Dissecting the heterogeneity of autism spectrum disorder with sensory behavior, brain, and epigenetic factors

Abstract

Autism spectrum disorder (ASD), a disorder with heterogeneous etiology, is characterized by abnormal behavioral responses to sensory inputs. However, there is still limited understanding of how brain and epigenetic factors, along with behavioral abnormality, contribute to ASD. After completing Adolescent-Adult Sensory Profile, a self-report questinnaire, 34 individuals with ASD and 72 controls underwent neuroimaging scans to measure brain structural (cortical and subcortical volume) and functional (thalamo-cortical resting-state functional connectivity) characteristics. For epigenetic measures, we computed DNA methylation values of the oxytocin receptor and arginine vasopressin receptor (AVPR) genes from the participants’ saliva. When sensory-related behavior was the default baseline, a machine learning algorithm demonstrated that the neuroimaging-epigenetic model outperformed the neuroimaging model or the epigenetic model. Thalamo-cortical hyperconnectivity and AVPR 1A epigenetic modification were found to be significant contributing factors in these models. By integrating neuroimaging and epigenetic biomarkers with behaviors, a more precise diagnosis of ASD can be achieved.

Introduction

Autism spectrum disorder (ASD) is known for its heterogeneity. By integrating neuroscience and epigenetics, remarkable breakthroughs have been made in identifying reliable biomarkers [1, 2]. Considering that atypical behavioral responses to sensory inputs could cascade into maladaptive social behaviors, the recent edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) added altered sensory behavioral response as a core symptomatic feature of ASD. The thalamus (THA), a relay station of sensory and motor information, has caught attention from ASD researchers. They have reported aberrant resting-state functional connectivity (rs-FC) between the THA and multiple cortical areas [3–6]: thalamic hyperconnectivity was observed with the precentral/postcentral gyri (PreCG/PoCG), superior parietal lobule, superior/middle temporal gyri (STG/MTG), and prefrontal cortex (PFC). Researchers have suggested that sensory processing anomalies in ASD may be linked to changes in the limbic structures [7]. Furthermore, it was proposed that structural changes related to sensory characteristics may be associated with the arginine vasopressin receptor gene 1A (AVPR1A) [8].

Recent studies have increasingly provided evidence of associations between epigenetic modifications and the autistic brain, although these effects are region-specific. DNA methylation, the addition of a methyl chemical group to the C-5 position of the DNA cytosine ring through DNA methyltransferases, is one of epigenetic mechanisms [9]. Compared to the controls, the autistic human brain showed different DNA methylation levels in the PFC, temporal, and anterior cingulate cortices [10–12] but not in the cerebellar and occipital areas [2].

Human research on ASD has identified genetic variants of the oxytocin receptor gene (OXTR) as a significant risk factor [13]. Although the results do not converge yet, OXTR epigenetic modulation has been associated not only with autistic social traits, but with the autistic brain [1, 13]. Nonetheless, OXTR hypermethylation was observed in the temporal cortex of the autistic human brain [14], and associated with rs-FC involved in the theory of mind, the reward circuitry [15], and the salience and attention networks [16].

Both oxytocin (OXT) and arginine vasopressin (AVP) are synthesized in the hypothalamus, released from the posterior pituitary gland, and known for their mediating role in social behaviors and sensory processing in mammals [17, 18]. The intranasal administration of both OXT [19] and AVP [20] improved the ASD behavioral symptoms. However, OXT has long been the primary focus and the effects of AVP and its receptor genes on ASD were largely understudied [18]. Demonstrating associations between AVPR1A hypomethylation and joint attention, a recent study on primates suggests that AVPR1A epigenetic modulation could serve as a promising biomarker of ASD [21].

ASD is principally defined by behavioral symptoms. Researchers have explored brain and epigenetic factors as potential biomarkers. However, the contribution of brain and epigenetic biomarkers to the classification of ASD, particularly in relation to atypical behaviors in response to sensory inputs, remains largely unknown. With the growing interest in computational approaches in psychiatry, machine learning has become popular for classification purposes [22]. This new tool is also useful for identifying meaningful features derived from various factors to classify ASD. Using the eXtreme Gradient Boosting (XGBoost) algorithm [23], this study aims to investigate, if characteristics of sensory behaviors are default baseline, the roles of brain (thalamo-cortical rs-FCs, volume) and epigenetic factors (DNA methylation values of OXTR, AVPR1A, AVPR1B) in predicting ASD. For default behavior measures, we used the Adolescent-Adult Sensory Profile (AASP), a self-report questionnaire [24]. To determine better ASD biomarker, we separately built three models, a neuroimaging-epigenetic model (behavior, brain, and epigenetic factors), a neuroimaging model (behavior and brain factors) and an epigenetic model (behavior and epigenetic factors). We hypothesized differences in predictive accuracy across the models and the superiority of the neuroimaging-epigenetic model over the other models. Since no study to date has investigated whether brain factor contributes more to ASD compared to epigenetic factor or vice versa, our null hypothesis is no difference in accuracy between the neuroimaging and epigenetic model.

