Manu Saraswat

Single-cell Deep Omics Regulatory Inference

A mechanistically constrained VAE that reconstructs enhancer-driven gene regulatory networks from single-cell ATAC–RNA data. Applied to >1M glioblastoma cells, scDoRI identified repressive transcription factor circuits regulating tumor plasticity—validated in vivo by slowing tumor growth and increasing survival in mouse models. This computation → wet-lab validation → therapeutic hypothesis pipeline demonstrates AI that doesn't just predict, but explains and enables intervention.

Personalized Gene Expression Prediction

A modular deep learning system predicting cell-type-specific gene expression from personalized genomes, combining genomic sequence with chromatin accessibility. Unlike Enformer/Borzoi (which predict from reference genomes), DeepGenoXcan is designed for individual genetic backgrounds—critical for precision medicine. Outperforms both linear TWAS models and deep neural network models on donor-level generalization.

Using Biology to Design Training Strategies

Demonstrated that training strategies informed by biological knowledge—Transcription Factor (TF) family relationships, DNA-binding domain similarity—systematically outperform naive transfer learning. Biology isn't just the application domain; it's a source of insight for how to train models better.

Explainable Neural Networks for Genomics

One of the first fully transparent deep learning architectures for genomics. Demonstrates that high predictive performance and complete interpretability can coexist, producing motif-level explanations of transcription factor binding and chromatin accessibility predictions.