Research Highlights

Supervised Deep Learning for Regulatory Genomics: Deep learning models have been shown to have high performance across a variety of genomic applications. But can these models lead to new biological discoveries? To investigate this question, I have worked on interpreting deep learning models trained on single-cell ATAC-seq data from the brain in order to unearth novel, cell-type specific epigenomic regulation.

Interpretable Deep Learning for Regulatory Genomics: Gradient-based attribution methods are popularly used to interpret the patterns learned by deep learning models. However, these attribution scores can be affected by noise emerging from the model learning arbitrary functions outside the probabilistic simplex inhabited by one-hot encoded DNA. To address this issue, I helped to demonstrate the efficacy of a statistical correction of input gradients to many regulatory genomic prediction tasks.