Functional Connectivity Subtypes in Autism Spectrum Disorder: A Comprehensive Review for Researchers and Therapeutics Development

Abstract

This article synthesizes current research on functional connectivity (FC) subtypes in Autism Spectrum Disorder (ASD), addressing the critical need to deconstruct the disorder's pronounced heterogeneity. We explore the neurobiological foundations of distinct ASD subtypes, defined by unique patterns of both hyper- and hypo-connectivity across major brain networks like the Default Mode Network and Frontoparietal Network. The review critically evaluates data-driven clustering methodologies, including semi-supervised and normative modeling approaches, for their robustness and clinical applicability. We further examine how these FC subtypes correlate with divergent behavioral phenotypes, sensory processing profiles, and treatment responses. Finally, the article discusses validation strategies and the translational potential of these findings, outlining a path toward biologically grounded diagnostics and personalized therapeutic interventions for ASD.

Deconstructing Heterogeneity: Foundational Neurobiological Subtypes of ASD

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by significant heterogeneity in both its biological underpinnings and behavioral manifestations. This diversity presents a fundamental challenge for developing effective biomarkers and therapeutic strategies [1]. The traditional diagnostic framework, which primarily relies on behavioral phenotyping including impairments in social communication and interaction alongside restricted, repetitive patterns of behavior, often obscures the distinct neurobiological pathways that may underlie these conditions [2] [1]. The emerging paradigm of neuro-subtyping seeks to address this challenge by moving beyond behavioral symptoms to identify biologically distinct subgroups within the autism spectrum. This approach leverages advanced neuroimaging technologies and machine learning algorithms to delineate subtypes based on distinct brain organization patterns, offering a transformative path toward personalized diagnostics and targeted interventions [3] [1].

Comparative Analysis of Neuro-Subtyping Approaches

Research conducted over the past decade has employed diverse methodological frameworks to identify neural subtypes within ASD. The table below provides a systematic comparison of three prominent subtyping approaches based on functional connectivity, neuroanatomy, and behavioral phenotyping.

Table 1: Comparative Analysis of ASD Neuro-Subtyping Approaches

Subtyping Approach	Primary Data Source	Identified Subtypes	Sample Size	Key Distinguishing Features
Functional Connectivity Subtyping [2] [1]	Resting-state fMRI (rs-fMRI)	2 subtypes: Hyper-connectivity & Hypo-connectivity	~1,800 individuals (across studies)	Opposite patterns of functional connectivity within and between major brain networks (DMN, FPN, visual, auditory).
Multidimensional Neuroanatomical Subtyping [3]	Structural MRI (sMRI)	3 biotypes: ASD-I, ASD-II, ASD-III	220 males from 3 sites	Distinct cortical organization patterns (thickness, surface area, tissue contrast, geodesic distance).
Behavioral Phenotyping via Machine Learning [4]	Clinical/Behavioral Assessments	2 overlying phenotypes (16 subgroups)	2,400 children with ASD	Unique behavioral deficit profiles and differential response to treatment.

Functional Connectivity Subtyping

Functional connectivity (FC) subtyping analyzes the temporal correlation of neural activity in different brain regions, providing insight into the functional organization of large-scale brain networks. A seminal 2025 study by Liu et al. analyzed 1,046 participants and identified two distinct neural ASD subtypes with unique functional brain network profiles, despite comparable clinical presentations [2] [5]. One subtype showed positive deviations in the occipital and cerebellar networks with negative deviations in the frontoparietal, default mode, and cingulo-opercular networks, while the other exhibited the inverse pattern [2].

Another study implementing a semi-supervised clustering method known as HYDRA on approximately 1,800 individuals confirmed this binary classification, revealing hyper-connectivity and hypo-connectivity subtypes [1]. The hyper-connectivity subtype shows increased connectivity within major large-scale networks, while the hypo-connectivity subtype shows the opposite. These subtypes also demonstrated varying correlations between connectivity patterns and core ASD symptoms, underscoring their potential clinical relevance [1].

Neuroanatomical Subtyping

Neuroanatomical subtyping focuses on structural brain variations. A framework integrating cortex-wide MRI markers of both vertical and horizontal cortical organization identified three distinctive anatomical subtypes [3]:

• ASD-I: Characterized by cortical thickening, increased surface area, and tissue blurring.
• ASD-II: Defined by cortical thinning and decreased geodesic distance.
• ASD-III: Marked by increased geodesic distance.

These biotypes were associated with differential symptom load and intrinsic connectivity anomalies, particularly in networks supporting communication and social cognition. The study further demonstrated that incorporating subtyping information significantly improved the prediction of Autism Diagnostic Observation Schedule (ADOS) scores in single subjects compared to subtype-blind approaches [3].

Experimental Protocols in Neuro-Subtyping Research

Protocol for Functional Connectivity Subtyping

The workflow for identifying functional connectivity subtypes typically involves a multi-stage analytical pipeline, as detailed in recent high-impact studies [2] [1] [6].

Diagram 1: Functional Connectivity Subtyping Workflow.

Participant Cohorts and Data Acquisition

Studies typically utilize large, multi-site datasets such as the Autism Brain Imaging Data Exchange (ABIDE I and II), which aggregate neuroimaging and phenotypic data from thousands of individuals with ASD and typical controls (TD) [2] [3] [1]. For instance, a 2025 study included a discovery cohort of 1,046 participants (479 ASD, 567 TD) from ABIDE and a separate validation cohort of 21 ASD individuals with additional eye-tracking data [2]. Inclusion criteria require available resting-state fMRI and T1-weighted structural images, while exclusion criteria typically involve excessive head motion and poor image quality [1].

Image Acquisition and Preprocessing

Participants undergo T1-weighted structural scans and resting-state fMRI (rs-fMRI) acquisitions across multiple sites with standardized protocols [3]. Preprocessing is typically performed using standardized pipelines like fMRIPrep [2] or the ABIDE Preprocessed Connectome Project pipeline [3], which includes steps for slice-time correction, head motion correction, skull stripping, intensity normalization, and registration to standard templates.

Multilevel Functional Connectivity Features

Multilevel functional connectivity is assessed through:

• Static Functional Connectivity (SFCS): Calculated using Pearson correlation between average BOLD signals from predefined regions of interest (e.g., Dosenbach 160 ROIs) [2].
• Dynamic Functional Connectivity (DFCS/DFCV): Assessed using dynamic conditional correlation for instant dynamic FC, measuring both strength and variability over time [2].

Normative Modeling and Clustering

Normative models are developed using TD data to establish standard functional developmental trajectories. Individual ASD participants are then mapped against these models to quantify deviations [2]. Advanced clustering techniques are applied to these deviation profiles:

• Semi-supervised methods like HYDRA (HeterogeneitY through DiscRiminative Analysis) incorporate diagnostic labels (ASD vs. TD) to guide the subtyping process [1].
• Unsupervised methods include Gaussian Mixture Models and Hierarchical Agglomerative Clustering [4].
• Hybrid approaches combine connectome-based gradient mapping with supervised random forest algorithms to inform clustering with diagnostic labels [6].

Protocol for Behavioral Phenotyping and Validation

Independent cohorts often undergo behavioral phenotyping to validate and characterize identified neural subtypes. A key validation method involves eye-tracking tasks focused on autism-sensitive social cues [2]:

• Face Emotion Processing Task: Participants observe static facial expressions (happy, angry, fearful, neutral) with defined areas of interest (eyes, nose, mouth).
• Joint Attention Task: Participants watch videos where individuals use eye-gaze or eye-gaze with finger pointing to direct attention toward target objects.

Eye-gaze data is acquired using systems like Tobii TX300 at 300 Hz sampling rate, with primary outcomes including z-scores for first fixation duration, fixation duration, and fixation count [2].

Table 2: Key Research Reagents and Solutions for Neuro-Subtyping Studies

Resource Category	Specific Tool/Resource	Function in Research	Example Use Case
Neuroimaging Datasets	ABIDE I & II (Autism Brain Imaging Data Exchange)	Provides large-scale, multi-site neuroimaging and phenotypic data for discovery and validation.	Primary data source for multiple subtyping studies [2] [3] [1].
Data Processing Pipelines	fMRIPrep, FreeSurfer, C-PAC (Configurable Pipeline for ASL/ABIDE)	Standardizes preprocessing of structural and functional MRI data across sites and studies.	Image preprocessing, cortical surface extraction [2] [3].
Clustering Algorithms	HYDRA, K-means, Gaussian Mixture Models (GMM), Hierarchical Clustering	Identifies data-driven subgroups within ASD populations based on neural features.	Identification of hyper/hypo-connectivity subtypes [1] and behavioral phenotypes [4].
Dimension Reduction Methods	OPNNMF, Connectome Gradient Mapping	Reduces high-dimensional neuroimaging data to meaningful, lower-dimensional features for clustering.	Feature extraction prior to HYDRA clustering [1] [6].
Behavioral Assessment Tools	Tobii Eye-Tracking Systems, ADOS, ADI-R, SRS	Quantifies behavioral phenotypes and validates neural subtypes against clinical measures.	Validation of neural subtypes via social attention patterns [2].

Implications for Targeted Interventions and Drug Development

The identification of neurobiological subtypes holds profound implications for advancing personalized therapeutic strategies in ASD. Different neural subtypes are likely to demonstrate differential treatment responses, which could explain the frequent failure of one-size-fits-all interventional trials [4]. For instance, research has shown that one ASD subtype exhibited a 61.5% response rate to chronic intranasal oxytocin treatment, while another subtype demonstrated only a 13.3% response [2]. This suggests that pre-segregating participants based on neurosubtyping could significantly enhance clinical trial sensitivity and lead to more effective, targeted therapies.

Furthermore, neuro-subtyping can guide treatment strategies by linking specific biological profiles to mechanism-based interventions. The FDA's recent action to approve leucovorin calcium for individuals with cerebral folate deficiency (CFD) who present with autistic features exemplifies this approach, representing a move toward treating specific biological subgroups within the autism spectrum [7].

The paradigm of neuro-subtyping represents a critical evolution in autism research, moving beyond descriptive behavioral phenotyping to dissect the neurobiological heterogeneity of ASD. Converging evidence from multiple large-scale studies confirms the existence of reproducible neural subtypes with distinct functional connectivity profiles, neuroanatomical features, and behavioral correlates. The standardization of experimental protocols—from data acquisition and preprocessing to advanced clustering methodologies—provides a robust foundation for validating these subtypes across independent cohorts. For researchers and drug development professionals, these advances offer a roadmap for developing personalized diagnostic tools and targeted interventions aligned with the distinct biological pathways underlying different ASD presentations. Future research focusing on longitudinal tracking of subtypes and their relationship to treatment outcomes will be essential for realizing the full clinical potential of this approach.

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by substantial heterogeneity in clinical presentation, underlying biology, and developmental trajectories. This heterogeneity has long challenged the identification of consistent biomarkers and the development of targeted interventions. The functional connectivity (FC) framework—specifically the dichotomy between hyper-connectivity and hypo-connectivity in large-scale brain networks—has emerged as a powerful approach to deconstruct this complexity into biologically meaningful subtypes. Research leveraging advanced analytical methods has consistently demonstrated that distinct patterns of neural over- and under-connectivity correlate with specific clinical profiles and genetic architectures, moving the field toward a precision medicine approach for ASD.

This comparison guide provides a systematic evaluation of the hyper-hypo connectivity framework for ASD subtyping, synthesizing evidence from recent neuroimaging studies to objectively compare subtype characteristics, methodologies, and clinical implications. We present standardized comparisons of experimental data and detailed protocols to facilitate research replication and integration across laboratories.

Comparative Analysis of ASD Connectivity Subtypes

Defining the Core Connectivity Subtypes

Research utilizing semi-supervised clustering methods on large datasets has consistently identified at least two primary connectivity-based subtypes of ASD. Wang et al. (2025) applied HeterogeneitY through DiscRiminative Analysis (HYDRA) to a sample of approximately 2,000 subjects, clearly delineating these subtypes based on distinct functional connectivity profiles [8]:

• Hyper-connectivity Subtype: Characterized by increased connectivity within major large-scale networks (default mode, frontoparietal, salience) and mixed between-network connectivity patterns, including hyper-connectivity between default mode and attention networks, but hypo-connectivity between default mode and visual/auditory networks [8].

• Hypo-connectivity Subtype: Exhibits essentially the inverse pattern, with widespread reductions in connectivity both within and between major brain networks [8].

These subtypes demonstrate high reliability and show differential relationships between connectivity patterns and core ASD symptoms, providing evidence that they represent clinically meaningful subdivisions of the autism spectrum [8].

Comparative Table: Hyper-connectivity vs. Hypo-connectivity Subtypes

Table 1: Characteristic differences between hyper-connectivity and hypo-connectivity ASD subtypes

Feature	Hyper-connectivity Subtype	Hypo-connectivity Subtype
Within-Network Connectivity	Increased within major large-scale networks [8]	Decreased within major large-scale networks [8]
Between-Network Connectivity	Mixed pattern: Hyper-connectivity between default mode and attention networks; Hypo-connectivity between default mode and visual/auditory networks [8]	Generally decreased, often inverse of hyper-connectivity pattern [8]
Developmental Trajectory	More characteristic of younger children with ASD [9]	Emerges in adolescence and persists into adulthood [9]
Primary Networks Affected	Default mode, frontoparietal, salience networks [8] [2]	Default mode, frontoparietal, salience networks [8] [2]
Connection Length Profile	Potential dominance of short-range connections [10]	Reduction in long-range connections [10]
Relationship to Symptoms	Distinct correlations with social communication and repetitive behaviors [8]	Different correlation patterns with core ASD symptoms [8]

Network-Specific Alterations Across Subtypes

Different large-scale brain networks show characteristic alterations across the connectivity subtypes. A comprehensive analysis published in Molecular Psychiatry (2025) identified distinct deviation patterns in two neural ASD subtypes despite comparable clinical presentations [2]:

• Subtype 1: Displayed positive deviations (increased connectivity) in the occipital network and cerebellar network, coupled with negative deviations (decreased connectivity) in the frontoparietal network, default mode network, and cingulo-opercular network [2].

• Subtype 2: Exhibited essentially the inverse pattern of functional deviations across these same networks [2].

These network-specific alterations were associated with different gaze patterns during autism-sensitive eye-tracking tasks, providing a crucial link between neural connectivity profiles and behavioral manifestations [2].

Developmental Trajectories of Connectivity Subtypes

The Hyper-to-Hypo Connectivity Shift

Emerging evidence suggests that connectivity subtypes follow distinct developmental courses. Shan et al. (2025) systematically mapped the developmental trajectory of intrinsic functional connectivity in 800 participants from the ABIDE dataset, revealing a predictable shift in connectivity patterns [9]:

• Early childhood: Hyper-connectivity is more characteristic of young children with ASD [9]
• Pre-adolescence: A shift from intrinsic hyper- to hypo-connectivity occurs [9]
• Adolescence and adulthood: Hypo-connectivity becomes predominant and persists into later developmental stages [9]

This developmental shift follows a specific temporal sequence across functional networks. Primary networks (somatomotor, auditory, visual) undergo the hyper-to-hypo connectivity shift earlier than higher-order networks (default mode, salience, frontoparietal), with the transition occurring in sequence throughout pre-adolescence [9].

Comparative Table: Developmental Trajectories Across Brain Networks

Table 2: Developmental timing of hyper-to-hypo connectivity shift across major brain networks

Brain Network	Network Type	Timing of Hyper-to-Hypo Shift	Functional Domain
Somatomotor Network (SMN)	Primary	Earliest shift	Sensory and motor processing
Auditory Network (AN)	Primary	Early shift	Auditory processing
Visual Network (VN)	Primary	Early shift	Visual processing
Default Mode Network (DMN)	Higher-order	Later shift	Social cognition, self-referential thought
Salience Network (SAN)	Higher-order	Later shift	Attention, detection of behaviorally relevant stimuli
Frontoparietal Network (FPN)	Higher-order	Latest shift	Executive function, cognitive control

Methodological Approaches for Connectivity Subtyping

Experimental Protocols and Analytical Frameworks

Several methodological approaches have been successfully employed to identify and validate connectivity-based subtypes in ASD:

Semi-Supervised Clustering (HYDRA Protocol) Wang et al. (2025) implemented a comprehensive protocol for neuro-subtyping [8]:

• Data Acquisition: Collected resting-state fMRI data from ∼2000 subjects across multiple sites
• Feature Extraction: Estimated functional connectivity matrices between brain regions
• Dimension Reduction: Applied multi-scale dimension reduction to high-dimensional input features
• Clustering: Implemented HYDRA, a semi-supervised clustering method guided by ASD/control labeling information
• Validation: Conducted systematic evaluation of clustering performance and compared with unsupervised approaches

Normative Modeling Approach The study by Molecular Psychiatry (2025) employed an alternative methodology [2]:

• Multi-level FC Features: Characterized both static functional connectivity strength (SFCS) and instant dynamic functional connectivity (DFCS and DFCV)
• Normative Modeling: Devised normative models based on multilevel FC features in typically developing groups
• Deviation Quantification: Quantified individual-level deviations in multilevel FC among ASD participants
• Clustering Analysis: Applied clustering to identify potential ASD subtypes based on deviation profiles
• Behavioral Correlation: Evaluated subtype differences in gaze patterns using eye-tracking tasks

Experimental Workflow Diagram

Diagram 1: Experimental workflow for connectivity subtyping

Key Research Reagent Solutions

Table 3: Essential materials and analytical tools for ASD connectivity subtyping research

Resource Category	Specific Tools/Platforms	Research Application
Neuroimaging Data	ABIDE I & II (Autism Brain Imaging Data Exchange)	Multi-site resting-state fMRI dataset with ASD and control participants [2] [11]
Analytical Frameworks	HYDRA (HeterogeneitY through DiscRiminative Analysis)	Semi-supervised clustering for neuro-subtyping [8]
Normative Modeling	Custom MATLAB/Python pipelines	Quantifying individual deviations from typical developmental trajectories [2]
Functional Parcellations	Dosenbach 160 ROIs, MIST_20 parcellation	Standardized brain atlases for connectivity analysis [2] [12]
Preprocessing Pipelines	fMRIPrep, Connectome Computation System	Standardized MRI data preprocessing [2] [11]
Dynamic FC Analysis	Dynamic Conditional Correlation (DCC)	Assessing instant dynamic functional connectivity [2]
Genetic Correlation Tools	SPARK cohort genetic data	Linking connectivity subtypes to genetic profiles [13]

Connectivity Patterns in Relation to Clinical and Genetic Subtypes

Integration with Genetically-Defined ASD Subtypes

Recent research has begun to integrate connectivity-based subtyping with genetic findings. A landmark study analyzing data from over 5,000 children in the SPARK cohort identified four clinically and biologically distinct subtypes of autism with distinct genetic profiles [13]:

• Social and Behavioral Challenges Group (37%): Core ASD traits without significant developmental delays; highest genetic predisposition for ADHD, anxiety, and depression [13]
• Moderate Challenges (34%): Core ASD behaviors but less strongly; generally no co-occurring psychiatric conditions [13]
• Mixed ASD with Developmental Delay (19%): Developmental milestones reached later; usually no anxiety, depression or disruptive behaviors [13]
• Broadly Affected (10%): Most severe difficulties including developmental delays, intellectual disability, and co-occurring psychiatric conditions [13]

These subtypes demonstrated distinct patterns of genetic variation, with the Broadly Affected group showing the highest proportion of damaging de novo mutations, while only the Mixed ASD with Developmental Delay group was more likely to carry rare inherited genetic variants [13].

Transdiagnostic Considerations: ASD and ADHD Connectivity Profiles

The hyper-hypo connectivity framework also helps differentiate ASD from frequently co-occurring conditions like ADHD. Norman et al. (2025) conducted a mega-analysis of 12,732 children and adolescents, revealing distinct neural signatures [14]:

• ASD Traits: Associated with weaker connections between the thalamus and putamen brain regions, as well as within salience networks [14]
• ADHD Traits: Characterized by hyperconnectivity between the default mode network (DMN) and dorsal attention network (DAN) [14]

These differential connectivity patterns align with the core symptoms of each condition, with ASD showing disruptions in sensory integration networks, while ADHD primarily affects attention regulation networks [14].

The hyper-hypo connectivity dichotomy provides a robust framework for deconstructing the heterogeneity of autism into biologically meaningful subtypes. Comparative analysis of these subtypes reveals distinct developmental trajectories, network involvement patterns, and clinical correlates that have direct implications for targeted interventions. The consistency of these findings across multiple large-scale studies and methodological approaches strengthens the validity of this framework.

Future research directions should focus on longitudinal tracking of connectivity subtypes across development, integration of genetic and connectomic data, and testing subtype-specific interventions. The tools and methodologies summarized in this guide provide a foundation for these next-generation studies, moving the field closer to truly personalized approaches for autism spectrum disorder.

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by deficits in social communication and interaction, alongside restricted, repetitive behaviors [15]. Rather than a single uniform disorder, ASD represents a highly heterogeneous condition with complex variations in behavioral symptoms and underlying neurobiology. Research over the past decade has increasingly focused on understanding the brain network alterations that underlie this heterogeneity, with particular emphasis on the Default Mode Network (DMN), Frontoparietal Network (FPN), and Visual System. These networks subserve distinct cognitive functions—social processing, executive control, and visual perception, respectively—that are frequently affected in ASD. This review synthesizes current functional connectivity research to compare how alterations in these key networks contribute to identifiable ASD subtypes, providing a framework for researchers and drug development professionals to target interventions more precisely.

The Default Mode Network: Social Cognition and Self-Referential Processing

Functional Anatomy and Role in Social Cognition

The Default Mode Network (DMN) is a strongly intrinsically interconnected network of brain structures that includes the posterior cingulate cortex (PCC), precuneus, medial prefrontal cortex (mPFC), temporoparietal junction (TPJ), and hippocampus [15]. This network is most active during passive resting states and engages in cognitive processes profoundly relevant to ASD symptomatology, including self-referential processing, mentalizing (theory of mind), and autobiographical memory. The PCC serves as a core functional hub with high baseline metabolic rate and connectivity, implicated in both self- and other-relevant processing. The mPFC monitors both one's own mental states and those of others, while the TPJ preferentially encodes 'other-relevant' information, including the mental states and beliefs of others [15].

DMN Dysfunction in ASD

Substantial evidence demonstrates DMN dysfunction in ASD, particularly during social cognitive tasks. During self-referential processing, individuals with ASD show reduced activation in the PCC and mPFC and reduced connectivity between these regions compared to neurotypical individuals [15]. Similarly, during theory of mind and mentalizing tasks, studies consistently report decreased recruitment of the TPJ and dorsal mPFC in adults with ASD [15]. These functional alterations correlate with behavioral measures, as the severity of social and communication deficits inversely correlates with FC magnitude in DMN core areas [16].

Table 1: Default Mode Network Alterations in ASD

DMN Component	Primary Functions	Alterations in ASD	Relationship to Symptoms
Posterior Cingulate Cortex (PCC)	Self-referential processing, autobiographical memory	Reduced activation during self-referential tasks; Insensitive to social semantic processing	Correlates with social affect severity
Medial Prefrontal Cortex (mPFC)	Monitoring self and other mental states	Reduced activation for self-related judgments; Atypical ventral-dorsal specialization	Linked to theory of mind deficits
Temporoparietal Junction (TPJ)	Mental state attribution, distinguishing self/other	Decreased recruitment during mentalizing; Reduced inter-hemispheric connectivity	Associated with difficulties inferring intentions

Subtyping Evidence

Data-driven clustering analyses reveal that DMN connectivity patterns can distinguish ASD subtypes. One prominent study identified two FC-based subtypes across ASD and typically developing participants: Subtype 1 showed stronger FC between the DMN and other networks but weaker within-network connectivity, while Subtype 2 exhibited the opposite pattern [17]. Importantly, these subtypes were not exclusively aligned with diagnostic status but potentially represent transdiagnostic neurobiological dimensions with implications for understanding the ASD spectrum.

The Frontoparietal Network: Executive Function and Cognitive Control

Network Organization and Cognitive Roles

The Frontoparietal Network (FPN), also known as the executive control network, comprises primarily the dorsolateral prefrontal cortex (dlPFC), inferior parietal lobule (IPL), and anterior cingulate cortex [18]. This network exercises top-down cognitive control necessary for adapting to changing goals and demands, including working memory, inhibition, task-switching, and attention regulation. The IPL associates primarily with maintenance of information in short-term memory, while the dlPFC involves in recognition of previously presented stimuli and information updating [19].

FPN Alterations in ASD

Meta-analytical evidence from 16 fMRI studies demonstrates that while both ASD and typically developing participants activate PFC regions during executive function tasks, individuals with ASD show differential recruitment of a wider frontoparietal network [18]. Specifically, ASD participants show greater activation in the right middle frontal gyrus and anterior cingulate cortex, but lesser activation in bilateral middle frontal gyri, left inferior frontal gyrus, right inferior parietal lobule, and precuneus compared to controls [18]. This suggests a constrained executive network in ASD that relies more heavily on prefrontal regions while under-recruiting parietal components.

Frequency-Specific Connectivity Abnormalities

Magnetoencephalographic studies during n-back tasks reveal frequency-specific FPN connectivity alterations in adults with ASD. While performing short-term and working memory tasks, adults with ASD exhibit atypical modulation of theta-band connectivity during recognition processes, but compensatory recruitment of alpha-band synchrony during maintenance periods [19]. This suggests that individuals with ASD may employ alternative neural oscillatory mechanisms to achieve behavioral performance levels comparable to neurotypical individuals.

