Application of deep learning algorithm on whole genome sequencing data uncovers structural variants associated with multiple mental disorders in African American patients

The study addresses the challenge of accurately diagnosing mental disorders—particularly in individuals with comorbid conditions and in young children where standardized assessments are difficult to administer. Mental disorders impose substantial personal and economic burden, and misdiagnosis rates remain high despite clinical guidelines. Genomic structural variation, including non-coding variants (e.g., lncRNAs, UTRs, introns), has been implicated in psychiatric conditions and may serve as biomarkers or therapeutic targets. Prior machine learning work often targets single disorders or non-diverse populations. This study investigates whether deep learning applied to whole-genome sequencing (WGS)–derived structural variant burdens can (1) distinguish mental disorder patients from controls and (2) multi-label patients across eight disorders (ADHD, depression, anxiety, autism, intellectual disabilities, speech/language disorders, developmental delays, and oppositional defiant disorder) in an African American cohort.

Existing research shows structural variants and dysregulation of lncRNAs can affect gene regulation and contribute to complex diseases, including neuropsychiatric disorders. Machine and deep learning have been used for mental health using clinical, genetic/genomic, behavioral, and social media features, but many focus on a single disorder (e.g., bipolar disorder, ADHD) and do not address comorbidity or minority populations. Prior genomic studies have prioritized susceptibility genes/pathways, and deep learning efforts have identified mutations linked to specific disorders. There is a gap in multi-label prediction across multiple psychiatric diagnoses, particularly in African American cohorts.

Cohort: 4179 African American individuals from the CHOP Center for Applied Genomics biobank with TOPMed WGS; 1384 had at least one of eight mental disorder diagnoses. Consent and IRB approvals were obtained. Mental health phenotypes were derived from de-identified EHR using ICD-9/ICD-10 codes. WGS processing: DNA from whole blood; libraries prepared with Illumina TruSeq DNA PCR-Free; sequencing on Illumina HiSeq X Ten (150 bp paired-end). bcl2fastq v2.15.0 used for FASTQ generation; alignments followed CCDG standardized pipeline. Common variants (MAF > 0.05 in African ancestry per ExAC) were removed. Feature construction: The GRCh38 genome was partitioned into 587 ~5 Mb segments. For each individual and genomic segment, burdens (occurrence counts) were computed for seven variant classes: nonsynonymous SNVs, frameshift SNVs, stop-gain SNVs, UTR SNVs, non-coding RNA SNVs, intronic SNVs, and intergenic SNVs. These segment-level burdens formed feature vectors. Feature reduction/weighting: A Random Forest (Sklearn 0.21.3) computed feature importance per segment (Gini impurity; 500 trees; min samples split=2; expand until leaves pure or <2 samples; max features=sqrt of total). Importances were normalized to sum to 1; zero-importance features were removed. Segments with highest weights were considered hotspots; drug-target genes in hotspots were explored via DGIdb (FDA-approved only). Modeling: A multi-layer perceptron (MLP) classifier (Sklearn 0.21.3) was trained separately for each variant class. Two tasks: (1) binary classification of mental disorder cases vs controls; (2) multi-label classification among eight disorders for patients with ≥1 diagnosis (phenotype as 1x8 binary vector). MLP hyperparameters (max iterations, L2 alpha, activation, solver, learning rate, number of layers/neurons) were optimized using Bayesian optimization (gp_minimize, scikit-optimize 0.7.2). Evaluation: Two-fold shuffle tests were run for 50 random rounds. For case-control, 1384 cases and 2795 controls were split 1:1 into train/test each round; selected features from training were used to train the MLP and evaluate on the held-out test. For multi-label, the 1384 cases were split 1:1 for train/test each round; predictions were 1x8 vectors. Metrics: accuracy for case-control, Hamming loss and exact match rate for multi-label, and per-disorder precision/recall. Additional analyses included weight distribution across chromosomes for coding vs non-coding classes and DAVID enrichment of hotspot genes. Computational considerations: feature extraction ~5 days on clusters; hyperparameter optimization ~3 days; model training under a day on standard hardware.

