data present ongoing hurdles [89] [90]. Future research should prioritize prospective validation studies, integration of multimodal biomarkers (including genetic and environmental factors), and the development of standardized analytical pipelines to facilitate comparison across studies. The establishment of robust, independently validated ASD subtypes will ultimately enable more targeted interventions and advance the field toward precision medicine approaches for this heterogeneous disorder.

The pursuit of reliable biomarkers for Autism Spectrum Disorder (ASD) is increasingly focused on disentangling its significant biological heterogeneity. This comparison guide evaluates the performance of emerging, subtype-informed classification models against traditional general ASD classification models. Evidence from recent neuroimaging research demonstrates that models accounting for distinct neural subtypes—often identified through semi-supervised clustering of structural (sMRI) and functional MRI (fMRI) data—consistently achieve superior diagnostic accuracy and reveal unique brain-behavior relationships compared to one-size-fits-all approaches. This guide provides a detailed benchmarking of their experimental protocols, performance metrics, and the essential reagents required to implement these advanced analytical frameworks.

Autism Spectrum Disorder (ASD) is characterized by profound clinical and neurobiological heterogeneity, which presents a major obstacle for developing precise diagnostic tools and effective, targeted interventions [4]. Traditional general classification models treat ASD as a single, homogeneous entity, often leading to models with limited predictive power and clinical applicability. The emerging paradigm of neuro-subtyping seeks to address this by first identifying biologically distinct subgroups within the autism spectrum before building predictive models.

This shift is largely driven by advances in neuroimaging, which have reliably identified at least two major neural subtypes: one characterized by widespread functional hyper-connectivity and another by functional hypo-connectivity in the brain [4]. These subtypes, which cut across traditional behavioral diagnoses, demonstrate unique functional connectivity profiles and correlate differently with core ASD symptoms. This guide directly benchmarks classification strategies that leverage these subtype discoveries against conventional general models, providing researchers with a clear framework for selecting and implementing the most effective approach for their specific goals.

Comparative Performance Analysis: Subtype-Informed vs. General Models

The tables below synthesize quantitative evidence from recent studies, directly comparing the performance and characteristics of subtype-informed and general classification models.

Table 1: Benchmarking Key Performance Metrics

Model Characteristic	General Classification Models	Subtype-Informed Models
Primary Objective	Distinguish ASD from neurotypical (NT) controls [93] [94].	First identify neural subtypes within ASD, then classify or predict outcomes [6] [4].
Typical Accuracy	Varies widely: 66% (3D CNN) to 100% (LR, AB) on behavioral data; ~98% on facial images [93] [94].	Provides nuanced profiles; identifies subgroups with inverse connectivity patterns (e.g., hyper vs. hypo-connectivity) [6] [4].
Diagnostic Power	Focuses on ASD vs. NT separation.	Enhances diagnostic prediction accuracy compared to general ASD classification [2].
Clinical Relevance	Provides a binary risk assessment.	Enables stratification for targeted interventions; shows distinct symptom correlations per subtype [4].

Table 2: Technical and Methodological Comparison

Aspect	General Classification Models	Subtype-Informed Models
Core Data Modality	Often single-modality (e.g., sMRI, fMRI, facial images) [5] [94].	Frequently multimodal (e.g., integrating sMRI and fMRI) or high-dimensional FC [2] [4].
Key Algorithms	CNNs (VGG19, ResNet), SVM, Logistic Regression, Random Forest [93] [94].	Semi-supervised clustering (HYDRA), normative modeling, multi-scale dimension reduction (OPNNMF) [6] [4].
Sample Size (Typical)	Can operate on smaller datasets (e.g., 100s-1,000s of images) [94].	Often requires larger cohorts (e.g., ~1,000+ participants) for robust subtype discovery [6] [4].
Interpretability	Often "black box"; provides limited neurobiological insight.	High; aims to elucidate distinct neurobiological substrates of each subgroup [2] [4].

Experimental Protocols for Subtype-Informed Classification

The superior performance of subtype-informed models hinges on sophisticated experimental workflows. The following section details the key methodologies cited in recent literature.

Semi-Supervised Clustering with HYDRA

This protocol identifies subtypes based on how individuals deviate from neurotypical norms.

