Digvijay Kumar Yarlagadda

I am a PhD student in the Tri-Institutional Program in Computational Biology & Medicine at Cornell University, Sloan Kettering Institute and Rockefeller University. I am comentored by Christina Leslie and Joan Massagué. My research interests lie at the interface of computer vision and biology. In my current research, I am focused on developing novel representation learning methods for spatial transcriptomics data and applying these methods to study cancer metastasis.

During summer 2024, I interned with Olga Troyanskaya at Flatiron Institute, Simons Foundation and Princeton University, working on vision transformer models for integrating microscopy and genomics datasets. Prior to joining PhD program, I was part of Thomas Fuchs Lab at Memorial Sloan-Kettering Cancer Center where I worked on developing computer vision and deep learning algorithms for computational pathology, which sparked my interest in exploring the challenging problem of cancer metastasis through the lens of machine learning.

I have a master's degree in computer science from University of Missouri-Kansas City, where I was advised by Praveen Rao and funded by dean's international scholar award. I also took on roles to launch products at early stage startups AgShift, Strayos and Zenjoi.

Email  /  Google Scholar  /  Github  /  LinkedIn   

I'm interested in computer vision and deep learning research, with applications in biology and medicine.

Spatial and single-cell profiling of ccRCC tumors reveals that cancer cells distribute along a continuous axis of embryonic nephrogenesis, and that immune checkpoint blockade enriches for the cells most committed to it. Nicheverse, a discrete representation learning framework for imaging-based spatial transcriptomics, shows that these persister cells reside in niches depleted of CD8+ T cells.

A spatial and single-cell atlas of the human pancreatic cancer perineural niche, built with a vision-centric model for imaging-based spatial transcriptomics coupled with pathology-guided snRNA-seq, showing that invaded nerve neighborhoods harbor classical cancer cells with myofibroblastic CAFs and lipid-associated macrophages, while non-invaded nerves harbor inflammatory CAFs and B and T lymphocytes.

A differentiable cellular representation for imaging based spatial transcriptomics that models each cell as an anisotropic 3D Gaussian, keeping geometry and per cell features differentiable through both volumetric reconstruction and soft transcript assignment.

Discrete representations can capture multi-scale spatial structures in spatial transcriptomics data.

Memory augmentated neural networks provides a better way to rememeber rare occurances of cancer in histology images.

Segmentation of histopathology images can be improved by training on patches from multiple magnifications.

Using Z-indexing creates more effictive platform for WSI storage and retrieval.

Computer vision algorithm to extract annotations on large molecular pathology datasets.

End to end part learning achieves state-of-the-art performance for multi-label prediction.