PREFMoDeL: A Systematic Review and Proposed Taxonomy of Biomolecular Features for Deep Learning

Abstract

Of fundamental importance in biochemical and biomedical research is understanding a molecule’s biological properties—its structure, its function(s), and its activity(ies). To this end, computational methods in Artificial Intelligence, in particular Deep Learning (DL), have been applied to further biomolecular understanding—from analysis and prediction of protein–protein and protein–ligand interactions to drug discovery and design. While choosing the most appropriate DL architecture is vitally important to accurately model the task at hand, equally important is choosing the features used as input to represent molecular properties in these DL models. Through hypothesis testing, bioinformaticians have created thousands of engineered features for biomolecules such as proteins and their ligands. Herein we present an organizational taxonomy for biomolecular features extracted from 808 articles from across the scientific literature. This objective view of biomolecular features can reduce various forms of experimental and/or investigator bias and additionally facilitate feature selection in biomolecular analysis and design tasks. The resulting dataset contains 1360 nondeduplicated features, and a sample of these features were classified by their properties, clustered, and used to suggest new features. The complete feature dataset (the Public Repository of Engineered Features for Molecular Deep Learning, PREFMoDeL) is released for collaborative sourcing on the web.

Keywords

• feature
• machine learning
• deep learning
• feature selection
• feature engineering
• drug discovery