Workflow Diagram: Semi-Supervised Clustering with HYDRA

Detailed Protocol:

• Data Input and Preprocessing: Acquire resting-state fMRI data from a large, multi-site cohort (e.g., ABIDE I/II). Preprocess using standardized pipelines (e.g., fMRIPrep) to correct for motion, normalize to a standard space, and extract time series from predefined brain atlases (e.g., Dosenbach 160 ROIs). Calculate a high-dimensional functional connectivity (FC) matrix for each participant [6] [4].
• Dimensionality Reduction: Apply Orthonormal Projective Non-Negative Matrix Factorization (OPNNMF) to the high-dimensional FC data. This step reduces feature dimensions while preserving the intrinsic structure of the data, preventing overfitting and creating a more representative feature space for clustering. The number of components (e.g., M=1195) is determined via parameter selection [4].
• Semi-Supervised Clustering (HYDRA): Implement the HeterogeneitY through DiscRiminative Analysis (HYDRA) algorithm. Unlike unsupervised methods, HYDRA uses the diagnostic labels (ASD vs. Neurotypical Controls) to guide the clustering process. It models each identified subtype as a separate linear classification boundary that distinguishes it from the neurotypical reference group. The optimal number of clusters (typically K=2) is determined through reliability and validity indices [4].
• Output and Validation: The output is two or more distinct ASD neural subtypes. These are validated by:
  • Assessing the reliability and distinctness of their FC profiles.
  • Establishing unique neuro-behavioral correlations between specific connectivity patterns (e.g., within the default mode network) and core ASD symptoms (e.g., from ADOS scores) for each subtype [4].

Structural and Functional Coupling for Subtyping

This protocol identifies subtypes based on the interaction between white matter structure and brain function.

Workflow Diagram: Structural and Functional Coupling for Subtyping

Detailed Protocol:

• Multi-Modal Data Acquisition: Collect both Diffusion Tensor Imaging (DTI) and resting-state fMRI data from participants with ASD and neurotypical controls [2].
• Data Integration via Skeleton Projection: Integrate the structural and functional data by projecting the fMRI signals onto a white matter (WM) skeleton derived from the DTI data. This creates a voxel-wise map of structure-function coupling [2].
• Feature Extraction and Clustering: Use WM low-frequency oscillations (LFOs) from the coupled data as input features for a clustering algorithm (e.g., K-means). This aims to categorize individuals with ASD into distinct neurological subgroups based on the coupled architecture of their brain [2].
• Subtype Characterization and Validation: Statistically compare the identified subtypes against controls on measures of white matter integrity (e.g., Fractional Anisotropy (FA) and Mean Diffusivity (MD)) and clinical scores (e.g., FIQ). Validate the subgroups by demonstrating that a classifier (e.g., SVM) achieves higher diagnostic accuracy when distinguishing between these subtypes than when classifying ASD as a single group [2].

Successful implementation of the aforementioned protocols requires a suite of specific data, software, and instrumentation.

Table 3: Essential Research Reagents and Solutions

Item Name	Type	Function/Application	Example Source
ABIDE Datasets	Data Repository	Provides large-scale, multi-site sMRI and fMRI data for discovery and validation of subtypes.	Autism Brain Imaging Data Exchange [5] [6] [2]
fMRIPrep	Software Pipeline	A robust, standardized tool for preprocessing of fMRI data, ensuring reproducibility and data quality.	https://fmriprep.org [6]
HYDRA	Algorithm	A semi-supervised clustering method that uses diagnostic labels to identify discrete neurobiological subtypes.	Varol et al., 2017 [4]
Dosenbach 160 Atlas	Brain Atlas	A predefined set of 160 regions of interest derived from meta-analyses of multiple cognitive domains, used for extracting FC features.	Dosenbach et al., 2010 [6]
OPNNMF	Algorithm	A dimensionality reduction technique used to extract meaningful, lower-dimensional features from high-dimensional FC matrices for clustering.	Wen et al., 2022 [4]

The empirical evidence overwhelmingly indicates that subtype-informed ASD classification models outperform general models in key areas of diagnostic accuracy, biological interpretability, and clinical relevance. While general models can achieve high accuracy on specific tasks like facial image classification, they fail to capture the fundamental neurobiological heterogeneity of autism. The paradigm shift towards first identifying robust neural subtypes—such as the hyper-connectivity and hypo-connectivity subgroups—using advanced semi-supervised clustering and multimodal integration provides a more powerful and nuanced framework. This approach not only improves classification metrics but also paves the way for developing personalized intervention strategies tailored to an individual's specific neurobiological profile, ultimately aligning ASD research with the core principles of precision medicine.

