Hierarchical clustering of chemical space using binary-encoded SMARTS for building data-driven chemical perception models

Molecular mechanics force fields require a chemical perception model to assign parameters to molecules. A recent advancement in force fields is the use of the SMARTS substructure query language as the perception model. Although it is straightforward to write SMARTS patterns to define new force field parameters, it is difficult to manually determine the particular patterns needed to improve performance. Here we present a scheme to automatically generate SMARTS patterns that partitions the data according to a given objective. From this operation we describe how hierarchies of SMARTS patterns can be generated from the input data. Such hierarchies can be used for building force fields from scratch or improve upon established ones. We test our methods by building SMARTS hierarchies that map SMARTS patterns to existing force fields such as GAFF and Sage which allows us to test that our approach can recognize the chemical substructures typically used by human-designed force field perception models. We then build a SMARTS hierarchy that clusters bonds based on quantum mechanically-derived bond lengths which demonstrates the possibility of creating novel force field chemical perception models.