Methods

Demographic information

This study included a total of N = 106 participants, 34 individuals with ASD (Male = 22; mean [SD] age = 26.0 [4.24] years; mean Full-Scale Intelligence Quotient (FSIQ) [SD] = 108.85 [12.59]) and 72 individuals with TD (Male = 33; mean [SD] age = 32.0 [12.73] years; mean FSIQ [SD] = 109.26 [11.72]). FSIQ was assessed using the Wechsler Adult Intelligence Scale IV. All participants were screened by experienced clinicians. The ASD group met the diagnostic criteria for the DSM-5, Clinical Version, and the Diagnostic Interview for Social and Communication Disorders. The TD group had neither intellectual disabilities nor any psychiatric disorders. Exclusion criteria for both groups were: (1) major physical illness or history of brain injury, (2) FSIQ < 70, and (3) any substance abuse or alcohol consumption problems. Prior to the experiment, all participants provided written informed consent for inclusion. This study was approved by the ethics committee at the University of Fukui and conformed to the Declaration of Helsinki (Assurance No. 20200081). The required sample size was calculated using the pwr package in R for a multiple regression analysis using 30 predictors, the number of input data for model development. Based on a medium-to-large effect size of (f² = 0.3), a significance level of α = 0.05, and a desired power of 0.8, the analysis determined N = 105 as a minimum sample size.

Data for sensory-related behavioral factor

We assessed levels of behavioral abnormalities in response to sensory stimuli using the self-report questionnaire, AASP. It characterizes four basic patterns (Low registration, Sensitivity, Sensation seeking, Avoidance), determined by the interaction between two levels of neurological thresholds (high or low) and two levels of self-regulation (active or passive), of behaviors involved in six sensory processing domains (Activity level, Auditory, Movement, Taste/smell, Touch, Visual) (Table 1).

Table 1.

Default features: demographic and sensory behavioral characteristics in TD and ASD groups.

	TD (median; mean; interquartile range)	ASD (median; mean; interquartile range)	U	P-value
Demographic base feature
Sample size (N)	72	34	-
Ages (in years)	24; 27.847; 10	27; 27.824; 9	1139.5	0.567
Sex (% Male)	48.61%	64.71%	-	0.146
FSIQ	110;109.263; 15.25	106.5; 108.853; 17.26	1316	0.54
Behavioral base features
Avoidance	33; 34.056; 12	43; 44.618; 16	550.5	< 0.0001
Low registration	26; 26.917; 11.25	37; 36.882; 15.25	527.5	< 0.0001
Sensitivity	33; 34.5; 11	42; 42.824; 13	671	< 0.0001
Sensation seeking	41; 40.514; 10	33; 32.824; 10.75	1852	< 0.0001
Activity level	26; 26.083; 6	33.5; 31.912; 8.76	504	< 0.0001
Auditory	22; 22.79; 8	31.5; 31.088; 8.25	452.5	< 0.0001
Touch	29; 29.278; 10	34; 33.765; 7.75	744.5	0.001
Smell	17; 16.917; 6	18; 17.676; 5.75	1057.5	0.259
Movement	17; 17.528; 6	17; 17.441; 4.5	1252	0.852
Visual	24; 23.389; 6.25	24; 25.265; 9.5	1022.5	0.173

FSIQ Full-Scale Intelligence Quotient.

Data for brain factors

Magnetic resonance imaging (MRI) data acquisition

Structural and resting-state functional MRI (rs-fMRI) data were acquired using a 3T PET/MR scanner (Signa; GE, Milwaukee, WI, USA) with an 8-channel head coil at the University of Fukui Hospital. High-resolution T1-weighted structural MR data were acquired with the following parameters: repetition time (TR) = 6.38 ms, echo time (TE) = 1.99 ms, flip angle (FA) = 11°, field of view (FoV) = 256 mm, matrix = 256 × 256, slice number = 172, voxel dimension = 1 × 1 × 1 mm³. Based on the conventional T2-weighted gradient-echo planar imaging sequence, we acquired blood-oxygen-level-dependent (BOLD) contrast rs-fMRI scans with the following parameters: 3.0 mm thickness, 220 volume, 0.5 mm gap, TR = 2000 ms, TE = 24 ms, FA = 80°, FoV = 192 × 192 mm², slice number = 180, in-plane resolution = 64 × 64 pixels, pixel dimension = 3 × 3 mm².

Structural data preprocessing

Structural MRI data were preprocessed for subcortical segmentation and cortical parcellation using FreeSurfer version 6.2 [25]. For the preprocessing, we followed the conventional recon-all procedures including brain extraction, bias field correction, motion and heterogeneity correction, Talairach transformation, intensity normalization, subcortical white and gray matter segmentation, and cortical parcellation. Trained researchers inspected automatically reconstructed subcortical segmentation and cortical parcellation, and corrected them manually, if necessary.