Conclusion

The integration of structural and functional MRI has irrevocably shifted the paradigm of autism research from a one-size-fits-all approach to a nuanced, subtype-driven framework. The consistent identification of distinct neurosubtypes—each with unique brain network deviations, genetic underpinnings, and clinical trajectories—provides a powerful roadmap for deconstructing ASD heterogeneity. These advances directly address the critical barriers that have long plagued drug development, offering a path to stratify patient populations for clinical trials and identify those most likely to respond to specific interventions. The future of ASD research lies in refining these subtypes with ever-larger, multi-omics datasets, validating them prospectively in clinical settings, and ultimately translating these discoveries into personalized diagnostic tools and targeted therapeutics that improve the quality of life for individuals across the autism spectrum.

References

• 1. The diagnosis of ASD with MRI: a systematic review and ... [https://www.nature.com/articles/s41398-024-03024-5]
• 2. The Autism Spectrum Disorder Subtypes Identification ... [https://pubmed.ncbi.nlm.nih.gov/40730718/]
• 3. Autism classified by magnetic resonance imaging [https://pmc.ncbi.nlm.nih.gov/articles/PMC7723195/]
• 4. Diagnosis-informed neuro-subtyping reveals subgroups of ... [https://www.sciencedirect.com/science/article/abs/pii/S0278584625002064]
• 5. Comparison of autism spectrum disorder subtypes based ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1440222/full]
• 6. Identifying brain functional subtypes and corresponding ... [https://www.nature.com/articles/s41380-025-03086-x]
• 7. Major autism study uncovers biologically distinct subtypes, ... [https://www.princeton.edu/news/2025/07/09/major-autism-study-uncovers-biologically-distinct-subtypes-paving-way-precision]
• 8. Polygenic and developmental profiles of autism differ by ... [https://www.nature.com/articles/s41586-025-09542-6]
• 9. New Autism Subtypes Identified Based on Genes, ... [https://www.brainfacts.org/diseases-and-disorders/childhood-disorders/2025/new-autism-subtypes-identified-based-on-genes-individual-traits-111125]
• 10. Subgrouping autism and ADHD based on structural MRI ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12135403/]
• 11. Structural and functional magnetic resonance imaging of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3465665/]
• 12. Decomposition of phenotypic heterogeneity in autism ... [https://www.nature.com/articles/s41588-025-02224-z]
• 13. New Study Reveals Subclasses of Autism by Linking Traits ... [https://www.simonsfoundation.org/2025/07/09/new-study-reveals-subclasses-of-autism-by-linking-traits-to-genetics/]
• 14. in Brain : development followed by... autism early overgrowth [https://pubmed.ncbi.nlm.nih.gov/15362165/]
• 15. in infants at high risk for Early ... | Nature brain development autism [https://www.nature.com/articles/nature21369?error=cookies_not_supported&code=cee7db63-690e-430a-a79e-d5945df374ba]
• 16. A Key Brain Difference Linked to Autism Is Found for the First ... [https://medicine.yale.edu/news-article/a-key-brain-difference-linked-to-autism-is-found-for-the-first-time-in-living-people/]
• 17. Age-varying distinct neuroanatomy in young children with ... [https://pubmed.ncbi.nlm.nih.gov/41238842/]
• 18. Atypical Development of Face and Greeble Recognition in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3071970/]
• 19. Comparison of autism spectrum disorder subtypes based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11486766/]
• 20. The Default Mode Network in Autism - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC5635856/]
• 21. Salience Network Connectivity in Autism Is Related to Brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4924541/]
• 22. Salience network connectivity is altered in 6-week-old ... [https://www.nature.com/articles/s42003-024-06016-9]
• 23. Altered functional and structural brain network organization in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3777708/]
• 24. Meta-analytic connectivity modelling of functional magnetic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10049951/]
• 25. Altered Functional Connectivity in a Triple-Network Model ... [https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2021.736755/full]
