Knowledge-based fragment binding prediction.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Target-based drug discovery requires screening numerous compounds.
• Fragment-based drug design reduces chemical search space but faces limitations in predicting protein-fragment interactions.
• Existing experimental and physics-based computational methods have drawbacks in accuracy and applicability.

Purpose of the Study:

• To present FragFEATURE, a novel machine learning approach for predicting small molecule fragments preferred by protein structures.
• To leverage high-resolution structural data for data-driven fragment binding prediction.
• To improve the efficiency and accuracy of identifying starting points for drug design.

Main Methods:

• Creation of a knowledge base of protein structural environments annotated with bound small molecule substructures (fragments).
• Utilizing FragFEATURE to compare target protein structural environments against the knowledge base.
• Retrieving statistically preferred fragments by merging information across diverse ligands with shared substructures.

Main Results:

• FragFEATURE achieved an average precision of 74% and recall of 82% in rediscovering fragments corresponding to known bound ligands.
• The method successfully identified high-scoring fragments that are substructures of known inhibitors for many protein targets.
• Demonstrated the capability of a data-driven approach to accurately predict fragment binding preferences.

Conclusions:

• FragFEATURE offers a powerful data-driven alternative to traditional methods for predicting protein-fragment interactions.
• The predicted fragments can be utilized as starting points for fragment-based drug design.
• FragFEATURE provides valuable criteria for refining compound libraries for experimental or computational screening.