Functional data preprocessing and thalamo-cortical rs-FC analysis

We preprocessed rs-fMRI data and conducted seed-to-voxel analysis implemented in the SPM12-based [26] CONN version 21 toolbox [27]. Following a standard preprocessing procedure, we discarded the first 10 volumes to reduce the effects of magnetic field inhomogeneity during the initial stage of MRI scan. The preprocessing procedure also included realignment to the mean image, slice-timing correction, removing volumes with a mean intensity > 1.5% of the mean global signal or 0.5 mm/repetition rate framewise displacement to reduce the effect of head movement, anatomical component-based noise correction to reduce physiological and other noise artifacts, structural segmentation and normalization, smoothing with a 4 mm full-width at half-maximum gaussian, and band-pass filtering of functional images (0.01–0.08 Hz).

For a seed-to-voxel analysis, we set the bilateral THA as seed regions of interest (ROIs) and calculated the functional correlation values between each of the two seed regions and each voxel in the whole brain. Based on FreeSurfer segmentation, the subject-specific ROIs of bilateral THA were imported into CONN. We created individual FC maps for all subjects by extracting the average time course of the BOLD signals of each seed region and correlating them with the BOLD signal time series of other voxels in the brain. The resulting Pearson correlation coefficients were converted to Fisher’s z-scores. Between-group analyses were performed to compare thalamo-cortical rs-FC changes using two-sample t-tests, with age and sex as nuisance variables. Thresholds of voxel-wise p < 0.001 uncorrected and cluster-level p < 0.05 family-wise error correction were applied.

Data for epigenetic factors

DNA methylation of AVPR1A, AVPR1B, and OXTR

Using the Oragene Discover OGR-500 kit (DNA Genotek, Ottawa, Canada), participants’ saliva samples were collected between 10:00 AM and 6:00 PM to minimize circadian variation in DNA methylation and we used a stabilizer to prevent bacterial contamination [28]. DNA were extracted from 500 μL saliva using the prepIT•L2P reagent (DNA Genotek) and mixed with a stabilizer solution. The procedure was in accordance with the manufacturer’s instruction. DNA quantification was performed using the Qubit dsDNA HS Assay Kit (Thermo Fisher Scientific, Pittsburgh, PA, USA).

In accordance with the manufacturer’ protocol, 500 ng of DNA was subjected to bisulfite conversion using the EZ DNA Methylation-Gold kit (Zymo Research, Irvine, CA, USA) and we conducted whole-genome amplification, fragmentation, and hybridization to a HumanMethylationEPIC BeadChip (Illumina, San Diego, CA). BeadChips were scanned using iSCAN (Illumina, San Diego, CA), and the methylation level (β value) was calculated for 21 CpG loci for AVPR1A, 15 CpG loci for AVPR1B, and 22 CpG loci for OXTR using the GenomeStudio Methylation Module software (Illumina, San Diego, CA), followed by the Psychiatric Genomics Consortium-Epigenome-Wide Association Studies quality control pipeline [29]. Samples with probe detection call rates of < 90% and those with an average intensity value of either < 50% of the experiment-wide sample mean or < 2000 arbitrary units were excluded using the CpGassoc package in R. Probes with a detection p > 0.001 or those based on < 3 beads were set as missing, as were probes cross-hybridizing between autosomes and sex chromosomes. Samples exhibiting missing CpG sites more than 10% in the dataset were excluded from the analysis. After the quality check, 787,680 probes were used for further analysis.

Development of predictive model

All available measurements were considered as potential predictive features. However, due to the high dimensionality and small sample size of our data, we had the specific feature selection procedure. First, we identified default sensory behavior, brain, and epigenetic predictive features for input data using Mann–Whitney U tests. It is because the data did not satisfy the assumptions normality that was assessed using the Shapiro-Wilk test. Second, we used an ensemble feature selection procedure that is based on feature occurrence frequency [30]. Third, we conducted our analyses using stratified ten-fold cross-validation to minimize fraud detection and to mitigate the risk of overfitting due to imbalanced small sample size [31, 32] (Fig. S1).

In specific, U tests were initially performed for each measurement to assess differences between ASD and TD at α = 0.05 (two-tailed) (Tables S1-S3). The default features were set with three demographic information (age, sex, FSIQ) and seven scores of sensory behaviors that exhibited significant group differences. A total of 20 features were selected from two biomarker, including 11 brain features and 9 epigenetic features, which passed the U tests. Our rationale was that these 20 selected baseline predictive features, in comparison to excluded ones, would have a greater impact on ASD classification.

At the initial stage for model development, we examined four neuroimaging-epigenetic models with 30 selected features using different machine learning algorithms including XGBoost, Random Forest, Support Vector Machine, and Logistic Regression (Table S4). We used the default parameters for each algorithm. XGBoost outperformed the other algorithms in terms of prediction accuracy. Therefore, we refined the model using XGBoost.