Table 2: Frontoparietal Network Alterations in ASD

FPN Component	Primary Functions	Alterations in ASD	Methodological Evidence
Dorsolateral Prefrontal Cortex (dlPFC)	Recognition, updating, manipulation	Mixed findings: Both hyper- and hypo-activation reported; Atypical developmental trajectory	fMRI, MEG
Inferior Parietal Lobule (IPL)	Maintenance, storage	Consistent under-recruitment across studies	Coordinate-based ALE meta-analysis
Anterior Cingulate Cortex	Performance monitoring, error detection	Hyperactivation in ASD relative to controls	fMRI meta-analysis
Frontoparietal Theta Synchrony	Recognition processes	Decreased connectivity during stimulus recognition	MEG phase synchrony
Frontoparietal Alpha Synchrony	Maintenance processes	Increased/compensatory recruitment	MEG phase synchrony

The Visual System: Perception and Attention

Visual Processing Alterations in ASD

Individuals with ASD frequently exhibit atypical visual perception, characterized by enhanced low-level visual processing and superior performance on visual search and embedded figures tasks [20]. Neuroimaging studies consistently show that during visual detection tasks, individuals with ASD activate posterior brain regions (V1, extrastriate cortex) more extensively, while showing less activity in frontal regions compared to neurotypical individuals [20]. This neurobiological profile aligns with the behavioral observation of locally oriented visual processing and weak central coherence in ASD.

Visual Network Connectivity and Social Symptoms

Recent evidence indicates that visual system connectivity extends beyond perceptual processing to influence core social symptoms in ASD. A 2025 study demonstrated that stronger functional connectivity between the visual and salience networks was associated with higher social affect scores at school-age, even after accounting for other behavioral domains [21]. This relationship was specific to social affect symptoms, with similar trends observed for restricted and repetitive behaviors. These findings suggest that the visual system plays a canalizing role in the emergence of ASD symptoms, potentially influencing how social information is processed from early development.

Developmental Cascade Framework

Longitudinal neuroimaging studies suggest that atypical visual system development in infancy may represent one of the earliest neural markers along the path to ASD [22]. According to the developmental cascade framework, early differences in visual processing could subsequently impact the development of attention, social communication, and broader cognitive domains. This positions the visual system as a potential building block in the pathogenesis of ASD, with implications for early identification and intervention.

Comparative Analysis of Network Alterations Across ASD

Table 3: Comparative Analysis of Three Key Networks in ASD

Network	Primary Cognitive Domains	Direction of FC Changes in ASD	Developmental Trajectory	Subtype Associations
Default Mode Network	Social cognition, self-referential thought, mentalizing	Predominantly decreased within-network FC; Subnetwork hypoconnectivity	Atypical developmental trajectory; Age-dependent activation patterns	Associates with social affect severity; Two FC-based subtypes identified
Frontoparietal Network	Executive function, working memory, cognitive control	Mixed: Prefrontal hyperactivation with parietal hypoconnectivity; Frequency-specific synchrony alterations	Decreased structural connectivity during adolescence vs. increase in TD	Stronger baseline connectivity predicts better outcome; Altered theta/alpha band synchrony
Visual System	Visual perception, attention, sensory processing	Enhanced local connectivity; Altered visual-salience network FC	Early developmental differences; Potential canalizing role	Connectivity with salience network predicts social affect symptoms

Experimental Protocols and Methodologies

Resting-State Functional Connectivity MRI

Protocol Description: Resting-state fMRI (rs-fMRI) measures spontaneous low-frequency fluctuations in the blood oxygenation level-dependent (BOLD) signal while participants lie at rest in the scanner. This method identifies functionally connected brain regions that exhibit synchronous activity patterns, providing maps of intrinsic functional networks without requiring task performance.

Key Applications in ASD Research:

• Identification of FC subtypes using data-driven clustering approaches [17] [12]
• Investigation of network-level connectivity (within-network and between-network) [17]
• Correlation of connectivity measures with behavioral symptoms (e.g., Social Responsiveness Scale scores) [21] [16]

Analysis Approaches:

• Independent Component Analysis (ICA) to identify naturally coherent networks [16]
• Seed-based correlation analysis to examine specific network connections
• Multivariate statistical analyses to characterize group differences [17]

Task-Based Functional MRI

Protocol Description: Task-based fMRI measures brain activity while participants engage in specific cognitive tasks, allowing researchers to identify regions involved in particular cognitive processes. In ASD research, this approach has been particularly valuable for examining neural correlates of social cognition and executive function.

Common Paradigms in ASD Research:

• Self-referential vs. other-referential judgment tasks to probe DMN function [15]
• Theory of mind tasks (e.g., reading the mind in the eyes, false belief stories) [15]
• N-back tasks to assess working memory and FPN recruitment [19]
• Embedded figures tasks and block design tasks to examine visual processing [20]

Analysis Considerations:

• Examination of both activation magnitude and task-modulated connectivity
• Careful matching of behavioral performance between groups
• Accounting for motion artifacts, particularly challenging in ASD populations

Diffusion Spectrum Imaging and Structural Connectomics

Protocol Description: Diffusion Spectrum Imaging (DSI) is an advanced diffusion MRI technique that maps the complex architecture of white matter pathways by sampling a full spectrum of diffusion directions. This enables reconstruction of whole-brain structural connectomes representing the physical connections between brain regions.

Longitudinal Application:

• Tracking developmental changes in structural connectivity over time (3-7 year intervals) [23]
• Reconstruction of weighted structural networks using deterministic fiber tracking [23]
• Identification of structural connectivity subtypes predictive of symptom trajectories

Analysis Pipeline:

• Quality assurance for head motion exclusion criteria
• Q-space diffeomorphic reconstruction (QSDR) for normalization to standard space
• Network-based statistics (NBS) to identify significant subnetworks [23]

Visualization of Experimental Workflows

Functional Connectivity Subtyping Pipeline

Diagram Title: Functional Connectivity Subtyping Workflow

Cross-Network Interactions in ASD

Diagram Title: Network Interactions and Symptom Expression in ASD

Table 4: Essential Methodologies and Analytical Tools for ASD Network Research

Methodology/Resource	Primary Application	Key Considerations for ASD Research
Resting-State fMRI	Mapping intrinsic functional connectivity; Identifying FC subtypes	Motion artifacts particularly problematic; Requires specialized preprocessing
Task-Based fMRI	Assessing network recruitment during specific cognitive processes	Task compliance challenges; Need for age-appropriate paradigms
Diffusion Spectrum Imaging	Reconstructing structural connectomes; Tracking white matter development	Advanced acquisition protocols; Computational intensive processing
Magnetoencephalography (MEG)	Measuring frequency-specific neural synchrony with high temporal resolution	Excellent for examining oscillatory dynamics in FPN; Less common for DMN studies
Independent Component Analysis	Data-driven identification of functional networks without a priori seeds	Effectively identifies DMN and other major networks; Useful for capturing heterogeneity
Network-Based Statistics	Identifying significant subnetworks while controlling family-wise error	Appropriate for connectome-wide analyses; Detects distributed group differences
k-Means/Hierarchical Clustering	Data-driven subtyping based on connectivity patterns	Reveals neurobiological subtypes crossing diagnostic boundaries
Autism Diagnostic Observation Schedule	Standardized diagnostic assessment and symptom severity measurement	Essential for correlating neural measures with behavioral symptoms

Research examining the Default Mode, Frontoparietal, and Visual networks in ASD has revealed complex patterns of both within-network and between-network connectivity alterations that contribute to the heterogeneity of the autism spectrum. The DMN shows prominent hypoconnectivity related to social cognition deficits, the FPN exhibits mixed patterns of prefrontal hyperactivation and parietal hypoconnectivity associated with executive dysfunction, and the Visual system demonstrates early developing alterations that may canalize later emerging symptoms. Rather than operating in isolation, these networks interact in ways that potentially give rise to identifiable ASD subtypes with distinct developmental trajectories and clinical outcomes.

Future research directions should include: (1) longitudinal studies tracking network development from infancy through adulthood; (2) multi-modal approaches combining functional, structural, and metabolic measures; (3) examination of how genetic risk factors influence network development; and (4) clinical trials targeting specific network alterations with neurostimulation or pharmacological interventions. For drug development professionals, these network-based subtypes offer potential biomarkers for stratifying clinical trials and measuring treatment response. The continuing refinement of ASD subtypes based on neurobiological measures rather than solely behavioral symptoms promises more targeted and effective interventions for this heterogeneous condition.

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in clinical symptoms, genetic underpinnings, and neurobiology, presenting a substantial challenge for developing effective, personalized interventions [2] [1]. This diversity suggests that ASD encompasses multiple distinct subtypes, each with unique neural signatures and behavioral manifestations. Refining the classification of ASD subtypes is therefore essential for advancing personalized intervention strategies beyond a standardized, one-size-fits-all approach [2] [5]. The integration of neuroimaging techniques like functional magnetic resonance imaging (fMRI) with behavioral tools such as eye tracking provides a powerful methodological framework for linking brain function to observable behavior. This guide objectively compares the primary methodological approaches and findings in this domain, detailing how different research protocols identify and characterize ASD subtypes through their functional connectivity profiles and corresponding behavioral correlates, including visual preference and sensory processing patterns.

Comparative Analysis of Neural Subtypes and Their Biomarkers

Research has consistently identified several neural and behavioral subtypes within ASD. The table below summarizes the key subtypes, their defining characteristics, and associated behavioral correlates.

Table 1: Comparative Overview of ASD Subtypes and Their Correlates

Subtype Category	Defining Characteristics	Associated Behavioral & Sensory Correlates
Functional Connectivity Subtype 1 (Hyper-connectivity)	Positive deviations in occipital and cerebellar networks; negative deviations in frontoparietal, default mode, and cingulo-opercular networks [2] [1].	Distinct gaze patterns in social cue tasks; differential symptom severity [2].
Functional Connectivity Subtype 2 (Hypo-connectivity)	Inverse pattern of Subtype 1: negative deviations in occipital/cerebellar networks; positive deviations in frontoparietal, default mode, and cingulo-opercular networks [2] [1].	Distinct gaze patterns in social cue tasks; differential symptom severity [2].
Geometric Preference (GeoPref) Subtype	Defined behaviorally by a strong visual preference for geometric images over social images (>69% fixation time) [24] [25].	Elevated symptom severity; lower cognitive, language, and adaptive behavior scores; fewer saccades to geometric images [24] [25].
Sensory Processing Sensitivity (SPS)	A biological trait characterized by greater awareness and reactivity to environmental stimuli [26] [27].	Increased sensitivity to medications; stronger reactions to emotional stimuli, lights, and sounds [26] [27].

Experimental Protocols for Identifying ASD Subtypes

Functional Connectivity Subtyping Using resting-state fMRI

Objective: To identify distinct subtypes of ASD based on patterns of brain functional connectivity (FC) derived from resting-state fMRI data.

Protocol Workflow:

Detailed Methodology:

• Participant Recruitment: Large, multi-site cohorts are essential. Studies often utilize public datasets like ABIDE-I and ABIDE-II, comprising over 1,000 individuals (ASD and typical development) [2] [1].
• Data Acquisition: Participants undergo resting-state fMRI and structural T1-weighted scans across multiple scanners with standardized protocols.
• Data Preprocessing: A standardized pipeline using software like fMRIPrep is used for motion correction, normalization, and other preprocessing steps to ensure data quality and comparability [2].
• Feature Extraction: Multilevel functional connectivity features are calculated. This includes not only static functional connectivity strength (SFCS) but also instant dynamic functional connectivity, such as dynamic functional connectivity strength and variance (DFCS and DFCV) [2]. Blood-oxygen-level-dependent (BOLD) signals are extracted from predefined brain atlases (e.g., Dosenbach's 160 regions of interest).
• Dimensionality Reduction and Clustering: High-dimensional FC features are reduced using methods like Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF). Subsequently, clustering algorithms, particularly semi-supervised methods like HYDRA which incorporate diagnostic labels (ASD vs. control), are applied to identify subtypes [1]. This approach has been shown to be superior to unsupervised methods like K-means.
• Validation: Identified subtypes are validated for reliability and distinctness. Their functional profiles are then correlated with clinical symptoms (e.g., from ADOS, SRS) and behavioral task performance [1].

Eye-Tracking-Based Subtyping

Objective: To identify ASD subtypes based on characteristic gaze patterns and visual preferences during stimulus viewing.

Protocol Workflow:

Detailed Methodology:

• Stimulus Paradigms: Multiple paradigms are employed:
  • Preferential Looking (GeoPref): Toddlers view a side-by-side movie of dynamic geometric images and social images (children interacting) [24] [25].
  • Social Scenes: Viewing of video clips or still images of faces, biological motion (point-light displays), or joint attention scenarios [2] [28].
  • Gap-Overlap Paradigm: Measures attentional disengagement and orienting by presenting central and peripheral stimuli with temporal overlaps [29].
• Data Collection: Eye movements are recorded using systems like Tobii TX300 or SMI RED250. Areas of Interest (AOIs) are defined for quantitative analysis (e.g., eyes, mouth, geometric shapes, target objects) [2] [28].
• Key Metrics:
  • Fixation Duration: Total time spent looking within an AOI.
  • Fixation Count: Number of discrete looks to an AOI.
  • Saccades: Number of rapid eye movements per second.
  • First Fixation Duration: Duration of the initial look at an AOI.
• Data Analysis: Proportion of fixation time on different AOIs is calculated. Subgroups can be defined using thresholds (e.g., >69% fixation on geometric images defines the GeoPref subtype) [25] or via machine learning classifiers like Support Vector Machines (SVM) to discriminate ASD from typical development [28].

Linking Neural Subtypes to Behavior

The true power of subtyping lies in connecting neural profiles to behavioral outcomes. The following diagram synthesizes the relationships between the identified neural subtypes and their corresponding behavioral and sensory profiles, providing an integrated model for understanding ASD heterogeneity.

The Scientist's Toolkit: Essential Research Reagents and Solutions

The following table catalogs key materials and tools essential for conducting research in ASD subtyping, as derived from the experimental protocols cited.

Table 2: Key Research Reagents and Solutions for ASD Subtyping Studies

Item	Specification / Example	Primary Function in Research
Eye Tracking System	Tobii TX300, SMI RED250 [2] [28]	High-precision, non-invasive measurement of gaze patterns and visual preference in response to controlled stimuli.
Stimulus Presentation Software	Tobii Pro Studio, E-prime 2.0 [2]	Precisely control the display of experimental paradigms (videos, images) and synchronize with eye-tracking data acquisition.
fMRI Scanner	3T MRI Scanner	Acquisition of high-resolution structural (T1-weighted) and functional (rs-fMRI) brain imaging data.
fMRI Preprocessing Pipeline	fMRIPrep [2]	Standardized, automated preprocessing of raw fMRI data to correct for motion, normalize space, and prepare for feature extraction.
Brain Atlas	Dosenbach 160 ROI Atlas [2]	A predefined map of brain regions used to extract average BOLD signals for functional connectivity analysis.
Clinical Assessment Tools	ADOS, ADI-R, SRS, CARS [2] [28]	Gold-standard instruments for diagnosing ASD and quantifying symptom severity across social communication and repetitive behavior domains.
Clustering Algorithm	HYDRA (semi-supervised) [1]	Advanced computational method for identifying distinct data-driven subgroups within a heterogeneous population (e.g., ASD) using high-dimensional features.

The research protocols outlined yield distinct yet complementary insights. The following table synthesizes key quantitative findings from the cited studies, allowing for a direct comparison of outcomes across subtyping approaches.

Table 3: Summary of Key Quantitative Findings from ASD Subtyping Studies

Study Focus	Key Metric	Result / Value	Implication
Functional Connectivity Subtyping [1]	Number of Robust Subtypes	2 (Hyper-connectivity & Hypo-connectivity)	Confirms a fundamental bipartite division in ASD neural architecture.
Geometric Preference Subtyping [25]	Diagnostic Specificity	98%	High confidence in identifying ASD when geometric preference is extreme.
Geometric Preference Subtyping [25]	Diagnostic Sensitivity	21%	Identifies a specific, more severe ASD subgroup, not the entire spectrum.
SPS & Medication Sensitivity [26]	Correlation Coefficient (r)	0.21 - 0.36	Establishes a significant, moderate link between a sensory trait and medication response.
Eye-Tracking Machine Learning [28]	Classification Accuracy (Toddlers)	80%	Demonstrates strong potential of eye-tracking as an early objective screening tool.

The evidence confirms that ASD can be stratified into biologically and behaviorally meaningful subtypes. Functional connectivity subtyping reveals two primary neural profiles with opposing connectivity patterns that are associated with distinct gaze behaviors, despite similar clinical presentations [2] [1]. Independently, eye-tracking identifies a GeoPref subtype, a behavioral subgroup with a well-defined and replicable biomarker—preferential fixation on geometric imagery—which is associated with significantly greater cognitive and symptom severity [24] [25]. Furthermore, the trait of Sensory Processing Sensitivity highlights that fundamental, biologically-based individual differences in environmental reactivity have measurable implications for areas like medication sensitivity, underscoring the need for personalized approaches in treatment beyond core ASD symptoms [26] [27].

For researchers and drug development professionals, these findings emphasize that subgroup stratification is not merely a theoretical exercise but a critical step towards precision medicine. Clinical trials and intervention studies may achieve greater efficacy by targeting specific ASD subtypes defined by these objective neural and behavioral markers. Future work should focus on integrating multi-modal data (fMRI, eye tracking, genetics) to further refine subtypes and develop robust, clinically applicable biomarkers.

The study of functional brain connectivity in Autism Spectrum Disorder (ASD) has long been characterized by a fundamental contradiction: numerous studies report conflicting evidence of both hyper-connectivity and hypo-connectivity across different brain networks and populations [30]. This inconsistency has hindered the identification of reliable biomarkers and the development of targeted interventions. The mesoscopic framework represents a paradigm shift that reconciles these contradictions by examining connectivity patterns at an intermediate scale—focusing on specific subsets of brain regions and their organized interactions—rather than seeking uniform whole-brain explanations [30]. This approach recognizes that ASD heterogeneity stems not from measurement error but from the existence of distinct neurobiological subtypes with unique functional connectivity profiles, each exhibiting characteristic patterns of both increased and decreased connectivity [2] [1].

Advanced analytical techniques now enable researchers to detect contrast subgraphs—maximally different mesoscopic connectivity structures between ASD and neurotypical populations—that coexist within the same individuals [30]. Furthermore, semi-supervised clustering methods leveraging large, multi-site datasets have successfully identified reproducible ASD subtypes characterized by distinct hyper-connected and hypo-connected profiles, despite comparable clinical presentations [2] [1]. This framework fundamentally transforms the interpretation of contradictory findings from problematic inconsistencies to expected manifestations of ASD's underlying neurobiological diversity.

Mesoscopic Connectivity Patterns in ASD Subtypes

Identified Neural Subtypes and Their Connectivity Signatures

Recent large-scale studies have consistently identified two primary ASD subtypes with opposing connectivity profiles. Analysis of 1,046 participants (479 with ASD, 567 typically developing) from ABIDE I and II datasets revealed two distinct neural ASD subtypes despite comparable clinical presentations [2] [5]:

• ASD Subtype 1: Characterized by positive deviations (hyper-connectivity) in the occipital network and cerebellar network, coupled with negative deviations (hypo-connectivity) in the frontoparietal network, default mode network, and cingulo-opercular network [2].

• ASD Subtype 2: Exhibited the inverse pattern, with hypo-connectivity in the occipital and cerebellar networks, and hyper-connectivity in the frontoparietal, default mode, and cingulo-opercular networks [2].

A separate study implementing semi-supervised clustering (HYDRA) on ∼1,800 individuals confirmed this fundamental division, identifying hyper-connectivity and hypo-connectivity subtypes with distinct between-network and within-network connectivity patterns [1]. The hyper-connectivity subtype shows hyper-connectivity within major large networks and mixed hyper/hypo-connectivity between networks, while the hypo-connectivity subtype displays the opposite connectivity patterns [1].

Developmental Trajectories of Mesoscopic Connectivity

The mesoscopic framework reveals that connectivity patterns evolve across development, with distinct profiles emerging in different age groups. Contrast subgraph analysis of children and adolescents with ASD has identified age-specific patterns of hyper- and hypo-connectivity [30]:

Table 1: Age-Specific Contrast Subgraph Patterns in ASD

Age Group	Hypo-Connectivity Patterns	Hyper-Connectivity Patterns
Children	• Frontal Lobe: Superior Frontal Gyrus (medial) with orbital regions and temporal areas• Temporal Lobe: Between temporal pole and inferior/middle temporal gyri	• Occipital Lobe: Middle Occipital Gyrus with Inferior Occipital Gyrus• Occipital-Parietal: Between calcarine, cuneus, lingual gyri and Superior Parietal Gyrus
Adolescents	• Fronto-Temporal: Inferior Frontal Gyri, insula, and temporal regions• Limbic: Amygdala and hippocampus• Posterior: Posterior Cingulate, precuneus, and cerebellar regions	• Occipital Lobe: Superior/Middle Occipital Gyri, lingual, calcarine, and cuneus• Cerebellar: Between cerebellar hemispheres (3, 8, 9, 10) and vermis (9, 10)

These findings demonstrate that while occipital hyper-connectivity and frontal-temporal hypo-connectivity represent core components of ASD connectivity architecture, their specific manifestations vary developmentally [30].

Methodological Framework for Mesoscopic Analysis

Experimental Protocols and Workflows

The mesoscopic framework employs sophisticated analytical pipelines that integrate multiple computational approaches. The following workflow illustrates the standardized protocol for contrast subgraph analysis:

Figure 1: Contrast Subgraph Analysis Workflow

The normative modeling approach for ASD subtyping follows a complementary pathway:

Figure 2: Normative Modeling and Subtyping Pipeline

Key Analytical Techniques

Contrast Subgraph Extraction

This technique identifies maximally different mesoscopic connectivity structures between ASD and typically developing groups through an optimization problem defined on a difference graph [30]. The algorithm detects multiple sets of regions that simultaneously show hyper-connectivity in one group and hypo-connectivity in the other, reframing previous contradictory findings as a complex interplay between both connectivity types [30].

Semi-Supervised Clustering (HYDRA)

HeterogeneitY through DiscRiminative Analysis (HYDRA) incorporates diagnosis labels (ASD vs. controls) to perform neuro-subtyping, demonstrating superior performance compared to unsupervised methods [1]. Combined with orthogonal projective non-negative matrix factorization (OPNNMF) for dimension reduction, this approach effectively handles high-dimensional functional connectivity data while avoiding overfitting [1].

Multilevel Functional Connectivity Assessment

Comprehensive connectivity characterization integrates both static functional connectivity strength (SFCS) via Pearson correlation and dynamic functional connectivity (DFCS and DFCV) via dynamic conditional correlation [2]. This multi-level approach captures both stable and time-varying connectivity properties, providing a more complete picture of functional network organization [2].

Quantitative Findings and Data Integration

Comparative Connectivity Profiles Across Studies

Table 2: Cross-Study Comparison of ASD Subtype Connectivity Patterns

Study	Sample Size	Method	Subtype 1 Findings	Subtype 2 Findings	Behavioral Correlations
Liu et al., 2025 [2]	1,046 participants (479 ASD)	Normative modeling + clustering	Hyper: Occipital, CerebellarHypo: Frontoparietal, DMN, Cingulo-opercular	Inverse pattern of Subtype 1	Distinct gaze patterns in eye-tracking tasks
Wang et al., 2025 [1]	~1,800 individuals	HYDRA (semi-supervised)	Hyper-connectivity within major networks; Mixed between networks	Opposite connectivity patterns	Varying correlations with core ASD symptoms
Contrast Subgraph Study [30]	137 males (57 ASD)	Contrast subgraph extraction	N/A (Group-level patterns)	N/A (Group-level patterns)	Classification accuracy: 75.8% (children), 78.2% (adolescents)

Validation Through Behavioral and Physiological Measures

The functional significance of mesoscopic connectivity subtypes is validated through distinct behavioral profiles. In an independent cohort of 21 ASD individuals, neural subtypes demonstrated distinct gaze patterns assessed by autism-sensitive eye-tracking tasks focusing on preference for social cues [2]. This crucial finding establishes a direct link between the identified neurobiological subtypes and meaningful behavioral differences in social attention.

Dynamic connectivity features also show significant clinical correlations. Studies using functional near-infrared spectroscopy (fNIRS) have revealed that children with ASD show reduced dwell time in specific brain states and fewer state transitions, patterns that negatively correlate with autism symptom severity and positively correlate with adaptive behavior and cognitive performance [31]. Furthermore, dynamic connectivity features achieved 74.4% accuracy in distinguishing ASD from typically developing children, underscoring their diagnostic potential [31].

Table 3: Key Research Reagents and Computational Tools for Mesoscopic Connectivity Analysis

Resource Category	Specific Tool/Resource	Function/Application
Data Resources	ABIDE I & II Datasets	Multi-site resting-state fMRI data from ASD and control participants [2] [1]
Computational Tools	Contrast Subgraph Algorithm	Identifies maximally different mesoscopic connectivity structures between groups [30]
Analytical Frameworks	HYDRA (Semi-supervised clustering)	Neuro-subtyping incorporating diagnostic labels for enhanced separation [1]
Connectivity Measures	Dynamic Conditional Correlation (DCC)	Quantifies instant dynamic functional connectivity strength and variability [2]
Validation Methods	Eye-tracking (Social cue tasks)	Provides behavioral validation of neural subtypes through gaze patterns [2]
Normative Modeling	Lifespan trajectories	Establishes normative developmental trajectories for functional connectivity [2]

Implications for Drug Development and Personalized Interventions

The mesoscopic framework directly addresses the pharmaceutical industry's challenge in developing treatments for ASD's heterogeneous population. By identifying biologically distinct subgroups, this approach enables:

• Targeted Participant Selection: Clinical trials can enroll specific ASD subtypes most likely to respond to mechanism-based treatments, reducing trial failure rates [2] [1].

