Supervised learning speeds up parametrization of chromatin simulations

The physical organization of the genome in three-dimensional space regulates many biological processes, including gene regulation and cell differentiation. Given the complexity of biological systems, molecular modeling of chromatin structure provides essential insight into the factors driving genome organization. A common modeling approach is to parametrize a polymer model using an experimental Hi-C contact map as a constraint. Existing approaches in this framework leverage the maximum entropy principle to optimize the polymer model parameters. The maximum entropy approach requires running simulations iteratively until the parameter values converge, which is computationally expensive. Here, we show that supervised machine learning enables dramatically faster parametrization of chromatin simulations. We train a graph neural network to predict polymer model parameters from experimental Hi-C contact maps by training on simulated contact maps with known parameters. Our approach is comparably accurate to the maximum entropy approach, but at a fraction of the runtime. We anticipate that our method will be particularly applicable to the single-cell setting, as single-cell Hi-C variants can generate thousands of Hi-C contact maps.