Functional Brain Asymmetry Reveals Heterogeneous Subtypes in Autism Spectrum Disorder Authors: Chae Yeon Kim, Sunghun Kim, Bo‐yong Park Published: 2026 ABSTRACT ABSTRACTHeterogeneity is a critical factor in understanding inter‐individual brain and behavioral variability in autism spectrum disorder (ASD). Since individuals with ASD exhibit atypical communication and social interaction skills closely linked to brain lateralization, this study investigated ASD heterogeneity using an asymmetry index that captures interhemispheric differences in brain function. Degree centrality was computed from the functional connectivity matrix, and its asymmetry was quantified for each brain region. Using an unsupervised clustering algorithm, we identified three distinct subtypes within the ASD population that showed significant differences in brain asymmetry in the somatomotor, frontoparietal, and default mode networks. Comparisons of symptom severity indicated that subtype 2, which displayed atypical asymmetry patterns along the sensory–default mode axis, exhibited higher clinical severity. Functional connectivity analyses further revealed that subtype 1, characterized by rightward asymmetry, demonstrated hypoconnectivity across sensory and heteromodal association regions, whereas subtype 3, characterized by leftward asymmetry, exhibited hyperconnectivity. Cognitive decoding suggested distinct neurocognitive profiles for each subtype: subtype 1 was associated with increased engagement of self‐referential and motivational processes (social‐motivational subtype), subtype 2 with impaired regulatory and executive functions (cognitive control/impulsive subtype), and subtype 3 with enhanced involvement of mnemonic and language‐related systems (memory‐language subtype). Together, these findings advance our understanding of functional brain heterogeneity in ASD and demonstrate that subtype identification based on hemispheric asymmetry offers a promising framework for elucidating topological and behavioral variability within the ASD population. Introduction Autism spectrum disorder (ASD) is a common neurodevelopmental condition with a global prevalence of approximately 1%–2% (Salari et al. 2022). It is characterized by impairments in social interactions and communication, accompanied by restricted and repetitive patterns of behavior (Redfield et al. 2014). A hallmark of ASD is its marked heterogeneity, which is reflected in the substantial inter‐individual variability in symptom presentation and severity (Lenroot and Yeung 2013; Masi et al. 2017). Accordingly, ASD is conceptualized as a spectrum that encompasses multiple coexisting neurobiological pathways (Lombardo et al. 2019; Masi et al. 2017). Recent neuroimaging studies have advanced the delineation of this heterogeneity by identifying ASD subtypes based on magnetic resonance imaging (MRI) features (Hong et al. 2023; Jiang et al. 2024; Wang et al. 2024). These studies revealed distinct neuroanatomical and functional profiles, including variations in cortical morphology (Hong et al. 2018), functional connectivity (Easson et al. 2019), and large‐scale network organization (Hahamy et al. 2015). These findings indicate that neuroimaging‐derived phenotypes can capture biologically meaningful dimensions of the autism spectrum. In parallel, recent fMRI‐based, data‐driven approaches have further characterized functional heterogeneity in ASD by identifying subtypes with distinct connectivity profiles that are not fully captured by traditional case–control contrasts (Rasero et al. 2023). These studies have revealed the presence of both hypo‐ and hyperconnected subgroups, highlighting substantial inter‐individual variability and supporting the existence of neurobiologically meaningful subtypes within ASD (Guo et al. 2022). Nevertheless, given the distributed and multifactorial nature of ASD‐related brain alterations, further exploration of neural features that reflect the fundamental organizational principles of the brain is warranted to deepen our understanding of its biological diversity. Thus, an analysis that quantifies individual differences while minimizing between‐subject variance within a given diagnostic group is required. Hemispheric asymmetry provides such an index, as it represents a foundational organizational property of the brain (Douglas et al. 2018). As a fundamental principle of human brain organization, hemispheric asymmetry is closely associated with higher‐order cognitive functions such as language, executive control, and