Combined Physics- and Machine-Learning-Based Method to Identify Druggable Binding Sites Using SILCS-Hotspots.

Area of Science:

• Computational chemistry and structural biology
• Machine learning in drug discovery
• Protein-ligand interactions

Background:

• Identifying druggable binding sites, especially cryptic or allosteric ones, is crucial but challenging.
• Existing methods may miss sites not evident in static protein structures.
• The Site-Identification by Ligand Competitive Saturation (SILCS) method uses simulations to capture protein flexibility and identify potential binding pockets.

Purpose of the Study:

• To develop a machine learning (ML) model for ranking protein binding sites identified by SILCS.
• To predict the likelihood of these sites accommodating drug-like molecules.
• To advance the discovery of novel therapeutic targets and drug candidates.

Main Methods:

• Utilized all-atom molecular simulations (SILCS) to model protein flexibility and solute distributions.
• Developed an ML model to rank SILCS-identified 'Hotspots' based on druggability.
• Validated the ML model on an independent set of enzymes and receptors.

Main Results:

• The ML model successfully recalled 67% and 89% of known ligand binding sites within the top 10 and 20 ranked Hotspots, respectively.
• The model's Decision Function proved effective in predicting binding sites and their druggability for new targets.
• Identified buried binding pockets missed by experimental structures.

Conclusions:

• The ML-enhanced SILCS approach significantly improves the identification of orthosteric and allosteric binding sites.
• This method facilitates the discovery of drug-like molecules targeting previously inaccessible protein sites.
• The developed tools represent a key advancement for ligand discovery and optimization in drug development.