- Dataset: 4179 African American individuals; 1384 with ≥1 of eight mental disorder diagnoses. Approximately 95% under 18 years; 22% under 3 years. - Case vs control classification (mean ± SD over 50 two-fold shuffles): • Nonsynonymous SNVs: 64.5 ± 1.2% accuracy (71.7 ± 1.34% when restricting to single-diagnosis cases vs controls). • Frameshift SNVs: 64.0 ± 1.4% (70.8 ± 1.61% single-diagnosis). • Stop-gain SNVs: 65.1 ± 0.97% (71.49 ± 1.69% single-diagnosis). • UTR SNVs: 65.5 ± 1.1% (72.4 ± 1.44% single-diagnosis). • ncRNA SNVs: 65.7 ± 1.3% (72.6 ± 1.52% single-diagnosis). • Intronic SNVs: 65.7 ± 1.1% (72.8 ± 1.29% single-diagnosis). • Intergenic SNVs: 64.6 ± 1.1% (73.1 ± 1.23% single-diagnosis). - Multi-label classification among 8 disorders: • Hamming loss: 0.28–0.31 across variant classes (i.e., ≥70% of labels correct). • Exact match rate: 7.2–9.3% (vs ~0.4% by random guess for 8 labels). • Per-disorder performance varied; relatively higher precision/recall for ADHD (~30–43%) and developmental delays (~28–37%) depending on variant class; lower for autism and ODD, consistent with smaller sample sizes. - Non-coding vs coding variants: Non-coding variant burdens (UTR, ncRNA, intronic, intergenic) achieved predictive performance comparable to coding variants. Weight distributions for non-coding features had narrower standard deviations and lacked distinct hotspots, suggesting they may serve as proxy biomarkers. - Hotspots and pathway enrichment (coding variants): Recurrent hotspots overlapped across case-control and multi-label models, notably chr19:50–55 Mb and chr11:55–60 Mb, with enrichment for immune response regulation, chemokine signaling, cytokine interactions, G-protein–coupled receptor signaling, olfactory receptor activity/transduction, antigen processing and presentation, osteoclast differentiation, and NK cell–mediated cytotoxicity. - Overall, models demonstrate moderate discrimination of cases vs controls (~65% accuracy) and meaningful multi-label performance (Hamming loss <0.3), providing proof-of-concept for genomic-ML–based adjunctive diagnostics.

The findings address two core questions: a deep learning model using WGS-derived structural variant burdens can moderately distinguish mental disorder patients from controls and can multi-label patients across eight psychiatric diagnoses better than chance by a large margin. Comparable performance of non-coding variants to coding variants and their uniform weight distributions imply non-coding burdens may act as proxy markers of underlying functional variation. Coding-region hotspots were consistent across tasks and enriched for immune, chemokine, GPCR, and antigen presentation pathways, aligning with proposed immune and signaling mechanisms in psychiatric disease. The results are clinically relevant given high misdiagnosis rates and challenges in diagnosing comorbid conditions and young children; such models could serve as adjunctive tools to inform clinical decision-making. Performance constraints likely reflect heterogeneity from aggregating multiple disorders, limited training data for some labels, and shared genetic architecture among psychiatric conditions. Nevertheless, exact match rates far exceed random expectation, supporting feasibility and motivating further refinement and validation.

This study demonstrates that deep learning applied to WGS structural variant burden—spanning both coding and non-coding regions—can aid in classifying mental disorder status and multi-labeling eight psychiatric diagnoses in an African American cohort. Non-coding variants are as informative as coding variants, and coding-region hotspots reveal pathway enrichments implicating immune and GPCR-related biology. These results provide a proof-of-concept for genomic-based adjunctive diagnostics in psychiatry, particularly valuable for complex comorbid presentations and in pediatric populations. Future work should expand cohort size (especially for underrepresented diagnoses), include diverse ancestries, integrate additional data modalities (e.g., gene expression, clinical features), explore more advanced architectures and representation learning, and validate in prospective clinical settings.

- Limited number of patients with multiple diagnoses (662 with >2; 274 with >3) reduced training signal for multi-label learning. - Small sample sizes for certain disorders (e.g., ODD, autism) led to lower per-disorder precision/recall. - Combining eight heterogeneous disorders as a single case group for case-control increases genetic heterogeneity, reducing accuracy compared to single-disorder models. - Shared genetic risk across psychiatric disorders may confound precise multi-label discrimination. - Results are derived from an African American cohort; external generalizability to other populations requires validation.