• Biomarker-Driven Endpoints: Functional connectivity measures provide objective neurobiological endpoints for treatment efficacy, complementing behavioral observations [31].

• Personalized Intervention Strategies: The framework underscores the importance of moving beyond one-size-fits-all approaches to interventions tailored to individual neural connectivity profiles [2].

Evidence for subtype-specific treatment response comes from a recent subtyping study that found one ASD subtype exhibited a 61.5% response rate to chronic intranasal oxytocin treatment, while the other subtype demonstrated only a 13.3% response [2]. This dramatic difference highlights the critical importance of subtype identification for pharmacological development.

The mesoscopic framework for understanding hyper- and hypo-connectivity in ASD represents a fundamental advance in neuropsychiatry research. By reconciling previously contradictory findings through the identification of distinct neural subtypes, this approach provides a more nuanced and clinically relevant understanding of ASD's neurobiology. The consistent identification of subtypes with opposing connectivity profiles across large, independent datasets suggests robust underlying biological mechanisms waiting to be fully characterized.

For researchers and drug development professionals, this framework offers concrete methodologies for subgroup identification, validated behavioral correlations, and a path toward personalized interventions. As these approaches continue to be refined and integrated with genetic and other molecular data, they hold significant promise for transforming how ASD is understood, diagnosed, and treated.

Methodological Frontiers: From Data Acquisition to Subtype Classification

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in clinical symptoms, brain organization, and developmental trajectories, presenting a substantial challenge for developing unified biological models and targeted interventions [2] [1]. This heterogeneity has prompted a paradigm shift from traditional case-control approaches toward data-driven subtyping frameworks that decompose the autism spectrum into more biologically homogeneous subgroups [1] [12]. The integration of multimodal data—particularly resting-state functional magnetic resonance imaging (rs-fMRI), eye-tracking, and clinical metrics—has emerged as a powerful approach for identifying robust ASD subtypes with distinct neurobiological profiles and behavioral manifestations [2] [32]. This comparison guide objectively evaluates the leading methodological frameworks for ASD subtyping, their technical performance, and clinical translatability for researchers, scientists, and drug development professionals.

Comparative Analysis of ASD Subtyping Approaches

Table 1: Performance Comparison of Primary ASD Subtyping Methodologies

Subtyping Approach	Core Methodology	Sample Size	Subtypes Identified	Key Distinguishing Features	Accuracy/Reliability Metrics
Semi-supervised Clustering (HYDRA)	Diagnosis-guided clustering with high-dimensional feature reduction [1]	847 ASD + 1030 TC	2: Hyper-connectivity & Hypo-connectivity	Distinct within- and between-network FC patterns; differential neurobehavioral correlations [1]	Superior to unsupervised methods; high test-retest reliability
Normative Modeling with Multilevel FC	Individual deviation from typical developmental trajectories [2]	479 ASD + 567 TD (Discovery); 21 ASD (Validation)	2: Occipital/Cerebellar+ & Frontoparietal/DMN- (and inverse) [2]	Unique functional brain network profiles despite comparable clinical symptoms [2]	Associated with distinct gaze patterns in eye-tracking tasks (independent cohort)
Unsupervised Clustering (K-means)	Data-driven partitioning without diagnostic labels [1]	Varies by study	2-4 subtypes typically reported	Moderate association with ASD diagnosis [12]	Lower performance compared to semi-supervised approaches [1]
Geometric Preference Eye-Tracking	Social vs. non-social visual preference [33]	49 toddlers with ASD (1-3 years)	2: Geometric-preferers & Social-preferers	Pronounced preference for geometric images (>69% viewing time) associated with greater symptom severity [33]	98% specificity for ASD; predicts long-term symptom severity

Table 2: Neurobiological Profiles of Identified ASD Subtypes

Subtype Classification	Functional Connectivity Profile	Behavioral & Clinical Correlates	Eye-Tracking Profile	Prognostic Implications
Hyper-connectivity Subtype [1]	Hyper-connectivity within major networks; Hyper-connectivity between DMN and attention; Hypo-connectivity between DMN and visual/auditory networks [1]	Distinct relationships between connectivity patterns and core ASD symptoms [1]	Not specifically assessed	Potential for individualized treatment targeting specific network dysregulation
Hypo-connectivity Subtype [1]	Inverse of hyper-connectivity pattern: Hypo-connectivity within major networks [1]	Different neurobehavioral relationships compared to hyper-connectivity subtype [1]	Not specifically assessed	May require different intervention approaches than hyper-connectivity subtype
Occipital/Cerebellar+ Subtype [2]	Positive deviations in occipital and cerebellar networks; Negative deviations in frontoparietal, DMN, and cingulo-opercular networks [2]	Comparable clinical presentation to other subtype but distinct neural basis [2]	Distinct gaze patterns in social cue tasks [2]	Personalized intervention strategies needed despite similar symptoms
Geometric-Preferring Subtype [33]	Not assessed in study	Greater ASD symptom severity; Fewer gaze shifts at school age [33]	>69% viewing time on geometric images [33]	Prognostic marker for more severe symptom trajectory

Experimental Protocols and Methodologies

Semi-Supervised Neuro-Subtyping Framework (HYDRA)

The HYDRA (HeterogeneitY through DiscRiminative Analysis) framework represents a advanced approach for identifying robust ASD subtypes by incorporating diagnostic labels into a clustering paradigm [1].

Participant Cohort:

• 847 individuals with ASD and 1030 typically controls from ABIDE I and II datasets
• Multicenter data with standardized inclusion/exclusion criteria
• Head motion exclusion: >2mm maximum displacement or >2° rotation [1]

Image Acquisition and Preprocessing:

• Resting-state fMRI and T1-weighted structural images
• Preprocessing: Standardized pipeline including head motion correction, spatial normalization, and global signal regression
• Functional connectivity matrices: Pearson correlation between regional time series [1]

Feature Reduction and Clustering:

• High-dimensional FC features reduced via Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF)
• Semi-supervised clustering using HYDRA with ASD-control labels as guiding information
• Optimal parameters determined through systematic evaluation (M=1195 components, K=2 clusters) [1]

Validation Approach:

• Test-retest reliability assessment
• Comparison with unsupervised K-means clustering with various feature reduction methods
• Examination of neurobehavioral correlations within subtypes [1]

Normative Modeling with Multilevel Functional Connectivity

This approach identifies ASD subtypes by quantifying individual deviations from typical developmental trajectories of functional connectivity [2].

Participant Cohorts:

• Discovery sample: 479 ASD, 567 typical development from ABIDE-I and ABIDE-II
• Validation sample: 21 ASD with rs-fMRI and eye-tracking data
• Multisite data harmonization with rigorous quality control [2]

Multilevel Functional Connectivity Features:

• Static Functional Connectivity Strength (SFCS): Pearson correlation between node time series
• Dynamic Functional Connectivity Strength (DFCS): Dynamic Conditional Correlation (DCC)
• Dynamic Functional Connectivity Variance (DFCV): Variability in instant connectivity [2]
• 160 regions of interest from Dosenbach atlas covering multiple cognitive domains

Normative Modeling Framework:

• Development of normative models using TD group multilevel FC features
• Quantification of individual-level deviations in ASD participants
• Clustering analysis of deviation profiles to identify subtypes [2]

Eye-Tracking Validation:

• Two autism-sensitive tasks: Face emotion processing and joint attention
• Areas of interest (AOI) defined for facial features (eyes, nose, mouth) and objects
• Tobii TX300 system at 300Hz sampling rate with 0.4° gaze accuracy [2]

Geometric Preference Eye-Tracking Paradigm

The GeoPref Test provides a behavioral subtyping approach based on fundamental differences in visual attention patterns [33].

Participant Characteristics:

• 49 toddlers with ASD (1-3 years old) followed longitudinally for 5-9 years
• School-age assessment of symptom severity, social functioning, and cognitive measures

Experimental Protocol:

• Stimuli: Competing dynamic geometric patterns and social scenes (children dancing)
• Presentation: 30-60 second trials with counterbalanced left-right positioning
• Viewing time quantification for social vs. geometric stimuli [33]

Subtype Classification:

• Geometric-preferring: >69% viewing time on geometric images
• Social-preferring: >69% viewing time on social images
• This cutoff established 98% specificity for ASD in previous studies [33]

Longitudinal Outcome Measures:

• Autism Diagnostic Observation Schedule (ADOS) for symptom severity
• Joint attention gaze shifts during eye-tracking tasks
• Social Responsiveness Scale (SRS), adaptive behavior, and IQ measures [33]

Visualization of Research Workflows

Diagram 1: Comprehensive Workflow for Multimodal ASD Subtyping. This diagram illustrates the integration of rs-fMRI, eye-tracking, and clinical data through three primary methodological approaches for identifying ASD subtypes with distinct neurobiological and behavioral profiles.

Table 3: Critical Research Resources for Multimodal ASD Subtyping Studies

Resource Category	Specific Tools & Measures	Primary Application	Key Advantages
Neuroimaging Datasets	ABIDE I & II (n=1046: 479 ASD/567 TD) [2]	Large-scale discovery and validation	Multisite data with phenotypic characterization
Eye-Tracking Paradigms	GeoPref Test (social vs. geometric) [33]	Behavioral subtyping and prognosis	98% ASD specificity; prognostic utility
Eye-Tracking Paradigms	Face emotion processing & Joint attention tasks [2]	Social attention assessment	Links neural subtypes to behavioral manifestations
Clinical Instruments	ADOS, ADI-R, SRS [2] [33]	Symptom characterization	Gold-standard ASD diagnostic measures
Analysis Frameworks	HYDRA semi-supervised clustering [1]	Neuro-subtyping	Incorporates diagnostic labels for improved sensitivity
Analysis Frameworks	Normative modeling [2]	Individual deviation quantification	Maps individual variation against typical development
Software Tools	fMRIPrep [2]	Neuroimaging preprocessing	Standardized, reproducible pipeline
Validation Approaches	Independent cohort replication [2]	Subtype validation	Ensures generalizability beyond discovery sample

The integration of resting-state fMRI, eye-tracking, and clinical metrics has substantially advanced the decomposition of ASD heterogeneity into biologically meaningful subtypes. Semi-supervised approaches like HYDRA demonstrate superior performance in identifying robust functional connectivity subtypes with distinct neurobehavioral profiles, while normative modeling effectively captures individual deviations from typical developmental trajectories [1] [2]. Eye-tracking paradigms, particularly the GeoPref Test, offer complementary behavioral subtyping with strong prognostic value [33]. The consistent identification of hyper/hypo-connectivity subtypes across studies suggests a fundamental organizational principle in ASD neurobiology, though differential network involvement highlights the need for network-specific analyses [1] [12]. For drug development professionals, these subtyping frameworks offer promising pathways for patient stratification in clinical trials and development of targeted interventions aligned with specific neurobiological profiles. Future research should prioritize prospective validation of these subtypes and their differential treatment responses to realize the potential of precision medicine in ASD.

The study of Autism Spectrum Disorder (ASD) is fundamentally challenged by the condition's significant clinical and biological heterogeneity. Identifying reproducible subtypes within the ASD population is crucial for advancing our understanding of its neurobiological underpinnings and moving toward personalized interventions [2] [34]. Neuroimaging, particularly functional connectivity (FC) analysis, has emerged as a powerful tool for probing this heterogeneity. This guide objectively compares three prominent computational approaches—HYDRA, Normative Modeling, and Hierarchical Clustering—for identifying ASD subtypes based on functional connectivity, summarizing their methodologies, performance, and applicability for researchers and drug development professionals.

The table below provides a high-level comparison of the three core methodologies.

Table 1: Core Methodological Comparison of HYDRA, Normative Modeling, and Hierarchical Clustering

Feature	HYDRA (Heterogeneity through Discriminative Analysis)	Normative Modeling	Hierarchical Clustering
Core Principle	Semi-supervised learning that separates patients from controls using a convex polytope, with each face defining a subtype [35].	Supervised framework that maps individual deviations from a normative trend of healthy brain development [2] [36].	Unsupervised learning that groups data points (subjects) based on similarity in functional connectivity patterns [37] [38].
Primary Objective	Integrated binary classification and disease subtype discovery [35] [39].	Quantification of individual-level deviations from a normative reference to identify extreme patterns [2].	Exploratory data analysis to find natural groupings or patterns without a priori labels [37].
Typical Input Features	Regional brain morphometry, neurite density, intracortical myelination [39].	Static/dynamic functional connectivity strength and variance [2].	Voxel-wise or region-wise temporal correlation matrices from rs-fMRI [37] [38].
Key Outputs	Discrete neuroanatomical or functional subtypes; Polytope faces representing subtype boundaries [35].	Continuous centile scores representing degree of deviation from normative trajectory; Subtypes based on deviation patterns [2] [36].	Dendrogram showing nested cluster relationships; Discrete cluster labels for subjects [37].

Experimental Protocols and Workflows

HYDRA (Heterogeneity through Discriminative Analysis)

HYDRA is a semi-supervised algorithm designed to simultaneously classify patients versus controls and identify disease subtypes within the patient group. Its workflow can be summarized as follows:

• Input Data Preparation: Features are derived from neuroimaging data, such as regional cortical thickness, surface area, and other measures of brain structure or function. These features are harmonized to account for multi-site acquisition differences [39].
• Model Training and Subtyping: HYDRA separates a patient group (e.g., with mood and anxiety disorders) from a healthy control group using a convex polytope—a geometric shape formed by multiple linear max-margin classifiers (hyperplanes). Each face of this polytope effectively defines a distinct patient subtype. Patients are clustered based on their association with a specific hyperplane [35] [39].
• Validation: The optimal number of clusters (subtypes) is determined via cross-validation using metrics like the Adjusted Rand Index (ARI). Significance is tested through permutation testing to ensure robustness [39].

Figure 1: HYDRA combines control and patient data to train a convex polytope classifier that outputs distinct subtypes.

Normative Modeling for Functional Connectivity

This approach characterizes heterogeneity by modeling how individuals deviate from a normative pattern of brain development.

• Normative Cohort: A large, typically developing (TD) cohort is used to build a model that predicts functional connectivity features (e.g., static/dynamic FC strength) based on age, capturing the non-linear trajectory of normal brain maturation [2] [36].
• Deviation Calculation: For each individual with ASD, their actual FC features are compared to the model's prediction. The output is a centile score that quantifies the degree and direction (positive or negative) of deviation [36].
• Subtype Identification: Individuals with ASD can then be clustered based on their spatial patterns of deviation across different brain networks. For instance, studies have identified subtypes characterized by opposite deviation patterns in networks like the frontoparietal network and default mode network [2].

Figure 2: Normative modeling maps developmental trajectories from a TD cohort to calculate centile deviations in ASD.

Hierarchical Clustering of Functional Connectivity

As a classic unsupervised method, hierarchical clustering seeks to find natural groupings in data without a pre-defined reference group.

• Similarity Matrix Construction: The process begins by computing a similarity matrix (e.g., using correlation coefficients) between the functional connectivity profiles of all subjects in the cohort, which can include only patients or a mixed group [37] [38].
• Tree Construction: A hierarchical tree (dendrogram) is built using a linkage algorithm (e.g., single link or nearest neighbor). This algorithm iteratively merges the most similar subjects or clusters until all data points are connected, visually representing the nested cluster relationships [38].
• Cluster Identification: The dendrogram is "cut" at a chosen level to yield discrete clusters, which are interpreted as potential subtypes. The number of clusters can be determined by statistical validation or domain knowledge [37].

Figure 3: Hierarchical clustering builds a dendrogram from all subject data, which is then cut to define clusters.

Performance and Application Data

The table below summarizes quantitative findings and applications of each method from key studies, providing a basis for performance comparison.

Table 2: Empirical Performance and Application in Neuroimaging Studies

Method	Cited Study & Population	Key Findings / Identified Subtypes	Performance / Validation Notes
HYDRA	Mood/Anxiety Youth (ABCD)N~1,931 patients, ~2,823 TD [39]	Identified 3 robust neuroanatomical subtypes: 1) Imbalanced cortical-subcortical maturation, 2) Delayed cortical maturation, 3) Atypical maturation with higher myelination. Subtypes differed in cognitive function and adversity burden.	Solution was robust to permutation testing (P < 0.05). Optimal cluster number (K=3) chosen via 5-fold cross-validation using Adjusted Rand Index.
Normative Modeling	ASD (ABIDE I/II)N=479 ASD, 567 TD [2]	Identified 2 distinct neural subtypes with unique FC deviation patterns despite similar clinical scores. One showed positive deviations in occipital/cerebellar networks, the other the inverse.	Subtypes were associated with different gaze patterns in an independent eye-tracking cohort (N=21), providing cross-modal validation.
Hierarchical Clustering	aMCI PatientsN=17 aMCI, 22 CN [37]	Identified functional clusters significantly different between aMCI and controls. The distribution of disconnected regions resembled altered memory networks from task-based fMRI.	Method was "guided" by group difference information (Z<-1.96, p<0.025) to ensure clusters reflected disease-related connectivity reductions.
Normative Modeling	ASD (ABIDE I/II & HCP-D)N=5-22 years [36]	Revealed a non-linear maturational trajectory in ASD: delayed hierarchy in childhood, "catch-up" in adolescence, and decline in young adulthood.	Used Generalized Additive Model for Location, Scale, and Shape (GAMLSS) for flexible normative modeling. Found persistent Default Mode Network segregation issues.

The Scientist's Toolkit: Essential Research Reagents and Materials

Successful implementation of these computational methods relies on specific data resources and software tools.

Table 3: Key Research Reagents and Resources for Functional Connectivity Subtyping

Resource / Solution	Type	Function in Research	Example Use Case
ABIDE I & II(Autism Brain Imaging Data Exchange)	Data Repository	Provides large-scale, publicly available aggregated datasets of resting-state fMRI, anatomical MRI, and phenotypic data for individuals with ASD and typical controls [2] [36].	Serves as the primary discovery and replication cohort for normative modeling and clustering studies in ASD [2] [34].
fMRIPrep	Software Tool	A robust, standardized pipeline for fMRI data preprocessing. Ensures consistency and reproducibility by handling tasks like motion correction, normalization, and noise removal [2].	Used as a standardized preprocessing protocol for studies employing normative modeling and other methods on ABIDE data [2].
HYDRA Algorithm(GitHub Repository)	Software Tool	A semi-supervised machine learning algorithm implemented in code for performing integrated classification and clustering [35] [39].	Applied to multimodal neuroimaging data (morphometry, myelination) to identify subtypes in youth with mood and anxiety disorders [39].
GAMLSS(Generalized Additive Model for Location, Scale, and Shape)	Statistical Model	Enables flexible normative modeling by capturing non-linear developmental trajectories and the variance around the mean, not just the average trend [36].	Used to model the non-linear trajectory of the principal functional connectome gradient in typical development and quantify deviations in ASD [36].
Dosenbach 160 Atlas	Brain Parcellation	A predefined set of 160 functional regions of interest (ROIs) derived from meta-analyses of various cognitive domains. Used to extract average BOLD signals for connectivity analysis [2].	Served as the ROI set for calculating multilevel functional connectivity features (static and dynamic) in a large-scale ASD subtyping study [2].

HYDRA, Normative Modeling, and Hierarchical Clustering offer distinct and complementary strengths for delineating ASD heterogeneity. HYDRA excels in jointly distinguishing patients from controls while identifying neurobiologically distinct subtypes in a semi-supervised framework. Normative Modeling is powerful for characterizing individual deviations from normal developmental trajectories, making it ideal for probing lifespan questions in ASD. Hierarchical Clustering remains a valuable, purely data-driven tool for initial exploratory analysis. The choice of method depends on the specific research question, data availability, and whether the goal is classification, dimensional assessment, or pure discovery. Future directions will likely involve integrating these approaches and linking the resulting subtypes more closely to genetic factors and treatment outcomes.

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by significant heterogeneity in clinical symptoms, underlying neurobiology, and behavioral presentations. This variability poses substantial challenges for developing effective, personalized interventions. The traditional diagnostic approach, which often treats ASD as a single entity, has limitations in addressing the diverse needs of individuals on the spectrum. Consequently, researchers have turned to neuroimaging-based subtyping to disentangle this heterogeneity by identifying more homogeneous subgroups based on distinct neural signatures.

Advanced feature extraction methods from resting-state functional magnetic resonance imaging (rs-fMRI) have emerged as powerful tools for parsing this neurobiological heterogeneity. These techniques enable researchers to quantify various aspects of brain function, organization, and dynamics that may distinguish ASD subtypes with different pathophysiological mechanisms, clinical trajectories, and treatment needs. By moving beyond conventional clinical assessments, these computational approaches offer the potential to identify biologically-based ASD subtypes that could inform more precise diagnostic and therapeutic strategies.

The most impactful feature extraction methodologies for ASD subtyping generally fall into three primary categories: static and dynamic functional connectivity analyses, amplitude-based measures of low-frequency fluctuations, and graph theory approaches that quantify brain network topology. Each method provides unique insights into brain organization and function, with a growing consensus that combining multiple approaches yields the most comprehensive understanding of ASD heterogeneity.

Comparative Analysis of Feature Extraction Methods

Table 1: Comparison of Primary Feature Extraction Methods for ASD Subtyping

Method Category	Specific Metrics	Biological Interpretation	ASD Subtyping Findings	Technical Considerations
Functional Connectivity (FC)	Static FC (sFC), Dynamic FC (dFC)	Strength and temporal variability of coordinated activity between brain regions	Two distinct subtypes: hyper-connected and hypo-connected patterns across major networks [1]	sFC assumes stationarity; dFC reveals temporal dynamics but requires careful parameter selection
Amplitude-Based Measures	ALFF, fALFF, dALFF, dfALFF	Regional intensity of spontaneous neural activity in low-frequency bands	Altered activity in medial temporal lobe, thalamus, and frontal regions; correlates with cognitive function [40] [41]	fALFF reduces noise sensitivity; dynamic versions capture temporal variability
Graph Theory	Nodal degree, betweenness centrality, path length	Architectural organization of brain networks (integration, segregation, efficiency)	Dissimilar alteration patterns in left vs. right TLE; limbic network showed 82.9% lateralization accuracy [42]	Provides both local (nodal) and global (network-wide) metrics
Hybrid Approaches	Normative modeling with multiple features	Individual deviations from typical neurodevelopmental trajectories	Identified two neural subtypes with opposite deviation patterns despite similar clinical presentations [2] [5]	Integrates multiple feature types for comprehensive subtyping

Table 2: Experimental Evidence for ASD Subtypes from Recent Studies

Study	Sample Size	Feature Extraction Methods	Subtypes Identified	Behavioral Correlations
Liu et al., 2025 [2]	1,046 participants (479 ASD)	Static/dynamic FC, normative modeling	Subtype 1: Positive deviations in occipital/cerebellar networks; negative in frontoparietal/DMN	Distinct gaze patterns in eye-tracking tasks despite similar clinical scores
Wang et al., 2025 [1]	~1,800 individuals	Semi-supervised clustering with FC	Hyper-connectivity vs. hypo-connectivity subtypes	Differential correlations between connectivity patterns and core ASD symptoms
Song et al., 2022 [40] [41]	57 unilateral TLE patients, 42 HCs	Six static IBA indicators + corresponding dynamic metrics	Not ASD-specific but demonstrated dynamic analysis reveals more pathological regions	Correlation with epilepsy duration and cognitive scores

Methodological Deep Dive: Protocols and Analytical Frameworks

Functional Connectivity Analysis Protocols

Functional connectivity analysis examines the temporal correlations between neural activity in different brain regions, providing insights into functional brain networks. The standard analytical pipeline begins with extensive fMRI preprocessing, including slice timing correction, head motion realignment, coregistration to structural images, normalization to standard stereotaxic space, and spatial smoothing. Nuisance signals (white matter, cerebrospinal fluid, global signal, and motion parameters) are typically regressed out, followed by temporal band-pass filtering (usually 0.008-0.09 Hz) to focus on spontaneous low-frequency fluctuations [43].

For static FC analysis, Pearson correlation coefficients are computed between the average blood-oxygen-level-dependent (BOLD) time series of predefined regions of interest (ROIs) across the entire scanning session, resulting in a symmetric connectivity matrix for each participant. The Dosenbach 160 ROI atlas is commonly used, encompassing networks involved in error processing, default mode, memory, language, and sensorimotor functions [2].

Dynamic FC analysis employs a sliding window approach to capture temporal variations in connectivity. The window length is a critical parameter, typically ranging from 30 seconds to 2 minutes. Within each window, correlation coefficients are computed, creating a time-varying estimate of connectivity strength (DFCS) and variability (DFCV) [2]. More advanced techniques like dynamic conditional correlation (DCC) can also be implemented to model instant dynamic FC [2].

Amplitude-Based Feature Extraction

Amplitude of low-frequency fluctuation (ALFF) measures the total power of spontaneous BOLD signal oscillations within the typical frequency range (0.01-0.1 Hz). Fractional ALFF (fALFF) represents the ratio of power in the low-frequency range to that in the entire frequency range detectable, which improves sensitivity and specificity by reducing noise from non-specific brain areas [40] [44].

The analytical protocol involves partitioning the preprocessed BOLD time series into multiple overlapping windows for dynamic analysis. For each window, ALFF is calculated as the square root of the power spectrum integrated across the low-frequency range, while fALFF is computed as the ratio of low-frequency power to the entire frequency power. Dynamic ALFF (dALFF) and dynamic fALFF (dfALFF) are then derived as the standard deviation or variability of these metrics across all windows [40] [41]. Research indicates that analyzing different frequency bands (slow-4: 0.027-0.08 Hz; slow-5: 0.01-0.027 Hz) can reveal frequency-dependent alterations in neurological populations [44].

