Comprehensive Polygenic Score Profiling Reveals Autism Spectrum Disorder Subgroups with Different Genetic Predisposition Related to High-Density Lipoprotein Cholesterol, Urea, and Body Mass Index

Abstract

Background: Autism spectrum disorder (ASD) is a complex and heterogeneous neurodevelopmental disorder. This study aims to demonstrate the potential of comprehensive polygenic scores (PGSs) as clinical biomarkers for stratifying individuals with ASD and for advancing the understanding of ASD’s heterogeneous etiology. Methods: We calculated 2602 PGSs—representing all publicly available, license-cleared PGSs in the PGS Catalog—for 75 individuals with ASD by utilizing the database of the Tohoku Medical Megabank Birth and Three-generation cohort study. Results: Unsupervised clustering revealed three ASD subgroups. We identified twenty PGSs with the most significant differences among these subgroups as distinctive PGSs for each subgroup. PGS set enrichment analysis associated these distinctive PGSs with different traits in each subgroup: high-density lipoprotein cholesterol measurements, urea measurement, and body mass index. Furthermore, distinctive PGSs indicated consistent genetic predisposition directions: lower high-density lipoprotein cholesterol levels in subgroup 1, higher urea levels in subgroup 2, and lower body mass index in subgroup 3. Conclusions: Comprehensive PGSs extending beyond psychiatry-related traits represent promising clinical biomarkers for identifying ASD subgroups with different genetic predispositions. Such stratification may enhance understanding of heterogenous genetic backgrounds and targeted drug development.

1. Introduction

Autism spectrum disorder (ASD) is a complex and heterogeneous neurodevelopmental disorder. Its global prevalence is estimated at approximately 1% [1]. This prevalence has increased likely due to changes in diagnostic criteria, improved diagnostic tools, and greater public awareness [2]. ASD is diagnosed using behavior criteria, focusing on core psychopathological symptoms such as deficits in social communication and the presence of restricted, repetitive patterns of behavior [3]. The heterogeneity of ASD encompasses its etiology, clinical manifestations, and treatment outcomes. This heterogeneity is difficult to characterize because it is influenced by multiple factors, including genetic predisposition, environmental exposures, biological sex, and the frequent co-occurrence of psychiatric and non-psychiatric comorbidities [4,5]. Such heterogeneity complicates accurate diagnosis, effective clinical management, and the development of individualized therapeutic approaches.

The heterogeneous etiology of ASD makes it challenging to develop effective pharmacological treatments for its core symptoms. Currently, no medications specifically target the core symptoms of ASD [6]. Most pharmacological interventions focus on managing associated symptoms such as irritability and aggression rather than addressing the core features themselves [7]. The limited efficacy observed in past clinical trials for investigational drugs for ASD may be partly explained by the etiological heterogeneity. Uncontrolled heterogeneity in patient pathophysiology and treatment response can mask true efficacies of potential therapy, which makes it difficult to develop a single pharmacological therapy for all individuals with ASD [8]. Therefore, elucidating the biological basis of this heterogeneity is crucial for optimizing treatments and for facilitating the development of novel drugs. In this context, precision medicine offers promising strategies: stratifying patients based on their underlying pathophysiology to deliver targeted therapies to specific subgroups [9,10].

Omics-based liquid biomarkers have emerged as powerful means of identifying biologically distinct subgroups of patients. Liquid biomarkers, such as those obtained from blood, are minimally invasive and thus potentially feasible diagnostic or monitoring tools in routine clinical practice [11]. Molecular profiling using liquid biomarkers, such as transcriptomic and proteomic analyses of blood samples, has revealed pathophysiological differences that help explain the heterogeneity of ASD [12,13,14]. For example, a transcriptome analysis of blood-derived lymphoblastoid cell lines has linked cholesterol biosynthesis and metabolism to ASD subtypes characterized by language impairment [13]. However, practical challenges limit the clinical application of ribonucleic acid (RNA)- and protein-based biomarkers, including their instability under storage conditions [15,16] and the presence of measurement batch effects [17]. These effects can make results from different analytical batches not directly comparable. Addressing these technical limitations is crucial for the clinical implementation of omics-based liquid biomarkers.

Polygenic scores (PGSs) possess attractive characteristics as clinical biomarkers for patient stratification. PGS is an estimate of an individual’s genetic liability to a trait or disease, calculated as a weighted sum of multiple genotypes of single nucleotide polymorphism (SNP) based on relevant genome-wide association study (GWAS) data [18]. By simultaneously integrating the effects of numerous genetic variants, PGS enables quantification of risk for multifactorial diseases (e.g., obesity) and quantitative traits (e.g., body mass index: BMI), surpassing traditional genomic approaches that focus on limited variants [19]. Thousands of PGS definitions, each comprising specific SNPs and their corresponding weights based on GWAS data are publicly available through databases such as the PGS Catalog [20,21]. Given SNP genotype data, the PGSs can be readily calculated for individuals. The stability of genomic data as a property of PGSs makes them suitable as clinical biomarkers. Genomic data are unaffected by sampling timing because an individual’s deoxyribonucleic acid (DNA) sequence remains constant throughout life [22]. Additionally, DNA’s chemical stability enables accurate genotyping even after prolonged storage under appropriate conditions [23]. Furthermore, PGSs are robust across different measurement batches because they are based on the presence or absence of specific SNPs, which are less susceptible to technical variability than quantitative molecular measurements such as RNA or protein levels [24].

