Integrating multi-omics data through deep learning with applications to autism risk gene prioritization

Autism spectrum disorder (ASD) is a highly heritable neurodevelopmental disorder. Advances in genetic research, particularly whole-exome sequencing studies, have identified an increasing number of ASD risk genes, especially through large-effect rare variant analyses. However, current rare-variant studies still have limited power to recover the full spectrum of ASD risk genes, leaving a substantial gap between known genes and the estimated genetic architecture of ASD. To address this challenge, we developed Deep-MOTIFs (Multi-Omics inTegratIon For autiSm), a Bayesian-guided positive-unlabeled deep learning framework for multi-omics integration that combines genetic, transcriptomic, and protein-interaction evidence for genome-wide ASD risk-gene prioritization. Deep-MOTIFs achieved stronger performance than existing methods across held-out ASD gene prediction, top-ranked prioritization, positive-label efficiency, and label-noise robustness analyses. Functional enrichment and genome-wide decile analyses further supported the biological relevance of the top-ranked candidates, showing convergence on ASD-related phenotypic annotations and synaptic, neurodevelopmental, and chromatin-regulatory mechanisms. Together, these results support Deep-MOTIFs as a robust framework for ASD risk-gene prioritization under incomplete supervision. Source code is available at: https://github.com/shan-team/Deep-MOTIFs.