Sampling and Ranking of Protein Conformations Using Machine Learning Techniques Do Not Improve the Quality of Rigid...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biochemistry

Background:

• Rigid docking is a primary method for predicting protein-protein interactions when experimental structures are unavailable.
• Protein structures in unbound (Apo) forms may differ substantially from their bound (Holo) states, impacting docking accuracy.
• Machine learning (ML) conformational sampling aims to generate functionally relevant protein structures closer to Holo forms.

Purpose of the Study:

• To evaluate the effectiveness of ML-based conformational sampling in improving rigid protein-protein docking.
• To assess the performance of various scoring functions in prioritizing ML-generated conformations.
• To identify limitations in current ML-enhanced rigid docking workflows.

Main Methods:

• Conformation sampling of protein subunits in 30 complexes from the PINDER dataset using two ML techniques.
• Evaluation of docking performance with physics-based, data-based, and ML-based scoring functions.
• Comparison of ML-generated conformations against experimental unbound (Apo) and bound (Holo) structures.

Main Results:

• ML-based conformational sampling infrequently generated structures closer to Holo conformations than Apo structures.
• Existing scoring functions failed to correctly prioritize or rank the generated conformations.
• The study identified critical limitations in current ML-enhanced rigid protein-protein docking.

Conclusions:

• ML-enhanced rigid docking workflows face significant challenges.
• Further research is needed to develop novel approaches for conformational generation and scoring in protein-protein docking.
• Current ML techniques do not consistently improve upon traditional rigid docking methods using Apo structures.