social cognition (Toga and Thompson 2003). During typical development, structural and functional asymmetries gradually emerge and strengthen from infancy to adolescence, facilitating efficient information processing and functional specialization across higher‐order domains (Bisiacchi and Cainelli 2022; Lee et al. 2024; Liu et al. 2022). Given that atypical communication and language abilities, which are the core symptoms of ASD, are closely related to hemispheric specialization (Lee et al. 2017; Senouci 2021; Vogindroukas et al. 2022), disrupted brain asymmetry may serve as a biologically grounded marker for capturing individual variability within the spectrum. In support of this notion, previous studies have reported reduced cortical thickness asymmetry (Kong et al. 2022) and widespread alterations in structural connectome asymmetry (Yoo et al. 2024). Furthermore, large‐scale evidence from the Enhancing Neuro‐Imaging Genetics through Meta‐Analysis (ENIGMA) consortium demonstrated globally reduced cortical thickness asymmetry in ASD (Postema et al. 2019), suggesting that aberrant asymmetry reflects altered neurodevelopmental processes that may underlie the diverse behavioral and cognitive phenotypes observed in ASD. Building on these insights, the present study investigated ASD heterogeneity using functional MRI‐derived brain asymmetry. Using multi‐site datasets, we first identified differences in asymmetry between individuals with ASD and typically developing (TD) controls. We then characterized the distinct asymmetry‐based ASD subtypes and examined their associations with symptom severity, brain network organization, and cognitive characteristics. Our study identified subtypes that exhibited distinct principles of hemispheric organization and network configuration beyond the traditional diagnostic boundaries. Using this approach, we aimed to establish functional brain asymmetry as a biologically meaningful dimension to characterize ASD heterogeneity. Methods Study Participants and Data Acquisition We analyzed T1‐weighted structural MRI and resting‐state functional MRI (rs‐fMRI) data obtained from the Autism Brain Imaging Data Exchange I initiative (Di Martino et al. 2014). The sample comprised 247 individuals with ASD (mean ± standard deviation [SD] age = 18.43 ± 8.86 years; 88.2% male) and 332 TD controls (17.61 ± 7.68 years; 82.2% male). The participants with failed cortical surface reconstruction, subcortical segmentation, or inaccurate registration to the standard template were excluded. Imaging data were acquired across 19 independent sites. Detailed demographic information and MRI data acquisition parameters are provided in Table S1. Data Preprocessing The imaging data were preprocessed using Micapipe v0.2.0 (https://github.com/MICA‐MNI/micapipe) (Cruces et al. 2022). Briefly, T1‐weighted MRI data were de‐obliqued and reoriented to the left posterior–inferior orientation. Intensity non‐uniformity correction, intensity normalization, and non‐brain tissue removal were subsequently performed. Subcortical structures, including the accumbens, amygdala, caudate, hippocampus, pallidum, putamen, and thalamus, were segmented using FSL‐FIRST (Patenaude et al. 2011). Cortical surfaces were reconstructed using FreeSurfer (Bruce Fischl et al. 2002; Dale et al. 1999; Fischl 2012; Fischl, Sereno, and Dale 1999; Fischl, Sereno, Tootell, and Dale 1999; Ségonne et al. 2007), and the mid‐thickness surface was generated by averaging the white matter and pial surfaces. The mid‐thickness surface was then used to generate a spherical surface, which was registered on a Conte69 template surface. The rs‐fMRI data were preprocessed as follows: The first five volumes were discarded to allow for magnetic field stabilization, and the remaining images were reoriented. Head motion correction was performed by realigning all volumes to the mean image over time. After non‐brain tissue removal, nuisance signals were removed using FMRIB's ICA‐based X‐noiseifier (Salimi‐Khorshidi et al. 2014), and motion outliers and spikes were regressed. The processed volumetric data were registered to the Conte69 template surface using MSMAll (Robinson et al. 2014), and the resulting time series were spatially smoothed with a 10 mm full‐width at half‐maximum Gaussian kernel. Functional Brain Asymmetry Functional brain asymmetry was evaluated using a functional connectivity (FC) matrix constructed by computing Pearson's correlations of the time series between brain regions defined by the Yan atlas, comprising 