PGSs have been utilized to investigate phenotypic heterogeneity in individuals with ASD. For example, Warrier et al. associated PGSs for ASD with co-occurring developmental disabilities in autistic individuals [25]. They also associated PGSs for schizophrenia, attention-deficit/hyperactivity disorder, and educational attainment with specific behavioral features in individuals with ASD. Similarly, Antaki et al. linked PGSs for ASD and educational attainment to behavioral traits in autistic populations [26]. Klein et al. investigated heterogeneous comorbid conditions with ASD through PGS analysis and revealed that a PGS for ASD may contribute to the occurrence of allergies and sensory processing issues in individuals with ASD [27]. However, these studies have focused on limited PGS sets, particularly those related to psychiatric disorders. In contrast, Albiñana et al. demonstrated that combining 937 PGSs improves ASD risk prediction accuracy compared to single PGS approaches [28]. This finding suggests the potential utility of employing a broad spectrum of PGSs to capture the complex genetic architecture underlying ASD. To our knowledge, no study has systematically investigated the stratification of an ASD population using such a comprehensive set of PGSs. This gap highlights the need for research to elucidate the potential of comprehensive PGSs in characterizing ASD heterogeneity.

This study aims to demonstrate the potential of comprehensive PGSs as clinical biomarkers for stratifying individuals with ASD and for advancing understanding of their heterogeneous etiology in a hypothesis-free manner. We have identified three subgroups of individuals with ASD based on profiles of 2602 PGSs. Enrichment analysis revealed that these subgroups have different genetic predispositions to high-density lipoprotein cholesterol (HDL-C) levels, urea levels, and BMI. These findings may contribute to realizing precision medicine and developing novel therapeutic strategies for ASD.

2. Materials and Methods

2.1. Study Population

This study utilized the database of the Tohoku Medical Megabank Birth and Three-generation (TMM BirThree) cohort study (Tohoku Medical Megabank Organization: ToMMo, Sendai, Japan) [29,30]. The TMM BirThree cohort study recruited 73,529 participants, including 32,086 children, from Miyagi Prefecture, Japan, between 2013 and 2017. The database contains genome data (e.g., SNP genotypes) and self-reported questionnaire data (e.g., birth year, sex, and self-reported medical history). For this study, available demographic and clinical characteristics were limited to birth year, sex, and self-reported medical history specifically ASD diagnosis status. We identified individuals with ASD among the 32,086 children based on the self-reported medical history. Specifically, we selected participants who reported an autism diagnosis in the questionnaire and were genotyped using an Affymetrix Axiom Japonica Array version 2 (Toshiba Corporation, Tokyo, Japan). We excluded participants with only a “suspected” autism diagnosis from the analysis.

This study has been approved by the ethical research practice committee of Daiichi Sankyo Co., Ltd. (Tokyo, Japan, Registration code 001119, approved on 1 September 2025) and the ethical committee of ToMMo in Tohoku University (Sendai, Japan, Registration code 2025-4-051-1, approved on 28 July 2025).

2.2. Polygenic Scores (PGS)

This study utilized SNP genotype data from the TMM BirThree Cohort Study database [30]. Genotyping was performed using the Affymetrix Axiom Japonica Array version 2 on DNA extracted from blood samples. The array was designed to assay 659,328 SNP markers. Following quality control, genotype imputation was conducted using the 1KJPN whole-genome reference panel, which comprises whole-genome sequencing data from 1070 Japanese individuals [31]. Imputed SNP genotype data are available in the TMM BirThree Cohort Study database.

We evaluated all publicly available, license-cleared PGS definitions in the PGS Catalog [20,21]. We distinguish between PGS definitions, which specify the SNP weights and effect sizes derived from GWAS summary statistics, and the resulting PGS values, which are calculated as a weighted sums of genotypes across multiple SNPs using the PGS definitions. We downloaded a total of 5022 PGS definitions from the PGS Catalog in December 2024. Of these, we included 3315 PGS definitions with no license restrictions [32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,190,191,192,193,194,195,196,197]. We first included PGS definitions registered with CC-BY 4.0 in the PGS Catalog. Then, for PGS definitions without explicit license information, we contacted the authors and included only those for which explicit permission was obtained. The permissions obtained for this study do not necessarily extend to other research contexts.