Graph Theory Applications

Graph theory provides mathematical tools for quantifying the topological organization of brain networks. In this framework, the brain is modeled as a graph consisting of nodes (brain regions) and edges (functional connections between them). The analysis begins with defining the network nodes, typically using anatomical or functional atlases, followed by constructing association matrices based on correlation coefficients between node time series [42] [45].

Key graph metrics include nodal degree (number of connections to a node), betweenness centrality (proportion of shortest paths passing through a node), clustering coefficient (measure of local interconnectedness), and path length (average shortest distance between nodes). These metrics capture different aspects of network organization, including integration, segregation, and efficiency of information transfer [42]. In subtyping applications, these metrics are often calculated for specific networks of interest, such as the default mode network (DMN), attention network (AN), limbic network (LN), sensorimotor network (SMN), and visual network (VN) [42].

Integrated and Normative Modeling Approaches

Recent advances in ASD subtyping have moved toward integrating multiple feature types and employing normative modeling frameworks. Normative modeling establishes typical developmental trajectories of brain features using large samples of typically developing individuals, then quantifies how individuals with ASD deviate from these expected patterns [2]. This approach can incorporate both static and dynamic functional connectivity features to delineate multi-level functional developmental trajectories.

Semi-supervised clustering methods like HeterogeneitY through DiscRiminative Analysis (HYDRA) incorporate diagnostic labels (ASD vs. controls) during the subtyping process, often demonstrating superior performance compared to unsupervised approaches [1]. These methods typically employ dimension reduction techniques like orthogonal projective non-negative matrix factorization (OPNNMF) to handle high-dimensional feature spaces before clustering [1].

Table 3: Key Research Reagents and Computational Tools for Feature Extraction

Resource Category	Specific Tools/Resources	Application in Research	Key Functionality
Analysis Packages	fMRIPrep, CONN toolbox, SPM12, GIFT	Data preprocessing, feature extraction, statistical analysis	Standardized preprocessing pipelines; FC, ALFF, and graph theory metrics calculation
Reference Atlases	Dosenbach 160, Power 264, AAL	Region of interest definition	Provide standardized parcellations for consistent ROI definition across studies
Clustering Algorithms	HYDRA, K-means, normative modeling	Identification of ASD subtypes	Group individuals based on neural feature similarity; semi-supervised approaches incorporate diagnostic labels
Validation Tools	Eye-tracking (Tobii TX300), ADOS, SRS	Behavioral correlation and subtype validation	Link neural subtypes to behavioral measures and clinical symptoms
Databases	ABIDE I & II, NKI/Rockland Sample	Data source for discovery and validation	Provide large-scale, multi-site neuroimaging data for robust subtyping

Implications for Research and Therapeutic Development

The identification of biologically-based ASD subtypes through advanced feature extraction methods has significant implications for both research and clinical practice. From a research perspective, these approaches enable a more nuanced understanding of the neurobiological mechanisms underlying ASD heterogeneity, potentially linking specific neural profiles to genetic markers, environmental factors, and developmental trajectories.

For therapeutic development, neuro-subtyping offers the promise of personalized intervention strategies. The distinct functional connectivity profiles observed in different ASD subtypes suggest they may respond differently to various treatments. For instance, a study cited in the search results found that one ASD subtype demonstrated a 61.5% response rate to chronic intranasal oxytocin treatment, while another subtype showed only a 13.3% response rate [2]. This highlights the potential for targeting interventions based on an individual's neural subtype rather than relying on a one-size-fits-all approach.

Furthermore, the temporal dynamics captured by dynamic feature extraction methods may serve as sensitive biomarkers for tracking treatment response and disease progression. The ability to quantify how brain dynamics change in response to interventions could provide valuable insights for optimizing therapeutic strategies and dosing regimens in clinical trials.

As these methodologies continue to evolve, we can anticipate more refined subtype classifications that integrate multimodal data, including genetic, neuroimaging, and clinical behavioral measures. This multidimensional approach will likely yield increasingly precise biomarkers for stratifying ASD populations and developing targeted therapies that address the specific pathophysiological processes underlying each subtype.

Autism Spectrum Disorder (ASD) is characterized by substantial phenotypic and neurobiological heterogeneity, which complicates diagnosis and treatment [2] [1]. Traditional categorical diagnoses are increasingly being supplemented by data-driven subtyping approaches that seek to identify more homogeneous subgroups. This guide compares two pivotal methodological paradigms in ASD functional connectivity (FC) research: normative modeling and semi-supervised clustering. These approaches exemplify the shift from seeking discrete diagnostic categories to mapping individuals onto continuous spectra of neural deviation, with significant implications for personalized intervention and drug development [2] [5] [1].

Comparative Analysis of Key Studies

The following tables summarize the core characteristics, findings, and implications of two major studies representing each methodological approach.

Aspect	Liu et al. (2025) - Normative Modeling Approach [2] [5]	Wang et al. (2025) - Semi-Supervised Clustering (HYDRA) [1]
Core Methodology	Normative modeling of multilevel FC against a typical development (TD) cohort.	Semi-supervised heterogeneity through discriminative analysis (HYDRA) using diagnosis labels.
Primary Dataset	ABIDE I & II (N=1046; ASD=479, TD=567) [2].	ABIDE I & II (N=1877; ASD=847, TD=1030) [1].
Validation Cohort	Independent cohort (UESTC; ASD=21) with eye-tracking [2].	Internal reliability and robustness analyses [1].
Key Feature	Quantifies individual deviations from a normative FC trajectory.	Directly clusters individuals into subgroups informed by diagnosis.
Conceptual Frame	Continuous Spectrum: Places individuals along axes of deviation.	Discrete Categories: Identifies distinct clusters (subtypes).

Table 2: Identified Subtype Profiles & Behavioral Correlates

Subtype	Neural Profile (vs. Norm)	Associated Behavioral/Task Performance
ASD Subtype 1 (Liu et al.) [2]	Positive deviations: Occipital & Cerebellar networks. Negative deviations: Frontoparietal, Default Mode, Cingulo-Opercular networks.	Distinct gaze patterns in social cue eye-tracking tasks (face emotion, joint attention) [2].
ASD Subtype 2 (Liu et al.) [2]	Inverse pattern of Subtype 1 across the same networks.	Different gaze pattern profile compared to Subtype 1 [2].
Hyper-Connectivity Subtype (Wang et al.) [1]	Hyper-connectivity within major networks; Mixed hyper-/hypo-connectivity between networks (e.g., DMN-Attention hyper, DMN-Visual hypo).	Distinct correlations between connectivity patterns and core ASD symptoms [1].
Hypo-Connectivity Subtype (Wang et al.) [1]	Opposite connectivity patterns to the hyper-connectivity subtype.	Different neurobehavioral relationships compared to the hyper-connectivity subtype [1].

Table 3: Experimental Outcomes & Translational Potential

Outcome Metric	Normative Modeling Study [2]	Semi-Supervised Clustering Study [1]
Number of Subtypes	2	2
Clinical Symptom Correlation	Subtypes had comparable clinical scores (ADOS, SRS) despite neural differences [2].	Subtypes showed varying correlations between FC and symptom severity [1].
Validation Method	Independent cohort with multimodal data (eye-tracking) [2].	Internal cluster reliability and comparison to unsupervised methods [1].
Personalized Intervention Insight	Suggests interventions targeting specific network deviations (e.g., gaze training for social attention subtypes) [2].	Indicates that treatment response may differ by connectivity subtype (e.g., hyper-vs. hypo-connectivity) [1].
Methodological Advantage	Captures individual-level, multidimensional deviation from a norm, ideal for continuous spectra.	Efficiently identifies robust, diagnosis-informed subgroups using high-dimensional data.

Detailed Experimental Protocols

Objective: To identify ASD subtypes based on individual deviations from typical brain functional development trajectories. Workflow:

• Data Acquisition & Preprocessing:
  • Dataset: Resting-state fMRI (rsfMRI) and phenotypic data from ABIDE I and II [2].
  • Participants: 1046 participants (479 ASD, 567 TD) after quality control (excessive head motion: mean FD > 0.3).
  • Preprocessing: Standardized pipeline using fMRIPrep, including normalization and nuisance regression.
• Feature Extraction:
  • Regions of Interest: Time-series extracted from Dosenbach's 160 ROI atlas.
  • Multilevel FC Features: Calculated for each participant:
    • Static Functional Connectivity Strength (SFCS): Pearson correlation.
    • Dynamic Functional Connectivity Strength (DFCS) & Variance (DFCV): Derived using Dynamic Conditional Correlation.
• Normative Model Construction:
  • Built separately for each FC feature using data from the TD group (N=567).
  • Models predicted expected FC based on age/sex, establishing a normative developmental trajectory.
• Deviation Quantification & Clustering:
  • For each ASD individual, a z-score was computed for each FC feature, representing deviation from the TD norm.
  • Clustering analysis (unspecified algorithm) was applied to the matrix of deviation scores across all features and ROIs to identify ASD subtypes.
• Validation with Eye-Tracking:
  • An independent cohort (21 ASD, 15 TD) underwent two eye-tracking tasks:
    • Face Emotion Processing: Observing static faces; AOIs: eyes, nose, mouth.
    • Joint Attention: Watching videos of gaze/pointing cues; AOIs: target vs. non-target objects.
  • Eye-tracking metrics (fixation duration, count) were compared between neural subtypes.

Objective: To discover robust ASD subtypes by incorporating diagnostic labels into the clustering of high-dimensional FC data. Workflow:

• Data Preparation:
  • Dataset: rsfMRI from ABIDE I and II.
  • Participants: 1877 subjects (847 ASD, 1030 TD) after rigorous motion exclusion (max displacement >2mm or rotation >2°).
  • Preprocessing: Included slice-timing correction, realignment, normalization, and smoothing.
• Feature Generation & Reduction:
  • FC Matrix: Full correlation matrices were computed for all brain regions (parcellation not specified).
  • Dimensionality Reduction: Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF) was applied to reduce high-dimensional FC features. The optimal number of components (M=1195) was determined via parameter selection.
• Semi-Supervised Clustering:
  • The reduced features and diagnosis labels (ASD vs. TD) were input into the HYDRA algorithm.
  • HYDRA performs discriminative analysis to find subtypes within the ASD group that are maximally distinct from the TD group and from each other.
  • The optimal number of clusters (K=2) was determined by evaluating clustering performance metrics.
• Cluster Characterization & Validation:
  • Reliability: Clustering robustness was assessed.
  • Profile Analysis: FC patterns (within- and between-network connectivity) were characterized for each subtype (hyper- vs. hypo-connectivity).
  • Neurobehavioral Correlation: Relationships between subtype-specific FC patterns and core ASD symptoms (e.g., from ADOS) were examined.

Visualizing Methodological Workflows

Normative Modeling Subtyping Pipeline

Semi-Supervised Clustering (HYDRA) Pipeline

The Scientist's Toolkit: Key Research Reagent Solutions

Category	Item / Solution	Function in ASD FC Subtyping Research
Data & Biobank	ABIDE Consortium Datasets (I & II) [2] [1]	Provides large-scale, multi-site rsfMRI and phenotypic data essential for training and validating subtyping models, ensuring statistical power and generalizability.
Neuroimaging Analysis	fMRIPrep Pipeline [2]	Standardized, reproducible preprocessing software for rsfMRI data, handling steps from motion correction to normalization, critical for ensuring data quality and comparability across sites.
Feature Extraction	Dosenbach 160 ROI Atlas [2]	A functionally defined parcellation scheme used to extract regional time-series, serving as the basis for calculating static and dynamic functional connectivity matrices.
Normative Benchmarking	Normative Modeling Framework [2]	A statistical toolset to model typical brain development trajectories (using TD data) and quantify individual deviations (z-scores), enabling a continuous spectrum view of neural atypicality.
Clustering Algorithm	HYDRA (HeterogeneitY through DiscRiminative Analysis) [1]	A semi-supervised clustering algorithm that uses diagnostic labels to guide the discovery of subgroups within a patient population, improving separation from controls and cluster robustness.
Dimensionality Reduction	Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF) [1]	A feature reduction technique applied to high-dimensional FC data before clustering. It extracts lower-dimensional, interpretable components while reducing noise and risk of overfitting.
Behavioral Phenotyping	Autism-Sensitive Eye-Tracking Tasks (e.g., Face Emotion, Joint Attention) [2]	Provides quantifiable, objective measures of social attention and information processing, used to validate neural subtypes by linking them to distinct behavioral phenotypes.
Clinical Assessment	Autism Diagnostic Observation Schedule (ADOS) & Social Responsiveness Scale (SRS) [2] [1]	Gold-standard clinical tools for assessing core ASD symptoms. Used to evaluate symptom profiles across subtypes and correlate neural features with behavioral severity.
Dynamic FC Analysis	Dynamic Conditional Correlation (DCC) Modeling [2]	A method for estimating time-varying functional connectivity, providing metrics like dynamic strength and variability, which capture a richer, multilevel picture of brain dynamics.
Validation Tool	Independent, Multimodal Validation Cohort [2]	A separately recruited sample with both neuroimaging and additional biomarker data (e.g., eye-tracking). Critical for testing the generalizability and real-world relevance of identified subtypes.

The search for reliable neurobiological subtypes within Autism Spectrum Disorder (ASD) is a critical step towards personalized medicine. The heterogeneity in symptoms, genetics, and brain connectivity has long hindered the development of precise biomarkers and targeted interventions [1]. Among various methodological approaches, semi-supervised clustering has emerged as a powerful tool for disentangling this complexity by leveraging diagnostic labels to guide the discovery of data-driven subgroups. This case study focuses on one such successful application: the use of the semi-supervised clustering method HYDRA (HeterogeneitY through DiscRiminative Analysis) to identify robust functional connectivity subtypes in ASD [1] [8] [46]. We will frame this within the broader context of ASD subtyping research, comparing HYDRA's performance and outcomes with other common methodological alternatives.

HYDRA Methodology and Workflow

The successful application of HYDRA for ASD neuro-subtyping followed a multi-stage, systematic protocol [1].

1. Data Acquisition and Preprocessing: The study utilized a large, multicenter dataset from the Autism Brain Imaging Data Exchange (ABIDE I and II), comprising resting-state functional MRI (rs-fMRI) data from approximately 2000 subjects (847 ASD, 1030 healthy controls) [1]. Key inclusion criteria were the availability of both T1-weighted and rs-fMRI images, with exclusions for excessive head motion (>2mm displacement or >2° rotation) or poor scan quality [1]. Preprocessing aimed to extract functional connectivity (FC) matrices, which served as the high-dimensional neural features for subtyping.

2. Feature Dimension Reduction: To address the high dimensionality of FC data and prevent overfitting, a multi-scale dimension reduction method called Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF) was implemented [1] [46]. This step transformed the high-dimensional FC features into a more robust and representative lower-dimensional feature space before clustering.

3. Semi-Supervised Clustering with HYDRA: The core analysis employed the HYDRA algorithm. Unlike purely unsupervised methods, HYDRA is a semi-supervised technique that uses the diagnostic labels (ASD vs. control) to inform the clustering process [1] [8]. It models the neuroimaging data of the control group as a single distribution, while modeling the ASD group as a mixture of several distinct distributions (subtypes), each representing a different pattern of deviation from the control norm.

4. Parameter Optimization and Validation: A three-layer procedure was conducted to determine the optimal number of dimension components (M) and the number of clusters (K). Validity and reliability indices were used to evaluate clustering performance. Based on this process, the parameters were set at M=1195 components and K=2 clusters [1]. The reliability of the identified subtypes was assessed using metrics like the silhouette score.

Performance Comparison: HYDRA vs. Alternative Methods

The study provided a direct comparison between the semi-supervised HYDRA approach and more traditional unsupervised methods, with quantitative results summarized in the table below.

Table 1: Comparative Performance of Clustering Methods for ASD Subtyping

Method	Clustering Type	Key Performance Metric	Reported Outcome	Identified Subtypes
HYDRA + OPNNMF [1]	Semi-supervised	Superior clustering performance & reliability	Demonstrated superior performance compared to unsupervised K-means. High reliability of clusters.	2 Subtypes: Hyper-connectivity & Hypo-connectivity
K-means Clustering [1]	Unsupervised	Baseline for comparison	Underperformed compared to the semi-supervised HYDRA approach.	Not specified (performance was inferior)
Normative Modeling + Clustering [2]	Unsupervised/Data-driven	Identified neural subtypes with distinct gaze patterns	Identified 2 neural subtypes with inverse FC deviation patterns, linked to different eye-tracking profiles.	2 Subtypes: Positive deviations in Occipital/Cerebellar nets & negative in FP/DMN/Cingulo-opercular nets, and the inverse pattern.
Clinical Feature-Based Clustering [47]	Unsupervised (on clinical scores)	Machine learning model accuracy for subtype classification	Achieved ~75% accuracy in classifying ASD into "more severe" (Cluster-1) and "moderate" (Cluster-2) subtypes based on rs-fMRI connectivity properties.	2 Subtypes: More severe impairment (Cluster-1) & Moderate impairment (Cluster-2)
Smile-GAN for ADHD [48] (Contextual Reference)	Semi-supervised, non-linear	Adjusted Random Index (ARI)	Identified 3 ADHD subtypes where HYDRA failed to show statistically significant results, highlighting method-specific sensitivities.	3 Subtypes: Under-developed, Over-developed, and Mixed cortical thickness patterns

Key Comparative Insights:

• HYDRA's Advantage: The primary study concluded that the semi-supervised HYDRA approach demonstrated superior performance in deriving distinct and reliable clusters compared to the commonly used unsupervised K-means method [1].
• Subtype Consistency: Notably, HYDRA's identification of two primary subtypes characterized by hyper-connectivity and hypo-connectivity finds echoes in other independent subtyping efforts. For instance, a normative modeling study also found two subtypes with opposing patterns of functional deviation across major brain networks [2].
• Methodological Divergence: The comparison with Smile-GAN, another semi-supervised method used in ADHD research, underscores that different algorithms may be suited to capturing different types of heterogeneity (e.g., linear deviations vs. non-linear patterns) [48].

Detailed Experimental Protocol for HYDRA-based Subtyping

For researchers seeking to replicate or build upon this work, the core experimental protocol is as follows:

• Cohort Assembly: Collect rs-fMRI and structural MRI data from a sufficiently large cohort of individuals with ASD and age-/sex-matched typically developing controls (TD). Publicly available resources like the ABIDE repository are suitable [1].
• Image Preprocessing: Process rs-fMRI data using standardized pipelines (e.g., fMRIPrep) for slice-timing correction, motion realignment, normalization to standard space (e.g., MNI152), and nuisance regression (e.g., global signal, white matter, CSF signals) [2].
• Functional Connectivity Matrix Generation: Extract time series from a predefined brain atlas (e.g., Dosenbach 160 ROIs). Compute pairwise Pearson correlation coefficients between all regional time series to create a subject-specific FC matrix [1] [2].
• Feature Reduction with OPNNMF: Apply Orthogonal Projective Non-Negative Matrix Factorization to the vectorized upper-triangular portions of the FC matrices from all subjects. This reduces the feature dimension while preserving discriminative information [1] [46].
• HYDRA Clustering: Input the reduced-dimension features and diagnostic labels (ASD/TD) into the HYDRA algorithm. Execute the model to determine the optimal number of subtypes (K) and assign each ASD participant to a subtype.
• Validation and Analysis:
  • Reliability: Assess clustering stability using internal validation indices (e.g., silhouette score) and test-retest reliability if data is available [1].
  • Characterization: Compare the FC profiles of the identified subtypes against each other and against the TD group. Analyze between-network and within-network connectivity patterns [1].
  • Neurobehavioral Correlation: Examine correlations between subtype-specific FC patterns and core ASD behavioral symptoms (e.g., from ADOS, SRS) to establish distinct brain-behavior relationships [1] [2].

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Materials and Tools for Functional Connectivity Subtyping Research

Item	Function / Description	Example/Reference
ABIDE Datasets (I & II)	Publicly shared neuroimaging repository providing rs-fMRI, sMRI, and phenotypic data for ASD and TD controls, enabling large-scale analysis.	Primary data source [1] [2].
fMRIPrep Pipeline	A robust, standardized software pipeline for automated preprocessing of diverse fMRI data, ensuring reproducibility.	Used for preprocessing in validation studies [2].
HYDRA Software Package	Implementation of the semi-supervised HeterogeneitY through DiscRiminative Analysis clustering algorithm.	Core clustering tool [1] [46].
OPNNMF Code	Implementation of Orthogonal Projective Non-Negative Matrix Factorization for feature dimension reduction.	Used in conjunction with HYDRA [1] [46].
Dosenbach 160 ROI Atlas	A predefined functional brain atlas comprising 160 regions of interest, used for extracting time series for FC calculation.	Used to define network nodes [2].
Clinical Phenotype Measures	Standardized assessment tools for quantifying ASD symptoms and cognitive abilities (e.g., ADOS, ADI-R, SRS, FIQ).	Essential for behavioral correlation and clinical stratification [1] [2] [47].
Graph Theoretical / GSP Toolboxes	Software libraries (e.g., Brain Connectivity Toolbox, custom GSP scripts) for calculating advanced network metrics beyond simple correlation.	Enables complementary analysis of network topology [49].

Visualization of Workflows and Findings

Navigating Methodological Challenges and Optimizing Subtyping Robustness

The quest to identify reproducible functional connectivity (FC) subtypes in Autism Spectrum Disorder (ASD) is fundamentally a data-intensive endeavor, necessitating the aggregation of neuroimaging datasets across multiple research sites [17] [12]. While this approach increases statistical power and enhances generalizability, it introduces significant technical and biological confounding factors that can obscure true brain-behavior relationships if left unaddressed [50] [51]. This comparison guide objectively evaluates the methodologies and tools for mitigating three primary confounds—site effects, head motion, and age—within the context of multicenter ASD FC subtyping research, providing a framework for robust and reproducible biomarker discovery.

Comparative Analysis of Confound Mitigation Strategies

The table below summarizes the core methodological approaches for handling each confounding factor, their underlying principles, key advantages, and limitations as evidenced by recent research.

Table 1: Comparison of Primary Methods for Addressing Confounds in Multicenter FC Studies

Confounding Factor	Method Category	Representative Technique(s)	Principle & Application	Key Advantages	Documented Limitations & Considerations
Site Effects	Batch Effect Harmonization	ComBat & CovBat [50]	Uses an empirical Bayes framework to adjust for additive and multiplicative site/scanner biases in derived features (e.g., FC strength, cortical thickness). Preserves biological variance of interest.	Effectively removes unwanted site variance; preserves covariate effects (e.g., age, diagnosis); widely adopted and validated.	Assumes linear site effects; may not handle site-by-covariate interactions well; risk of over-correction [50].
		Normative Modeling [50]	Models the expected distribution of a brain feature (e.g., FC) given covariates like age and sex across a healthy reference cohort. Individual deviations from the norm are calculated, reducing site-specific baseline differences.	Provides a personalized, deviation-based measure; naturally accounts for non-linear covariate effects.	Requires a large, high-quality normative sample; performance depends on reference cohort selection [50].
	Data-Driven Regression	Covariate Regression [17] [12]	Site and age are treated as nuisance variables and regressed out from the FC data (e.g., connectivity matrices) prior to downstream analysis like clustering.	Simple and straightforward to implement; directly controls for known variables.	Assumes linear relationships; may remove variance shared with signals of interest; less effective for complex, non-linear site effects [17].
	Deep Learning	Generative Adversarial Networks (GANs) [50]	Learn to translate neuroimaging data from one site to appear as if acquired at another, creating a harmonized feature space.	Potential to correct for complex, non-linear site differences at the image level.	"Black-box" nature reduces interpretability; requires large training datasets; validation frameworks are still evolving [50].
Head Motion	Quality Control & Censoring	Framewise Displacement (FD) & DVARS Thresholding	Identifies and removes (scrubs) individual volumes where head motion exceeds a predefined threshold (e.g., FD > 0.2mm).	Directly removes data segments most corrupted by motion artifact.	Reduces data quantity; can introduce bias if motion is systematic across groups [12].
	Regression-Based Correction	Motion Parameter Regression	The estimated 6-parameter rigid-body head motion timeseries are regressed out from the BOLD signal as a first- or second-level confound.	Standard part of preprocessing; reduces motion-related variance.	Incomplete correction; residual motion artifacts often remain correlated with neural signals.
	PCA-Based Waveform Correction [52]	Uses Principal Component Analysis to identify and subtract motion-induced waveform bias from kinematic or timeseries data in post-processing.	Effective for systematic biases; demonstrated in multicenter motion analysis [52].	More commonly applied in biomechanics; application to fMRI timeseries is an area of active research.
Age	Statistical Control	Covariate Regression [17] [12]	Age is regressed out from FC features to examine diagnosis- or subtype-related effects independent of developmental changes.	Essential for case-control matching; clarifies effects not attributable to typical development.	May over-control if age-related effects are part of the disorder's pathophysiology.
	Developmental Stratification	Within-Age-Band Analysis [53]	Conducts separate analyses for distinct developmental stages (e.g., children vs. adolescents) to account for non-linear FC changes.	Acknowledges and explores differential brain-behavior relationships across development.	Reduces sample size within each analysis; requires careful age-band definition.
	Normative Modeling [50]	As described for site effects, models the expected trajectory of a brain feature with age. Individual's deviation from their age-expected norm is used for analysis.	Accounts for non-linear developmental trajectories; provides a context for individual differences.	Same as above: requires a large normative sample.

Detailed Experimental Protocols from Key Studies

The following section details the specific protocols employed in seminal studies that successfully addressed these confounds in ASD FC research.

