Benchmarking Active Learning Protocols for Ligand-Binding Affinity Prediction

Abstract

Active learning (AL) has become a powerful tool in computational drug discovery, enabling the identification of top binders from vast molecular libraries. To design a robust AL protocol, it is important to understand the influence of AL parameters, as well as the features of the data sets on the outcomes. We use four affinity data sets for different targets (TYK2, USP7, D2R, Mpro) to systematically evaluate the performance of machine learning models [Gaussian process (GP) model and Chemprop model], sample selection protocols, and the batch size based on metrics describing the overall predictive power of the model (R2, Spearman rank, root-mean-square error) as well as the accurate identification of top 2%/5% binders (Recall, F1 score). Both models have a comparable Recall of top binders on large data sets, but the GP model surpasses the Chemprop model when training data are sparse. A larger initial batch size, especially on diverse data sets, increased the Recall of both models as well as overall correlation metrics. However, for subsequent cycles, smaller batch sizes of 20 or 30 compounds proved to be desirable. Furthermore, adding artificial Gaussian noise to the data up to a certain threshold still allowed the model to identify clusters with top-scoring compounds. However, excessive noise (<1σ) did impact the model’s predictive and exploitative capabilities.

Optimal Molecular Design: Generative Active Learning Combining REINVENT with Absolute Binding Free Energy Simulations

“…We have shown in this article that AL can be an effective method to improve compound optimization for a particular target. It is important to point out how the term "active learning" is being used as we follow here the convention which has been adopted in at least part of the community (Crivelli-Decker et al, 2023;Gorantla et al, 2024;Gusev et al, 2023;Khalak et al, 2022;Knight et al, 2021;Konze et al, 2019;Mohr et al, 2022;Thompson et al, 2022). However, we have also pointed out in the Introduction that we are not principally interested in creating a surrogate model for the purpose of optimally finding new labels i.e., binding affinities.…”

“…In AL a, typically, computationally expensive method like docking or MD simulation is used to assign new labels e.g., a free energy of binding (also known as the binding affinity) or a docking score as proxy for the affinity. AL approaches that efficiently optimize compounds for binding affinity with RBFE (relative binding free energy) methods, (also referred to as FEP=free energy perturbation) have only appeared recently in the literature (Crivelli-Decker et al, 2023;de Oliveira et al, 2023;Gorantla et al, 2024;Gusev et al, 2023;Khalak et al, 2022;Knight et al, 2021;Konze et al, 2019;Mohr et al, 2022;Thompson et al, 2022) but have also been used to optimize an RBFE protocol itself (de Oliveira et al, 2023). Combining Generative AI with active learning (GAL) has only started very recently (Filella-Merce et al, 2023) including a proofof-concept study with peptides (Hernandez-Garcia et al, 2023) and an application with ABFE (absolute binding free energy) (Eckmann et al, 2024).…”