ESPResSo Summer School: Dr. Tristan Bereau

Machine learning in multiscale modeling for molecular discovery and backmapping

Advanced statistical methods are rapidly impregnating many scientific fields, offering new perspectives on long-standing problems. In materials science, data-driven methods are already bearing fruit in various disciplines, such as hard condensed matter or inorganic chemistry, while comparatively little has happened in soft matter. I will describe how we use multiscale simulations to leverage data-driven methods in soft matter. We aim at establishing structure-property relationships for complex thermodynamic processes across the chemical space of small molecules. Akin to screening experiments, we devise a high-throughput coarse-grained simulation framework. Coarse-graining is an appealing screening strategy for two main reasons: it significantly reduces the size of chemical space and it can suggest a low-dimensional representation of the structure-property relationship. I will describe a biological application of our methodology that led to the discovery of in vivo active compounds. Finally, I will focus on the problem of backmapping, i.e., reconstructing atomistic details from coarse-grained representations. I will present a generative approach conditional on the coarse-grained degrees of freedom.