Protocol 1: Data-Driven FC Subtyping with Confound Regression

• Study Context: Easson et al. (2019) used k-means clustering to identify FC-based subtypes across ASD and typically developing (TD) participants [17].
• Confound Handling: Prior to clustering, the effects of age and acquisition site were regressed out of the individual FC matrices. This step was critical, as clustering on unadjusted data led to subtypes significantly confounded by site distribution (X² = 78.60, p < 0.001). After regression, subtypes differed only in age (t(264)=2.50, p=0.01), which had been intentionally preserved as a variable of interest for later dimensional analysis [17].
• Outcome: The study identified two robust FC subtypes that cut across diagnostic labels, characterized by divergent within- vs. between-network connectivity patterns. This approach revealed unique brain-behavior relationships within each subtype that were obscured in the whole-group analysis [17].

Protocol 2: The Two-Step Phantom Framework for Multicenter DTI

• Study Context: Zhu et al. (2012) proposed a quality assurance framework for multicenter Diffusion Tensor Imaging (DTI) studies, applicable to assessing site effects [51].
• Step 1 - Outlier Identification: Compute voxel-wise median maps of tensor-derived metrics (e.g., Fractional Anisotropy) across all phantom scans from all sites and time points. Subtract each individual scan from the median map. Scans showing large, systematic deviations are flagged as outliers for investigation or exclusion [51].
• Step 2 - Variance Component Analysis: Calculate intrasite variance (mean of variances within each site over time) and intersite variance (variance of the mean values across different sites). A significant intersite variance indicates a systematic bias between scanners that requires harmonization before pooling subject data [51].
• Application: This protocol allows for continuous monitoring of scanner stability and provides quantitative metrics to justify the use of post-hoc harmonization tools like ComBat.

Protocol 3: Diagnosis-Informed Neurosubtyping with Gradient Analysis

• Study Context: Hong et al. (2022) developed a supervised-unsupervised hybrid clustering framework to find ASD subtypes based on connectome gradient profiles [6].
• Confound Handling: The study employed functional random forest, a supervised algorithm, to select connectivity gradient features that were most informative for ASD vs. TD classification. This "diagnosis-informed" feature selection inherently prioritizes patterns related to core ASD anomalies over potential noise or site-specific artifacts. The resulting subtypes showed reproducible symptom profiles across independent datasets [6].
• Outcome: This method demonstrates how incorporating diagnostic labels into the feature engineering step can guide subtyping towards clinically relevant and reproducible neurobiological dimensions, potentially making the results more robust to unmeasured confounds.

Visualization of Methodological Workflows

Title: Overall Pipeline for Confound-Corrected FC Subtyping

Title: Decision Path for Site Effect Correction Methods

The Scientist's Toolkit: Essential Reagents & Solutions

Table 2: Key Research Tools for Multicenter ASD FC Studies

Tool Category	Specific Solution/Reagent	Primary Function in Context
Data Harmonization Software	NeuroComBat (Python/R)	Implements the ComBat algorithm for harmonizing site effects in neuroimaging features [50].
	CovBat	Extension of ComBat that additionally harmonizes covariance patterns across sites [50].
Preprocessing Pipelines	fMRIPrep / HCP Pipelines	Automated, standardized preprocessing of fMRI data, reducing pipeline variability as a source of site effects.
	DPABI / CONN	Toolboxes that integrate preprocessing, QC, and connectivity analysis, often including motion correction utilities.
Quality Control Metrics	Framewise Displacement (FD)	Quantifies volume-to-volume head motion; used for scrubbing and inclusion/exclusion criteria [12].
	Distance-Dependent Correlation	Diagnostic plot to identify residual motion artifact in FC matrices.
Phantom Devices	ACR MRI Phantom	Standardized phantom used for longitudinal monitoring of scanner stability (e.g., signal-to-noise, geometric distortion) [51].
Clustering & Analysis	Scikit-learn (Python)	Provides implementations of k-means, hierarchical clustering, and other algorithms used for FC subtyping [17] [12].
Normative Reference Data	UK Biobank / HCP-Aging	Large-scale, publicly available datasets that can serve as reference populations for normative modeling of age and other effects [50].

The integration of rigorous confound mitigation protocols—from phantom-based scanner monitoring [51] to advanced harmonization techniques [50]—is non-negotiable for advancing ASD FC subtyping from descriptive clustering to the discovery of biologically grounded, clinically actionable subgroups. The comparative data presented here underscore that no single method is universally superior; rather, a layered approach tailored to the specific confounds and research question is essential for generating reproducible and meaningful results.

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in both clinical symptoms and neurobiological underpinnings. This variability presents a substantial challenge for developing effective, personalized interventions and reliable biomarkers. Neurosubtyping based on functional connectivity (FC) has emerged as a promising approach to decompose this heterogeneity into more homogeneous subgroups. Currently, two predominant methodological frameworks guide this effort: discrete subtyping, which categorizes individuals into distinct groups, and continuous subtyping, which characterizes individuals based on their position along a multidimensional spectrum. This review objectively compares the stability and clinical utility of these two paradigms within ASD functional connectivity research, synthesizing evidence from recent large-scale neuroimaging studies to inform future research and therapeutic development.

Theoretical Frameworks and Definitions

Discrete Subtyping in Neuroscience

Discrete subtyping is a categorical approach that assigns individuals to distinct, mutually exclusive subgroups based on their functional connectivity profiles. This method typically employs clustering algorithms—such as k-means, hierarchical clustering, or semi-supervised models—to partition a population into groups believed to share common neurobiological features [17] [1]. The core assumption is that a heterogeneous population, like individuals with ASD, comprises multiple more homogeneous subgroups, each potentially with distinct etiologies, neural circuitry, and treatment responses. For example, a common finding is the distinction between "hyper-connectivity" and "hypo-connectivity" subtypes, which show widespread differences in within- and between-network connectivity [1].

Continuous Subtyping and Dimensional Approaches

In contrast, continuous subtyping treats neurobiological variation as a spectrum, quantifying where an individual falls along one or more continuous dimensions. This approach includes methods like normative modeling, which quantifies individual deviations from a normative neurotypical trajectory, and continuous assignment to subtypes based on similarity metrics like spatial correlation [2] [12]. Rather than imposing hard category boundaries, continuous frameworks aim to preserve the granular, often non-categorical, nature of neurobiological variation. They fully leverage continuous biomarker measures without the information loss that accompanies categorization [54].

Fundamental Distinctions in Data Representation

The difference between these approaches mirrors the fundamental statistical distinction between discrete/categorical and continuous variables. Discrete variables represent counts and take on distinct, separable values, while continuous variables represent measurable amounts and can take on any value within a given range [55] [56]. In subtyping, this translates to whether researchers conceptualize ASD heterogeneity as comprising distinct "types" (discrete) or as varying along seamless "dimensions" (continuous).

Figure 1: Conceptual workflow comparing discrete versus continuous subtyping frameworks in ASD research.

Methodological Comparison of Experimental Protocols

Discrete Subtyping Protocols

Discrete subtyping methodologies typically follow a standardized workflow. First, high-dimensional functional connectivity data is extracted from resting-state fMRI using predefined brain atlases. Common protocols involve computing correlation matrices between regions of interest, often followed by dimension reduction techniques to mitigate the curse of dimensionality.

K-means clustering was applied to FC matrices from 266 participants after regressing out effects of age and acquisition site. The optimal cluster number (k=2) was determined using the elbow point criterion, validated via a bootstrapping procedure that confirmed this optimal number in 500/500 bootstrap samples [17].

Semi-supervised clustering with HYDRA (HeterogeneitY through DiscRiminative Analysis) represents a more advanced discrete approach. This method incorporates diagnostic labels (ASD vs. controls) to guide the clustering process while handling high-dimensional input features. Researchers applied Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF) for dimension reduction before HYDRA implementation, determining optimal parameters (M=1195 components, K=2 clusters) through systematic evaluation of clustering performance [1].

Hierarchical agglomerative clustering has been used to identify FC subtypes for individual brain networks. One protocol defined FC subtypes based on two criteria: average spatial dissimilarity within a subtype <1 and a minimum of 20 individuals per subtype. This identified 87 FC subtypes across 18 seed networks, with 97% of individuals assigned to a subtype [12].

Continuous Subtyping Protocols

Continuous approaches employ different methodological frameworks focused on quantifying individual positions within a multidimensional space.

Normative modeling protocols involve building a model of normal functional connectivity development using typically developing controls. This model is then applied to individuals with ASD to quantify their deviation from the normative trajectory. One comprehensive analysis incorporated static functional connectivity strength (SFCS), dynamic functional connectivity strength (DFCS), and dynamic functional connectivity variance (DFCV) as predictive variables within normative models to delineate multi-level functional developmental trajectories [2].

Continuous assignment methods avoid discrete cluster assignments altogether. Instead, individuals are characterized by their continuous similarity to reference profiles. One approach computed the spatial correlation between each individual's seed-FC map and the average FC subtype maps, creating continuous assignments that reflect degree rather than kind of subtype membership [12].

Cause-specific Cox proportional hazards models for continuous biomarkers represent another continuous approach, originally developed for cancer, but with applicability to neurodevelopment. This method models how exposure-disease associations change across continuous biomarker levels using restricted cubic spline functions, fully leveraging continuous data without categorization [54].

Figure 2: Experimental workflow for discrete versus continuous subtyping approaches using functional connectivity data.

Comparative Performance Analysis

Stability and Robustness

The stability of subtyping approaches fundamentally impacts their reliability and potential for clinical translation.

Table 1: Stability Comparison of Discrete vs. Continuous Subtyping

Metric	Discrete Subtyping	Continuous Subtyping
Assignment Reliability	Moderate discrete assignment stability [12]	Higher robustness of continuous assignments [12]
Replication Performance	Identified subtypes generalize to independent data [1]	Continuous deviations show replicable brain-behavior relationships [2]
Boundary Sensitivity	Sensitive to clustering parameters and sample characteristics [57]	Less dependent on arbitrary categorization thresholds [54]
Handling of Intermediate Cases	Forces all individuals into categories, potentially misrepresenting intermediates	Naturally accommodates intermediate and mixed presentations

Evidence directly indicates that "continuous assignments are more robust than discrete ones" [12]. Discrete assignments can be sensitive to methodological choices like clustering algorithms and dimension reduction techniques, whereas continuous approaches avoid arbitrary categorical thresholds that may not reflect neurobiological reality.

Clinical and Behavioral Correlations

Both subtyping approaches have demonstrated associations with clinically relevant variables, though through different mechanistic pathways.

Table 2: Clinical Utility Comparison of Subtyping Approaches

Clinical Feature	Discrete Subtyping Findings	Continuous Subtyping Findings
Symptom Severity	Hyper-connectivity and hypo-connectivity subtypes show distinct correlation patterns with core ASD symptoms [1]	Individual deviations in FC predict symptom severity [17]
Cognitive Profiles	Subtypes with identical clinical profiles show significant differences in cognitive ability [58]	Continuous brain-behavior relationships exist on a spectrum from ASD to controls [17]
Social Communication	Two neural subtypes associated with distinct gaze patterns in social eye-tracking tasks [2]	Relationship between FC and social communication exists as a continuum [17]
Treatment Response	Not directly assessed but hypothesized to differ between subgroups	Not directly assessed but dimensional approaches may better predict treatment trajectories

Discrete approaches have successfully identified subgroups with "distinct gaze patterns assessed by two autism-sensitive eye-tracking tasks focused on preference for social cues" despite comparable clinical presentations [2]. This suggests discrete subtypes may capture neurobehavioral subgroups that transcend conventional diagnostic boundaries.

Conversely, continuous approaches reveal that "dimensional analyses of FC patterns with behavioral measures revealed unique information about brain-behavior relations" that exist along a continuum [17]. These dimensional relationships may more accurately reflect the complex, multifactorial nature of neurodevelopmental conditions.

Neurobiological Distinctness

The neurobiological profiles identified by each approach offer complementary insights into ASD heterogeneity.

Discrete subtyping consistently identifies two primary neurobiological profiles across multiple studies:

• Hyper-connectivity subtype: Characterized by increased within-network connectivity and specific between-network patterns, such as hyper-connectivity between default mode and attention networks, and hypo-connectivity between default mode and visual/auditory networks [1].
• Hypo-connectivity subtype: Shows the opposite connectivity pattern, with decreased within-network connectivity and inverse between-network profiles [1].

Continuous approaches instead identify patterns of deviation across multiple brain systems. One study found one ASD subtype 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," while the other subtype showed the inverse pattern [2]. These dimensional deviations provide more granular information about an individual's specific neurobiological profile.

Practical Implementation Considerations

Research Reagent Solutions

Table 3: Essential Methodological Components for Functional Connectivity Subtyping

Research Component	Function/Purpose	Example Implementations
fMRI Preprocessing Tools	Standardize and clean raw imaging data	fMRIPrep [2]
Functional Parcellations	Define regions for connectivity analysis	Dosenbach 160 ROIs [2], MIST_20 [12]
Clustering Algorithms	Identify discrete subgroups	K-means [17], HYDRA [1], Hierarchical clustering [12]
Normative Modeling Frameworks	Quantify individual deviations from typical development	Python or R implementations [2]
Dynamic FC Metrics	Capture time-varying connectivity properties	Dynamic Conditional Correlation (DCC) [2]
Validation Measures	Assess stability and generalizability	Bootstrap procedures [17], independent cohort replication [2]

Analytical Recommendations

For researchers selecting between discrete and continuous approaches, consider these evidence-based recommendations:

• For categorical intervention development: Discrete subtyping may be preferable when developing targeted interventions for specific neurobiological profiles. The identification of two robust ASD subtypes with distinct connectivity patterns supports this approach [1].

• For dimensional biomarker discovery: Continuous approaches are superior for identifying granular brain-behavior relationships that exist along a spectrum. Continuous assignments demonstrate higher robustness [12].

• For handling sample heterogeneity: Semi-supervised discrete methods (like HYDRA) outperform unsupervised approaches when diagnostic information is available [1].

• For maximizing information preservation: Continuous methods avoid the information loss inherent in categorization, potentially providing more sensitive measures for tracking change over time [54].

Discrete and continuous subtyping approaches offer complementary yet distinct frameworks for decomposing ASD heterogeneity. Discrete methods effectively identify categorically distinct subgroups with potential for targeted interventions, while continuous approaches more robustly capture the dimensional nature of neurobiological variation. The choice between these paradigms should be guided by specific research goals: discrete subtyping for categorical questions of group differences, and continuous approaches for dimensional questions of individual differences. Future research would benefit from hybrid models that leverage the clinical interpretability of discrete subtypes while preserving the dimensional precision of continuous approaches, ultimately advancing personalized diagnostic and therapeutic strategies for ASD.

A significant challenge in autism spectrum disorder (ASD) research is the replication of findings across studies. This review objectively compares contemporary methodological approaches for defining ASD subtypes based on functional connectivity (FC), evaluating their experimental protocols, and presenting quantitative data to inform robust, generalizable biomarker discovery.

Comparative Analysis of Subtyping Methodologies

The table below summarizes the core design and output of three distinct subtyping approaches, highlighting the source and nature of their findings.

Study Focus	Methodological Approach	Sample Size (ASD/TD)	Identified Subtypes	Key Algorithm/Tool
Phenotypic-Genetic Integration [13] [59] [60]	Person-centered phenotypic clustering followed by genetic analysis	5,392 ASD (from SPARK cohort)	1. Social and Behavioral Challenges (37%)2. Mixed ASD with Developmental Delay (19%)3. Moderate Challenges (34%)4. Broadly Affected (10%)	General Finite Mixture Modeling
Brain-Functional Subtyping [2] [5]	Normative modeling of multilevel functional connectivity (static & dynamic)	Discovery: 479 ASD / 567 TDValidation: 21 ASD	1. Positive deviations in occipital/cerebellar networks; negative in frontoparietal/DMN*2. Inverse pattern of network deviations	Normative Modeling, Clustering Analysis
Semi-Supervised Neuro-Subtyping [1]	Semi-supervised clustering using diagnosis labels to guide high-dimension reduction	~847 ASD / 1030 TD (from ABIDE I & II)	1. Hyper-connectivity subtype2. Hypo-connectivity subtype	HYDRA (HeterogeneitY through DiscRiminative Analysis), OPNNMF

*DMN: Default Mode Network

Detailed Experimental Protocols

To ensure methodological transparency and facilitate replication, here are the detailed workflows for the key cited experiments.

Protocol 1: Person-Centered Phenotypic-Genetic Subtyping

This protocol outlines the methodology for identifying subtypes based on comprehensive trait data and linking them to distinct genetic profiles [13] [59].

• Cohort & Data Collection: Utilize a large-scale cohort (e.g., SPARK) with deep phenotypic and genotypic data. Collect data on over 230 traits per individual, spanning core autism features, developmental milestones, and co-occurring conditions [13] [59].
• Phenotypic Clustering: Apply a general finite mixture model to perform person-centered clustering. This model handles mixed data types (yes/no, categorical, continuous) and calculates the probability of each individual belonging to a specific subgroup based on their entire trait profile [59].
• Genetic Analysis: Subsequent to phenotypic clustering, analyze the genetic data within each subtype.
  • Variant Type Analysis: Assess the burden of damaging de novo mutations (non-inherited) and rare inherited variants in each subgroup [13].
  • Pathway Analysis: Identify biological pathways (e.g., neuronal action potentials, chromatin organization) enriched for genetic mutations in each subtype. Analyze the developmental timings of gene expression for the implicated genes [59] [60].
• Validation: Replicate the identified subtypes and their associated genetic patterns in an independent sample [60].

Protocol 2: Normative Modeling of Functional Connectivity

This protocol describes the process of identifying ASD subtypes by comparing an individual's brain connectivity to a normative model built from typically developing controls [2] [5].

• Data Acquisition & Preprocessing: Acquire resting-state fMRI (rs-fMRI) data from multi-site datasets (e.g., ABIDE I/II). Preprocess data using a standardized pipeline (e.g., fMRIPrep) to account for site-specific differences and control for head motion [2].
• Feature Extraction: Calculate multilevel functional connectivity features for each participant.
  • Static FC (SFCS): Use Pearson correlation to compute the strength of connectivity between brain regions over the entire scan.
  • Instant Dynamic FC: Use dynamic conditional correlation (DCC) to compute the strength (DFCS) and variability (DFCV) of connectivity over time [2].
• Normative Model Construction: Build a model that predicts the expected multilevel FC features across the brain based on age, using data from the typically developing (TD) control group only. This establishes a normative trajectory of brain development [2].
• Deviation Mapping: For each individual with ASD, calculate their deviation from the normative model, creating an individual-level map of functional atypicality [2].
• Subtype Identification: Apply clustering analysis (e.g., k-means) to the deviation maps to identify subgroups of individuals with ASD who share similar patterns of brain network dysfunction [2].

Protocol 3: Semi-Supervised Clustering with HYDRA

This protocol employs a machine learning technique that uses diagnostic labels to guide the discovery of neurologically distinct subtypes [1].

• Input Feature Preparation: Extract high-dimensional resting-state functional connectivity (FC) matrices from a large, multi-site dataset (e.g., ABIDE I/II) [1].
• Dimensionality Reduction: Apply a multi-scale dimension reduction method, Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF), to the high-dimensional FC data. This step reduces noise and extracts a lower-dimensional, more representative feature space for clustering [1].
• Semi-Supervised Clustering: Use the HYDRA algorithm. Unlike unsupervised methods, HYDRA is informed by the diagnostic labels (ASD vs. TD controls). It identifies subtypes by finding directions in the feature space that maximize the separation between the combined ASD group and the TD control group, thereby directly linking subtype definitions to neurodivergence [1].
• Reliability and Validation: Systematically evaluate clustering performance and reliability. Compare the results against common unsupervised methods (e.g., K-means) to demonstrate superior performance in identifying robust subtypes [1].

Research Reagent Solutions

Essential computational tools and datasets that form the backbone of modern, reproducible ASD subtyping research.

Reagent / Resource	Type	Primary Function in Research
ABIDE I & II [2] [1]	Data Repository	Publicly shared consortium data providing aggregated rs-fMRI and phenotypic data from multiple sites, enabling large-scale analyses.
SPARK Cohort [13] [59]	Data Repository	Largest study of autism in the U.S., providing deeply phenotyped data and genotypic data for thousands of individuals.
fMRIPrep [2]	Software Tool	Standardized, robust pipeline for preprocessing of fMRI data, mitigating site-specific variability and enhancing reproducibility.
General Finite Mixture Model [59]	Statistical Algorithm	Integrates mixed data types (binary, categorical, continuous) for person-centered, phenotypic clustering.
HYDRA [1]	Machine Learning Algorithm	Semi-supervised clustering method that uses diagnosis labels to guide the discovery of neurologically distinct subtypes.
Normative Modeling [2]	Statistical Framework	Quantifies individual-level deviations in brain metrics (e.g., FC) from a normative model built on typical development.

Experimental Workflow Visualization

The diagram below illustrates the core logical differences between the two predominant paradigms in ASD subtyping research.

Optimizing Feature Selection and Dimensionality Reduction to Prevent Overfitting

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in phenotypic clinical symptoms, neurobiology, and underlying brain connectivity patterns. This heterogeneity presents substantial challenges for developing effective biomarkers and interventions, as individuals with ASD display diverse functional and structural brain profiles despite similar clinical presentations [2]. Research indicates that ASD encompasses a series of neurodevelopmental disorders affecting social, behavioral, and communication abilities, traditionally divided into subtypes including autism, Asperger's, and pervasive developmental disorder-not otherwise specified (PDD-NOS) [11].

The analysis of high-dimensional neuroimaging data, particularly functional connectivity metrics derived from resting-state functional magnetic resonance imaging (rs-fMRI), has become instrumental in identifying ASD subtypes. However, these datasets present significant analytical challenges due to the "curse of dimensionality," where the number of features (e.g., brain region connections) vastly exceeds the number of participants [61] [62]. This imbalance increases the risk of overfitting, where models memorize noise and spurious correlations rather than learning generalizable patterns, ultimately compromising their clinical applicability and reliability.

This guide provides a comprehensive comparison of feature selection and dimensionality reduction techniques for optimizing ASD functional connectivity research, with specific emphasis on preventing overfitting while enhancing model interpretability and biological validity.

Core Concepts and Distinctions

Feature selection and dimensionality reduction represent two distinct strategic approaches for addressing high-dimensional data challenges in ASD research:

• Feature Selection identifies and retains the most informative original features from the dataset while discarding irrelevant or redundant ones [63] [64]. This approach preserves the original feature meanings, enhancing interpretability for domain experts. Methods include filter, wrapper, and embedded approaches that select features based on statistical measures, model performance, or integrated selection during model training [62] [64].

• Dimensionality Reduction transforms the original features into a new, lower-dimensional space by creating combinations of original variables [63] [64]. Unlike feature selection, these techniques generate entirely new features that may not have direct biological interpretations but often better capture complex relationships in the data.

The Overfitting Challenge in ASD Connectivity Research

In ASD research, overfitting manifests when models trained on high-dimensional connectivity data fail to generalize to new populations or sites. This problem is particularly acute given the typically small sample sizes relative to the number of extracted features. Microstructural processing pipelines often generate data from over 200 distinct brain regions per participant, creating a "wide data" problem where the number of predictors challenges statistical power and reproducibility [61]. For example, one study with 213 participants faced dimensionality challenges with millions of data points, creating ideal conditions for overfitting without appropriate dimensionality management [61].

Methodological Comparison for ASD Subtype Identification

Dimensionality Reduction Techniques

Dimensionality reduction methods transform functional connectivity data into compressed representations that retain essential information while reducing feature space.

Table 1: Dimensionality Reduction Methods in ASD Research

Technique	Type	Key Mechanism	ASD Application Example	Advantages	Limitations
Principal Component Analysis (PCA)	Linear	Creates new uncorrelated components that maximize variance [63]	Integrated with ML to handle genetic and microstructural data complexity [61]	Reduces dimensionality while preserving variance; Computationally efficient [64]	Limited to linear relationships; May not capture biologically relevant variance [64]
t-Distributed Stochastic Neighbor Embedding (t-SNE)	Non-linear	Preserves local structure and reveals clusters in high-dimensional data [63]	Visualizing high-dimensional brain connectivity patterns [63]	Excellent for cluster visualization; Captures complex non-linear patterns [63] [64]	Computational intensive; Results sensitive to parameter choices [63]
Linear Discriminant Analysis (LDA)	Supervised	Maximizes between-class variance while minimizing within-class variance [63]	Classification tasks with labeled ASD subtypes [63]	Enhances class separability; Effective for classification tasks [63] [64]	Requires labeled data; Assumes normal distribution and equal class covariances [64]
Autoencoders	Non-linear	Neural network with encoder-decoder structure learns compressed representation [64]	Extracting complex patterns from brain connectivity data [64]	Captures hierarchical, non-linear relationships; Handles noisy data well [64]	Computationally demanding; Requires large datasets; Black box nature [64]
Uniform Manifold Approximation and Projection (UMAP)	Non-linear	Preserves both local and global data structure [63]	Handling large-scale brain imaging datasets [63]	Faster than t-SNE; Better preservation of global structure [63]	Relatively newer; Less established in neuroimaging [63]

Feature Selection Approaches

Feature selection methods identify the most discriminative connectivity features for distinguishing ASD subtypes while maintaining biological interpretability.

Table 2: Feature Selection Methods in ASD Research

Technique	Category	Selection Mechanism	ASD Application Example	Advantages	Limitations
Recursive Feature Elimination (RFE)	Wrapper	Iteratively removes least important features based on model performance [65]	Identifying informative taxa in metabarcoding data [65]	Model-agnostic; Considers feature interactions [65]	Computationally intensive; Risk of overfitting to specific dataset [65]
Lasso Regression (L1 Regularization)	Embedded	Performs feature selection during model training via L1 penalty [63]	Selecting relevant neural features while regularizing models [63]	Built-in regularization; Feature selection integrated with modeling [63]	May exclude correlated informative features; Tuning sensitive [62]
Random Forest Importance	Embedded	Ranks features by their contribution to node impurity reduction [65]	Benchmark analysis showing robustness without explicit feature selection [65]	Handles non-linear relationships; Robust to outliers [65] [66]	May bias toward high-cardinality features; Computational cost with many trees [65]
Hybrid Metaheuristics (TMGWO, ISSA)	Hybrid	Optimization algorithms inspired by natural processes [66]	Medical dataset classification including neurological disorders [66]	Effective for high-dimensional spaces; Balances exploration and exploitation [66]	Complex parameter tuning; Computationally demanding [66]

Experimental Protocols in ASD Subtype Research

Normative Modeling with Functional Connectivity Deviations

Objective: Identify ASD subtypes based on individual deviations from typical functional connectivity development trajectories [2].

