Invitation Only
Speaker: Christina Leslie, Ph.D., Memorial Sloan Kettering Cancer Center
Title: AI Models for Regulatory Genomics and Spatial Biology
We will present new ML/AI models for single-cell multiome (scRNA+ATAC-seq) and imaging-based spatial transcriptomics with applications to decoding immune cell state dynamics and the organization of the tumor microenvironment. First, we will present a generative neural ordinary differential equation (ODE) model for single-cell multiome data called DynaVelo, which we use to learn a functional form of the dynamics — corresponding to learned RNA velocity and transcription factor (TF) motif velocity vector fields — of wildtype and mutant germinal center B cells. We show that Jacobian analysis or in silico perturbations can recover dynamic regulatory networks governing cell state transitions in the germinal center. We also show how to predict the impact of genetic loss-of-function mutations on cell dynamics, and how to predict TF perturbations that rescue loss-of-function phenotypes. Second, we will present a new discrete representation learning model for annotating neighborhoods and cells in imaging-based spatial transcriptomics (ST) by exploiting pixel-level transcript information, without using cell segmentation or converting ST data to a cell-by-gene matrix. We apply our model to ST data profiling pancreatic ductal adenocarcinoma (PDAC) tumors from Memorial Sloan Kettering patients to decode the local microenvironment of perineurally invaded versus non-invaded nerves.