300 symmetric cortical parcels (Yan et al. 2023) and 14 individually parcellated bilateral subcortical regions. The resulting correlation coefficients were Fisher's r‐to‐z transformed, and row‐wise thresholding was applied by retaining the top 10% of the connections for each region. Degree centrality (DC), defined as the sum of the weights of all links connected to a given node, was calculated from the thresholded FC matrix (Rubinov and Sporns 2010). To ensure the robustness of the findings across different levels of matrix sparsity, we additionally calculated DC using alternative thresholds (5%, 20%, and 30%). The interhemispheric asymmetry of the DC was calculated as follows: asymmetry = (left—right)/{(left + right)/2}, where left and right denote the DC values in the left and right hemispheres, respectively. Positive values indicate leftward asymmetry (i.e., a higher DC in the left hemisphere relative to the right), whereas negative values indicate rightward asymmetry. Between‐group differences in DC asymmetry were assessed using two‐sample t‐tests. To account for potential confounding effects, site‐related variability was harmonized using NeuroComBat (Fortin et al. 2018), which corrects for site effects while preserving biological variation. Age and sex were included as covariates in the group comparisons. To address the data imbalance between the ASD and TD groups, individuals with TD were randomly subsampled 100 times, and t‐statistics were computed for each iteration. Significant regions were identified using Fisher's method, with a p‐value < 0.05 (Whitlock 2005), and multiple comparisons across brain regions were corrected using a false discovery rate (FDR) < 0.05 (Benjamini and Hochberg 1995). Heterogeneity Modeling To characterize heterogeneity within the ASD group, we applied the neuralization‐propagation (NEON) clustering algorithm to the asymmetry index (Kauffmann et al. 2021). NEON provides an efficient and robust approach for identifying clusters, while offering feature‐level interpretability by assigning relative contribution weights to each feature during clustering (Kauffmann et al. 2021). This algorithm builds upon the traditional k‐means framework to enhance the interpretability of the cluster outcomes. Given the sensitivity of k‐means clustering to the random initialization of cluster centroids, we assessed stability across multiple random seeds and selected robust clusters for further analyses. The optimal number of clusters was determined by evaluating several clustering validity metrics, including the Elbow method (Thorndike 1953), the Silhouette coefficient (Rousseeuw 1987), and the Calinski‐Harabasz score (Caliński and Harabasz 1974), for cluster numbers ranging from two to nine. To assess the stability of the clustering, we repeatedly sampled 90% of the subjects (1000 iterations) and computed clustering solutions for k = 2–9. For each k, we evaluated clustering quality using the metrics and quantified their variability across iterations using standard deviations. Each metric was min‐max normalized across k and rescaled so that higher values indicated better clustering quality. Additionally, hierarchical clustering with Ward's linkage (Ward 1963) was performed to visualize the hierarchical relationships among the clusters using a dendrogram. Between‐Subtype Differences in Brain Asymmetry To examine differences in DC asymmetry across ASD subtypes, we performed a one‐way analysis of variance (ANOVA), including age and sex as covariates of no interest. Post hoc pairwise comparisons were conducted using the Games‐Howell test to identify differences between subtype pairs. Multiple comparisons were corrected using an FDR < 0.05 (Benjamini and Hochberg 1995). We conducted several sensitivity analyses. First, we performed the same analyses with handedness included as a covariate of no interest. Second, head motion measured using framewise displacement (FD) was included in the model. Finally, we included handedness, FD, and intelligence quotient (IQ) as covariates of no interest. Phenotypic Representations of ASD Subtypes We characterized the phenotypic profiles of ASD subtypes across three domains: (1) symptom severity, (2) functional brain organization, and (3) cognitive function. First, the Autism Diagnostic Observation Schedule (ADOS) total score and sub‐scores (communication, social cognition, and repetitive behavior) were compared across subtypes. Between‐group differences were evaluated using two‐sample t‐tests, and multiple comparisons were