Assessing interaction recovery of predicted protein-ligand poses

TL;DR

This paper highlights the importance of evaluating protein-ligand interaction fingerprints in pose prediction models, revealing that neglecting these interactions can lead to overestimated performance, especially in cofolding models.

Contribution

It introduces the assessment of interaction recovery as a crucial metric for evaluating protein-ligand pose prediction accuracy, emphasizing its importance over traditional pose accuracy measures.

Findings

Ignoring interaction fingerprints can overestimate model performance.

Recent cofolding models often fail to recapitulate key interactions.

Interaction-based evaluation provides more realistic performance assessment.

Abstract

The field of protein-ligand pose prediction has seen significant advances in recent years, with machine learning-based methods now being commonly used in lieu of classical docking methods or even to predict all-atom protein-ligand complex structures. Most contemporary studies focus on the accuracy and physical plausibility of ligand placement to determine pose quality, often neglecting a direct assessment of the interactions observed with the protein. In this work, we demonstrate that ignoring protein-ligand interaction fingerprints can lead to overestimation of model performance, most notably in recent protein-ligand cofolding models which often fail to recapitulate key interactions.