Inferring molecular inhibition potency with AlphaFold predicted structures.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning in drug discovery

Background:

• In silico drug discovery methods often struggle with predicting interactions for novel or unassessed protein targets.
• Ligand-based approaches are limited when target structures are unknown or targets are novel.

Purpose of the Study:

• To develop a machine learning model that integrates 3D structural information from predicted protein structures (AlphaFold 2) to improve drug-target interaction prediction.
• To enable accurate predictions for previously untested molecules and protein targets, including orphan targets.

Main Methods:

• Extraction of 3D structural protein fingerprints from AlphaFold 2 predicted structures.
• Integration of protein structural fingerprints with ligand structural data.
• Training a single machine learning model to capture proteo-chemometric relationships.
• Validation using a dataset of 144 Human G-protein Coupled Receptors (GPCRs) and over 140,000 inhibition constants (Ki).

Main Results:

• The proposed approach demonstrates performance comparable to state-of-the-art ligand-based methods.
• In a separate test set, the model correctly predicted interactions for 73% of targets, with explained variance > 0.50 in 22% of cases.
• Models built using predicted protein structures showed no statistical difference compared to those using experimentally determined structures.

Conclusions:

• The proteo-chemometric approach effectively leverages predicted protein structural data for drug screening.
• This method provides a simple, scalable way to predict protein-molecule interactions, even for orphan targets.
• The integration of structural data enhances the predictive power of machine learning models in drug discovery.