corrected using FDR (Benjamini and Hochberg 1995). Second, DC was computed for each individual's FC matrix and compared across subtype pairs. Two‐sample t‐tests were conducted for each brain region, and multiple comparisons across regions were corrected using FDR (Benjamini and Hochberg 1995). Finally, cognitive decoding was performed using NeuroSynth (Yarkoni et al. 2011) implemented in the BrainStat toolbox (Larivière et al. 2023). NeuroSynth is a meta‐analytic framework that identifies the cognitive terms associated with brain maps. In the present study, we applied cognitive decoding to the asymmetry map of each ASD subtype to infer putative cognitive associations. To improve interpretability, we excluded terms corresponding to anatomical labels, numeric descriptors, and general or non‐specific expressions. Detailed classifications and correlation coefficients are provided in Data S1. The remaining terms were ranked according to their correlation coefficients for visualization. Results Between‐Group Differences in Brain Asymmetry We calculated brain asymmetry using the DC values derived from the FC matrices (Figure 1A) and visualized the asymmetry maps for the ASD and TD groups (Figure 1B). Both groups exhibited leftward asymmetry in the frontoparietal and default mode networks and rightward asymmetry in the somatomotor and visual cortices (Figure 1B). Significant between‐group differences in DC asymmetry were observed across multiple brain regions (Figure 1C). Specifically, individuals with ASD showed left‐dominant FC in the ventromedial prefrontal cortex, angular gyrus, and superior temporal pole, and right‐dominant asymmetry in the visual cortex, dorsomedial prefrontal cortex, and pallidum. In contrast, TD controls exhibited left‐dominant asymmetry in the dorsolateral prefrontal cortex, inferior parietal lobule, precuneus, lateral temporal region, accumbens, and right‐dominant asymmetry in the medial paracentral lobule and visual association cortex. These findings suggest that atypical asymmetry patterns in individuals with ASD are predominantly localized within the default mode and executive control networks. When we repeated computing DC maps across different thresholds, highly consistent patterns were observed relative to the 10% threshold (r = 0.990 at 5%, r = 0.982 at 20%, and r = 0.928 at 30%), indicating robustness. Functional connectivity‐based brain asymmetry in ASD. (A) Schematic illustration of the computation of brain asymmetry. (B) Degree centrality asymmetry maps for the ASD and TD groups. (C) T‐statistics for regions showing significant (FDR < 0.05) between‐group differences in asymmetry are shown on brain surfaces. Histograms depict the distribution of t‐statistics in the identified regions for the ASD and TD groups, with directional shifts indicated by arrows. ASD, autism spectrum disorder; FDR, false discovery rate; L, left; R, right; TD, typically developing. Identification of ASD Subtypes Using Functional Brain Asymmetry According to the Elbow method, the Silhouette coefficient, and the Calinski‐Harabasz score (Figure 2A), the optimal number of clusters was suggested to be two. However, the average variability across the three metrics exhibited that k = 3 showed the lowest variability across iterations (Table S2), indicating that the three‐cluster solution demonstrates slightly greater stability than the two‐cluster solution. Importantly, the three‐cluster solution partially preserved the leftward and rightward subtypes identified by the two‐cluster solution (Figure S1A), suggesting that the addition of a third cluster does not compromise the overall validity or interpretability of the results. Thus, we adopted a three‐cluster solution to capture the heterogeneity among individuals with ASD (Figure 2B). Demographic information for each subtype is summarized in Table S3. Subtype 1 showed overall rightward asymmetry, whereas subtype 3 showed leftward asymmetry. Subtype 2 was characterized by leftward asymmetry in the default mode and frontoparietal networks, and rightward asymmetry in the somatomotor and visual cortices (Figure 2B). A comparison of asymmetry values across subtypes revealed significant differences in the primary motor cortex, prefrontal cortex, and visual association areas (Figure 2C). Network‐wise analysis demonstrated that the somatomotor, frontoparietal, and default mode networks exhibited strong subtype‐related differences. Pairwise comparisons revealed distinct asymmetry patterns across subtypes, with subtype 2 showing