PASSer

PASSer predicts and ranks protein pockets by their likelihood of being allosteric sites to support identification of targets for allosteric modulation.

Key Features:

• Learning to Rank (LTR) approach: Employs a Learning to Rank model to prioritize protein pockets according to their relevance to known allosteric sites.
• Performance metrics: Achieved F1 scores of 0.662 on the Allosteric Database (ASD) and 0.608 on CASBench, with Matthews correlation coefficients of 0.645 and 0.589, respectively.
• Validation across datasets: Trained and validated on ASD and CASBench, ranking true allosteric pockets within the top three positions for 83.6% of ASD test proteins and 80.5% of CASBench test proteins.
• Ensemble modeling of structural features: Integrates ensemble learning approaches that capture physical properties and topological information relevant to allosteric site prediction.

Scientific Applications:

• Allosteric site identification for drug discovery: Identifies potential allosteric pockets to aid the design of drugs that modulate protein activity via allosteric sites.

Methodology:

Uses a Learning to Rank framework integrating ensemble learning methods, including eXtreme gradient boosting and graph convolutional neural networks, to model physical properties and topological information for allosteric site prediction.