We maintained the proportion of TD versus ASD in the data and split it into 80% for training data and 20% for testing. Subsequently, 80% of the training data were used for validation at each iteration. The neuroimaging-epigenetic, neuroimaging, and epigenetic models were developed separately. To achieve robust feature subsets, we applied an ensemble feature selection procedure that involved two steps. First, various feature sets were iteratively created for each training model. Then, the outputs of each model were aggregated based on the feature occurrence frequency method. Using Optuna library, we further refined each training model with automatically optimized hyperparameters. Specifically, learning_rate and n_estimators were set at 0.03 and 1000, respectively. The ranges of tuning parameters included: booster = [“booster”, “gbtree”], alpha = [1e-8, 1.0], max_depth = [3,20], eta = [0.01, 0.1], gamma = [1e-8, 1.0] and grow_policy = [“grow_policy”, “depthwise”, “lossguide”].

Since neuroimaging and epigenetic models have a reduced feature number compared to the neuroimaging-epigenetic model, all three models were built on the top 15 of 30 features (10 default feature and 20 baseline predictive features) that survived the above-described feature selection procedure. We standardized the features using a standard scaler function of the sci-kit learn for training and test data. The top 15 features with the highest discriminative power and optimal hyperparameters were fed into testing. The testing model was also based on a stratified ten-fold cross-validation. The model performance was evaluated with the ROC-AUCs, which were estimated 1000 times. We tested the differences in estimated ROC-AUCs across the neuroimaging-epigenetic, neuroimaging and epigenetic models using Kruskal–Wallis ANOVA and conducted Dunn’s test for post-hoc analysis. Lastly, we separately calculated the weighted average of feature importance for the neuroimaging-epigenetic, neuroimaging and epigenetic models. Each model consisted of 15 features that were estimated 1000 times (Table 2). We present more detail for associations across the features and predictive model development process in the Supplement.

Table 2.

Weighted feature importances for neuroimaging-epigenetic, neuroimaging, and epigenetic models.

Model	domain	feature	Weighted feature importance
Neuroimaging-epigenetic	Brain function	L.MTG-L.THA	139.75
	Sensory behavior	Activity level	115.46
	Sensory behavior	Sensation seeking	115.03
	AVPR1A	cg23085912	110.72
	Brain structure	L.CAU	95.68
	Brain function	R.PreCG-R.THA	79.05
	Sensory behavior	Auditory	78.6
	Brain function	R.MTG-L.THA	76.68
	AVPR1A	cg10862431	56.25
	AVPR1A	cg09208611	54.09
	AVPR1B	cg27169129	49.59
	Brain function	R.PoCG-L.THA	46.57
	AVPR1A	cg12807275	46.29
	AVPR1B	cg09631259	42.95
	Brain function	L.MTG-R.THA	40.95
Neuroimaging	Sensory behavior	Sensation seeking	185.02
	Brain structure	L.CAU	171.54
	Brain function	L.MTG-L.THA	149.33
	Sensory behavior	Auditory	137.14
	Brain function	R.PreCG-R.THA	132.98
	Brain function	R.MTG-L.THA	117.26
	Brain function	L.MTG-R.THA	107.93
	Brain structure	R.ITG	97.78
	Demographic	Age	95.56
	Sensory behavior	Activity level	88.05
	Demographic	FSIQ	64.16
	Brain function	R.PoGG-R.THA	61.2
	Sensory behavior	Avoidance	54.12
	Brain function	R.IOG-R.THA	53.02
	Brain function	R.PoCG-L.THA	52.33
Epigenetic	Sensory behavior	Activity level	297.5
	AVPR1A	cg10862431	203.38
	OXTR	cg03987506	136.54
	Sensory behavior	Auditory	134.23
	Sensory behavior	Sensitivity	132.86
	AVPR1A	cg12807275	126.09
	AVPR1A	cg23085912	105.19
	AVPR1A	cg09208611	104.33
	Demographic	FSIQ	103.91
	OXTR	cg04523291	103.81
	AVPR1B	cg27169129	98.86
	Demographic	Age	91.55
	Sensory behavior	Avoidance	70.87
	AVPR1B	cg09631259	64.91
	Sensory behavior	Touch	51.35

AVPR arginine vasopressin receptor gene, CAU caudate, FSIQ Full-Scale Intelligence Quotient, L left, MTG middle temporal gyrus, IOG inferior occipital gyrus, ITG inferior temporal gyrus, OXTR oxytocin receptor gene, PHIG parahippocampal gyrus, PoCG postcentral gyrus, PreCG precentral gyrus, R right, THA thalamus.

Results

Default features: demographic and sensory behavior characteristics

This study included ASD (N = 34) and typical development (TD) (N = 72) individuals. ASD and TD did not significantly differ in age (U = 1139.5, p = 0.567) and FSIQ (U = 1316, p = 0.54). The result of Fisher’s exact test to examine independence of sex and group (ASD versus TD) was not statistically significant (p = 0.146).