significant differences relative to the other subtypes, particularly along the sensory–default mode hierarchical axis (Figure 2C). To further interpret the clustering results, we examined feature contributions to identify brain regions that disproportionately influenced cluster formation (Figure S2). Notably, the default mode and frontoparietal networks showed the highest contribution values, suggesting that asymmetry within these networks is particularly important for distinguishing between ASD subtypes. These results remained consistent after controlling for potential confounding factors. Differences in DC asymmetry across subtypes remained robust when controlling for handedness (Figure S4) or head motion (Figure S5). Particularly, when we simultaneously controlled for handedness, FD, and IQ, largely consistent findings were observed (F‐statistic maps: r = 0.759; subtype 1 vs. 2 t‐statistic maps: r = 0.868; subtype 1 vs. 3: r = 0.833; subtype 2 vs. 3: r = 0.886; all p < 0.001; Figure S6). To further assess the potential influence of age differences between the subtypes, we performed an additional analysis after removing the two oldest participants from Subtype 3, which rendered the between‐group age difference non‐significant (p = 0.12). The resulting maps remained highly consistent with the original findings (F‐statistic maps: r = 0.998; subtype 1 vs. 2 t‐statistic maps: r = 1.000; subtype 1 vs. 3: r = 0.996, subtype 2 vs. 3: r = 0.998; all p < 0.001; Figure S7), indicating that the subtype‐specific asymmetry patterns are not driven by age differences. The results of the two‐cluster solution are shown in Figure S1. Subtype identification within ASD. (A) Hierarchical clustering dendrogram showing three clusters delineated by the red line (left). Line plots display the Elbow method, Silhouette coefficient, and Calinski‐Harabasz score for various cluster numbers (right). Error bars represent the subsampling‐based means and standard deviations across iterations. (B) Degree centrality asymmetry maps for each ASD subtype are visualized on brain surfaces. (C) F‐statistic values showing regions with significant between‐subtype differences in asymmetry are shown on brain surfaces, stratified by functional brain networks and subcortical regions (left). T‐statistic maps from pairwise subtype comparisons are shown on the right. ASD, autism spectrum disorder; DAN, dorsal attention network; DMN, default mode network; FDR, false discovery rate; FPN, frontoparietal network; L, left; LN, limbic network; R, right; SMN, somatomotor network; VAN, ventral attention network; VN, visual network. Behavioral and Brain Phenotype Differences Between the Subtypes Pairwise comparisons of the ADOS scores revealed significant differences, particularly for subtype 2 (Figure 3A). Subtype 2 exhibited significantly higher ADOS total, communication, and social scores than the other subtypes, suggesting that subtype 2, which is characterized by atypical asymmetry patterns along the sensory‐default mode hierarchy, was associated with greater symptom severity. When comparing DC across subtypes, subtype 2 demonstrated significantly higher hubness in the ventrolateral prefrontal cortex and inferior parietal lobule, as well as reduced connectivity within the somatomotor region, relative to the other subtypes (Figure 3B). These findings indicate that subtype 2 shows pronounced alterations along the sensory‐heteromodal axis. Moreover, subtype 3 exhibited significantly higher DC values than did subtype 1 in the ventrolateral prefrontal cortex, inferior parietal lobule, and somatomotor cortex (Figure 3B). These results suggest that rightward asymmetry (subtype 1) is associated with hypoconnectivity, whereas leftward asymmetry (subtype 3) is associated with hyperconnectivity in sensory and heteromodal association regions. To further characterize the cognitive relevance of these asymmetric patterns, we performed meta‐analytic decoding for each subtype (Figure 3C). Subtype 1 was associated with terms related to motivation and self‐referential processing; subtype 2 with cognitive control and inhibitory processes, patterns consistent with previous findings in autism; and subtype 3 with memory and language functions, reflecting mnemonic and semantic processing. These decoding results indicate that the identified subtypes are not only structurally distinct but also functionally dissociable, with each preferentially engaging in distinct higher‐order cognitive domains. In summary, subtype 1 can be described as a