We calculated PGS values using PLINK v2.00a3LM [198]. Quality control included SNPs with a minor allele frequency >5% and genotyping rate >95%, referenced against an allele frequency panel of 3552 Japanese individuals (54KJPN-SNV/INDEL) [199]. SNPs absent in our genotype data or failing quality control thresholds were excluded from PGS calculation. We also excluded PGS definitions for which >30% of all samples had identical values. This threshold was selected to ensure adequate statistical power for subsequent subgroup-based analyses: PGS definitions exceeding this threshold in three equal-sized subgroups (~33% each) would result in nearly entire subgroups sharing identical PGS values, compromising the validity of statistical comparisons. PGS values were standardized to a mean of zero and a standard deviation of one to ensure comparability of scales across different PGS definitions.

2.3. Stratification of Individuals with Autism Spectrum Disorder (ASD) Using PGS Profiles

We conducted heatmap pattern-based clustering to stratify individuals with ASD into subgroups. This methodology, previously applied to stratify schizophrenia patients based on microRNA expression profiles [200,201], uses visual inspection of PGS patterns in heatmaps to identify clusters while excluding low-information regions. We performed hierarchical clustering analysis using the unweighted pair group method with arithmetic mean and Euclidean distance to generate a dendrogram. The number of subgroups and their assignments were determined on the basis of the dendrogram and heatmap patterns. We also performed principal component analysis (PCA) on all the PGSs to visualize similarities among individuals within each subgroup. Specifically, we calculated the first and second principal components and plotted the individuals on this two-dimensional space to assess subgroup patterns. To evaluate robustness of these subgroups, we additionally performed k-means clustering as a complementary validation. To assess dimensional stability, we visualized PCA scores using heatmaps to test whether the clustering results in original PGS space persisted after dimensionality reductions.

Distinctive PGSs in each subgroup were defined as those showing significant differences in that subgroup compared with the remaining subgroups. We compared each subgroup against the remaining subgroups using a two-tailed unpaired t-test. The Benjamini–Hochberg method was applied for global false discovery rate (FDR) correction, with the total number of comparisons calculated as the number of PGSs multiplied by the number of subgroups. Significant PGS enrichment was defined as Benjamini–Hochberg corrected p-value (q) < 0.05. For each comparison, we calculated Cohen’s d as a measure of effect size and computed 95% confidence intervals for mean differences. To minimize the potential overlap of significant PGSs across multiple subgroups—where the same PGS could demonstrate significant differences in subgroups—we selected the twenty PGSs with the lowest Benjamini–Hochberg corrected p-values as distinctive PGSs in each subgroup. All PGS analysis were performed on Python 3.12.3 (Python Software Foundation, Wilmington, DE, USA).

2.4. Enrichment Analysis for Distinctive PGSs in Each Subgroup

We performed PGS set enrichment analysis using an over-representation analysis. The PGS definitions were classified into 569 traits in the PGS Catalog, according to the Experimental Factor Ontology [202]. For example, the trait ‘BMI’ includes 75 PGS definitions (e.g., PGS000841 and PGS000910). To assess the enrichment of distinctive PGSs for specific traits in each subgroup, we performed hypergeometric tests using the following parameters: (1) total number of the PGSs in the whole PGS profiles, (2) number of trait-associated PGSs in the whole PGS profiles, (3) number of the distinctive PGSs in each subgroup, and (4) number of trait-associated PGSs among the distinctive PGSs in each subgroup. The Benjamini–Hochberg method was applied for global FDR correction across all enrichment tests, with the total number of tests calculated as the number of traits multiplied by the number of subgroups. Significant enrichment was defined as q < 0.05.

To assess the robustness of enrichment findings, we conducted sensitivity analyses accounting for PGS correlation within traits. We removed highly correlated PGS pairs (|r| > 0.8, where r is Pearson’s correlation coefficient). From this pruned PGS set, we re-identified distinctive PGSs for each subgroup by selecting the twenty PGSs with the lowest Benjamini–Hochberg corrected p-values. All enrichment analyses were repeated using this refined set of distinctive PGSs. To visualize PGS redundancy within trait categories, we generated correlation heatmaps displaying pairwise Pearson’s correlation coefficients among all PGSs within each trait showing enrichment in the original analysis.

3. Results

3.1. Polygenic Score Profiles Revealed Three Subgroups of Individuals with ASD

We identified 75 individuals with ASD from the TMM BirThree cohort in the study (

Table 1. Characteristics of individuals with autism spectrum disorder (ASD).

Characteristics	n = 75
Birth year, mean ± S.D.	2013.9 ± 2.0
Male, n (%)	55 (73.3)
Ethnicity: Japanese, n (%)	75 (100)

). After quality control for PGS calculation, PGS values were obtained for 2602 of 3315 available PGS definitions (Supplementary Figure S1 and Supplementary Table S1) for individuals with ASD. In detail, PGS values were initially calculated for 2876 of 3315 PGS definitions. Then, we excluded 274 PGS definitions that had identical values in more than 30% of samples, indicating insufficient variability. Of the 2602 PGS definitions included in the final analysis, 2550 (97.9%) included European ancestry in their training GWAS, whereas only 87 (3.3%) included East Asian ancestry.

