Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Assessing interaction recovery of predicted protein-ligand poses

David Errington, Constantin Schneider, Cédric Bouysset, Frédéric A. Dreyer

TL;DR

It is demonstrated 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.

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.

Assessing interaction recovery of predicted protein-ligand poses

TL;DR

It is demonstrated 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.

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.
Paper Structure (13 sections, 5 figures, 1 table)

This paper contains 13 sections, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Protein-ligand interaction fingerprint
  4. Classical docking algorithms
  5. ML docking algorithms
  6. Protein-ligand cofolding
  7. Data and Metrics
  8. Results
  9. PoseBusters benchmark
  10. Interaction recovery rates
  11. Recovery of different interaction types
  12. Discussion
  13. Correlation betweeen PLIF recovery and RMSD

Figures (5)

  • Figure 1: Left: Two-dimensional representation of the ligand EZO and its four interactions with the crystal structure 6M2B. Basic residues are shown in blue and residues containing a sulfur atom are shown in yellow. Right: Docked poses generated with GOLD, DiffDock-L and RosettaFold-AllAtom showing the calculated interactions for each model, with the ground truth ligand in grey.
  • Figure 2: The ratio of predicted protein-ligand complex structures for each model passing checks on ligand positioning (RMSD$\leq$2Å), physicality (PoseBuster-valid) and interaction recovery (PLIF-valid).
  • Figure 3: Recovery of protein-ligand interaction fingerprint for each model. The distribution of PLIF recovery among poses that pass the RMSD and PoseBuster test are shown in dashed and dotted lines.
  • Figure 4: Ratio to the ground truth of calculated and correctly recovered (recall) interactions shown separately for each interaction types.
  • Figure 5: PLIF recovery rate and RMSD, highlighting data points which are PoseBuster-valid. Note that we use a modified definition of PB-validity that excludes ligand RMSD. The red line indicates a ligand RMSD of 2Å.