For AASP, the ASD group had a significantly lower score in Sensation seeking (U = 1852, p < 0.0001), but higher scores in Avoidance (U = 550.5, p < 0.0001), Low registration (U = 527.5, p < 0.0001), Sensitivity (U = 671, p < 0.0001), Activity level (U = 504, p < 0.0001), Auditory (U = 452.5, p < 0.0001), and Touch (U = 744.5, p = 0.001) compared to TD group.

Brain features

The ASD group showed significant hyperconnectivity between the left thalamus (L.THA) seed and the bilateral MTG (U = 556.5, p < 0.0001 for left; U = 655.5, p < 0.0005 for right) and PoCG (U = 702, p < 0.0005 for left; U = 729.5, p < 0.001 for right). The right thalamus (R.THA) in the ASD group had hyperconnectivity with the left MTG (U = 561.5, p < 0.0001), right inferior occipital gyrus (IOG) (U = 628, p < 0.0001), right PreCG (U = 670, p < 0.0001), and right PoCG (U = 711, p < 0.001).

Structurally, ASD group showed significantly larger left caudate (L.CAU; U = 879.5, p = 0.02), left parahippocampal gyrus (L.PHIG; U = 873.5, p = 0.018), and right inferior temporal gyrus (R.ITG; U = 902, p = 0.03), compared to TD group (Fig.1).

Fig. 1. Differences in brain features between ASD and TD groups.

a significantly different rs-FCs of L.THA and 4 cortical regions. b significant group difference in rs-FCs with L.THA seed (top), L.THA seed-based positive and negative connectivity in TD (middle) and in ASD (bottom). c significantly different rs-FCs with of R.THA and 4 cortical regions. d significant group difference in rs-FCs with R.THA seed (top) as well as R.THA seed-based positive and negative connectivity in TD (middle) and in ASD (bottom). e significant group difference in volumes of L.CAU, L.PHIG (left) and R.ITG (right). CAU caudate, L left, MTG middle temporal gyrus, IOG inferior occipital gyrus, ITG inferior temporal gyrus, ParaCG paracentral gyrus, PHIG parahippocampal gyrus, PoCG postcentral gyrus, PreCG precentral gyrus, R right, rs-FC resting state functional connectivity, SMA supplementary motor area, STG superior temporal gyrus, THA thalamus.

Epigenetic feature

For AVPR1A, we observed significant hypermethylation in cg12807275 (U = 916, p = 0.037) but hypomethylation in cg09208611 (U = 1601, p = 0.011), cg10862431 (U = 1594, p = 0.012) and cg23085912 (U = 1722, p < 0.001) in the ASD group compared to the TD group. For AVPR1B, the ASD group showed significant hypermethylation in cg09631259 (U = 923, p = 0.042) and cg26596278 (U = 905, p = 0.031) but hypomethylation of cg27169129 (U = 1568, p = 0.02). For OXTR, ASD group showed significant hypermethylation of cg04523291 (U = 896, p = 0.027) but hypomethylation of cg03987506 (U = 1611, p = 0.009) compared to TD group (Fig. 2).

Fig. 2. Significant differences in levels of DNA methylation between individuals with TD (N = 72) and individuals with ASD (N = 34).

a 4CpG loci for AVPR1A (top). b 3 CpG loci for AVPR1B (bottom left). c 2 CpG loci for OXTR (bottom right). AVPR arginine vasopressin receptor gene, OXTR oxytocin receptor gene.

Classification performance evaluation across neuroimaging-epigenetic, neuroimaging, and epigenetic models

The Kruskal–Wallis ANOVA revealed a significant difference in F1 score (χ²(2) = 375.95, p < 0.0001, η² = 0.125), the receiver operating characteristic area under the curves (ROC-AUCs) (χ²(2) = 383.58, p < 0.0001, η² = 0.127), and precision-recall area under the curves (PR-AUCs) (χ²(2) = 184.99, p < 0.0001, η² = 0.112) across the three models. The neuroimaging-epigenetic model (mean = 0.841, median = 0.84, interquartile range = 0.1 for F1; mean = 0.84, median = 0.857, interquartile range = 0.107 for ROC-AUC; mean = 0.859, median = 0.87, interquartile range = 0.146 for PR-AUC) performed better in predicting ASD with higher sensitivity and specificity compared to the neuroimaging model (mean = 0.783, median = 0.79, interquartile range = 0.11 for F1; mean = 0.776, median = 0.786, interquartile range = 0.107 for ROC-AUC; mean = 0.818, median = 0.833, interquartile range = 0.151 for PR-AUC) and epigenetic model (mean = 0.769, median = 0.77, interquartile range = 0.11 for F1; mean = 0.765, median = 0.768, interquartile range = 0.107 for ROC-AUC; mean = 0.788, median = 0.803, interquartile range = 0.167 for ROC-AUC).