Dataset: 1,046 participants (479 ASD, 567 typical development) from ABIDE-I and ABIDE-II; independent validation cohort of 21 ASD individuals with eye-tracking data [2].

Methodological Workflow:

• Feature Extraction: Calculate multilevel functional connectivity features including static functional connectivity strength (SFCS), dynamic functional connectivity strength (DFCS), and dynamic functional connectivity variance (DFCV) using the Dosenbach 160 ROI atlas [2].
• Normative Modeling: Develop normative models of brain maturation using TD participants to quantify individual-level functional connectivity deviations in ASD participants [2].
• Clustering Analysis: Apply clustering algorithms to the deviation profiles to identify potential ASD subtypes [2].
• Validation: Correlate neural subtypes with clinical symptoms and eye-tracking measures of social attention [2].

Key Findings: The study identified two distinct neural ASD subtypes with unique functional brain network profiles despite comparable clinical presentations, which were also associated with different gaze patterns during social attention tasks [2].

ASD Subtyping Workflow: This diagram illustrates the comprehensive methodology for identifying autism subtypes through normative modeling of functional connectivity data, culminating in validation with behavioral measures.

Tensor Decomposition for Brain Community Feature Extraction

Objective: Systematically compare three common ASD subtypes (autism, Asperger's, and PDD-NOS) using functional and structural brain features [11].

Dataset: 152 patients with autism, 54 with Asperger's, and 28 with PDD-NOS from ABIDE I [11].

Methodological Workflow:

• Feature Extraction: Extract four types of brain features: (1) brain patterns via tensor decomposition of fMRI data, (2) amplitude of low-frequency fluctuation (ALFF), (3) fractional ALFF (fALFF), and (4) gray matter volume (GMV) from structural MRI [11].
• Tensor Decomposition: Apply tensor decomposition to capture different brain communities in ASD subtypes, addressing the high dimensionality of fMRI data combining brain regions, time, and patients [11].
• Statistical Analysis: Use statistical tests to identify significant differences between ASD subtypes in both functional and structural features [11].

Key Findings: Found significant dissimilarities between subtypes, with impairments in the subcortical network and default mode network of autism representing major differences from the other two subtypes [11].

PCA with Supervised Machine Learning Framework

Objective: Navigate the complex multidimensional space created by combining genetic and microstructural data modalities for ASD classification [61].

Dataset: 213 participants (113 autistic, 100 non-autistic) from Wave 1 of an NIH-sponsored Autism Centers for Excellence network [61].

Methodological Workflow:

• Unsupervised Dimensionality Reduction: Apply PCA to genetic and microstructural data to reduce dimensionality while retaining within-class variation [61].
• Supervised Classification: Integrate PCA components into traditional classification machine learning algorithms [61].
• Model Evaluation: Assess classification performance while monitoring for overfitting through cross-validation and hold-out testing [61].

Key Findings: This integrated approach addressed the "curse of dimensionality" without relying on traditional feature selection methods, retaining generalizability to unseen data [61].

Performance Comparison and Experimental Data

Benchmark Analysis of Method Efficacy

Table 3: Performance Comparison of Feature Selection and Dimensionality Reduction Methods

Method	Dataset	Performance Metrics	Overfitting Mitigation	Key Findings
Random Forest without FS	13 metabarcoding datasets [65]	High accuracy in regression and classification	Robust without additional feature selection [65]	Feature selection more likely to impair than improve performance for tree ensemble models [65]
Recursive Feature Elimination with Random Forest	13 metabarcoding datasets [65]	Enhanced performance across various tasks	Selective feature reduction [65]	Effectively identifies relevant features while maintaining model performance [65]
PCA with Supervised ML	213 participants (ASD vs controls) [61]	Improved classification accuracy	Addressed curse of dimensionality without traditional feature selection [61]	Retained generalizability to unseen data while managing high-dimensional space [61]
Hybrid Algorithms (TMGWO)	Medical datasets including neurological data [66]	96% accuracy with only 4 features	Reduced model complexity significantly [66]	Outperformed Transformer-based approaches (TabNet: 94.7%, FS-BERT: 95.3%) using fewer features [66]
Normative Modeling with FC	1,046 participants, 2 ASD subtypes [2]	Identified distinct neural subtypes	Individual-level deviation quantification from normative trajectories [2]	Subtypes showed different gaze patterns despite similar clinical symptoms [2]

Impact on Model Generalization and Stability

Research indicates that the choice between feature selection and dimensionality reduction significantly impacts model stability and reliability. One comprehensive comparison of feature selection algorithms found substantial variation in stability metrics across methods, with some algorithms demonstrating higher sensitivity to data perturbations than others [62]. This stability is crucial for developing reproducible biomarkers in ASD research, where findings must generalize across diverse populations and sites.

Dimensionality reduction techniques like PCA generally provide more stable feature sets than many filter-based selection methods, though at the cost of biological interpretability [62]. Embedded feature selection methods such as Lasso regularization typically offer a favorable balance between stability and interpretability, making them particularly valuable for ASD connectivity research where both reliability and biological plausibility are essential.

Table 4: Research Reagent Solutions for ASD Connectivity Analysis

Resource	Type	Function	Example Implementation
ABIDE I/II Datasets	Data Resource	Publicly available neuroimaging datasets aggregating fMRI, structural MRI, and phenotypic data from multiple sites [11] [2]	Provides standardized datasets for method development and validation across diverse ASD populations [11] [2]
fMRIPrep	Software Tool	Robust preprocessing pipeline for functional MRI data ensuring standardized processing across studies [2]	Used in recent subtype identification studies for consistent data preprocessing [2]
Dosenbach 160 Atlas	Parcellation Scheme	Functional ROI atlas derived from meta-analyses across multiple cognitive domains [2]	Standardized brain parcellation for feature extraction in functional connectivity analyses [2]
Normative Modeling Framework	Analytical Approach	Statistical framework quantifying individual deviations from normative neurodevelopmental trajectories [2]	Identifies ASD subtypes based on individualized patterns of functional connectivity deviations [2]
Tensor Decomposition	Algorithm	Method for extracting compressed feature sets from high-dimensional fMRI data [11]	Analyzes complex fMRI data structures combining brain regions, time, and patients [11]
scikit-learn	Python Library	Comprehensive machine learning library implementing various feature selection and dimensionality reduction methods [62] [64]	Accessible implementation of algorithms like PCA, RFE, and embedded methods [62]

Integrated Workflow for Optimal Method Selection

Method Selection Guide: A decision framework for choosing between feature selection and dimensionality reduction techniques based on research objectives, data characteristics, and interpretability requirements.

Optimizing feature selection and dimensionality reduction represents a critical step in advancing ASD subtype research beyond correlation-based findings toward clinically meaningful biomarkers. The comparative analysis presented in this guide demonstrates that method selection should be guided by specific research objectives:

For hypothesis-driven research requiring biological interpretability, feature selection methods like recursive feature elimination or Lasso regression provide optimal balance between performance and explanatory power. For exploratory subtype discovery, dimensionality reduction techniques like normative modeling or PCA enable identification of patterns not captured by a priori hypotheses.

The most robust ASD research implementations increasingly adopt hybrid approaches, combining unsupervised dimensionality reduction to manage data complexity with supervised feature selection to refine biological insights. This strategic integration, coupled with rigorous validation across independent cohorts, represents the most promising path toward developing generalizable, clinically relevant biomarkers for ASD heterogeneity.

Regardless of methodological approach, successful implementation must prioritize prevention of overfitting through appropriate cross-validation, external validation, and careful monitoring of model complexity. By adopting these optimized approaches, researchers can enhance both the statistical robustness and biological validity of their findings, accelerating progress toward personalized interventions for ASD.

Clustering analysis serves as a fundamental tool in unsupervised machine learning, playing a pivotal role in identifying hidden patterns within complex datasets without prior labeling. In autism spectrum disorder (ASD) research, clustering techniques have become indispensable for parsing the substantial heterogeneity inherent in the disorder, leading to the identification of neurobiologically distinct subtypes based on functional connectivity profiles [1]. The effectiveness of these approaches depends critically on robust validation methodologies, as the choice of clustering algorithms and validation indices directly impacts the reliability and interpretability of the resulting subtypes.

This guide provides a comprehensive benchmarking framework for clustering performance in the context of ASD neurosubtyping research. We objectively compare the effectiveness of various cluster validity indices (CVIs) and clustering approaches, supported by experimental data from recent studies. The insights presented aim to equip researchers with evidence-based methodologies for selecting optimal clustering strategies that can uncover biologically meaningful subgroups within the ASD population, ultimately advancing personalized diagnostic and therapeutic approaches.

Cluster Validity Indices: A Comparative Analysis

Cluster validity indices are mathematical criteria used to evaluate the quality of clustering results and, in automatic clustering algorithms, to determine the optimal number of clusters. They primarily assess intra-cluster cohesion (how similar data points are within the same cluster) and inter-cluster separation (how distinct different clusters are from one another) [67]. In metaheuristic-based automatic clustering algorithms, CVIs serve as fitness functions that guide the optimization process toward optimal cluster configurations [68].

Performance Benchmarking of Common Validity Indices

Different CVIs exhibit varying performance characteristics depending on dataset structure. A recent large-scale multivariate comparison of 68 cluster validity indices revealed that indices based on the min/max decision rule generally provide more reliable results [69]. Another benchmarking study within an Evolutionary K-means framework evaluated 15 internal validity indices across diverse synthetic and real-life datasets [68].

Table 1: Performance Comparison of Key Cluster Validity Indices

Validity Index	Primary Measurement Focus	Performance Characteristics	Best Suited Data Structures
Calinski-Harabasz (CH)	Between-cluster vs within-cluster variance	Consistently outperforms others; reliable performance [68]	Well-separated, compact clusters
Silhouette Index	Pairwise distance within vs between clusters	Superior performance; robust to noise [68]	Clusters of similar density
Davies-Bouldin Index	Similarity between clusters	Moderate performance; commonly used [70]	Spherical cluster distributions
Xie-Beni Index	Fuzzy cluster separation	Varying performance; data-dependent [68]	Overlapping cluster structures
S_Dbw Index	Density-based separation	Lower reliability in benchmarks [68]	Non-spherical distributions

The experimental results indicate that the Calinski-Harabasz (CH) and Silhouette indices consistently deliver superior performance across diverse dataset types, making them particularly suitable for the complex, high-dimensional data characteristic of neuroimaging studies [68]. This reliability is crucial in ASD functional connectivity research, where the underlying cluster structure is often unknown a priori.

Clustering Methodologies in ASD Research: Experimental Protocols

In ASD functional connectivity research, multiple clustering approaches have been employed, each with distinct experimental protocols and performance characteristics.

Semi-Supervised Clustering with HYDRA

A recent study implemented a semi-supervised clustering method called HYDRA (HeterogeneitY through DiscRiminative Analysis) guided by diagnosis labels (ASD vs. controls) to identify neurosubtypes [1].

Table 2: Experimental Protocol for Semi-Supervised ASD Subtyping

Protocol Component	Specification	Purpose/Rationale
Dataset	ABIDE I & II (∼2000 subjects) [1]	Large sample size for reliable subtyping
Input Features	Resting-state functional connectivity matrices [1]	Captures network-level neural interactions
Dimension Reduction	Orthogonal Projective Non-Negative Matrix Factorization (OPNNMF) [1]	Handles high-dimensional feature space
Clustering Algorithm	HYDRA (semi-supervised) [1]	Leverages diagnostic labels for guidance
Validation Approach	Systematic reliability assessment [1]	Ensures robustness of identified subtypes
Optimal Parameters	K=2 clusters, M=1195 components [1]	Determined through validation procedures

The protocol successfully identified two distinct ASD subtypes with reliable functional connectivity profiles: a hyper-connectivity subtype showing increased connectivity within major large-scale networks, and a hypo-connectivity subtype displaying the opposite pattern [1]. This semi-supervised approach demonstrated superior performance compared to unsupervised K-means clustering, particularly in capturing clinically relevant neural signatures.

Unsupervised Clustering Comparison

A comprehensive comparison of four unsupervised machine learning methods was conducted using polysomnographic data from 865 individuals, providing insights into the relative performance of different unsupervised approaches [70]:

• K-means clustering: Demonstrated the best overall clustering performance across most evaluation criteria [70]
• Fuzzy C-means: Showed strong agreement with K-means (κ = 0.87) and greatest potential for handling overlapping clusters [70]
• Gaussian Mixture Models: Moderate performance with weakest agreement with Agglomerative Hierarchical Clustering (κ = 0.51) [70]
• Agglomerative Hierarchical Clustering: Produced distinct cluster formations with moderate agreement with other methods [70]

The study revealed that the selection of clustering methods directly impacts the formation and characteristics of identified clusters, with K-means and Fuzzy C-means showing particularly strong performance for complex biomedical data [70].

Visualization of Methodological Workflows

Semi-Supervised Neurosubtyping Pipeline

ASD Subtyping Workflow

Cluster Validation Decision Framework

CVI Selection Guide

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Resources for Clustering Analysis in ASD Connectomics

Research Resource	Specification/Function	Application Context
ABIDE Dataset	Preprocessed neuroimaging data from >2000 individuals [1]	Large-scale ASD subtyping studies
HYDRA Algorithm	Semi-supervised clustering leveraging diagnostic labels [1]	Neurosubtyping with clinical guidance
OPNNMF	Orthogonal Projective Non-Negative Matrix Factorization [1]	High-dimensional feature reduction
Calinski-Harabasz Index	Variance ratio-based cluster validity index [68]	Optimal cluster number determination
Silhouette Index	Pairwise distance-based validation metric [68]	Cluster quality assessment
Enhanced FA-K-means	Firefly Algorithm hybrid with K-means [68]	Automatic clustering optimization
Contrast Subgraph Mining	Network comparison technique [53]	Identifying differential connectivity patterns

Discussion and Performance Implications

The benchmarking data presented reveals several critical considerations for clustering applications in ASD functional connectivity research:

Validity Index Selection significantly impacts clustering outcomes. The consistent superiority of Calinski-Harabasz and Silhouette indices across multiple benchmarking studies [69] [68] suggests these should be preferred for robust ASD subtyping, particularly when ground truth labels are unavailable.

Semi-supervised approaches like HYDRA demonstrate clear advantages for clinical neurosubtyping by incorporating diagnostic information to guide cluster separation [1]. The resulting ASD subtypes (hyper-connectivity and hypo-connectivity) not showed distinct functional connectivity profiles but also correlated with differential clinical symptom patterns, enhancing their potential biological validity.

The performance trade-offs between clustering methods indicate that while K-means excels in many scenarios [70], Fuzzy C-means offers particular advantages for handling the overlapping cluster boundaries likely present in complex neurodevelopmental conditions like ASD.

Future directions in ASD subtyping should leverage multi-scale validation approaches, combining internal validity indices with external clinical correlates to ensure that identified subtypes reflect both statistical robustness and clinical relevance. The integration of genetic data with neuroimaging-based clusters, as suggested by studies showing different polygenic architectures associated with early- and later-diagnosed ASD [71], represents a promising avenue for developing comprehensive ASD stratification frameworks.

Validation and Comparative Analysis: From Neural Profiles to Clinical Translation

Autism Spectrum Disorder (ASD) is characterized by substantial heterogeneity in both clinical symptoms and underlying neurobiology, complicating the development of effective interventions and biomarkers. The identification of reproducible subtypes through functional connectivity (FC) research represents a crucial step toward precision medicine in autism. Cross-validation in independent cohorts ensures that identified subtypes reflect robust neurobiological phenomena rather than cohort-specific artifacts, directly impacting the reliability of conclusions drawn for drug development and clinical applications. This guide objectively compares methodological approaches and evaluates the generalizability of ASD subtype classifications across key studies, providing researchers with a framework for assessing replication validity in neuroimaging-based subtyping research.

Comparative Analysis of ASD Subtyping Approaches

Table 1: Key Studies on ASD Functional Connectivity Subtypes and Their Replication Status

Study Reference	Subtypes Identified	Sample Size (ASD/TD)	Validation Approach	Replication Outcome	Primary Biomarkers Used
Molecular Psychiatry 2025 [2]	2 neural subtypes	479/567 (Discovery) + 21 ASD (Validation)	Independent cohort with eye-tracking	Successfully replicated in independent cohort	Static/dynamic FC in ON, CerN, FPN, DMN, CON
Nature Genetics 2025 [72]	4 phenotypic classes	5,392 (SPARK)	Simons Simplex Collection	Strong replication of feature patterns	Behavioral profiles & genetic programs
elife 2020 [12]	Multiple FC subtypes (3-6 per network)	388 (Mixed ASD/TD)	Independent dataset	Moderate association with diagnosis	Network-specific FC patterns
Frontiers in Neuroscience 2024 [73]	3 clinical subtypes (Autism, Asperger's, PDD-NOS)	234 ASD only	Cross-dataset analysis	Significant functional differences	ALFF/fALFF, GMV, tensor decomposition

Table 2: Methodological Comparison of Cross-Validation Approaches

Methodological Component	Molecular Psychiatry 2025 [2]	Nature Genetics 2025 [72]	elife 2020 [12]	Frontiers in Neuroscience 2024 [73]
Statistical Framework	Normative modeling + clustering	Generative mixture modeling	Hierarchical clustering	Tensor decomposition + statistical testing
Primary Features	Multilevel FC (static + dynamic)	239 phenotypic features	Network-specific FC	ALFF/fALFF, GMV, brain patterns
Validation Cohort	UESTC (n=21 ASD)	Simons Simplex Collection (n=861)	Independent dataset	ABIDE I dataset
Outcome Measures	Eye-tracking patterns + clinical symptoms	Feature enrichment patterns	Diagnostic association	Network dissociation patterns
Replication Success	High (despite small N)	High	Moderate	Limited to specific networks

Detailed Experimental Protocols

Normative Modeling with Multilevel Functional Connectivity

Protocol Overview: The approach outlined in Molecular Psychiatry 2025 [2] establishes normative developmental trajectories of functional brain organization using typically developing (TD) individuals, then quantifies individual ASD deviations from these trajectories to identify subtypes.

Step-by-Step Methodology:

• Participant Selection and Inclusion Criteria:

  • Utilize large, multi-site datasets (ABIDE-I and ABIDE-II) with stringent quality control
  • Apply exclusion criteria: excessive head motion (mean FD > 0.3), poor spatial normalization, missing demographic information
  • Final sample: 1046 participants (479 ASD, 567 TD) across 23 sites

• Multilevel Functional Connectivity Feature Extraction:

  • Acquire resting-state fMRI data using standardized protocols
  • Preprocess data using fMRIPrep with normalization to standard space
  • Extract average BOLD signals from Dosenbach 160 ROIs
  • Calculate three distinct FC metrics:
    • Static Functional Connectivity Strength (SFCS): Pearson correlation between regional time series
    • Dynamic Functional Connectivity Strength (DFCS): Instantaneous connectivity using dynamic conditional correlation
    • Dynamic Functional Connectivity Variance (DFCV): Variability in connectivity over time

• Normative Model Construction:

  • Train models on TD participants to establish typical functional connectivity trajectories across development
  • Quantify individual ASD deviations from normative trajectories for each FC metric

• Subtype Identification via Clustering:

  • Apply clustering algorithms to deviation patterns
  • Identify subtypes based on distinct neurobiological profiles

• Validation in Independent Cohort:

  • Recruit independent cohort with both rsfMRI and eye-tracking data
  • Validate subtype stability and association with behavioral measures (eye-tracking patterns)

Diagram 1: Normative Modeling Workflow (23 chars)

Generative Mixture Modeling for Phenotypic Class Discovery

Protocol Overview: Nature Genetics 2025 [72] employs a person-centered approach to identify robust phenotypic classes through generative finite mixture modeling (GFMM) of comprehensive phenotypic data, subsequently validating classes in independent cohorts and linking them to genetic programs.

Step-by-Step Methodology:

• Phenotypic Feature Selection:

  • Collect 239 item-level and composite phenotype features from standardized diagnostic questionnaires
  • Include: Social Communication Questionnaire-Lifetime (SCQ), Repetitive Behavior Scale-Revised (RBS-R), Child Behavior Checklist (CBCL)
  • Incorporate developmental milestone history from background forms

• Model Training and Class Identification:

  • Apply GFMM to accommodate heterogeneous data types (continuous, binary, categorical)
  • Train models with 2-10 latent classes, selecting optimal model using Bayesian information criterion (BIC) and clinical interpretability
  • Assign features to seven clinically relevant categories for interpretation

• Class Validation and Replication:

  • Validate classes using medical history questionnaire data not included in original model
  • Test replication in independent Simons Simplex Collection cohort
  • Apply GFMM trained on SPARK data to SSC test set
  • Independently train GFMM on SSC data to confirm class stability

• Genetic Association Analysis:

  • Associate phenotypic classes with patterns of common genetic variation using polygenic scores
  • Analyze de novo and rare inherited variation across classes
  • Identify class-specific gene expression patterns during development

Diagram 2: Phenotypic Class Discovery (24 chars)

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Research Reagent Solutions for ASD Subtyping Studies

Reagent/Material	Specific Function	Example Implementation in Studies
ABIDE I & II Datasets	Multi-site resting-state fMRI and phenotypic data for discovery and validation	Primary data source for Molecular Psychiatry 2025 [2] and elife 2020 [12]
fMRIPrep Pipeline	Standardized fMRI data preprocessing for consistency across sites	Used in Molecular Psychiatry 2025 for reproducible preprocessing [2]
Dosenbach 160 ROI Atlas	Defined functional regions for connectivity analysis	Feature extraction in Molecular Psychiatry 2025 [2]
Social Communication Questionnaire (SCQ)	Assesses core autism symptoms in social communication	Phenotypic feature in Nature Genetics 2025 study [72]
Repetitive Behavior Scale-Revised (RBS-R)	Quantifies restricted and repetitive behaviors	Component of phenotypic profiling in Nature Genetics 2025 [72]
Tobii Eye-Tracking Systems	Measures visual attention patterns to social stimuli	Validation in independent UESTC cohort in Molecular Psychiatry 2025 [2]
Dynamic Conditional Correlation Algorithms	Captures instant dynamic functional connectivity	Multilevel FC feature in Molecular Psychiatry 2025 [2]

Critical Assessment of Replication Success and Limitations

The comparative analysis reveals varying success in replication across methodologies. The phenotypic approach demonstrated particularly strong replication, with the Nature Genetics 2025 study [72] successfully replicating four phenotypic classes in an independent cohort with highly similar feature enrichment patterns across all seven phenotype categories. Similarly, the FC-based normative modeling approach in Molecular Psychiatry 2025 [2] identified two neural subtypes that replicated in an independent cohort and showed distinct gaze patterns in eye-tracking tasks.

However, significant challenges remain in FC-based subtyping. As noted in NeuroImage 2022 [74], cross-cohort replicability of connectivity-based prediction patterns is often limited, achieving only "moderate similarity" across cohorts. Furthermore, feature weight reliability in connectivity-based predictive models is generally poor (ICC < 0.3) according to Tian et al. [75], highlighting fundamental methodological challenges in achieving reproducible brain-behavior associations.

The convergence of findings across independent studies using different methodologies strengthens the validity of ASD subtypes. The identification of subtypes with distinct clinical trajectories and genetic profiles provides a foundation for developing targeted interventions and personalized treatment approaches in autism spectrum disorder.

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in both clinical presentation and underlying neurobiology, posing substantial challenges for research and therapeutic development. Establishing clear correlations between distinct neural subtypes and their behavioral manifestations is crucial for advancing personalized interventions. This review synthesizes findings from recent large-scale studies that employ diverse methodological approaches to delineate ASD subtypes based on functional connectivity (FC) profiles, genetic underpinnings, and their relationship to behavioral symptoms. By comparing these subtyping frameworks, we provide researchers and drug development professionals with a comprehensive overview of the current landscape in ASD stratification, highlighting how different neurobiological signatures correlate with specific symptom profiles and clinical outcomes.

Comparative Analysis of ASD Subtyping Frameworks

Functional Connectivity-Based Subtyping

Table 1: Functional Connectivity ASD Subtypes from Recent Studies

Study	Sample Size	Subtyping Method	Identified Subtypes	Key Neural Characteristics	Behavioral Correlations
Wang et al., 2025 [76] [2]	1,046 participants (479 ASD)	Normative modeling of static/dynamic FC	Subtype 1	Positive deviations: occipital and cerebellar networks. Negative deviations: frontoparietal, DMN, cingulo-opercular networks [76].	Distinct gaze patterns in eye-tracking tasks despite comparable clinical scores [76].
			Subtype 2	Inverse pattern of Subtype 1 across the same networks [76].	Distinct gaze patterns in eye-tracking tasks despite comparable clinical scores [76].
Wang & Chen, 2025 [1]	~1,877 participants (847 ASD)	Semi-supervised HYDRA clustering	Hyper-connectivity Subtype	Increased within-network connectivity; mixed between-network connectivity (e.g., hyper-connectivity between DMN-attention) [1].	Varying correlations between connectivity patterns and core ASD symptoms [1].
			Hypo-connectivity Subtype	Opposite connectivity pattern to hyper-connectivity subtype [1].	Varying correlations between connectivity patterns and core ASD symptoms [1].
Easson et al., 2019 [17]	266 participants (145 ASD)	K-means clustering of FC	Subtype 1	Stronger between-network FC, particularly DMN with other networks [17].	No significant differences in SRS or ADOS scores between ASD subgroups [17].
			Subtype 2	Stronger within-network FC, weaker between-network FC [17].	No significant differences in SRS or ADOS scores between ASD subgroups [17].

