(361d) Pareto Optimization to Accelerate Multi-Property Virtual Screening

Molecular discovery is a constrained multi-objective optimization problem that aims to identify one or more molecules that balance multiple, often competing, properties. One approach to molecular design is to evaluate the properties of all molecules in a virtual library (“virtual screening”), and methods based on active learning and Bayesian optimization have previously been proposed to accelerate virtual screening [1, 2]. However, these methods typically consider single-objective optimization only or, in the case of multi-objective optimization, combine properties of interest into a scalarized objective [3]. A promising alternative to scalarization is Pareto optimization, which can elucidate the trade-offs between objectives and does not rely on assumptions about the relative importance of objectives. In the context of molecular discovery, Pareto optimization identifies the molecules that optimally balance multiple desired properties, revealing the best that is possible in the objective space.

We combine Pareto optimization and Bayesian optimization methods to efficiently search a library of millions of molecules and identify those that optimally balance multiple objectives. Using graph neural networks as the surrogate model architecture, we apply this methodology to identify (1) dual inhibitors which minimize docking scores to two targets and (2) selective small molecule inhibitors which minimize docking scores to an on-target and maximize those to an off-target. We compare optimization performance between Pareto and scalarization acquisition functions and find that multi-objective acquisition functions outperform scalarization. This improvement is greater for the selective inhibitor task due to stronger competition between the objectives. Finally, we find that encouraging molecular and functional diversity during acquisition does not improve the hypervolume of acquired molecules but does increase the diversity of the acquired set.

[1] Gentile, F.; Agrawal, V.; Hsing, M.; Ton, A.-T.; Ban, F.; Norinder, U.; Gleave, M. E.; Cherkasov, A. Deep Docking: A Deep Learning Platform for Augmentation of Structure Based Drug Discovery. ACS Cent. Sci. 2020, 6 (6), 939–949.

[2] Graff, D. E.; Shakhnovich, E. I.; Coley, C. W. Accelerating High-Throughput Virtual Screening Through Molecular Pool-Based Active Learning. Chem. Sci. 2021, 12 (22), 7866–7881.

[3] Mehta, S.; Goel, M.; Priyakumar, U. D. MO-MEMES: A Method for Accelerating Virtual Screening Using Multi-Objective Bayesian Optimization. Frontiers in Medicine 2022, 9, 916481.