The Dunn’ Test found significant differences in F1 score between the neuroimaging-epigenetic and neuroimaging model (Z = 15.27, p < 0.0001, Bonferroni-corrected), between the neuroimaging-epigenetic and epigenetic model (Z = 17.98, p < 0.0001, Bonferroni-corrected), and between the neuroimaging and epigenetic models (Z = 2.71, p = 0.0067, Bonferroni-corrected). The neuroimaging-epigenetic model had significantly different ROC-AUC and PR-AUC from the neuroimaging model (Z = 15.72, p < 0.0001, Bonferroni-corrected for ROC-AUC; Z = 15.72, p < 0.0001, Bonferroni-corrected for PR-AUC) than the epigenetic model (Z = 17.98, p < 0.0001, Bonferroni-corrected for ROC-AUC; Z = 17.98, p < 0.0001, Bonferroni-corrected for PR-AUC). However, the neuroimaging model did not differ from the epigenetic model (Z = 2.25, p = 0.072, Bonferroni-corrected for ROC-AUC; Z = 2.25, p = 0.073, Bonferroni-corrected for PR-AUC) (Fig. 3).

Fig. 3. Comparison of classification performance across the neuroimaging-epigenetic model, neuroimaging model and epigenetic model.

Significant model difference in F1 score (a), ROC-AUC (b) and PR-AUC (c). ROC-AUC receiver operating characteristic area under the curve, PR-AUC precision-recall area under the curve.

Predictive feature importance

We computed the mean weighted feature importances of the three models in the training phase. The 15 features of the neuroimaging-epigenetic model are present in either the neuroimaging model or epigenetic model. For the neuroimaging-epigenetic model, the Kruskal-Wallis ANOVA indicated a significant difference in feature importance across the three feature domains (χ²(3) = 235.95, p < 0.0001). Pairwise comparison was conducted using the Wilcoxon rank-sum test with Bonferroni correction. The significant difference was observed between behavior and neuroimaging domain (p < 0.0001), between behavior and epigenetic domain (p < 0.0001), between neuroimaging and epigenetic domain (p < 0.0001). For the neuroimaging model, the Wilcoxon signed-rank test revealed that features from the behavioral domain contributed more significantly to classification than those from the neuroimaging domain (V = 1301.5 p < 0.0001). Additionally, for the epigenetic model, the features from the neuroimaging domain contributed more significantly to classification than those from the epigenetic domain (V = 62, p < 0.0001) (Fig. 4).

Fig. 4. Comparison in feature importance between neuroimaging-epigenetic, neuroimaging, and epigenetic training models.

a weighted features of brain and epigenetic domain in addition to default sensory-behavior and demographic characteristics used in all the three models. b weighted features of the neuroimaging-epigenetic model. c significant domain difference in feature importance. d weighted features of the neuroimaging model. e significant difference in feature importance between neuroimaging domain and behavior domain. f weighted features of the epigenetic model. g significant difference in feature importance between epigenetic domain and behavior domain. CAU caudate, FSIQ Full-Scale Intelligence Quotient, MTG middle temporal gyrus, IOG inferior occipital gyrus, ITG inferior temporal gyrus, L left, PHIG parahippocampal gyrus, PoCG postcentral gyrus, PreCG precentral gyrus, R right, THA thalamus.

Discussion

This study demonstrated a comprehensive and generalizable prediction of ASD by using features from sensory behavior, brain and epigenetic biomarkers. The XGboost-based classification of ASD improved prediction accuracy by including both brain and epigenetic factors as biomarkers (neuroimaging-epigenetic model), instead of relying either on brain (neuroimaging model) or epigenetic features (epigenetic model). These findings seem to confirm the heterogeneous nature of the disorder. Removing brain-related features (epigenetic model) had a more significant impact on decreasing predictive accuracy than removing epigenetic features (neuroimaging model). Supporting this, in the neuroimaging-epigenetic model, L.MTG-L.THA connectivity had the greatest feature importance and other brain features had larger feature importance compared to the epigenetic ones. This indicates that brain features may play a relatively more important role in predicting ASD than epigenetic ones. This could be attributed to the fact that brain features, particularly from brain function, are more closely linked to the default features of sensory behavior when compared to epigenetic features. It is interesting to note that the neuroimaging-epigenetic model did not include any OXTR epigenetic features. Furthermore, in the epigenetic model, AVPR1A features had greater feature importances than OXTR features. Considering that sensory behaviors were default features in the present study, we speculate that AVPR genes may be more closely associated with sensory behavior, while OXTR may be more strongly linked to social behavior.

A growing body of research suggests that thalamo-cortical dysconnectivity is considered a core feature of ASD [3, 4, 6]. Specifically, it has been proposed that hyperconnectivity between THA and the temporal and somatosensory/motor cortices in the autistic brain may indicate abnormal sensory processing [33]. In consistent with this, our study found that brain features, retained in both the neuroimaging-epigenetic and neuroimaging models included the thalamo-temporal and thalamo-sensorimotor networks, in addition to the Auditory and Activity level among behavior features.