We first applied a univariate approach, analyzing each PGS individually, to explore potential subgroups among the individuals with ASD. We plotted distributions of PGSs for traits previously associated with behavioral features in individuals with ASD [25,26], including schizophrenia, educational attainment, and attention-deficit/hyperactivity disorder (Figure 1). However, no distinct subgroups were identified based on the distribution of an individual PGS.

We then employed a multivariate approach, simultaneously evaluating all PGSs, to explore subgroups among individuals with ASD because comprehensive PGSs have the potential to capture the complex ASD characteristics more effectively than a single PGS [28]. We visualized PGS profiles as a heatmap, where each PGS profile consisted of PGS values for all 2602 PGS definitions (Figure 2A). Heatmap pattern-based clustering on the whole PGS profiles revealed three subgroups (1, 2, and 3): subgroup 1 (n = 43 out of 75) exhibited high PGS values in the fifth decile from the left and low PGS values in the eighth decile from the left, subgroup 2 (n = 17 out of 75) showed high PGS values in the fifth and eighth deciles from the left, and subgroup 3 (n = 15 out of 75) had low PGS values in the fifth and eighth deciles from the left. K-means clustering derived similar subgroups (designated a, b, and c). PCA using two principal components (12.1% of total variance) also separated the individuals into these three subgroups (Figure 2B). No apparent bias in birth year and sex was observed among the subgroups (Figure 2C). To assess dimensional stability, we examined these subgroups in PCA-reduced space using 20 and 50 principal components (51.6% and 84.4% of total variance, respectively); however, the characteristic patterns observed in the original PGS space (Figure 2A) were less distinct in PCA-reduced dimensions (Supplementary Figure S2).

3.2. Distinctive PGSs in Each Subgroup Enriched High-Density Lipoprotein Cholesterol (HDL-C) Measurements, Urea Measurement, and Body Mass Index (BMI)

To identify distinctive PGSs in each subgroup, we selected twenty PGSs with the most significant differences among subgroups (Figure 3A and Supplementary Table S2). Subgroup 1 had higher values for nine distinctive PGSs (e.g., PGS003816) and lower values for eleven distinctive PGSs (e.g., PGS003777) compared with the other subgroups. Subgroup 2 had higher values for seventeen distinctive PGSs (e.g., PGS002198) and lower values for three distinctive PGSs (e.g., PGS003273) compared with the other subgroups. Subgroup 3 had lower values for all twenty distinctive PGSs (e.g., PGS004985) compared with the other subgroups.

To investigate the traits represented by the distinctive PGSs, we performed enrichment analysis on the distinctive PGSs in each subgroup (

Table 2. Enriched traits by PGS set enrichment analysis.

PGS Set	Enriched Trait	Nmapped/Nall	q-Value
Twenty distinctive PGSs in subgroup 1	HDL-C measurement	6/35	8.62 × 10−5
Twenty distinctive PGSs in subgroup 2	Urea measurement	3/4	6.93 × 10−4
Twenty distinctive PGSs in subgroup 3	BMI	10/69	1.78 × 10−8

N_mapped, number of mapped PGSs in each trait; N_all, number of all the PGSs in each trait; q-value, Benjamini–Hochberg corrected p-value.

). The distinctive PGSs enriched different traits across subgroups. Enrichment analysis revealed that distinctive PGSs in subgroup 1 were associated with ‘HDL-C measurement’, those in subgroup 2 with ‘urea measurement’, and those in subgroup 3 with ‘BMI’.

To assess the stability of these findings, we conducted sensitivity analyses accounting for PGS correlation. The pruning procedure removed 535 highly correlated PGSs (20.6%), resulting in 2067 independent PGSs. By trait, pruning removed: 37 PGSs from ‘BMI’ (69 to 32; 53.6% reduction), 1 PGS from ‘urea measurement’ (4 to 3; 25.0% reduction), and 10 PGSs from ‘HDL-C measurement’ (35 to 25; 28.6% reduction). We re-identified distinctive PGSs for each subgroup from the pruned set and repeated enrichment analyses. The results showed different robustness (Supplementary Table S3): ‘HDL-C measurement’ enrichment in subgroup 1 remained significant (q = 7.68 × 10⁻⁵), whereas ‘urea measurement’ enrichment in subgroup 2 (q = 0.089) and ‘BMI’ enrichment in subgroup 3 (q = 0.540) lost significance (q > 0.05). Correlation heatmaps for the three trait categories (Supplementary Figure S3) visually confirmed different redundancy patterns: ‘BMI’ exhibited the most extensive within-trait PGS correlation (53.6% reduction), followed by ‘HDL-C measurement’ (28.6% reduction), whereas ‘urea measurement’ exhibited the least redundancy (25.0% reduction).

