Deep learning-based protein–ligand docking

Thesis

This thesis investigates deep learning (DL) approaches for docking small molecules into protein receptors, a central challenge in structure-based drug discovery. While DL has revolutionised protein structure prediction, DL-based docking methods have yet to achieve a comparable breakthrough. My work demonstrates how better evaluation, data, and modelling assumptions might close this gap.

First, I introduced PoseBusters, a test suite and an accompanying benchmark, both developed by me, that define chemical consistency and physical plausibility checks; protein-ligand complexes and structures that pass all tests are said to be “PB-valid”. Applying PoseBusters to DL-based and classical docking methods showed that, once physical and stereochemical validity was required, current DL methods did not outperform classical docking, and that a simple molecular mechanics energy minimisation step could correct failures, indicating that important docking-relevant physics has not been learnt by DL models.

I also showed that commonly used time-based train–test splits overestimate generalisation capabilities; stratification by receptor sequence, pocket similarity, ligand chemistry, and interaction patterns would give a more realistic view of model performance. Using PoseBusters to analyse widely used training and test sets revealed that the training data often contained systematic structural and stereochemical problems sharing some failure modes with the models, whereas curated benchmarks were substantially cleaner. However, training data issues alone could not explain the observed error rates.

Next, I mined the Protein Data Bank for pairs of similar ligands bound to the same protein pocket to assess the hypothesis that the structure of a smaller molecule can be used to guide the prediction of the structure of a larger elaborated molecule. The data show that binding-mode conservation statistically increases as the chemical similarity of the ligands increases; this relationship was particularly strong for elaborations, confirming the experimental basis for both fragment-based and drug discovery and providing substructure constraints.

Last, I present GuideDock, which augments the diffusion-based docking model DiffDock-L with physics- and knowledge-based guidance. By incorporating force-field terms alongside user-specified pocket, substructure, and interaction constraints, GuideDock enables pocket-specific and constrained docking while improving physical plausibility and accuracy under substructure constraints.

These studies provide a comprehensive assessment of DL-based protein–ligand docking, highlight the central role of robust metrics and carefully curated data, and show how diffusion guidance can inject physical priors and user knowledge into generative docking models, thus informing the development of a new class of protein–ligand docking and co-folding methods

Authors: Buttenschoen, M

• DOI: 10.5287/ora-jb49j6ee0
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: deep learning; docking
• Subjects: Deep learning; Docking