social‐motivational subtype, subtype 2 as a cognitive control/impulsive subtype, and subtype 3 as a memory‐language subtype. Additionally, when each ASD subtype was compared with the TD group and the resulting t‐statistic maps were decoded (Figure S3), subtype 2 showed associations with terms such as signal task and response inhibition, reflecting control and inhibitory functions. Subtypes 1 and 3 exhibited similar decoding profiles, although subtype 3 exhibited stronger associations with memory‐ and language‐related terms. Phenotypic characterization of ASD subtypes. (A) Violin plots showing the distributions of ADOS scores across ASD subtypes, with significant pairwise differences indicated by asterisks. (B) Whole‐brain degree centrality values for each subtype are displayed on brain surfaces, and violin plots on the right depict the values in regions showing significant between‐subtype differences. (C) Cognitive decoding of asymmetry patterns for each ASD subtype. ADOS, autism diagnostic observation schedule; ASD, autism spectrum disorder. Discussion Heterogeneity is a core feature of ASD, and understanding this inter‐individual variability is crucial for elucidating the underlying neurobiological mechanisms and advancing personalized diagnostic and therapeutic approaches. In this study, we investigated hemispheric asymmetry in functional brain organization to identify distinct subtypes within an ASD cohort. Beyond revealing markedly different spatial patterns of asymmetry in language and social‐cognitive networks between individuals with ASD and TD controls, we also observed pronounced heterogeneity in asymmetry patterns within the ASD population. These findings suggest that ASD does not arise from a single neurobiological pathway but instead encompasses diverse patterns of functional hemispheric specialization. Specifically, cognitive and sensory processing characteristics appear to be differentially organized across ASD subtypes. Taken together, our results provide insights into the phenotypic diversity and topological organization of the brain in ASD. Analyses of ASD and TD groups revealed hemispheric asymmetry, with notable leftward and rightward patterns. Individuals with ASD demonstrated pronounced left‐dominant connectivity in the ventromedial prefrontal cortex, angular gyrus, superior temporal pole, and somatomotor cortex, areas vital for social cognition, language comprehension, and motor planning and imitation. In particular, lateralization of the ventromedial prefrontal cortex has been closely associated with self‐referential processing, social interaction, and theory of mind, and atypical lateralization of these regions has been reported in individuals with ASD and social‐cognitive impairments (Gotts et al. 2012; Nielsen et al. 2014). Abnormal asymmetry in the angular gyrus and the superior temporal pole has been associated with deficits in language comprehension, verbal and nonverbal communication, and emotional cue interpretation (Li et al. 2023; Wan et al. 2023). Moreover, altered lateralization in the somatomotor cortex may reflect disrupted feedback integration of motor information, which is consistent with the sensorimotor hyperconnectivity previously observed in ASD (Benkarim et al. 2021; Cerliani et al. 2015; Ilioska et al. 2023). Conversely, right hemispheric dominance was observed in the visual cortex, dorsomedial prefrontal cortex, and pallidum, which are regions implicated in social cognition (Wittmann et al. 2025) and sensory processing (Cardinale et al. 2013; Wan et al. 2023). Rightward asymmetry in these areas may reflect atypical attentional shifts (Eickhoff et al. 2016) and altered sensory integration (Cardinale et al. 2013; Wan et al. 2023). In summary, individuals with ASD exhibit atypical or even reversed hemispheric lateralization across the social‐cognitive, language, and sensorimotor networks, in contrast to the typical left–right specialization observed in TD controls. These findings highlight that social and language impairments in ASD are closely linked to disrupted hemispheric organization of functional brain networks. Functional brain asymmetry delineated three distinct subtypes of ASD. Notably, subtypes (subtype 1 and 3) exhibited pronounced rightward and leftward asymmetries, respectively. The left hemisphere is primarily associated with language and analytic processing (Gupta et al. 2025; Olson et al. 2023), whereas the right hemisphere is involved in self‐referential, social, and emotional‐motivational functions (Gainotti 