Efficient computational implementation of polymer physics models to explore chromatin structure

Abstract

The development of novel experimental technologies able to map genome-wide chromatin contacts, as Hi-C, GAM or SPRITE, allowed to derive detailed information about the spatial structure of chromosomes in the cell nucleus. They revealed that the genome has a complex spatial organisation, which is highly connected with its activity. In the last years, such an abundance of experimental data prompted the development of quantitative models based on Polymer Physics to describe the chromatin architecture, clarifying many aspects about the molecular mechanisms underlying genome folding. Efficient algorithms are thus fundamental to perform massive numerical simulations for testing the accuracy of these models and provide a good description for small genomic regions or for whole chromosomes. Here, we consider the performances of Molecular Dynamics (MD) implementation of commonly used polymer physics models. Such models can be combined with Machine Learning approaches informed with experimental data to produce more accurate descriptions of real genomic regions. However, the execution times increase as a power-law with the size of the input data, which ultimately reflects the complexity of the investigated system. The best strategy is therefore a convenient trade-off between the accuracy in the description and the availability of computational resources. The combination of innovative experimental data and polymer physics theories allow to reconstruct the 3D genome structure. This is achieved by the use of machine learning approaches and massive parallel computing. Efficient algorithms and computational resources are then fundamental to produce models of increasingly high accuracy.