Regarding the Auditory, the neuroimaging-epigenetic model and the neuroimaging model adopted brain features indicating hyperconnectivity of the thalamo-temporal network (L.MTG-L.THA, R.MTG-L.THA, L.MTG-R.THA). It is important to note that the MTG cluster encompassed the superior temporal sulcus (STS). Considering that the STS and MTG play roles in auditory processing and multisensory integration, our findings indicate that dysregulation of the excitation/inhibition balance in the THA and MTG pathways disrupt auditory processing and multisensory integration in ASD [34]. That is, the hyperconnectivity may reflect both disrupted bottom-up and top-down processing among individuals with ASD. Regarding altered bottom-up processing, a deficit in thalamic inhibition could affect the filtering of redundant or irrelevant auditory inputs of all environmental stimuli [5]. Additionally, altered top-down processing may be due to the modulation of the MTG inhibitory system. An electrocorticography study showed that the MTG neurons exhibited highly selective inhibitory responses to understandable speech [35].

Regarding Activity level, a behavior feature indicating the level of disposition toward involvement in movement activities, our neuroimaging-epigenetic and neuroimaging models adopted brain features of the thalamo-sensorimotor networks (R.PreCG-R.THA, R.PoCG-R.THA, R.PoCG-L.THA; L.PoCG-L.THA). This is consistent with previous reports on the maladaptive thalamo-sensorimotor connectivity, which is characterized by right dominant asymmetry of hyperconnectivity in ASD [3, 5]. Ayub et al. [3] suggested that asymmetrical hyperconnectivity in the thalamo-sensorimotor network may be responsible for alterations in motor movement and somatosensory perception in ASD [3]. Neurochemical imaging studies have suggested that an imbalance in the excitatory and inhibitory in neurotransmitter systems may be linked to abnormalities in the thalamo-sensorimotor network in ASD. In children with ASD there is a sex-dependent unilateral alteration in serotonin synthesis between the cerebellum and thalamocortical pathway. Specifically, the increased serotonin level in the cerebellum was associated decreased serotonin in the thalamocortical pathway [36]. The serotonin imbalance may affect the excitability of the thalamo-cortical pathway [36], and this alteration in serotonin appears to indicate low serotonin levels in subcortical and cortical regions [37]. In addition to serotonin, researchers have proposed a role for the inhibitory neurotransmitter GABA in ASD. Sensory over-responsivity is associated with a lower concentration of GABA in the THA and a higher concentration of the excitatory neurotransmitter glutamate in the right PoCG among children with ASD [38]. The greater excitatory tone in the thalamo-temporal and thalamo-sensorimotor networks in individuals with ASD indicates changes in bottom-up and top-down processing via interactions of several neurotransmitter systems. Lastly, enhanced bottom-up sensory networks (thalamo-cortical hyperconnectivity) may suppress the top-down theory of mind, attention, and salience networks (cortico-cortical hypoconnectivity) that are necessary for social cognition.

Structurally, the autistic brain is characterized by its abnormal regional growth patterns [39]. Both the neuroimaging-epigenetic model and the neuroimaging model identified the left CAU as a valuable feature. In addition, the right ITG was included in the neuroimaging model. In line with previous reports, we noted significantly enlarged left CAU and right ITG in the ASD group compared to the TD group. As a sign of overgrowth in the autistic brain due to abnormal neuronal pruning, the enlarged CAU is associated with restricted repetitive behaviors [40] and impaired problem-solving abilities among children with ASD [41]. Both CAU and Touch were selected as predictive features in the neuroimaging model. A recent diffusion tensor imaging study presented that the white matter fibers connected to the CAU may be involved in tactile processing [42]. Enlarged volume of the ITG has been observed in children and adolescents with ASD. Specifically, the enlargement of the right ITG has been suggested for one of the core pathophysiological features of ASD [43]. A postmortem study found that the autistic brain has an enlarged ITG with disorganized lamina structures, particularly noticeable in layers IV and V. This enlargement is a result of abnormalities in the expression of ASD candidate genes, failure of cell migration and the accumulation of such cells [44]. The neuroimaging model included both Sensitivity and the ITG as significant features. It is worth noting that Sensitivity is significantly linked to the cortical thickness of the ITG in healthy participants in a genotype specific manner. Notably, significant reductions in the ITG thickness were observed in A-allele carriers of an AVPR1A polymorphism (rs1042615) [8]. Given that rs1042615 is a potential risk factor for ASD [45], it indicates that ASD may have disrupted triangular connections involving Sensitivity, the ITG, and the AVPR1A.

