We will present several recent machine learning methods that exploit single-cell chromatin accessibility (scATAC-seq), single-cell multiome, and 3D genomics to tackle problems in regulatory genomics. We will first present ChromaFold, a deep learning model that predicts the Hi-C contact map from scATAC-seq alone, allowing the prediction of gene regulatory interactions and the deconvolution of bulk Hi-C data into cell-type-specific contact maps. We will also describe a generative neural ODE model called DynaVelo for learning cellular dynamics from single-cell multiome data. We apply DynaVelo to resolve the complex dynamics of wild type and mutant murine germinal center B cells, and we show how in silico gene (and in particular transcription factor) perturbations allow both the prediction of cell dynamics under loss-of-function genetic mutations and the identification of transcription factor perturbations to rescue loss-of-function dynamic phenotypes. Finally, we will turn to a recent collaborative cancer immunology project, where single-cell analyses together mouse genetic studies establish the presence of two regulatory T (Treg) populations with opposing roles in colon cancer.