Multi-omics data integration:

Machine learning methods are applied to extract biomarkers and build predictive models from multi-omics data. Multiple Kernel Learning (MKL) is used recently to integrate DNA methylation and gene expression data to predict cancer stages. Further, deep learning approaches are applied to extract features and develop models from multi-omics data of cancer. Methods for interpretability of deep learning models are also being explored.

Computer aided drug design:

Intrepretable deep learning models for prediction of drug properties such as binding affinity and toxicity are being developed. Other datasets for use in autoencoders along with traditional docking calculations are being developed.

Single-cell transcriptomic analysis:

The focus is to develop methods to learn network-level features for clustering and to learn pseudotime trajectory from single cell data. This will be used for characterizing cell types/subpopulations and to study the inter- and intra-variability in gene expression in health and disease.

Quantum chemical energies and forces:

Deep neural network based models along with novel molecular feature vectors are being developed for accurate prediction of quantum mechanical energies and DFT level atomic forces for use in molecular modeling and molecular dynamics simulations.

Computer aided material design:

Modern machine learning methods such as conditional variational autoencoder and reinforcement learning are used for de novo design of materials with tailored properties.

Deep learning approaches to analyze digital pathology images:

Convolutional neural network models are developed to detect cancer subtypes and map morphological features to predict survival from whole-slide pathological images. Genotype to phenotype correlation is also being examined.