To assess genetic predisposition direction for enriched traits, we compared the PGSs mapped to these traits across subgroups (Figure 3B). Subgroup 1 had lower values for all the distinctive PGSs mapped to ‘HDL-C measurement’. Additionally, subgroup 1 consistently exhibited higher values for the three distinctive PGSs related to ‘triglyceride’ (PGS003816, PGS003811, and PGS003806), although ‘triglyceride’ was not significantly enriched in the enrichment analysis. Subgroup 2 had higher values for all the distinctive PGSs mapped to ‘urea measurement’. Similarly, subgroup 2 consistently exhibited lower values for the three distinctive PGSs related to ‘glomerular filtration rate/creatine measurement’ (PGS003273, PGS003258, and PGS003314), although ‘glomerular filtration rate/creatine measurement’ was not significantly enriched. Subgroup 3 had lower values for all the distinctive PGSs mapped to ‘BMI’. Box-and-swarm plots confirmed that each subgroup showed significant differences in PGS values for representative single PGSs corresponding to the enriched trait (Figure 3C).

4. Discussion

4.1. Polygenic Scores Are a Potential Biomarker to Stratify Individuals with ASD into Subgroups of Different Genetic Predispositions

This study demonstrated the potential of PGSs to stratify individuals with ASD into subgroups based on different genetic predispositions and to reveal the heterogeneity of genetic risk factors in ASD. In particular, utilizing comprehensive PGSs rather than a single PGS improved the ability to investigate heterogeneity of ASD. Single PGSs for schizophrenia, educational attainment, and attention-deficit/hyperactivity disorder failed to stratify individuals (Figure 1), despite previous reports that the PGSs for these traits are associated with behavioral features in individuals with ASD [25,26]. This result is consistent with the expected normal distribution of PGSs in a homogeneous population due to the central limit theorem [203]. In contrast, the PGS profiles consisting of 2602 PGS definitions identified three subgroups of the individuals with ASD (Figure 2). Such high-dimensional PGS profiles are expected to contain comprehensive genetic information that extends beyond risks of psychiatric disorders. To evaluate the robustness of the identified subgroups, we examined their structure in PCA-reduced space. The characteristic clustering patterns were less distinct in PCA-reduced dimensions (Supplementary Figure S2). This observation reflects a fundamental difference between our approach and global dimensionality reduction: heatmap pattern-based clustering identifies subgroups based on specific patterns of correlated PGSs that define distinctive profiles, whereas PCA aggregates correlated variables into composite dimensions, prioritizing global variance rather than profile distinctiveness. This distinction highlights the advantage of our clustering method, which is sensitive to PGS profiles that reflect complex genetic architecture in ASD. Additionally, k-means clustering on the original PGS data yielded subgroups that were concordant with those identified by heatmap pattern-based clustering, supporting the robustness and biological plausibility of the three-subgroup structure. Albiñana et al. further demonstrated the utility of this comprehensive PGS approach by showing improved accuracy in ASD risk prediction when leveraging multiple PGSs compared with using single PGSs [28]. This PGS profile-based approach is not limited to ASD. For example, Sandling et al. identified four subgroups of systemic lupus erythematosus patients based on 35 pathway-specific PGSs, where those subgroups showed significantly different PGS levels for the antigen processing and presentation pathway and Th17 cell differentiation [204]. Thus, using comprehensive PGSs as profiles offers a powerful means to explore the genetic backgrounds underlying heterogeneity within specific diseases or disorders.

PGS-based stratification of individuals with ASD may advance the precision medicinal strategies in ASD. Distinctive PGSs in each subgroup clearly distinguished one subgroup from the other subgroups (Figure 3), suggesting the subgroups reflect different genetic predispositions. Those genetic predispositions may include traits related to drug responses; however, this study did not incorporate reference data on drug responses and therefore could not directly evaluate this aspect. Previous studies in psychiatric disorders have demonstrated the utility of PGS-based patient stratification for predicting drug response [205]. For example, schizophrenia patients who had higher PGS for schizophrenia had poorer response to antipsychotic treatment [206,207]. Also, bipolar disorder patients who had higher PGSs for schizophrenia and major depressive disorder showed poorer response to lithium treatment [208,209]. For ASD, there are currently no approved pharmacological treatments targeting core symptoms, highlighting the urgent need for novel drug development [6]. If, as in other psychiatric conditions, PGS could be used to distinguish ASD subgroups with varying drug responsiveness, this would support the design of clinical trials that stratify participants based on their PGS profiles. Such clinical studies could evaluate drug efficacy within genetically defined subgroups, potentially facilitating the development of therapeutics tailored to specific ASD populations. Furthermore, investigating the genetic backgrounds underlying each ASD subgroup could enhance the precision medicinal strategies, leading to more effective and individualized interventions for ASD.

