Protein language model-embedded geometric graphs power inter-protein contact prediction

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Accurate prediction of contacting residue pairs is crucial for understanding protein-protein interactions and their structural characterization.
• Despite recent advancements, improving the accuracy of inter-protein contact prediction remains an active area of research.

Purpose of the Study:

• To introduce PLMGraph-Inter, a novel deep learning method designed to enhance the accuracy of inter-protein contact prediction.
• To demonstrate the effectiveness of PLMGraph-Inter compared to existing state-of-the-art methods and its complementary role with AlphaFold-Multimer.
• To showcase the utility of PLMGraph-Inter's predictions in improving protein-protein docking for complex structure prediction.

Main Methods:

• Utilizing rotationally and translationally invariant geometric graphs derived from interacting protein structures.
• Integrating multiple protein language models (PLMs) transformed by graph encoders, including geometric vector perceptrons and residual networks.
• Employing dimensional hybrid residual blocks for successive transformations to predict inter-protein contacts.

Main Results:

• PLMGraph-Inter significantly outperforms five leading inter-protein contact prediction methods (DeepHomo, GLINTER, CDPred, DeepHomo2, DRN-1D2D_Inter) on multiple test sets.
• Predictions from PLMGraph-Inter were shown to complement the results generated by AlphaFold-Multimer.
• Using PLMGraph-Inter's predicted contacts as constraints dramatically improved the performance of protein-protein docking for predicting complex structures.

Conclusions:

• PLMGraph-Inter represents a significant advancement in inter-protein contact prediction accuracy.
• The method offers valuable complementary information to existing tools like AlphaFold-Multimer.
• PLMGraph-Inter has the potential to substantially improve the prediction of protein complex structures through enhanced protein-protein docking.