• 26. A Review of Default Mode Network Connectivity and Its ... [https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2020.00614/full]
• 27. Altered Resting-State Functional Activity in Patients With ... [https://www.frontiersin.org/journals/neurology/articles/10.3389/fneur.2018.00556/full]
• 28. Contrast subgraphs catch patterns of altered functional ... [https://www.nature.com/articles/s41598-025-04932-2]
• 29. Common and unique multimodal covarying patterns in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7673101/]
• 30. Drug development for Autism Spectrum Disorder (ASD) [https://www.sciencedirect.com/science/article/pii/S0924977X2100239X]
• 31. Multimodal Approaches to Functional Connectivity in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5897150/]
• 32. Linking functional and structural brain organisation with ... [https://molecularautism.biomedcentral.com/articles/10.1186/s13229-023-00564-3]
• 33. Convergent Findings of Altered Functional and Structural ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0067329]
• 34. Twinned neuroimaging analysis contributes to improving ... [https://www.nature.com/articles/s41598-024-71174-z]
• 35. Parsing the heterogeneity of brain structure and function in ... [https://pubmed.ncbi.nlm.nih.gov/39966244/]
• 36. rTMS-induced neuroimaging changes measured with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12095301/]
• 37. Multi-Modal Machine Learning for Precision Medicine: Luo Lab [https://labs.feinberg.northwestern.edu/luolab/research/research-project-asd.html]
• 38. A Framework for Comparison and Interpretation of Machine ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11912182/]
• 39. Reproducible comparison and interpretation of machine ... [https://www.medrxiv.org/content/10.1101/2024.09.04.24313055v1.full-text]
• 40. ABIDE [https://fcon_1000.projects.nitrc.org/indi/abide/]
• 41. Identification of autism spectrum disorder using deep ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5635344/]
• 42. SPARK: Sparsity-based analysis of reliable k-hubness and ... [https://pubmed.ncbi.nlm.nih.gov/27046111/]
• 43. AI Model Detects Autism With 98% Accuracy [https://www.technologynetworks.com/diagnostics/news/deep-learning-model-reaches-98-accuracy-in-autism-detection-404935]
• 44. A meta-analytic investigation of 25 years of voxel-based ... [https://www.sciencedirect.com/science/article/pii/S0149763424002604]
• 45. A Comprehensive Framework for Differentiating Autism ... [https://www.sciencedirect.com/science/article/pii/S1071909120300115]
• 46. GM-VGG-Net: A Gray Matter-Based Deep Learning ... [https://www.mdpi.com/2075-4418/15/11/1425]
• 47. Quantifying individual differences in brain morphometry ... [https://www.nature.com/articles/s41598-019-45774-z]
• 48. Profiling brain morphology for autism spectrum disorder ... [https://www.nature.com/articles/s42003-025-08573-z]
• 49. Changes in grey matter development in autism spectrum ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3695319/]
• 50. Biomarkers in Autism Spectrum Disorder - PubMed Central - NIH [https://pmc.ncbi.nlm.nih.gov/articles/PMC6526849/]
• 51. Dissecting Static and Dynamic Functional Connectivity [https://pmc.ncbi.nlm.nih.gov/articles/PMC6545633/]
• 52. Connectivity dynamics in typical development and its ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6045960/]
• 53. Dynamic Functional Connectivity Alterations and Their ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2021.794151/full]
• 54. Altered resting state dynamic functional connectivity of ... [https://www.sciencedirect.com/science/article/pii/S0165032722006632]
• 55. Functional connectivity signatures in fMRI-derived ... [https://www.sciencedirect.com/science/article/pii/S2949921625000195]
• 56. Relationship between static and dynamic brain functional ... [https://archive.ismrm.org/2017/1720.html]
• 57. Fractionating autism based on neuroanatomical normative ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7648836/]
• 58. Big data approaches to decomposing heterogeneity across ... [https://www.nature.com/articles/s41380-018-0321-0]
• 59. Precision neurodiversity: personalized brain network ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12647089/]
