Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CompassDock: Comprehensive Accurate Assessment Approach for Deep Learning-Based Molecular Docking in Inference and Fine-Tuning

Ahmet Sarigun, Vedran Franke, Bora Uyar, Altuna Akalin

TL;DR

This work tackles the limitation of relying on RMSD as a sole measure of docking success by revealing substantial PCB noise in the PDBBind dataset and proposing CompassDock, a two-mode framework that integrates a PCB-focused Compass module (PoseCheck + AA-Score) with a diffusion-based docking method (DiffDock). In inference mode, Compass analyzes binding affinity, ligand strain energy, and steric clashes to provide a richer evaluation of docked poses; in fine-tuning mode, a LAN-MSE-based Compass Score regularizer biases training toward physically and biologically favorable conformations. The results show that fine-tuning with Compass Score yields limited RMSD improvements but enhances PCB-favorability, while standard fine-tuning without PCB features can degrade such properties. Overall, CompassDock offers a practical pathway to more physics- and bioactivity-aware docking models with publicly available code for reproducibility and extension.

Abstract

Datasets used for molecular docking, such as PDBBind, contain technical variability - they are noisy. Although the origins of the noise have been discussed, a comprehensive analysis of the physical, chemical, and bioactivity characteristics of the datasets is still lacking. To address this gap, we introduce the Comprehensive Accurate Assessment (Compass). Compass integrates two key components: PoseCheck, which examines ligand strain energy, protein-ligand steric clashes, and interactions, and AA-Score, a new empirical scoring function for calculating binding affinity energy. Together, these form a unified workflow that assesses both the physical/chemical properties and bioactivity favorability of ligands and protein-ligand interactions. Our analysis of the PDBBind dataset using Compass reveals substantial noise in the ground truth data. Additionally, we propose CompassDock, which incorporates the Compass module with DiffDock, the state-of-the-art deep learning-based molecular docking method, to enable accurate assessment of docked ligands during inference. Finally, we present a new paradigm for enhancing molecular docking model performance by fine-tuning with Compass Scores, which encompass binding affinity energy, strain energy, and the number of steric clashes identified by Compass. Our results show that, while fine-tuning without Compass improves the percentage of docked poses with RMSD < 2Å, it leads to a decrease in physical/chemical and bioactivity favorability. In contrast, fine-tuning with Compass shows a limited improvement in RMSD < 2Å but enhances the physical/chemical and bioactivity favorability of the ligand conformation. The source code is available publicly at https://github.com/BIMSBbioinfo/CompassDock.

CompassDock: Comprehensive Accurate Assessment Approach for Deep Learning-Based Molecular Docking in Inference and Fine-Tuning

TL;DR

This work tackles the limitation of relying on RMSD as a sole measure of docking success by revealing substantial PCB noise in the PDBBind dataset and proposing CompassDock, a two-mode framework that integrates a PCB-focused Compass module (PoseCheck + AA-Score) with a diffusion-based docking method (DiffDock). In inference mode, Compass analyzes binding affinity, ligand strain energy, and steric clashes to provide a richer evaluation of docked poses; in fine-tuning mode, a LAN-MSE-based Compass Score regularizer biases training toward physically and biologically favorable conformations. The results show that fine-tuning with Compass Score yields limited RMSD improvements but enhances PCB-favorability, while standard fine-tuning without PCB features can degrade such properties. Overall, CompassDock offers a practical pathway to more physics- and bioactivity-aware docking models with publicly available code for reproducibility and extension.

Abstract

Datasets used for molecular docking, such as PDBBind, contain technical variability - they are noisy. Although the origins of the noise have been discussed, a comprehensive analysis of the physical, chemical, and bioactivity characteristics of the datasets is still lacking. To address this gap, we introduce the Comprehensive Accurate Assessment (Compass). Compass integrates two key components: PoseCheck, which examines ligand strain energy, protein-ligand steric clashes, and interactions, and AA-Score, a new empirical scoring function for calculating binding affinity energy. Together, these form a unified workflow that assesses both the physical/chemical properties and bioactivity favorability of ligands and protein-ligand interactions. Our analysis of the PDBBind dataset using Compass reveals substantial noise in the ground truth data. Additionally, we propose CompassDock, which incorporates the Compass module with DiffDock, the state-of-the-art deep learning-based molecular docking method, to enable accurate assessment of docked ligands during inference. Finally, we present a new paradigm for enhancing molecular docking model performance by fine-tuning with Compass Scores, which encompass binding affinity energy, strain energy, and the number of steric clashes identified by Compass. Our results show that, while fine-tuning without Compass improves the percentage of docked poses with RMSD < 2Å, it leads to a decrease in physical/chemical and bioactivity favorability. In contrast, fine-tuning with Compass shows a limited improvement in RMSD < 2Å but enhances the physical/chemical and bioactivity favorability of the ligand conformation. The source code is available publicly at https://github.com/BIMSBbioinfo/CompassDock.
Paper Structure (39 sections, 24 equations, 6 figures, 1 table)

This paper contains 39 sections, 24 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Motivation
  4. CompassDock
  5. DiffDock
  6. PoseCheck
  7. AA-Score
  8. Inference Mode
  9. Fine-Tuning Mode
  10. Loss Regularizer with Compass Score for Fine-Tuning
  11. Log Absolute Normalized - Mean Square Error (LAN-MSE)
  12. Compass Score
  13. Loss Function
  14. Experiments
  15. PCB Quality Analysis
  16. ...and 24 more sections

Figures (6)

  • Figure 1: (A) Inference Mode: Provides comprehensive PCB feature quality assessment during DL-Based Molecular Docking Inference runtime using integrated Compass workflow. (B) Fine-Tuning Mode: Introduces a new fine-tuning method for DL-Based Molecular Docking with a PCB-based approach, utilizing the Integrated Compass Score.
  • Figure 2: Example of RMSD Fails. Green: Ground Truth 1a46 Ligand's Conformation; Yellow: Inference Prediction of 1a46 Ligand's Conformation; Gray: Ground Truth & Inference Prediction of 1a46 Protein Pocket
  • Figure 3: Distribution of PCB properties and quality within the PDBBind ground truth dataset. (A): Distribution of binding affinity across the PDBBind dataset. (B): Distribution of the number of steric clashes between protein-ligand pairs in PDBBind. (C): Distribution of strain energy in ligands from the PDBBind dataset.
  • Figure 4: Benchmarking DiffDock models: pre-trained model (DiffDock-L), fine-tuned model (DiffDock-L+FineTuning), and fine-tuning with Compass Score (CompassDock (ours)), based on RMSD and favorable PCB properties.
  • Figure 5: Comparison of PDB ID: 5c28 Ligand Conformation predictions during fine-tuning. Green Structures: Ground Truth Conformation of Ligand; Cyan Structures: Ground Truth Structures of Binding Pocket. Ground Truth Binding Affinity, Strain Energy, and Number of Clashes indicated as green in the first column
  • ...and 1 more figures