2019; Palomero‐Gallagher and Amunts 2022). Considering this hemispheric specialization, the rightward asymmetry observed in subtype 1 may reflect the atypical lateralization of self‐referential and motivational systems. Our meta‐analytic decoding analysis supports this interpretation, showing that subtype 1 is associated with terms related to motivation and self‐referential processing. This suggests that subtype 1 is characterized by an aberrant organization of networks underlying self‐perception and social motivation (Burrows et al. 2016; Kana et al. 2017; Padmanabhan et al. 2017; Tusche et al. 2023). This pattern aligns with previous findings of altered self‐representation and reduced social motivation in individuals with ASD (Chevallier et al. 2012; Cygan et al. 2014; Uddin 2011). Collectively, subtype 1 may represent a subgroup characterized by heightened self‐focused cognition and diminished social engagement. In contrast, subtype 3 exhibited overall left‐dominant functional asymmetry, particularly in memory and language‐related regions. Given the typical functional specialization of the left hemisphere (Heyrani et al. 2022; Hodgson et al. 2016), this leftward dominance may reflect the atypical preservation or compensatory reorganization of language and memory networks. Moreover, the memory‐ and language‐related terms identified in the meta‐analysis further suggested that subtype 3 is characterized by altered mnemonic and semantic processing profiles. Unlike the other subtypes, subtype 2 exhibited leftward dominance in the default mode and frontoparietal networks and rightward dominance in the somatomotor and visual networks. These mixed asymmetric patterns suggest an imbalance in interactions between cognitive control and sensory information processing. Consistent with this, subtype 2 showed significantly higher ADOS total, communication, and social scores than other subtypes, suggesting a potential neurobiological basis for greater social and communication impairments. The identified ASD subtypes were further characterized based on their functional hub properties. Compared with the other subtypes, subtype 2 exhibited greater hubness in the ventrolateral prefrontal cortex and inferior parietal lobule, accompanied by reduced connectivity in the somatomotor cortex. These findings suggest that subtype 2 displays a distinct dissociation between the sensory and heteromodal regions. Moreover, the observation that subtype 3 showed higher DC values than subtype 1 indicated that right hemispheric asymmetry was associated with hypoconnectivity, whereas left hemispheric asymmetry was linked to hyperconnectivity. The hypoconnectivity observed in subtype 1 may reflect the reduced efficiency of network integration underlying self‐referential processing and social motivation, whereas the hyperconnectivity in subtype 3 may indicate excessive recruitment of networks related to language and executive functions (Williams et al. 2013). Overall, these results suggest that functional asymmetry in ASD reflects not only structural lateralization but also an imbalance in cross‐network interactions between sensory and heteromodal association regions, where the direction of asymmetry may determine divergent patterns of connectivity strength. ASD subtypes, defined based on functional brain asymmetry, can be interpreted as follows: subtype 1 represents the social‐motivational type, subtype 2 represents the cognitive control/impulsive type, and subtype 3 represents the memory‐language type. This interpretation was derived from the (i) patterns of functional brain asymmetry, (ii) whole‐brain hub profiles, and (iii) meta‐analytic cognitive decoding results. Collectively, these findings provide insight into the underlying neurophenotypic diversity of ASD. However, this study had several limitations. First, the dataset used in this study was multi‐site, inherently involving variability in acquisition protocols across sites. Although ComBat harmonization was applied to minimize site‐related effects, residual inter‐site differences may have influenced the results. Second, we were unable to determine how the asymmetric patterns and subtype characteristics varied across developmental stages. Age, which could influence subtype identification, was not explicitly controlled in the current clustering analysis. Developing clustering frameworks that incorporate continuous variables and longitudinal data is necessary to assess developmental trajectories and the stability of subtype clustering. Finally, although