Interpreting Chemical Words of a Data-driven Segmentation Method as Protein Family Pharmacophores and Functional Groups

Machine learning models have found numerous successful applications in computational drug discovery. A large body of these models represents molecules as sequences since molecular sequences are easily available, simple, and informative. The sequence-based models often segment molecular sequences into pieces called chemical words and then apply advanced natural language processing techniques for tasks such as de novo drug design, property prediction, and binding affinity prediction. However, the fundamental building blocks of these models, chemical words, have not yet been studied from a chemical perspective so far, and it is unknown whether they capture chemical information. This raises the question: do chemical-word-based drug discovery models rely on chemically meaningful building blocks or arbitrary chemical subsequences? To answer this question, we first propose a novel pipeline to highlight the key chemical words for strong binding to a protein family and then study the substructures designated by the key chemical words for three protein families. For all three families, we find extensive evidence in the literature that chemical words can designate pharmacophores and functional groups, and thus chemical-word-based models, indeed, rely on chemically meaningful building blocks. Our findings will help shed light on the chemistry captured by the chemical words, and by machine learning models for drug discovery at large.