Chromatin structure (how genes are organized in 3-D space) exists at an intermediate level between a gene’s DNA sequence, the regulatory states of cells, the identities of cells, and how cells respond to chemical perturbations. Direct measurements of these structures with single-cell Hi-C and similar 3-D genome mapping techniques have been conducted on occasion; however, such measurements have typically occurred infrequently due to high costs. Additionally, most 3-D genome mappings have been conducted on relatively few cells compared to many other types of single-cell omics measurements. In this SuperUROP, we ask a focused yet bold question: Can ChromoGen be developed into a benchmarkable conditional generation engine using genomic sequence inputs for generating realistic 3-D ensembles of chromatin conformations? Furthermore, can we use small-molecule perturbation labels and/or additional single-cell omics regulatory readouts as conditions to determine when the condition has successfully influenced, failed to influence, or leaked into the output? In building a benchmarkable conditional generation engine to produce plausible 3-D chromatin conformations based upon genomic sequence and/or additional single-cell regulatory readouts under one or more conditions, the objective of this project is not to create a novel foundational model from scratch. Instead, the goal is to develop a clean and reproducible framework to support the current expansion of ChromoGen, including preprocessing of the input data, determining which branches to select, designing the conditions applied to the model, generating the coordinates of the 3-D chromatin ensembles, validating the structural integrity of the generated 3-D chromatin ensembles, and quantifying/analyzing uncertainty/failure in addition to attributing features within the 3-D chromatin ensembles to specific inputs.

I am incredibly excited to take ChromoGen-Engine to it’s next level! I am sure we will not even stop at version two.