Abstract
Active learning is an emerging paradigm used to help accelerating drug discovery, but most prior applications seek solely to optimize potency, whereas multiple properties influence a compound's utility as a drug candidate. We introduce a method for multiobjective ligand optimization, which is able to efficiently handle distinct molecular properties that are expensive to compute, such as binding affinities with respect to multiple protein targets. We validate this protocol retrospectively using docking scores, showing an improved retrieval of the top 0.04-0.4% binders from the dataset with our method compared to greedy acquisition, owing to a better distribution of the compute budget between different properties. Our results also suggest that fitting individual properties separately leads to a better rank correlation of the resulting predictions. This workflow addresses the needs of pharmaceutical research for improving the efficiency of hit-to-lead and lead optimization by considering binding to multiple targets. Our code is freely available on Github: https://github.com/MobleyLab/active-learning-notebooks/blob/main/MultiobjectiveAL.ipynb.