A Comparison of Different Compound Representations for Drug Sensitivity Prediction

Abstract

Deep learning (DL) has become increasingly popular in the field of drug discovery. A large variety of end-to-end DL methods for chemical compounds have recently been proposed in the literature, potentially eliminating the need for expert-designed compound representations. This study aims to determine which types of representations and DL algorithms are most suitable for the specific problem of anti-cancer drug response prediction. A newly developed chemoinformatics package called DeepMol was used to benchmark 12 different compound representation methods on 5 anti-cancer drug sensitivity datasets. We found that DL models that are able to learn compound representations directly from SMILES strings or molecular graphs can perform as well as or even better than models trained on molecular fingerprints, even on smaller datasets. We also conclude that popular molecular fingerprints might not always be the best choice and less well-known fingerprints might be worth exploring in future drug response prediction studies.