the number of clusters was determined using multiple validity indices, the degree of subjectivity in defining the optimal number of clusters could not be entirely excluded. Author Contributions C.K. and B.P. designed the study, analyzed the data, and wrote the manuscript. S.K. aided experiments and reviewed the manuscript. B.P. is the corresponding author of this study and is responsible for the integrity of data analysis. All authors reviewed and approved the final version of the manuscript for publication. Funding This study was supported by the Institute for Information and Communications Technology Planning and Evaluation (IITP) funded by the Korean Government (MSIT) (No. 2022‐0‐00448/RS‐2022‐II220448, Deep Total Recall: Continual Learning for Human‐Like Recall of Artificial Neural Networks), National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS‐2025‐20253035), and Basic Science Research Program through the NRF funded by the Ministry of Education (RS‐2025‐25422106). Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: Site‐specific imaging parameters of study participants. Table S2: Stability evaluation of clustering based on a subsampling approach. Standard deviations of min‐max normalized metrics across different numbers of clusters (k) are reported. The last row shows the average of the three metrics, and the lowest value is shown in bold. Table S3: Demographic information of ASD subtypes. Means and standard deviations are reported where applicable. Figure S1: Two‐cluster solution results. (A) Asymmetry patterns for each ASD subtype. (B) Feature contribution values used in the clustering process. (C) Between‐subtype comparisons, highlighting regions with significant differences. ASD, autism spectrum disorder; L, left; R, right; VN, visual network; SMN, somatomotor network; DAN, dorsal attention network; VAN, ventral attention network; LN, limbic network; FPN, frontoparietal network; DMN, default mode network; FDR, false discovery rate. Figure S2: Feature contributions for the three‐cluster solution. Feature contributions indicating the relative importance of cortical and subcortical regions are visualized on brain surfaces (left). Warmer colors represent higher contribution values. Contribution values are further stratified by functional brain networks (middle) and subcortical regions (right). VN, visual network; SMN, somatomotor network; DAN, dorsal attention network; VAN, ventral attention network; LN, limbic network; FPN, frontoparietal network; DMN, default mode network. Figure S3: Cognitive decoding of differences in brain asymmetry between each ASD subtype and the TD group. T‐statistic maps comparing each ASD subtype with the TD group are shown on brain surfaces (left). The corresponding t‐statistic values are summarized by functional brain networks and subcortical regions (middle). Cognitive decoding results are presented on the right. ASD, autism spectrum disorder; TD, typically developing; VN, visual network; SMN, somatomotor network; DAN, dorsal attention network; VAN, ventral attention network; LN, limbic network; FPN, frontoparietal network; DMN, default mode network. Figure S4: Between‐subtype differences in DC asymmetry after controlling for handedness. F‐statistic maps showing regions with significant between‐subtype differences in asymmetry are shown on brain surfaces (top). T‐statistic maps from pairwise subtype comparisons are shown (bottom). Figure S5: Between‐subtype differences in DC asymmetry after controlling for head motion. F‐statistic maps showing regions with significant between‐subtype differences in asymmetry are shown on brain surfaces (top). T‐statistic maps from pairwise subtype comparisons are shown (bottom). Figure S6: Between‐subtype differences in DC asymmetry after controlling for handedness, head motion, and intelligence quotient. F‐statistic maps showing regions with significant between‐subtype differences in asymmetry are shown on brain surfaces (top). T‐statistic maps from pairwise subtype comparisons are shown (bottom). Figure S7: Between‐subtype differences in DC asymmetry after removing two oldest participants from subtype 3. F‐statistic maps showing regions with significant between‐subtype differences in asymmetry are shown on brain surfaces (top). T‐statistic maps from pairwise subtype comparisons are shown (bottom). Data S1: The NeuroSynth decoding results, including the full terms for each subtype, their corresponding correlation coefficients, and labels (anatomy, cognition, or general).