Work

Trust in Machine Learning-Aided Docking: A Systematic Review of Good Practices for Chemically Reliable Predictions

Year: 2026
Type: review
Abstract: The massive growth of chemical libraries presents both an opportunity and a challenge in large-scale virtual screening. Machine learning (ML) algorithms have successfully accelerated this process with surrogate docking models that learn patterns between molecular representations and docking scores for ranking undocked molecules. Deep learning (DL), which represents the majority of ML-boosted docking approaches, has a sophisticated capability of finding nonlinear, complex relationships in voluminous data. However, DL algorithms operate as black boxes, face difficulties in generalization and raise concerns on whether correct predictions are, in fact, made for causal relationships between chemical structure and biological activity. Candidate molecules cannot be confidently tested if predictions are not understandable in human terms. The lack of understanding of relevant chemical properties impairs compound discovery, optimization and novelty, reinforcing the high-risk and high-cost scenario of drug design. To offer insights for expanding the recent breakthroughs achieved with the docking/ML integration, this review included 57 ML-aided docking papers from January 2020 to August 2026. We assessed how ML predictions are generated, refined and chemically interpreted to reliably reduce docking simulations and ideally increase experimental hit rates by data-driven and expert-driven decisions on compound selection. On this matter, the proper implementation of eXplainable Artificial Intelligence (XAI) techniques is informative to elucidate if and how trustworthy are predicted hits from a chemical perspective, assuring a safer and more effective decision-making process for finding true hits. Nonetheless, only four (7.0%) out of 57 papers applied XAI techniques, mostly involving feature attribution methods, but not the state-of-the-art contrastive explanations method. Therefore, the implementation of XAI in ML-boosted docking is still scarce and needs formalization to avoid misinterpreting results. We review the various opportunities and limitations for integrating ML-aided docking with XAI techniques alongside complementary trust-enhancing strategies present in the included studies, such as sampling strategies, target diversity, ensemble and consensus docking, active learning, ensemble models, physicochemical and pharmacophore-based filtering, multiobjective optimization of molecular properties, confidence metrics of predefined acceptable errors, and expected model improvement along iterations. Given the broad applications of XAI across scientific domains and proof-of-concept studies in computational chemistry, we believe benchmarking and combining current and novel trust-enhancing strategies in ML-aided docking can go beyond finding initial hits. It supports both the statistical assessment including XAI in predictions and the mechanistic understanding of chemically relevant drug candidates for transparent and credible model and drug development.
Source: ACS Omega
Language: English
Subfield: Drug Discovery
Open Access status: gold
Estimated APC paid: $1,935
OpenAlex ID: W7237250076
FWCI:
Cites: 98
Cited by:
Related to: