Machine learning-based parametrization of a chromatin polymer model

The physical organization of the genome in three-dimensional space regulates many biological processes, including gene expression and cell differentiation. In order to fully understand the complexity of these processes, sophisticated computational models of chromatin structure will be essential. Given an experimental Hi-C contact map, we seek to estimate the interaction parameters of a polymer model such that the polymer model reproduces the underlying chromatin structure. We show that a graph neural network (GNN) can learn a mapping from an experimental contact map to a set of interaction parameters that is empirically competitive with parameters estimated using traditional methods but can be computed an order of magnitude faster. We develop an approach to train the GNN using exclusively simulated data from the polymer model, avoiding the need for large quantities of experimental data. We demonstrate that our approach generalizes across experimental cell lines and is robust to the range of sequencing read depths seen in bulk Hi-C data. We anticipate our method to be particularly useful in the single-cell setting, as single-cell Hi-C variants can generate tens of thousands of contact maps.