Evaluating molecular fingerprint-based models of drug side effects against a statistical control.

ABSTRACT

There are many machine learning models that use molecular fingerprints of drugs to predict side effects. Characterizing their skill is necessary for understanding their usefulness in pharmaceutical development. Here, we analyze a statistical control of side effect prediction skill, develop a pipeline for benchmarking models, and evaluate how well existing models predict side effects identified in pharmaceutical documentation. We demonstrate that molecular fingerprints are useful for ranking drugs by their likelihood to cause a given side effect. However, the predictions for one or more drugs overall benefit only marginally from molecular fingerprints when ranking the likelihoods of many possible side effects, and display at most modest overall skill at identifying the side effects that do and do not occur.