In terms of epigenetics, we found the greater importance of AVPR1A over OXTR in the neuroimaging-epigenetic model and the epigenetic model. No OXTR features was survived in neuroimaging-epigenetic model. AVPR1A features were relatively ranked higher than OXTR features in the epigenetic model. Our finding of the importance of AVRP1A is in line with a recent study that found associations between chimpanzees’ joint attention task performance and AVPR1A DNA methylation levels at multiple CpG loci rather than with OXTR DNA methylation values [21]. Using blood samples, Hopkins et al. (2023) demonstrated that chimpanzees with AVPR1A hypermethylation (cg09208611, cg10862431, cg12807275) had poor performance in the joint attention task. Considering the association between impaired joint attention ability and ASD, our findings are partially consistent. In the present study, individuals with ASD had lower AVPRA methylation at chimpanzees’s hypermethylation sites (cg09208611, cg10862431) except for one CpG location, cg12807275. This discrepancy might be due to differences between the two studies, such as the species (chimpanzee versus human), the type of sample used (blood versus saliva), the experimental design (task-based), and the characteristics of the participants (e.g., ASD). In the neuroimaging-epigenetic and epigenetic models, both AVPR1A and Auditory features play a significant role in predicting ASD. This suggests that AVPR1A may be involved in the acoustic circuitry. AVPR1A is known to be highly expressed in the medial geniculate nucleus, the auditory THA, in mice that have developed vocal communication [17]. Similarly, non-human primates showed that AVPR1A was prominently expressed in the medial geniculate nucleus, while OXTR was more abundant in the lateral geniculate nucleus, which is primarily responsible for processing visual information [46]. The epigenetic model found that Sensitivity and Avoidance were contributing factors. These features indicate a low neurological threshold for sensory inputs and inability to filter irrelevant sensory information, as described by the sensory processing framework of AASP. The AVPR1A promoter region is known for its association with auditory pre-pulse inhibition. This phenomenon is characterized by a suppressed autonomic startle response to a stronger auditory stimulus that follows a weak stimulus [47]. Given that pre-pulse inhibition not only indicates sensorimotor gating but also measures an individual’s ability to filter out distracting inputs and attend to target stimuli, reduced pre-pulse inhibition has been considered an endophenotype of ASD [48].

Two AVPR1B features (cg09631259, cg26596278) were found to be present in both the neuroimaging-epigenetic and epigenetic models. AVPR1B is prominent in the mouse hippocampal CA2 region, which is known to play a role in aggression and social memory [49]. In humans, AVPR1B variants have been linked to aggression [50], empathy [51], and ASD [52]. Based on the findings, the current study suggests that epigenetic modification of AVPR1B can serve as a reliable biomarker for ASD.

In the epigenetic model, two features from OXTR (cg03987506, cg04523291) were found to be contributing factors to ASD prediction. Epigenetic modifications of OXTR have been linked to autistic traits related to social behaviors, and OXTR hypermethylation is associated with reduced FC [15, 16, 53]. However, several studies have reported hypomethylation in ASD [54]. A review by Moerkerke et al. (2021) revealed these contradictory results, attributing not only to differences in sex, age, and CpG sites, but also to the inclusion of other psychological factors, such as anxiety and stress. In line with these inconsistent findings, the current study retained two features that exhibited hypomethylation at cg03987506 and hypermethylation at cg04523291. This suggests that fine-tuning of DNA methylation of the three genes (OXTR, AVPR1A, AVPR1B), rather than either hypermethylation or hypomethylation of the genes plays a significant role in ASD. The epigenetic model incorporates Touch as a valuable feature, in addition to OXTR features. In a study that focused on human mother-infant pairs, it was found that infants had significant changes in OXTR methylation during their early sensitive period. These changes were observed because of maternal engagement, particularly through affectionate touch [55]. Caregiving touch was suggested for its protective effect on epigenetic mechanisms [56].

The current study has several limitations. Above all, the sample of the current study was not large enough to account for the high dimensionality of the obtained data. To prevent overlearning problems due to small sample size, for each diagnostic model, we reduced the number of features for input data using rigorous feature selection procedure at the early stage of model development (see Method). This may have excluded important features for predicting ASD. Additionally, the validation of the three diagnostic models developed in this study is required using larger data sets. Second, the collected data has class imbalance. Therefore, we employed a stratified sampling strategy and optimally tuned hyperparameters. Nonetheless, the error rate of the classifier can be affected by a class imbalance [32]. Third, DNA methylation values were computed from the participants’ saliva samples, which were neither cell-specific nor directly related to the brain tissue.

The present study dissected the interplay of brain and epigenetic factors in association with sensory processing to demonstrate that a precise classification of ASD could be achieved by combining these heterogeneous factors in a machine learning framework [22]. Considering that accurate diagnosis of ASD is critical, this study provides a new approach to improve diagnostic precision.

Author contributions

YC, JB, HK, and MJ designed research; HO, HK, and MJ performed research; YC, JB, SL, JR, and MJ analyzed data; and YC, JB, SL, JR, HO, HK, and MJ wrote the manuscript.

Data availability

Data are available upon reasonable request to corresponding authors.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-025-03566-2.