Virtual Cell Approaches to Understand Human Disorders
Autism Spectrum Disorder (ASD) is a highly heritable neurodevelopmental condition characterized by substantial genetic heterogeneity. While large-scale genomic studies have identified a growing number of risk genes, interpreting their biological significance and discovering novel candidates beyond statistically significant thresholds remains a fundamental challenge. Classical genomic approaches require massive sample sizes and struggle to bridge the gap between genetic findings and underlying neurobiology. To address these limitations, we present an AI-based virtual cell framework that leverages single-cell foundation models trained on human brain organoid data. By constructing a large-scale, harmonized organoid cell atlas encompassing diverse perturbation and disease conditions, we trained tissue-specific foundation models capable of predicting transcriptional responses to gene perturbations at single-cell resolution. This in silico perturbation approach enabled us to systematically map functional gene networks in a human cortical context without requiring matched experimental data for every gene of interest, effectively expanding the reach of perturbation biology beyond what is currently feasible in the laboratory. Applying this framework to ASD genetics, we demonstrate that single-cell foundation models can identify convergent neurodevelopmental programs underlying the disorder's genetic heterogeneity and prioritize novel candidate risk genes that share functional characteristics with established ASD genes. We further show that organoid-based virtual cell approaches can be extended to assess developmental maturity of organoid samples, model combinatorial perturbation responses, and probe cell-fate dynamics relevant to disease mechanisms. Together, these findings highlight the potential of single-cell foundation models trained on organoid data as a scalable and biologically grounded strategy for understanding human neurodevelopmental disorders, offering a complementary and powerful approach to traditional genomic and experimental methods.
