Reliable prediction of protein-protein binding affinity changes upon mutations with Pythia-PPI

Area of Science:

• Computational Biology
• Biochemistry
• Genomics

Background:

• Protein-protein interactions (PPIs) are vital for biological processes.
• Predicting how mutations affect PPI binding affinity is crucial for genetic variation studies and protein engineering.
• Limited experimental data hinders accurate prediction of binding affinity changes.

Purpose of the Study:

• To develop an accurate machine learning model for predicting protein-protein binding affinity changes caused by mutations.
• To overcome data limitations in predicting binding affinity through multitask learning and self-distillation.
• To create a valuable tool for analyzing the fitness landscape of protein-protein interactions.

Main Methods:

• Employed multitask learning, incorporating mutation stability prediction.
• Utilized self-distillation with a large, self-generated dataset of mutation effects.
• Developed the Pythia-PPI model and web server.

Main Results:

• Achieved state-of-the-art accuracy on the SKEMPI dataset (Pearson's correlation increased from 0.6447 to 0.7850).
• Significantly improved predictions on a viral-receptor dataset (Pearson's correlation increased from 0.3654 to 0.6050).
• Experimental validation identified high-affinity mutations for an antibody-receptor complex, with one mutant showing a 2-fold binding affinity increase.

Conclusions:

• Pythia-PPI effectively overcomes data limitations for accurate binding affinity prediction.
• The model demonstrates significant improvements over existing methods.
• Pythia-PPI is a valuable tool for analyzing protein interaction fitness landscapes and has been made accessible via a web server.