4.2. ASD Subgroups with Different Genetic Predisposition Toward Obesity May Warrant Investigation of Obesity-Targeted Interventions

Subgroups 1 and 2 demonstrate a genetic predisposition toward higher BMI. Enrichment analysis revealed that distinctive PGSs in subgroup 3 were associated with ‘BMI’ (Table 2), and these PGSs showed significantly lower values in subgroup 3 compared with subgroups 1 and 2 (Figure 3), indicating a genetic predisposition for lower BMI in subgroup 3. Conversely, subgroups 1 and 2 are characterized by enrichment of higher BMI-associated PGSs. Since BMI is a widely accepted index for obesity [210], these genetic enrichment patterns may suggest different obesity susceptibility across subgroups. However, measured BMI data are unavailable in this cohort, preventing direct validation of whether these genetic predispositions manifest as meaningful differences in actual BMI or obesity status.

The genetic enrichment patterns observed in this study align with epidemiological evidence of elevated obesity risk in ASD populations. Individuals with ASD have been associated with an increased risk for obesity. Multiple meta-analyses have demonstrated that the prevalence of obesity is significantly higher in individuals with ASD than in controls [211,212]. Consistent with these meta-analyses, 80% of participants (n = 60 out of 75) were classified into subgroups 1 and 2, suggesting that a majority of this cohort may possess a genetic predisposition to obesity. Previous research has further indicated that among children with ASD, those with more severe ASD symptoms are more likely to be classified as obese compared with those with milder symptoms [213]. Although ASD symptom severity scores were unavailable in the present study, the genetic predisposition toward higher BMI in subgroups 1 and 2 may potentially be associated with more severe ASD symptomatology, though this remains speculative without measured BMI.

If genetic predispositions are confirmed by measured BMI in future studies, subgroups 1 and 2 may warrant consideration for obesity-targeted interventions. A precision psychiatry approach for obesity-targeted interventions involves identifying individuals with ASD who possess a high genetic predisposition to obesity at early stage, followed by implementation of intensive and systematic preventive interventions against obesity [214]. For example, tailored dietary guidance, exercise programs, and lifestyle modifications can help prevent obesity in ASD [215]. From a precision medicine perspective, individuals with ASD carrying high genetic risk for obesity may respond better to anti-obesity pharmacotherapy. Makin et al. have suggested that anti-obesity drugs can be effective in treating autistic patients with obesity [216], not only by promoting weight reduction but also by addressing food-related behavioral problems and aggressive behaviors, as demonstrated in several case reports [217,218]. Therefore, subgroups 1 and 2 identified in this study may warrant particular consideration for such preventive interventions and for evaluating the efficacy of anti-obesity medications. Nonetheless, the translation of these findings into clinical practice will necessitate validation in independent cohorts with objectively measured BMI and comprehensive phenotypic characterization.

4.3. ASD Subgroup with Different Genetic Predisposition Toward Dyslipidemia May Warrant Investigation of Lipid Metabolism-Targeted Interventions

Subgroup 1 demonstrates a genetic predisposition toward lower HDL-C as well as higher BMI. Enrichment analysis revealed that distinctive PGSs in this subgroup were associated with ‘HDL-C measurements’ (Table 2), and these PGSs showed lower values in subgroup 1 (Figure 3), indicating an increased genetic risk for lower HDL-C levels. Additionally, subgroup 1 consistently exhibited high values for three distinctive PGSs related to ‘triglyceride measurement’. Since low HDL-C and high triglyceride levels are key indicators of dyslipidemia [219], these genetic enrichment patterns may suggest elevated dyslipidemia susceptibility in subgroup 1. However, measured lipid profiles (e.g., HDL-C and triglyceride levels) are unavailable in this cohort, preventing direct validation of whether these genetic predispositions manifest as meaningful differences in actual lipid metabolism or dyslipidemia status.

The genetic enrichment patterns observed in subgroup 1 align with epidemiological evidence of elevated dyslipidemia risk in ASD populations. Similarly to obesity, individuals with ASD show increased dyslipidemia risk. A meta-analysis demonstrated that dyslipidemia prevalence is significantly higher in individuals with ASD than in those without ASD, with ASD individuals exhibiting significantly higher triglyceride levels and lower HDL-C levels [220]. The elevated triglyceride concentrations in ASD were also associated with a greater food interest, including emotional overeating. These findings highlight the importance of considering lipid abnormalities in ASD clinical management [221].

