The Human Omnibus of Targetable Pockets

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Predicting ligand binding pockets is crucial for drug discovery, but a comprehensive human proteome-wide dataset is lacking.
• Existing pocket-finding methods have limitations and excel in specific use cases.
• Ensemble approaches combining multiple strategies offer potential for large-scale, diverse pocket prediction.

Purpose of the Study:

• To create a comprehensive, human proteome-wide dataset of predicted ligand binding pockets.
• To develop and validate a novel pocket scoring method for identifying druggable pockets.
• To provide a freely available resource for the research community.

Main Methods:

• Assembled the Human Omnibus of Targetable Pockets (HOTPocket) dataset using seven pocket-finding methods on PDB and AlphaFold2 structures.
• Applied a novel scoring method, hotpocketNN, to filter and curate over 2.4 million predicted pockets.
• Validated hotpocketNN against known pockets and benchmark datasets (Astex Diverse Set, PoseBusters).

Main Results:

• The HOTPocket dataset contains over 2.4 million predicted pockets across the human proteome.
• hotpocketNN successfully recovered known ligand binding pockets, including novel ones.
• hotpocketNN outperformed constituent methods like P2Rank and Fpocket in precision assessments.
• Identified druggable pockets in KRAS and the mu opioid receptor.

Conclusions:

• The HOTPocket dataset and hotpocketNN method represent a significant advancement in proteome-wide druggable pocket identification.
• hotpocketNN demonstrates superior performance and generalizability in predicting druggable pockets.
• These freely available resources will accelerate drug discovery efforts by providing comprehensive pocket information.