Coarse graining (CG) is used in molecular dynamics to approach larger spacial and temporal scales which are inaccessible with all-atom models. Two main challenges associated with developing CG models: selecting a suitable CG mapping and selecting CG interactions. A CG mapping projects the structural features of the all atom system to a latent space. However, ways of generating CG mappings become combinatorially large with the increasing number of atoms in a system. Hence, manual selection of CG mappings using human chemical intuition is the state-of-the-art method. We present a  graph neural network (GNN) based model for mapping prediction called Deep Supervised Graph Partitioning model (DSGPM). CG mapping prediction is analogous to partitioning a molecular graph, with some wild assumptions.  Spectral clustering is a baseline graph clustering method which models pairwise affinity between nodes comparatively better. We use a GNN to learn this affinity matrix from learned atom features. This enables to learn a model where the embeddings of similar data are separated by short distances (non-cut pair loss) and dissimilar data are separated by long distances (cut triple loss). We use a dataset of more than 1000 human annotated mappings of small molecules to train the DSGPM model. In table 1, our results are evaluated against other baseline methods. Results show that DSGPM outperforms state-of-the-art methods and the predictions are indistinguishable from human predictions. Furthermore, we show that DSGPM can be applied to predict CG mappings of arbitrarily large structures which addresses a main challenge in CG modeling.  

Predicting coarse-grained (CG) mappings using graph neural networks: Applications in CG molecular dynamics

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.