If genetic predispositions are confirmed by measured lipid profiles in future studies, subgroup 1 may warrant consideration for lipid metabolism-targeted interventions. A precision psychiatry approach involves identifying ASD individuals with elevated genetic risk for dyslipidemia and providing lipid metabolism-targeted interventions. Lipid-based interventions are gaining attention as potential therapeutic strategies for ASD. These interventions include dietary modifications such as omega-3 fatty acid supplementation, statins for cholesterol regulation, and other agents targeting lipid synthesis or metabolism pathways (e.g., fibrates or PCSK9 inhibitors) [222]. However, due to the broad heterogeneity of ASD manifestations and underlying biological pathways, accurately evaluating the effectiveness of such lipid-based therapies remains challenging [223]. Further clinical studies are necessary to determine which lipid interventions benefit which subgroups. Identifying ASD subgroups with increased genetic risk for dyslipidemia could help stratify individuals for tailored nutritional and pharmacological treatments. Specifically, PGS-based subgroup 1, characterized by elevated lipid-related genetic risk, could derive significant benefit from preventive approaches and lipid-targeted interventions, potentially improving both metabolic health and neuropsychiatric outcomes. However, such clinical applications would require validation in cohorts with measured lipid profiles and detailed phenotypic characterization.

4.4. ASD Subgroup with Different Genetic Predisposition Toward Impaired Renal Function Provide Insights into Potential Therapeutic Targets Common to Both Kidney Diseases and ASD

Subgroup 2 demonstrates a genetic predisposition toward increased urea levels as well as higher BMI. Enrichment analysis revealed that distinctive PGSs in this subgroup were associated with ‘urea measurement’ (Table 2). Consistently, all distinctive PGSs mapped to ‘urea measurement’, where all PGS definitions were derived from serum urea, showed significantly higher values in subgroup 2 (Figure 3), indicating increased risk for higher serum urea levels in this subgroup. Additionally, subgroup 2 consistently exhibited low values for the three distinctive PGSs related to ‘glomerular filtration rate’. Since high serum urea and low glomerular filtration rate are key indicators of impaired renal function [224], these genetic enrichment patterns may suggest elevated susceptibility to impaired renal function in subgroup 2. However, measured renal function markers (e.g., serum urea and glomerular filtration rate) are not available in this cohort, preventing direct validation of whether these genetic predispositions manifest as meaningful differences in actual renal function.

The genetic enrichment patterns observed in subgroup 2 may provide insights into potential therapeutic targets shared by kidney disease and ASD. Kidney disease and ASD frequently co-occur in various multisystem genetic disorders [225]. For example, tuberous sclerosis complex, which is characterized by hyperactivation of the mammalian target of rapamycin (mTOR) pathway, is associated with ASD in approximately 10–60% of patients and with kidney diseases such as renal angiomyolipomata in 50–80% of patients [226,227,228]. The mTOR inhibitor everolimus demonstrated efficacy in Phase 3 clinical trials for patients with tuberous sclerosis complex, showing benefits for renal angiomyolipomas (e.g., reduction in lesion volume and preservation of kidney function) as well as a trend toward improvement in ASD symptoms [229,230]. If the genetic predisposition toward impaired renal function in subgroup 2 is confirmed by measured renal function data in future studies, this subgroup may warrant particular consideration for evaluating the efficacy of mTOR inhibitors or other drugs targeting shared pathways in kidney disease and ASD.

4.5. This Study Has Strengths and Limitations

One major strength of this study is that we demonstrated the potential of PGSs, derived from genetic variant information, as robust clinical biomarkers for stratifying individuals with ASD. Unlike other liquid biomarkers such as blood-based RNAs or proteins, which may fluctuate depending on the patient’s condition (e.g., pre- or post-treatment) and reflect state rather than trait characteristics, PGSs provide stable, trait-level information. Additionally, a practical challenge of using RNA- and protein-based biomarkers is their instability during storage and processing, as both are prone to degradation, compromising their reliability [15,16]. In contrast, once an individual’s SNP profile is measured and PGSs are calculated, the result remains valid throughout the individual’s lifetime, as genetic variants do not change over time. This stability enables consistent and lifelong application of PGSs for the same individual. Furthermore, comprehensive SNP profiling provides information not only for stratifying individuals with ASD but also for assessing genetic predispositions to other conditions, such as schizophrenia, using the same SNP data. Thus, comprehensive SNP analysis can yield valuable insights into a wide range of genetic risks, highlighting the versatility and clinical utility of SNP-based PGSs as clinical biomarkers.

Another strength is the use of comprehensive PGSs, which improved the accuracy of patient stratification and enabled trait-based interpretation of each subgroup. By analyzing comprehensive PGSs—including those beyond psychiatric disorders—we performed hypothesis-free profiling. Our results suggest that, within the ASD cohort, there are subgroups with distinct genetic predispositions to traits such as HDL-C levels, urea levels, and BMI. This methodology is applicable not only to ASD and psychiatric disorders but also to other conditions; the same approach—evaluating comprehensive PGSs in a given population to stratify individuals and interpret genetic predispositions—could be applied to other disease cohorts.

A major limitation of this study lies in the dataset: the small sample size (n = 75), limited participant diversity, lack of detailed clinical information such as symptom severity scores and comorbidities, and the absence of healthy control subjects for comparison. The small sample size may be insufficient to generalize our findings. Additionally, we were unable to assess the influence of demographic factors such as ethnicity and comorbidities on subgroup identification. These confounding factors may introduce bias. Furthermore, this study relied on imputed genotype data derived from SNP array genotyping rather than whole-genome sequencing. Genotype imputation introduces potential inaccuracies because imputed variants represent statistical predictions. These errors could propagate through PGS calculations, affecting the precision of PGS values and the reliability of subgroup classifications.

