PRBB-CRG Sessions Christina Leslie

03/10/202512:00MARIE CURIEPRBB-CRG SessionsChristina LeslieMemorial Sloan Kettering Cancer Center"Machine learning models for single-cell and regulatory genomics"Host: Fanny Mollandin & Jorge FerrerAbstract:We will present new deep learning models that exploit single-cell multiome (scRNA+ATAC-seq) and 3D genomics data in order to link enhancers and non-coding variants to genes and to decipher gene regulatory networks. First, in collaboration with the Danwei Huangfu lab, we developed a model called GraphReg+ to integratively model (pseudobulk) multiome and HiCAR data at multiple steps of guided in vitro pancreatic differentiation of pluripotent human stem cells towards the late endocrine cell stage. GraphReg+ is a graph attention network that passes information from enhancers to promoters along HiCAR ‘edges’ to predict the expression of genes. We used GraphReg+ to identify novel cell-type-specific enhancers of diabetes genes and to functionally annotate non-coding genetic variants associated with type 2 diabetes. Next, we will present a neural ordinary differential equation (ODE) model called DynaVelo that learns cell dynamics at single-cell resolution from multiome data. We applied DynaVelo to model the in vivo dynamics of wildtype and mutant germinal center B cells, in collaboration with the Ari Melnick lab. This analysis enabled the identification of dynamic gene regulatory networks throughout the germinal center reaction. Further, in silico transcription factor (TF) perturbations allowed both the prediction of cell dynamics under loss-of-function genetic mutations and the identification of TF perturbations to rescue loss-of-function phenotypes.