• 60. Biological subtyping of autism via cross-species fMRI - PubMed [https://pubmed.ncbi.nlm.nih.gov/40093106/]
• 61. A Systematic Review of Diffusion Tensor Imaging Studies ... [https://www.mdpi.com/2076-3425/15/8/824]
• 62. Multi-Slice Generation sMRI and fMRI for Autism Spectrum ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10670036/]
• 63. Attention-Enhanced Fusion of Structural and Functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11512738/]
• 64. Identifying Autism from Resting-State fMRI Using Long ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5669262/]
• 65. rs-fMRI and machine learning for ASD diagnosis [https://www.nature.com/articles/s41598-022-09821-6]
• 66. Semi-Supervised Learning in Medical MRI Segmentation [https://pmc.ncbi.nlm.nih.gov/articles/PMC8228966/]
• 67. Lost in translation: the valley of death across preclinical and ... [https://transmedcomms.biomedcentral.com/articles/10.1186/s41231-019-0050-7]
• 68. Opportunities and challenges in translational science - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC8504824/]
• 69. Drug development for Autism Spectrum Disorder (ASD) [https://pmc.ncbi.nlm.nih.gov/articles/PMC10062405/]
• 70. A cautionary tale for autism drug development [https://www.thetransmitter.org/spectrum/guest-blog-a-cautionary-tale-for-autism-drug-development/]
• 71. Developmental changes of cortical white–gray contrast as ... [https://www.nature.com/articles/s41398-018-0296-2]
• 72. Altered Gray-White Matter Boundary Contrast in Toddlers at ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8248433/]
• 73. Advanced Quantitative Microstructure Imaging in Autism [https://pmc.ncbi.nlm.nih.gov/articles/PMC12661291/]
• 74. Unsupervised data-driven stratification of mentalizing ... [https://www.nature.com/articles/srep35333]
• 75. Mapping the structure of biomarkers in autism spectrum disorder [https://pmc.ncbi.nlm.nih.gov/articles/PMC11671500/]
• 76. Early biomarker for autism spectrum disorder unveiled [https://pubmed.ncbi.nlm.nih.gov/40481745/]
• 77. Flexible nonlinear modeling reveals age-related differences in ... [https://pubmed.ncbi.nlm.nih.gov/40234995/]
• 78. Functional connectivity subtypes associate robustly with ... [https://elifesciences.org/articles/56257]
• 79. Data-guided neuroimaging and visualization [https://apertureneuro.org/article/140427-data-guided-neuroimaging-and-visualization-from-functional-decomposition-to-dynamic-fusion]
• 80. Bridging the gap: explainable ai for autism diagnosis and ... [https://www.nature.com/articles/s41598-025-24662-9]
• 81. Early diagnosis of autism across developmental stages ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12164165/]
• 82. Subgrouping autism and ADHD based on structural MRI ... [https://molecularautism.biomedcentral.com/articles/10.1186/s13229-025-00667-z]
• 83. Improving brain difference identification in autism spectrum ... [https://www.nature.com/articles/s42003-025-07928-w]
• 84. Evaluation of motion artefact reduction depending on the ... [https://www.nature.com/articles/s41598-023-35794-1]
• 85. Reproduction of Motion Artifacts for Performance Analysis ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3674114/]
• 86. How the Autism Diagnosis Has Evolved Over Time [https://azaunited.org/blog/how-the-autism-diagnosis-has-evolved-over-time]
• 87. Subtyping the Autism Spectrum Disorder: Comparison of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6331497/]
• 88. The changing faces of autism - PubMed Central - NIH [https://pmc.ncbi.nlm.nih.gov/articles/PMC9365987/]
• 89. Cross-cohort replicability and generalizability of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9611632/]
• 90. Evidence from comprehensive independent validation ... [https://www.nature.com/articles/s41598-024-64487-6]
• 91. A cross-cohort replicable and heritable latent dimension ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9701210/]
• 92. Automated Color-Coding of Lesion Changes in Contrast ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8985679/]
• 93. A comprehensive review of autism detection models [https://www.sciencedirect.com/science/article/pii/S2773186325000659]
• 94. Application of artificial intelligence in modern healthcare for ... [https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2025.1569464/full]