This study lacks measurements of BMI, metabolic markers (e.g., HDL-C), or renal function indicators (e.g., serum urea and glomerular filtration rate). While PGS enrichment analyses revealed associations between subgroups and these traits based on genetic predispositions, we could not validate whether these genetic predispositions manifest as meaningful differences in measured phenotypes within our cohort. Therefore, all clinical interpretations regarding metabolic or renal predispositions (e.g., obesity risk, dyslipidemia susceptibility, and impaired renal function) remain speculative and require validation in cohorts with clinical measurements. Although PGSs are promising research tools, their diagnostic and prognostic utility for ASD in clinical practice has not yet been established. Therefore, the PGS-based subgroups identified in this study should not be interpreted as clinically actionable classifications. Future studies incorporating measurements of BMI, lipid profiles, and renal function would be essential to confirm the biological and clinical relevance of these classifications.

ASD identification in this study relied on self-reported questionnaire responses, introducing potential selection bias and diagnostic uncertainty. It is unclear whether diagnoses were clinician-confirmed, which diagnostic criteria were applied, or when diagnoses were made. This uncertainty may result in case misidentification, potentially affecting subgroup identification. This study did not include healthy control subjects due to database limitation. For instance, some ‘non-ASD’ participants who did not self-report ASD up to three years of age could be diagnosed with ASD after four years of age. Increasing the sample size, including more diverse populations, and incorporating comprehensive clinical information such as severity scores would enhance the robustness and generalizability of subgroup identifications and their interpretation.

Another limitation concerns PGS interpretation. Most PGSs were defined based on Western (predominantly European) populations, whereas our population is Japanese. Our PGSs, derived predominantly from European ancestry training data (2550 of 2602 PGSs, 97.9%), show reduced transferability to the Japanese cohort. Specifically, European-derived PGSs show 30–50% reduction in variance explained (R²) when applied to East Asian populations compared with European populations [231,232]. Consequently, we cannot verify whether PGS values calculated in this study accurately predict each trait or disease risk in our Japanese cohort. However, the use of PGS Catalogue standards ensures reproducibility, and as multi-ancestry GWAS data expand, future re-evaluation with improved multi-ancestry models will enable more accurate predictions in East Asian populations.

Additionally, trait representation and the number of PGSs in enrichment analysis show biases. For example, for ‘BMI,’ multiple PGS definitions were derived from the same GWAS dataset using different significance thresholds (e.g., p < 0.0001, 0.001, 0.01, 1 × 10⁻⁶, 5 × 10⁻⁸). Since these PGS are based on the same underlying GWAS data, their levels may be correlated and could collectively influence the enrichment analysis results. To address this limitation, we conducted sensitivity analyses by removing highly correlated PGS pairs (|r| > 0.8) and repeating enrichment analyses. Notably, enrichments for ‘urea measurement’ (subgroup 2) and ‘BMI’ (subgroup 3) lost statistical significance after removing correlated PGSs, confirming these enrichments were substantially driven by redundant PGS definitions. In contrast, enrichment for ‘HDL-C measurement’ (subgroup 1) remained robust, suggesting a stable genetic signal. These findings underscore the importance of accounting for PGS redundancy when interpreting enrichment results. Conversely, for some traits such as ‘addictive behavior,’ only a single PGS was available, limiting robust enrichment analysis.

An additional limitation concerns cluster determination methodology. Our analysis employed visual inspection rather than automated clustering algorithms to identify subgroups. While this approach incorporates data-specific characteristics that automated methods may miss, it precludes direct calculation of quantitative cluster validity indices (e.g., silhouette width) and cluster stability analyses (e.g., bootstrapped Jaccard similarity). These indices are typically computed following automatic cluster assignment, which our visual inspection methodology does not facilitate. Although quantitative cluster validity measures are widely recognized as objective evaluation standards, their application is constrained by manual cluster delineation. Consequently, subgroup classification reproducibility relies on consistency and interpretability of distinctive PGS profiles rather than on formal quantitative stability metrics. External validation in an independent cohort would be essential to establish the reproducibility and generalizability of the three-cluster classification.

5. Conclusions

Comprehensive PGSs extending beyond psychiatry-related traits represent promising clinical biomarkers for identifying ASD subgroups with different genetic predispositions. Our analysis demonstrated that PGS profiling can stratify individuals with ASD according to their genetic risk for traits such as HDL-C levels, urea levels, and BMI. Given ASD heterogeneity, such stratification may enhance understandings of diverse genetic backgrounds and facilitate precision psychiatry approaches, enabling more personalized treatment strategies and targeted drug development.