Genetic and Phenotypic Subtyping

Table 2: Clinico-Biological ASD Subtypes from the SPARK Cohort Study

Subtype Name	Prevalence	Core Clinical Features	Developmental Trajectory	Genetic Profile
Social and Behavioral Challenges [13] [72]	~37%	Core ASD traits + ADHD, anxiety, depression, OCD [13].	Milestones similar to non-ASD peers; later gene expression [13].	Highest genetic predisposition for ADHD, anxiety, depression [13].
Moderate Challenges [13] [72]	~34%	Core ASD behaviors less pronounced; few co-occurring conditions [13].	Milestones on track with non-ASD peers [13].	Rare variants in genes active during fetal/neonatal stages [13].
Mixed ASD with Developmental Delay [13] [72]	~19%	Developmental delays (language, motor); lower rates of anxiety/depression [13].	Early developmental delays; early diagnosis [13].	Carries rare inherited genetic variants [13].
Broadly Affected [13] [72]	~10%	Severe, wide-ranging challenges: DD, social/communication deficits, RRBs, co-occurring psychiatric conditions [13].	Early developmental delays; intellectual disability [13].	Highest burden of damaging de novo mutations [13].

Experimental Protocols for ASD Subtyping

Functional Connectivity Analysis Pipeline

Protocol 1: Resting-State fMRI Processing and Feature Extraction

• Data Acquisition: Multi-site resting-state fMRI (rsfMRI) and T1-weighted structural data are collected from large cohorts such as ABIDE I/II [76] [17] [1]. Consistent parameters across sites are ideal.
• Preprocessing: Standard pipelines like fMRIPrep are employed. Steps include motion correction, slice-timing correction, normalization to standard space (e.g., MNI152), and band-pass filtering (typically 0.01–0.1 Hz). Global signal regression and censoring for excessive head motion (e.g., FD > 0.3) are critical [76] [11].
• Feature Calculation:
  • Static Functional Connectivity (SFC): Calculated using Pearson correlation between the average BOLD time series of predefined brain regions (e.g., Dosenbach's 160 ROIs) [76] [2].
  • Dynamic Functional Connectivity (DFC): Quantified using Dynamic Conditional Correlation (DCC) to assess moment-to-moment changes in connectivity strength (DFCS) and variability (DFCV) over time [76].
• Normative Modeling: For studies using this approach, a normative model of brain development is built using the TD group's multilevel FC features. Individual ASD participants are then assessed as deviations from this normative trajectory [76].

FC Subtyping Workflow: Diagram outlining the key stages in functional connectivity-based subtyping of ASD, from data acquisition to validation.

Genetic and Phenotypic Subtyping Protocol

Protocol 2: Person-Centered Phenotypic and Genetic Class Identification

• Cohort and Data Collection: Large, deeply phenotyped cohorts like SPARK (n=5,392) are utilized. Data encompasses over 230 item-level and composite features from standardized instruments (e.g., SCQ, RBS-R, CBCL), developmental milestones, and medical history [13] [72].
• Phenotypic Modeling: A Generative Finite Mixture Model (GFMM) is applied to the heterogeneous phenotypic data (continuous, binary, categorical). This person-centered approach identifies latent classes based on individuals' holistic trait combinations rather than analyzing traits in isolation [72].
• Class Validation and Replication: Identified classes are validated using data not included in the model (e.g., medical diagnoses of co-occurring conditions). Generalizability is tested by replicating the model in an independent cohort (e.g., Simons Simplex Collection) [72].
• Genetic Analysis: Class-specific genetic profiles are investigated by examining polygenic scores, de novo mutations, and rare inherited variants. Differences in the biological pathways affected and the developmental timing of gene expression are analyzed [13] [72].

The Scientist's Toolkit: Key Research Reagents and Materials

Table 3: Essential Resources for ASD Subtyping Research

Resource Category	Specific Item/Instrument	Primary Function in Research
Data Repositories	ABIDE I & II (fMRI)	Provides large-scale, multi-site neuroimaging and phenotypic data for discovery and validation cohorts [76] [17] [1].
	SPARK Cohort (Genetics/Phenotype)	Supplies extensive genotypic and broad phenotypic data for person-centered, large-N analyses [13] [72].
Clinical Phenotyping	ADOS/ADIR/SRS	Gold-standard instruments for assessing core ASD symptom severity and social responsiveness [76] [17].
	RBS-R / SCQ / CBCL	Measures restricted/repetitive behaviors, social communication deficits, and associated behavioral/emotional problems [72].
Neuroimaging Tools	fMRIPrep	Standardized, automated pipeline for robust fMRI data preprocessing, enhancing reproducibility [76] [2].
	Dosenbach 160 Atlas	Predefined parcellation scheme for extracting time series from functional brain networks [76] [2].
Computational Methods	HYDRA (HYDRA)	Semi-supervised clustering algorithm that uses diagnostic labels to identify neurosubtypes from high-dimensional data [1].
	Normative Modeling	Statistical framework to quantify individual deviations from a typical neurodevelopmental trajectory [76].
Behavioral Assays	Eye-Tracking (e.g., Tobii)	Provides objective measures of social attention (e.g., gaze to eyes vs. mouth) for linking neural subtypes to behavior [76] [2].

Subtype Characteristics Map: A conceptual diagram comparing the defining features and implications of the two primary subtyping frameworks.

The classification of Autism Spectrum Disorder (ASD) has undergone a substantial paradigm shift, moving from categorical diagnostic systems toward dimensional, neurobiologically-informed frameworks. This transition reflects growing recognition of the extensive heterogeneity in ASD's etiology, neurobiology, and phenotypic presentation [77]. Historically, the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) conceptualized autism-related conditions as distinct Pervasive Developmental Disorders including Autistic Disorder, Asperger's Disorder, and Pervasive Developmental Disorder Not Otherwise Specified (PDD-NOS) [78]. This categorical approach aimed to differentiate conditions based primarily on behavioral observations and developmental histories. The subsequent DSM-5 consolidation into a single Autism Spectrum Disorder diagnosis in 2013 represented a fundamental reconceptualization, emphasizing core symptom domains across a continuum of severity [79] [77]. While this unified spectrum improved diagnostic reliability, it also created new challenges for research and clinical practice by encompassing tremendous biological and behavioral diversity under a single diagnostic umbrella.

Concurrently, advances in neuroimaging and data science have catalyzed the emergence of data-driven subtyping approaches that seek to decompose ASD heterogeneity using quantitative biological measures. These methods typically apply unsupervised machine learning algorithms to neuroimaging data—particularly resting-state functional magnetic resonance imaging (rs-fMRI)—to identify subgroups of individuals with shared patterns of functional brain organization [17] [12] [2]. Unlike DSM categories defined by clinical phenomenology, data-driven subtypes are derived from patterns of functional connectivity (FC) between brain regions and networks, offering a more direct window into the neurobiological mechanisms underlying ASD. This article provides a comprehensive comparison of these traditional and data-driven subtyping frameworks, with particular emphasis on their methodologies, neurobiological correlates, and implications for research and therapeutic development.

Traditional DSM Subtyping: Framework and Clinical Utility

Historical Development and Diagnostic Criteria

The DSM-IV (1994) and its text revision DSM-IV-TR (2000) established a multi-category classification system for autism spectrum conditions that would influence diagnosis and research for nearly two decades. This framework recognized several distinct disorders under the umbrella term "Pervasive Developmental Disorders":

• Autistic Disorder: Characterized by significant impairments in social interaction and communication, alongside restricted, repetitive patterns of behavior, interests, or activities, with symptoms typically apparent before age three [78]. This category aligned most closely with "classic autism" presentations.

• Asperger's Disorder: Distinguished from Autistic Disorder by the absence of clinically significant language or cognitive delays [78]. Individuals with this profile demonstrated preserved linguistic and intellectual abilities while still exhibiting core social communication challenges and restricted interests.

• Pervasive Developmental Disorder Not Otherwise Specified (PDD-NOS): Functioned as a subthreshold category for individuals who exhibited significant autistic features but did not meet full criteria for other specific disorders [78]. This heterogeneous grouping served as a catch-all diagnosis for atypical or milder presentations.

• Childhood Disintegrative Disorder and Rett Syndrome: Represented rare conditions with specific developmental trajectories and, in the case of Rett Syndrome, identified genetic etiology [78].

The consolidation of these categories into the single diagnosis of Autism Spectrum Disorder in DSM-5 reflected accumulating evidence that previously distinct categories lacked clear boundaries and demonstrated poor diagnostic reliability in practice [77]. Research consistently showed more similarities than differences between conditions like Autistic Disorder and Asperger's Disorder, supporting a dimensional rather than categorical approach [78].

Limitations of Traditional Subtyping for Neurobiological Research

While the DSM-IV framework offered clinicians familiar diagnostic categories, it presented significant limitations for neurobiological research and therapeutic development:

• Behavioral Classification Without Biological Validation: DSM categories were defined exclusively based on behavioral observations rather than neurobiological measures, resulting in groupings that might not align with underlying pathophysiological mechanisms [77].

• Heterogeneity Within Categories: Even within specific DSM diagnoses, individuals displayed substantial variation in symptom presentation, cognitive abilities, and developmental trajectories, complicating efforts to identify consistent neurobiological correlates [12] [77].

• Comorbidity and Boundary Issues: High rates of comorbidity with conditions like ADHD and anxiety disorders, alongside unclear boundaries between diagnostic categories, suggested overlapping rather than distinct biological mechanisms [80] [77].

These limitations motivated the search for alternative subtyping approaches that could better capture the neurobiological diversity within the autism spectrum.

Data-Driven Subtyping Approaches: Methodologies and Workflows

Foundational Neuroimaging Methods

Data-driven subtyping approaches typically begin with the acquisition and processing of resting-state functional MRI data, which measures spontaneous fluctuations in blood oxygen level-dependent (BOLD) signals across the brain. These signal variations provide indirect information about neural activity and functional coordination between brain regions [17] [12]. Standard preprocessing pipelines include procedures to minimize confounding influences from head motion, physiological cycles, and scanner-specific artifacts. Subsequent analytic steps focus on characterizing patterns of functional connectivity across the brain.

Table 1: Core Functional Connectivity Measures in Data-Driven Subtyping

Measure	Description	Biological Interpretation	Analysis Methods
Static Functional Connectivity	Pearson correlation between BOLD time series from different brain regions	Strength of functional communication between brain areas at rest	Correlation matrices, network-based statistics
Dynamic Functional Connectivity	Time-varying changes in connectivity strength	Flexibility and stability of functional interactions across time	Sliding window correlations, dynamic conditional correlation
Local Connectivity	Synchronization between neighboring brain regions	Regional functional specialization	Regional homogeneity (ReHo)
Remote Connectivity	Coordination between distant brain regions	Global integration of information	Voxel-mirrored homotopic connectivity (VMHC), graph theory

Following data preprocessing, researchers extract features for subtyping analyses, which may include static functional connectivity strength (SFCS), dynamic functional connectivity strength (DFCS), dynamic functional connectivity variance (DFCV), and graph-theoretical properties of brain networks [2]. These features collectively capture both stable and time-varying aspects of brain organization that may differentiate ASD subtypes.

Clustering Algorithms and Validation Approaches

The core of data-driven subtyping involves applying unsupervised machine learning algorithms to identify subgroups of individuals with similar functional connectivity profiles. Several approaches have emerged as particularly prominent in ASD research:

• K-means Clustering: Partitions individuals into k clusters by minimizing within-cluster variance in functional connectivity patterns [17]. The optimal number of clusters (k) is typically determined using criteria such as the elbow method or gap statistic.

• Hierarchical Clustering: Creates a nested tree of clusters (dendrogram) that can be divided at different similarity thresholds, allowing flexibility in determining the appropriate level of granularity for subtyping [12].

• Community Detection Algorithms: Identify naturally occurring groups in data based on the density of connections within versus between groups, potentially revealing more complex subgroup structures [80].

These methods are typically applied to data from mixed samples including both ASD participants and typically developing controls, allowing identification of connectivity patterns that transcend traditional diagnostic boundaries [17] [12]. Validation approaches often include demonstrating that identified subtypes show differential patterns of clinical symptoms, cognitive profiles, or treatment responses [2]. More rigorous validation tests whether subtypes replicate in independent datasets and show stability across different clustering methods [12].

The following diagram illustrates a typical workflow for data-driven subtyping studies:

Key Findings: Neurobiological Signatures of Data-Driven Subtypes

Consistently Identified Subtype Patterns

Despite methodological variations across studies, several consistent patterns of functional connectivity subtypes have emerged in the ASD literature:

• Within vs. Between-Network Connectivity Subtypes: Multiple studies have identified subtypes characterized by divergent patterns of connectivity within and between large-scale brain networks. One commonly observed subtype shows increased within-network connectivity coupled with decreased between-network connectivity, while another subtype demonstrates the opposite pattern [17] [12]. These connectivity profiles potentially reflect different balances between functional segregation and integration in brain networks.

• Transdiagnostic Subtypes: Data-driven approaches frequently identify subtypes that cross traditional diagnostic boundaries, with similar functional connectivity patterns observed in individuals with ASD, ADHD, and typical development [80] [12]. This suggests that some neurobiological dimensions underlying ASD may represent continuous traits rather than categorical pathologies.

• Executive Function Profiles: One large study (N=1,012) identified three transdiagnostic subtypes characterized by distinct patterns of executive function strengths and weaknesses: (1) difficulties with flexibility and emotion regulation, (2) primary impairments in inhibition, and (3) weaknesses in working memory, organization, and planning [80]. These cognitive profiles were more strongly associated with specific patterns of frontal-parietal brain activation than traditional diagnostic categories.

Clinical and Cognitive Correlates

Data-driven subtypes show promising associations with clinically relevant variables, supporting their potential utility for personalized intervention approaches:

• Symptom Severity: Subtypes defined by increased within-network connectivity and reduced between-network connectivity often correlate with more severe ASD symptoms, particularly in social communication domains [17] [47].

• Treatment Response: Preliminary evidence suggests that data-driven subtypes may predict differential response to interventions. One study found that chronic intranasal oxytocin treatment yielded a 61.5% response rate in one ASD subtype compared to only 13.3% in another subtype [2].

• Eye-Tracking Profiles: Recent research has linked neuroimaging-based subtypes to differences in visual attention patterns during social tasks, with distinct gaze patterns toward social stimuli observed across subtypes despite similar clinical presentations [2].

Table 2: Representative Data-Driven ASD Subtypes from Recent Studies

Study	Sample Size	Subtypes Identified	Key Neurobiological Features	Clinical Correlates
Vaidya et al. (2020) [80]	1,012	3 EF subtypes	Distinct frontal-parietal engagement patterns	Specific executive function profiles transcending DSM diagnoses
Easson et al. (2019) [17]	266	2 FC subtypes	Subtype 1: stronger between-network FCSubtype 2: stronger within-network FC	Differences in ADOS communication scores
elife (2020) [12]	388	Multiple network-specific subtypes	Compression of primary gradient of functional organization	Moderate association with ASD diagnosis
Molecular Psychiatry (2025) [2]	1,046	2 neural subtypes	Type A: Positive deviations in occipital/cerebellar networks, negative in FP/DMN/COType B: Inverse pattern of Type A	Distinct gaze patterns in eye-tracking tasks
ScienceDirect (2023) [47]	989	2 clinical severity subtypes	Cluster-1: Increased local, decreased remote connectivityCluster-2: Similar to TD with specific MCC impairment	Cluster-1: More severe impairmentCluster-2: Moderate impairment

Comparative Analysis: Traditional vs. Data-Driven Approaches

Methodological and Conceptual Differences

The distinction between traditional DSM-based subtyping and contemporary data-driven approaches reflects fundamental differences in both methodology and underlying conceptual frameworks:

• Basis for Classification: DSM subtyping relies exclusively on observable behaviors and developmental histories, while data-driven approaches classify based on neurobiological measures, particularly patterns of functional brain connectivity [77] [17].

• Handling of Heterogeneity: The DSM-5 spectrum approach acknowledges heterogeneity but provides limited tools for decomposing it, whereas data-driven methods explicitly aim to identify homogeneous subgroups within the broader spectrum [12] [2].

• Dimensional vs. Categorical Structure: DSM categories represent discrete diagnostic entities, while data-driven subtypes often emerge as points along continuous neurobiological dimensions that may extend into typically developing populations [80] [12].

Neurobiological Specificity and Validation

Data-driven subtypes generally demonstrate superior neurobiological specificity compared to traditional diagnostic categories:

• Distinct Neural Signatures: Data-driven subtypes show differential patterns of brain network engagement, particularly in frontoparietal control networks, default mode network, and subcortical structures [80] [2]. These neural differences are often more pronounced than those observed between DSM-defined groups.

• Gradient Alterations: Recent evidence suggests that some ASD connectivity subtypes reflect systematic alterations in the primary gradient of functional brain organization, representing a fundamental shift in how brain networks are spatially organized [12].

• Genetic Associations: While still preliminary, some data-driven subtypes show promising associations with specific genetic variants and molecular pathways, potentially offering insights into underlying pathophysiological mechanisms [77].

The following diagram illustrates key conceptual differences between these approaches:

Implications for Research and Therapeutic Development

Advancing Precision Medicine in ASD

The emergence of robust data-driven subtypes holds significant promise for advancing precision medicine approaches in autism:

• Stratified Clinical Trials: By enrolling participants based on neurobiological subtypes rather than behavioral diagnoses alone, clinical trials may achieve greater homogeneity and increased power to detect treatment effects [2] [77].

• Targeted Interventions: Different connectivity subtypes may benefit from distinct therapeutic approaches. For example, subtypes characterized by specific patterns of visual network connectivity might respond preferentially to different types of social skills training or sensory integration therapies [2].

• Prognostic Refinement: Data-driven subtypes show potential for predicting developmental trajectories and long-term outcomes, enabling earlier implementation of appropriate supports for individuals at risk for particular challenges [80].

Methodological Considerations and Best Practices

Implementing data-driven subtyping in research and clinical contexts requires careful attention to methodological considerations:

• Feature Selection: Studies must strategically select from among various functional connectivity measures (static, dynamic, local, remote) based on specific research questions and participant characteristics [2].

• Multi-Site Harmonization: Combining datasets across multiple imaging sites increases sample sizes and enhances generalizability but requires careful handling of scanner-specific effects through harmonization techniques [12] [2].

• Validation Standards: Identified subtypes should be validated through multiple approaches including replication in independent samples, demonstration of clinical utility, and association with external biological measures [12].

Table 3: Key Methodological Components in Data-Driven ASD Subtyping Research

Component	Representative Examples	Function in Research
Neuroimaging Databases	ABIDE-I, ABIDE-II	Provide large-scale, multi-site datasets with both imaging and phenotypic data for discovery and validation analyses
Preprocessing Tools	fMRIPrep, DPARSF	Standardize data quality control, artifact removal, and initial feature extraction
Functional Parcellations	Dosenbach 160, MIST_20	Define regions of interest for calculating functional connectivity between brain networks
Clustering Algorithms	K-means, Hierarchical Clustering, Community Detection	Identify subgroups with similar connectivity patterns in unsupervised framework
Validation Metrics	Bootstrap ratios, split-half reliability, cross-classification accuracy	Quantify robustness and generalizability of identified subtypes
Behavioral Measures	ADOS, SRS, ADI-R, eye-tracking	Assess clinical symptoms and cognitive profiles associated with neurobiological subtypes

The integration of data-driven subtyping approaches represents a paradigm shift in autism research, moving beyond behaviorally-defined categories toward neurobiologically-informed classification systems. While traditional DSM subtyping provided a important foundation for clinical practice and research, accumulating evidence suggests that functional connectivity-based subtypes offer superior neurobiological specificity and potential for personalized intervention [2] [77]. The consistent identification of subtypes characterized by distinct patterns of within- and between-network connectivity across multiple independent datasets provides compelling evidence for systematic alterations in large-scale brain network organization in ASD.

Future research directions should include:

• Longitudinal Studies: Tracking the stability of neurobiological subtypes across development and their relationship to changing clinical presentations [2].

• Multi-Modal Integration: Combining functional connectivity with structural imaging, genetic data, and other biological measures to create more comprehensive subtyping frameworks [81].

• Clinical Translation: Developing accessible methods for assigning individuals to neurobiological subtypes in clinical settings and testing subtype-specific interventions in randomized trials [77].

In conclusion, while the DSM's unified spectrum represented an important advance over previous categorical systems, data-driven approaches to subtyping offer unprecedented opportunities to decompose ASD heterogeneity and advance toward truly personalized approaches to research and clinical care. The continued refinement of these neurobiologically-informed classification systems holds particular promise for drug development, potentially enabling more targeted therapies for specific ASD subtypes and ultimately improving outcomes for individuals across the autism spectrum.

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in clinical presentation, neurobiology, and developmental trajectories, complicating diagnostic precision and intervention strategies. This comparative guide examines subtype-specific functional network alterations in ASD, synthesizing current research to inform targeted therapeutic development. The classification of ASD has evolved from behaviorally-defined discrete subtypes to a spectrum model, though neuroimaging research continues to reveal distinct neurobiological subtypes that may correspond to specific etiological pathways and treatment responses. Functional magnetic resonance imaging (fMRI) has emerged as a critical tool for delineating these subtypes by capturing intrinsic network organization and connectivity patterns that underlie behavioral manifestations [73] [12]. Recent studies have consistently demonstrated that functional connectivity subtypes are reproducible across datasets and show meaningful associations with clinical profiles, despite similar behavioral presentations [2] [12]. This guide systematically compares methodological approaches, empirical findings, and clinical implications across major studies to provide researchers and drug development professionals with a structured framework for understanding ASD heterogeneity.

Methodological Approaches in ASD Subtyping Research

Data Acquisition and Preprocessing Standards

Current ASD subtyping research employs rigorous standardized protocols for data acquisition and preprocessing to ensure reproducibility across sites. The Autism Brain Imaging Data Exchange (ABIDE) consortium has been instrumental in providing large-scale, multi-site datasets that enable robust subtype identification [2] [73]. Typical preprocessing pipelines include slice-time correction, realignment, normalization to standard space (e.g., MNI152), and nuisance regression (removing head motion, white matter, and cerebrospinal fluid signals) [73]. Studies increasingly implement global signal regression and band-pass filtering (0.01-0.1 Hz) to isolate neural signals of interest [73]. The fMRIPrep pipeline has emerged as a standardized tool for ensuring reproducible preprocessing across research sites [2].

Analytical Frameworks for Subtype Identification

Table 1: Methodological Approaches in ASD Subtyping Studies

Study	Sample Size	Analytical Approach	FC Features	Validation Method
Liu et al. (2025) [2]	1,046 participants (479 ASD, 567 TD)	Normative modeling + clustering	Static FC strength, dynamic FC strength/variability	Independent cohort (n=21) with eye-tracking
Frontiers in Neuroscience (2024) [73]	234 participants (152 autism, 54 Asperger's, 28 PDD-NOS)	Tensor decomposition + statistical comparison	FC, ALFF, fALFF, GMV	Cross-validation within ABIDE I
eLife (2020) [12]	388 participants (ASD + controls)	Hierarchical clustering	Seed-based FC across 18 networks	Replication in independent dataset

Three primary analytical approaches dominate ASD subtyping research:

• Normative Modeling: This approach characterizes individual-level deviations from typical neurodevelopmental trajectories using large reference datasets (typically developing controls). Liu et al. (2025) employed this method to quantify multilevel functional connectivity deviations in ASD participants, then applied clustering algorithms to identify subtypes [2].

• Data-Driven Clustering: Unsupervised machine learning techniques like hierarchical clustering identify subgroups based solely on neuroimaging features without diagnostic preconceptions. This approach has demonstrated robust subtype associations with ASD diagnosis that generalize to independent samples [12].

• Multimodal Feature Integration: Advanced studies integrate multiple neuroimaging modalities (functional connectivity, structural MRI, eye-tracking) to identify subtypes with distinct neurobehavioral profiles [2] [73]. Tensor decomposition methods have been applied to capture different brain communities across ASD subtypes [73].

Diagram 1: Experimental workflow for ASD subtyping studies illustrating sequential stages from data acquisition to validation.

Comparative Analysis of ASD Subtypes

Neural Subtypes Based on Functional Connectivity

Table 2: Functional Connectivity Subtypes in ASD

Subtype Characteristic	Subtype 1 Profile	Subtype 2 Profile	Associated Behavioral Correlates
Frontoparietal Network	Negative deviations	Positive deviations	Executive function, cognitive control
Default Mode Network	Negative deviations	Positive deviations	Social cognition, self-referential thought
Cingulo-Opercular Network	Negative deviations	Positive deviations	Salience processing, attention
Occipital Network	Positive deviations	Negative deviations	Visual processing
Cerebellar Network	Positive deviations	Negative deviations	Motor coordination, cognitive processing
Eye-tracking Patterns	Distinct gaze patterns in social cue tasks	Different social attention profile	Social affect, joint attention

Liu et al. (2025) identified two primary neural subtypes in ASD through normative modeling of functional connectivity. Both subtypes showed comparable clinical presentations but divergent functional network organization. Subtype 1 demonstrated positive deviations in the occipital and cerebellar networks coupled with negative deviations in frontoparietal, default mode, and cingulo-opercular networks. Subtype 2 exhibited the inverse pattern [2]. These neural subtypes were associated with distinct gaze patterns during eye-tracking tasks assessing social cue preference, confirming their behavioral relevance [2] [5].

