On the emergence of machine-learning methods in bottom-up coarse-graining

Abstract

Machine-learning methods have gained significant attention in the computational chemistry community as a viable approach to molecular modeling and analysis. Recent successes in utilizing neural networks to learn atomistic force-fields which 'coarse-grain' electronic structure have inspired similar applications to the thermodynamic coarse-graining of chemical and biological systems. In this review, we discuss the current viability and challenges in using machine-learning methods to represent coarse-grained force-fields, as well as the utility of machine-learning in various aspects of coarse-grained modeling.

Figure 1.

Current and future directions for ML in CG modeling. a) ML for optimal CG mappings, adapted from Ref. An autoencoding formalism is adopted in which the latent space of the autoencoder constitutes the CG configurational space (above); discrete optimization is performed to obtain a CG mapping (below). b) Modern ML CG models utilize equivariance and graph NNs to construct potential energies. c) Error analysis between AA and CG model (above) of the actin protein using ML-derived CVs (below), adapted from Ref. d) ML is utilized to enable backmapping from a CG configuration to an AA representation, adapted from Ref.

Figure 2.

Future challenges for ML in CG modeling. (Left) Addressing transferability of CG FFs across thermodynamic state points, in which explicit dependence on thermodynamic variables must be learned. (Middle) Investigation of mode additivity for CG FFs. The combining rule protocol for interactions between Lennard-Jones particles (above) is a paradigmatic example of modularity in FFs, similar protocols for ML FFs have yet to be investigated. (Right) Application of ML FFs to biological systems. Ideally, CG models can be synthesized directly from protein sequence and resulting interactions with bioactive molecules can be accurately assessed. The HIV capsid protein assembly complex and inositol hexakisphosphate (IP6) polyanion are shown as an example.