The study employed both static functional connectivity strength (SFCS) and dynamic functional connectivity (DFC) measures including strength (DFCS) and variability (DFCV) derived from dynamic conditional correlation. This multilevel approach captured both stable and time-varying connectivity properties, offering a more comprehensive characterization of network dysfunction [2].

Traditional Diagnostic Subtypes and Neurobiological Correlates

Research comparing historically defined ASD subtypes (autism, Asperger's, PDD-NOS) has identified distinct neurobiological profiles. A 2024 study found that autism subtype showed significant impairments in the subcortical network and default mode network compared to Asperger's and PDD-NOS [73]. These findings suggest that traditional diagnostic categories may reflect underlying neurobiological differences, despite their removal from DSM-5 in favor of a unified spectrum approach.

The study employed multiple feature extraction methods including amplitude of low-frequency fluctuation (ALFF), fractional ALFF (fALFF), and gray matter volume (GMV) to compare subtypes. Tensor decomposition methods effectively captured distinct brain community patterns across the three subtypes [73].

Network-Level Alterations Across Subtypes

Converging evidence indicates that specific functional networks show consistent alterations across ASD subtypes:

• Default Mode Network (DMN): Multiple studies report DMN dysfunction across ASD subtypes, potentially reflecting common social cognition deficits [2] [73] [12]. The DMN shows both hyperconnectivity and hypoconnectivity patterns across different subtypes, suggesting diverse pathophysiological mechanisms [2].

• Frontoparietal Network (FPN): This executive control network demonstrates subtype-specific alteration patterns, potentially underlying cognitive heterogeneity in ASD [2].

• Salience and Cingulo-Opercular Networks: These networks involved in attention and salience processing show distinct deviation patterns across subtypes, possibly contributing to sensory processing differences [2].

• Visual and Cerebellar Networks: Subtype-specific alterations in these networks may reflect sensory and motor coordination differences in ASD [2].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Materials and Analytical Tools

Tool Category	Specific Solution	Function in ASD Subtyping Research
Data Resources	ABIDE I & II Datasets	Provide large-scale, multi-site neuroimaging data for robust subtype identification
Preprocessing Tools	fMRIPrep	Standardized, reproducible fMRI preprocessing pipeline
	Connectome Computation System	Integrated preprocessing and analysis pipeline
Analytical Frameworks	Normative Modeling	Quantifies individual deviations from typical development
	Hierarchical Clustering	Data-driven subtype identification without a priori diagnostic categories
	Tensor Decomposition	Extracts brain patterns and communities from high-dimensional fMRI data
Feature Extraction	Static & Dynamic FC	Captures stable and time-varying connectivity properties
	ALFF/fALFF	Measures spontaneous brain activity intensity
	Gray Matter Volume	Assesses structural brain differences between subtypes
Validation Approaches	Independent Cohort Replication	Tests generalizability of identified subtypes
	Eye-tracking Tasks	Provides behavioral validation of neurobiological subtypes

Implications for Research and Therapeutic Development

The identification of biologically-based ASD subtypes has profound implications for research and therapeutic development. Clinical trial stratification using neurobiological subtypes may enhance treatment effect detection by reducing heterogeneity within experimental groups [2]. Evidence suggests distinct subtypes may show differential treatment responses, as demonstrated by oxytocin response studies where one subtype showed 61.5% response versus 13.3% in another subtype [2].

For drug development professionals, these findings highlight the importance of target engagement biomarkers based on subtype-specific network alterations. Therapeutic interventions might selectively target specific network deviations, such as DMN-modulating approaches for subtypes with prominent DMN alterations [2]. The correlation between functional connectivity subtypes and eye-tracking measures offers potential for cost-effective screening biomarkers in clinical trials [2] [5].

Future research directions include longitudinal studies to establish subtype stability over development, genetic and molecular correlates of functional subtypes, and intervention studies stratified by neurobiological subtype. Advanced analytical approaches like the Multi-criteria Quantitative Graph Analysis (MQGA) may further refine hub-level analyses of network dysfunction [82].

Diagram 2: Logical pathway from ASD heterogeneity to precision medicine applications showing key stages and components.

The comparative analysis of subtype-specific functional patterns reveals consistent network alterations across ASD subtypes despite methodological variations. The dual subtype model identified through normative modeling and the traditional diagnostic subtypes both demonstrate distinct neurobiological foundations with potential therapeutic implications. Convergence across studies suggests that default mode, frontoparietal, and salience networks represent core networks for understanding ASD heterogeneity. Future research integrating multimodal data, genetic information, and longitudinal designs will further refine these subtypes and accelerate the development of personalized interventions for ASD.

Autism Spectrum Disorder (ASD) is characterized by significant heterogeneity in clinical presentation, underlying biology, and treatment response, posing substantial challenges for therapeutic development. The recognition that ASD likely represents a collection of distinct neurobiological conditions rather than a single entity has catalyzed a paradigm shift toward precision medicine. This transition from a one-size-fits-all approach to targeted interventions necessitates the identification of biologically based subtypes with distinct functional connectivity profiles and predictable treatment outcomes. Groundbreaking research utilizing large-scale datasets and advanced computational methods has begun to delineate these subtypes, offering unprecedented opportunities for matching specific biological profiles with optimized therapeutic strategies. The emerging framework for ASD subtyping integrates multiple data modalities—including neuroimaging, genetic sequencing, eye-tracking, and immune profiling—to define homogeneous subgroups with shared pathophysiological mechanisms. This comparative analysis examines the most current and evidence-based ASD subtyping frameworks, their associated experimental protocols, and their profound implications for developing personalized interventions that align with an individual's unique neurobiological signature.

Comparative Analysis of Major ASD Subtyping Frameworks

Research consortiums have employed diverse methodologies to identify ASD subtypes, resulting in several complementary classification systems. The table below summarizes three prominent frameworks based on functional connectivity, genetics/clinical phenotypes, and treatment response biomarkers.

Table 1: Comparison of Major ASD Subtyping Frameworks

Subtyping Framework	Subtype Classification	Defining Characteristics	Associated Biological Mechanisms	Treatment Response Implications
Functional Connectivity (FC) [2] [1]	Subtype 1: Hyper-connectivitySubtype 2: Hypo-connectivity	- Subtype 1: Positive deviations in occipital/cerebellar networks; negative deviations in frontoparietal/DMN/cingulo-opercular networks [2]- Subtype 2: Inverse FC pattern of Subtype 1 [2]- Distinct gaze patterns in eye-tracking tasks [2]	- Altered within-network and between-network connectivity [1]- Excitation-inhibition imbalance hypothesis	- Suggests differential response to neuromodulation- Predicts variations in social information processing
Genetic & Clinical Phenotype [83] [13]	1. Social/Behavioral (37%)2. Mixed ASD with Developmental Delay (19%)3. Moderate Challenges (34%)4. Broadly Affected (10%)	- Group 1: Core ASD traits + ADHD/anxiety/OCD, no significant DD [83]- Group 2: Developmental delays, language impairments, less psychiatric issues [13]- Group 3: Milder symptoms [83]- Group 4: Severe challenges across domains [83]	- Group 4: Highest burden of rare, high-impact de novo mutations [83]- Group 2: Mix of de novo and inherited rare mutations [13]- Group 1: Common variants linked to ADHD/depression [83]	- Enables targeting of specific biological pathways (e.g., prenatal vs. postnatal brain development)- Informs prognosis and supportive care needs
Immuno-Behavioural Covariation [84]	1. Best Responders (21.5%)2. Medium Responders (55.7%)3. Least Responders (22.8%)	- Classification based on response to bumetanide (NKCC1 inhibitor) [84]- Distinct covariation between cytokine changes and symptom improvement [84]	- Association with interferon-γ, MIG, IFN-α2 cytokine patterns [84]- Immune system interaction with GABAergic signaling [84]	- Baseline cytokine levels predict bumetanide response (AUC=0.832) [84]- Identifies candidates for GABA-targeted therapy

Detailed Experimental Protocols for Subtype Identification

Functional Connectivity Subtyping Using Normative Modeling and Semi-Supervised Clustering

Objective: To identify ASD subtypes based on individual deviations from typical functional connectivity development trajectories [2] [1].

Participants: Large, multi-site datasets such as ABIDE-I and ABIDE-II (e.g., 1,046 participants: 479 with ASD, 567 typical development) [2] and independent validation cohorts (e.g., 21 ASD individuals) with additional eye-tracking data [2].

Image Acquisition and Preprocessing:

• Data Collection: Conduct T1-weighted structural scans and resting-state functional MRI (rs-fMRI) scans [2].
• Preprocessing: Utilize standardized pipelines like fMRIPrep for spatial normalization, head motion correction, and nuisance signal regression [2].
• Quality Control: Exclude participants for excessive head motion (mean Framewise Displacement > 0.3), poor normalization, or missing data [2].

Functional Connectivity Feature Extraction:

• Parcellation: Extract average Blood-Oxygen-Level-Dependent (BOLD) signals from a predefined atlas (e.g., Dosenbach's 160 regions of interest) [2].
• Multilevel FC Metrics: Calculate three types of functional connectivity for each participant [2]:
  • Static Functional Connectivity Strength (SFCS): Using Pearson correlation.
  • Dynamic Functional Connectivity Strength (DFCS): Using dynamic conditional correlation (DCC).
  • Dynamic Functional Connectivity Variance (DFCV): Using DCC.

Normative Modeling and Subtyping Analysis:

• Normative Model Construction: Build a model of typical functional connectivity development using data from the typically developing control group [2].
• Deviation Quantification: Calculate individual-level deviations in multilevel FC for each ASD participant from the normative model [2].
• Clustering: Apply clustering algorithms (e.g., HYDRA - HeterogeneitY through DiscRiminative Analysis) to the deviation scores to identify distinct ASD subtypes [1].
• Validation: Correlate subtypes with clinical symptoms and validate findings in an independent cohort using additional measures like eye-tracking during social tasks [2].

Genetic and Clinical Phenotype Subtyping

Objective: To decompose phenotypic heterogeneity in ASD by integrating large-scale genetic and clinical data [83] [13].

Participants: Large cohorts such as the SPARK study (over 5,000 autistic individuals) [83] [13].

Data Collection:

• Clinical Phenotyping: Collect data on over 230 traits encompassing social interaction, repetitive behaviors, developmental milestones, co-occurring psychiatric conditions (ADHD, anxiety), and cognitive function [83] [13].
• Genetic Sequencing: Perform whole-genome or exome sequencing to identify both common genetic variants and rare mutations (de novo and inherited) [83].

Computational Analysis:

• Trait Integration: Use computational models (e.g., a "person-centered" approach) to group individuals based on their unique combinations of traits, rather than analyzing single traits in isolation [13].
• Class Discovery: Apply advanced clustering algorithms to the integrated clinical data to identify naturally occurring subgroups [83].
• Genetic Validation: Link the identified clinical subtypes to distinct genetic profiles, including burden of rare de novo mutations, inherited rare variants, and polygenic risk scores for psychiatric conditions [83] [13].
• Pathway Analysis: Conduct biological pathway analysis on the genes associated with each subtype to identify disrupted biological processes (e.g., synaptic transmission, chromatin remodeling) [13].

Immuno-Behavioural Subtyping for Treatment Response

Objective: To identify immune-based biomarkers that predict response to bumetanide treatment in ASD [84].

Participants: Young children with ASD (e.g., aged 3-10) participating in a clinical trial of bumetanide [84].

Intervention and Assessment:

• Treatment Protocol: Administer a stable dose of bumetanide (e.g., 0.5mg twice daily) for a defined period (e.g., 3 months) without concomitant psychoactive medications or intensive behavioral interventions [84].
• Behavioral Assessment: Measure symptom severity using standardized tools like the Childhood Autism Rating Scale (CARS) and the Autism Diagnostic Observation Schedule (ADOS) at baseline and post-treatment [84].
• Blood Sampling and Cytokine Measurement: Collect blood samples at baseline and post-treatment. Measure serum levels of a panel of cytokines (e.g., 48 analytes) using multiplex immunoassays [84].

Statistical Analysis for Subtyping:

• Covariation Analysis: Calculate the covariation between improvements in core symptoms (e.g., change in CARS score) and changes in cytokine levels [84].
• Sparse Canonical Correlation Analysis (sCCA): Use sCCA to identify a robust multivariate pattern of cytokine changes that covaries with symptom improvement [84].
• Clustering: Perform clustering analysis based on the identified immuno-behavioural covariation to distinguish responder groups (e.g., best, medium, least) [84].
• Predictive Modeling: Test if baseline cytokine levels can predict treatment response group membership using classifiers (e.g., SVM, Random Forest) and evaluate performance using Area Under the Curve (AUC) [84].

Signaling Pathways and Experimental Workflows

Functional Connectivity Subtyping Workflow

The following diagram illustrates the comprehensive workflow for identifying functional connectivity subtypes in ASD, from data acquisition to clinical validation.

GABA-Immune Interaction Pathway in Bumetanide Response

This diagram outlines the proposed biological mechanism linking immune function to GABA signaling, which underlies differential response to bumetanide treatment.

Table 2: Key Research Resources for ASD Subtyping Studies

Resource Category	Specific Tool/Reagent	Research Application
Neuroimaging	Resting-state fMRI (rs-fMRI)	Measuring functional connectivity between brain regions at rest [2] [1]
Computational Tools	Normative Modeling Framework	Quantifying individual deviation from typical neurodevelopmental trajectories [2]
Clustering Algorithms	HYDRA (HeterogeneitY through DiscRiminative Analysis)	Semi-supervised clustering incorporating diagnostic labels for subtyping [1]
Genetic Analysis	Whole Genome/Exome Sequencing	Identifying rare de novo and inherited mutations, common risk variants [83] [13]
Immunoassays	Multiplex Cytokine Panels	Measuring levels of multiple cytokines in serum/plasma to profile immune function [84]
Behavioral Assessment	ADOS, CARS, SRS	Standardized measurement of core ASD symptoms and treatment response [2] [84]
Eye-Tracking	Tobii TX300 System	Quantifying visual attention patterns during social tasks (e.g., face emotion, joint attention) [2]
Data Resources	ABIDE I/II, SPARK Consortium	Large, shared datasets enabling discovery and validation of subtypes in large samples [2] [13]

Discussion and Future Directions

The convergence of evidence across multiple subtyping approaches confirms that ASD encompasses biologically distinct conditions with diverse underlying mechanisms and treatment needs. The functional connectivity framework provides a direct window into brain network organization, while genetic subtyping reveals fundamental etiological differences, and immuno-behavioural profiling offers practical biomarkers for treatment selection. Critically, these approaches are complementary rather than contradictory, together forming a multidimensional matrix for understanding ASD heterogeneity.

The path forward requires increased sample sizes through multi-site consortia, standardized protocols to enhance reproducibility, and longitudinal studies to track subtype stability and developmental trajectories. Most importantly, future clinical trials must adopt a stratified design, enrolling participants based on their biological subtype to truly test the efficacy of personalized interventions. This precision psychiatry approach promises to transform outcomes for individuals with ASD by moving beyond generic treatments to therapies targeted to their specific neurobiological profile.

Conclusion

The stratification of ASD into functional connectivity subtypes represents a paradigm shift from symptom-based description to neurobiologically informed classification. The convergence of evidence points to reproducible subtypes, primarily characterized by distinct patterns of hyper- and hypo-connectivity that are not fully explained by clinical presentation alone. Methodologically, semi-supervised and normative modeling approaches show superior performance in deriving robust and clinically relevant subtypes, though the field must prioritize replication and address the discrete-versus-continuous nature of these groupings. Crucially, these subtypes demonstrate divergent neuro-behavioral relationships and may underpin differential treatment responses, as suggested by studies on oxytocin sensitivity. The future of ASD research and drug development lies in leveraging these subtypes to stratify clinical trials, develop circuit-specific therapeutics, and ultimately deliver on the promise of precision medicine for this heterogeneous population.

References

• 1. Diagnosis-informed neuro-subtyping reveals subgroups of ... [https://www.sciencedirect.com/science/article/abs/pii/S0278584625002064]
• 2. Identifying brain functional subtypes and corresponding ... [https://www.nature.com/articles/s41380-025-03086-x]
• 3. Multidimensional Neuroanatomical Subtyping of Autism ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7190887/]
• 4. Identification and analysis of behavioral phenotypes in ... [https://www.sciencedirect.com/science/article/pii/S1386505618308669]
• 5. Identifying brain functional subtypes and corresponding ... [https://pubmed.ncbi.nlm.nih.gov/40542132/]
• 6. Diagnosis-informed connectivity subtyping discovers ... [https://www.sciencedirect.com/science/article/pii/S1053811922003366]
• 7. FDA Takes Action to Make a Treatment Available for ... [https://www.fda.gov/news-events/press-announcements/fda-takes-action-make-treatment-available-autism-symptoms]
• 8. Diagnosis-informed neuro-subtyping reveals subgroups of ... [https://pubmed.ncbi.nlm.nih.gov/40681079/]
• 9. to Hypo-Connectivity Occurring at Pre-Adolescence in ... [https://pubmed.ncbi.nlm.nih.gov/40948193/]
• 10. Contracted functional connectivity profiles in autism [https://molecularautism.biomedcentral.com/articles/10.1186/s13229-024-00616-2]
• 11. Comparison of autism spectrum disorder subtypes based ... [https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1440222/full]
• 12. Functional connectivity subtypes associate robustly with ... [https://elifesciences.org/articles/56257]
• 13. 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]
• 14. ASD Subtypes Study: Autism Brain Connectivity Patterns ... [https://www.additudemag.com/autism-spectrum-disorder-subtypes-adhd-asd-brain-connectivity/?srsltid=AfmBOor9NAq_35915OEI9oxtMYGhnDCiLmCCHa18yHbYrBzxC3tmUim0]
• 15. The Default Mode Network in Autism - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC5635856/]
• 16. Abnormal functional connectivity of default mode sub-networks in autism spectrum disorder patients - PubMed [https://pubmed.ncbi.nlm.nih.gov/20621638/]
• 17. Functional connectivity-based subtypes of individuals with and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6370474/]
• 18. Frontoparietal Network in Executive Functioning in Autism ... [https://pubmed.ncbi.nlm.nih.gov/33016005/]
• 19. Frontoparietal Network Connectivity During an N-Back ... [https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2020.551808/full]
• 20. Visual Perception in Autism Spectrum Disorder: A Review of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7350544/]
• 21. Functional connectivity between the visual and salience ... [https://pubmed.ncbi.nlm.nih.gov/40295911/]
• 22. The developing visual system: A building block on the path ... [https://www.sciencedirect.com/science/article/pii/S1878929325000428]
• 23. Development of frontoparietal connectivity predicts ... [https://www.nature.com/articles/s41398-019-0418-5]
• 24. The geometric preference subtype in ASD: identifying a ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5861622/]
• 25. Eye Tracking Reveals Abnormal Visual Preference for ... [https://www.sciencedirect.com/science/article/pii/S0006322315003108]
• 26. The relationship between sensory processing sensitivity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10826854/]
• 27. Sensory Processing and Medication Sensitivity Connection [https://sensitivityresearch.com/unravelling-the-link-between-sensory-processing-sensitivity-and-medication-sensitivity/]
• 28. Different Eye Tracking Patterns in Autism Spectrum ... [https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2022.899521/full]
• 29. Eye Tracking Reveals Impaired Attentional Disengagement ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5359772/]
• 30. Contrast subgraphs catch patterns of altered functional ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12234876/]
• 31. Linking connectivity dynamics to symptom severity and ... [https://pubmed.ncbi.nlm.nih.gov/41006060/]
• 32. An analytics-driven model for identifying autism spectrum ... [https://www.sciencedirect.com/science/article/pii/S2772442525000280]
• 33. Identifying Prognostic Markers in Autism Spectrum Disorder ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7166165/]
• 34. The Autism Spectrum Disorder Subtypes Identification ... [https://pubmed.ncbi.nlm.nih.gov/40730718/]
• 35. HYDRA: revealing Heterogeneity of imaging and genetic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5408358/]
• 36. Atypical maturation of the functional connectome hierarchy in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11948645/]
• 37. A clustering-based method to detect functional connectivity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3342474/]
• 38. Hierarchical clustering to measure connectivity in fMRI ... [https://www.sciencedirect.com/science/article/abs/pii/S0730725X02005039]
• 39. Neuroimaging profiling identifies distinct brain maturational ... [https://www.nature.com/articles/s41380-022-01925-9]
• 40. Static and temporal dynamic changes in intrinsic brain activity ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9471141/]
• 41. Static and temporal dynamic changes in intrinsic brain ... [https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2022.971062/full]
• 42. Graph theory application with functional connectivity to ... [https://www.sciencedirect.com/science/article/abs/pii/S092012112030499X]
• 43. Effects of Alterations in Resting-State Neural Networks on the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10376439/]
• 44. Disrupted Strength and Stability of Regional Brain Activity ... [https://www.sciencedirect.com/science/article/abs/pii/S0306452221003304]
• 45. Evolution of spatial and temporal features of functional brain ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC6613816/]
• 46. Semi-supervised clustering for neuro-subtyping of autism ... [https://consensus.app/papers/details/2a65cc5536875947aaa65e1f96081430/]
• 47. Stratifying ASD and characterizing the functional ... [https://www.sciencedirect.com/science/article/pii/S0166432823001766]
• 48. Distinct neuroimaging subtypes of ADHD among ... [https://www.nature.com/articles/s41398-025-03662-3]
• 49. Spectral feature modeling with graph signal processing for ... [https://www.nature.com/articles/s41598-025-06489-6]
• 50. Site effects how-to and when - PubMed Central - NIH [https://pmc.ncbi.nlm.nih.gov/articles/PMC9661923/]
• 51. A framework for the analysis of phantom data in multicenter ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3458186/]
• 52. Correcting waveform bias using principal component ... [https://www.sciencedirect.com/science/article/abs/pii/S0966636216306105]
• 53. Contrast subgraphs catch patterns of altered functional ... [https://www.nature.com/articles/s41598-025-04932-2]
• 54. Utility of Continuous Disease Subtyping Systems for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8997600/]
• 55. What is the difference between discrete and continuous ... [https://www.scribbr.com/frequently-asked-questions/discrete-vs-continuous-variables/]
• 56. Discrete vs. Continuous Variables: Differences Explained [https://articles.outlier.org/discrete-vs-continuous-variables]
• 57. Evaluating Discrete Time Methods for Subgrouping ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10873483/]
• 58. Parsing Clinical and Neurobiological Sources of ... [https://www.sciencedirect.com/science/article/pii/S0006322325011862]
• 59. 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/]
• 60. Four New Autism Subtypes Link Genes to Children's Traits [https://www.scientificamerican.com/article/four-new-autism-subtypes-link-genes-to-childrens-traits/]
• 61. Unsupervised Dimensionality Reduction Techniques for the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12262183/]
• 62. Analysis and comparison of feature selection methods ... [https://www.sciencedirect.com/science/article/abs/pii/S0957417424005335]
• 63. Dimensionality Reduction vs Feature Selection [https://medium.com/biased-algorithms/dimensionality-reduction-vs-feature-selection-504e47051f89]
• 64. Dimensionality Reduction in Machine Learning for ... [https://www.guvi.in/blog/dimensionality-reduction-in-machine-learning/]
• 65. A benchmark analysis of feature selection and machine ... [https://www.sciencedirect.com/science/article/pii/S2001037025001382]
• 66. Optimizing high dimensional data classification with a ... [https://www.nature.com/articles/s41598-025-08699-4]
• 67. Cluster validity indices for automatic clustering [https://www.sciencedirect.com/science/article/pii/S2405844025003330]
• 68. Benchmarking validity indices for evolutionary K-means ... [https://www.nature.com/articles/s41598-025-08473-6]
• 69. Extended multivariate comparison of 68 cluster validity ... [https://www.sciencedirect.com/science/article/pii/S0169743924000571]
• 70. A comparative analysis of unsupervised machine‐learning ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12069737/]
• 71. Polygenic and developmental profiles of autism differ by ... [https://www.nature.com/articles/s41586-025-09542-6]
• 72. Decomposition of phenotypic heterogeneity in autism ... [https://www.nature.com/articles/s41588-025-02224-z]
• 73. Comparison of autism spectrum disorder subtypes based ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11486766/]
• 74. Cross-cohort replicability and generalizability of ... [https://www.sciencedirect.com/science/article/pii/S105381192200684X]
• 75. Machine learning prediction of cognition from functional ... [https://pubmed.ncbi.nlm.nih.gov/34673248/]
• 76. Identifying brain functional subtypes and corresponding task ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12532700/]
• 77. Should the Autism Spectrum Be Split Apart? | DW [https://www.disabled-world.com/disability/publications/journals/asd-sub-division.php]
• 78. : Understanding the... - Autism Parenting Magazine Subtypes Autism [https://www.autismparentingmagazine.com/understanding-autism-subtypes/]
• 79. -IV DSM . Autism -IV-TR vs - What's the... | This DSM . That Autism vs [https://thisvsthat.io/dsm-iv-autism-vs-dsm-iv-tr-autism]
• 80. - Data identification of driven of executive function across... subtypes [https://pubmed.ncbi.nlm.nih.gov/31509248/]
• 81. A holistic review of autism spectrum disorder mechanisms ... [https://www.sciencedirect.com/science/article/pii/S0969996125004048]
• 82. MQGA: A quantitative analysis of brain network hubs using ... [https://www.sciencedirect.com/science/article/pii/S1053811924004105]
• 83. New Autism Study Uncovers Four Biological Subtypes [https://www.psychiatrist.com/news/new-autism-study-uncovers-four-biological-subtypes/]
• 84. The immuno-behavioural covariation associated with ... [https://www.nature.com/articles/s41398-022-01987-x]