Meta-Learning Enables Complex Cluster-Specific Few-Shot Binding Affinity Prediction for Protein-Protein Interactions.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in drug discovery

Background:

• Accurate prediction of protein-protein interaction (PPI) binding affinities is crucial for understanding biological processes and for developing targeted peptide- or protein-based drugs.
• Existing geometric models often struggle with adaptability to novel protein complex clusters, limiting their application in predicting binding affinities for unseen interactions.

Purpose of the Study:

• To introduce MCGLPPI++, a meta-learning framework designed to enhance the adaptability of pretrained geometric models for predicting PPI binding affinities in unseen protein complex clusters.
• To improve the robustness and accuracy of binding affinity predictions, particularly for challenging biological systems like T-cell receptor (TCR)-peptide-MHC (pMHC) interactions.

Main Methods:

• Developed MCGLPPI++, a meta-learning framework incorporating three novel training sample cluster splitting patterns based on protein interaction interfaces to inject prior intersample distribution knowledge.
• Integrated an independent energy component within MCGLPPI++ to explicitly model interface nonbonded interaction energies, which are critical for PPI strengths.
• Curated a new dataset featuring a challenging test cluster of TCR-pMHC interactions for validation.

Main Results:

• Geometric models enhanced with the MCGLPPI++ framework demonstrated significantly more robust binding affinity predictions after fine-tuning on a few samples from the novel TCR-pMHC cluster.
• The enhanced models outperformed their vanilla counterparts, showcasing improved adaptability and predictive power on unseen protein complex data.
• The explicit modeling of interface nonbonded interaction energies contributed to the improved prediction accuracy.

Conclusions:

• MCGLPPI++ effectively improves the adaptability of geometric models for predicting PPI binding affinities in novel protein complex clusters.
• The framework's ability to generalize to new interaction types, as demonstrated with TCR-pMHC complexes, highlights its potential for accelerating drug discovery and biological research.
• The integration of meta-learning strategies and explicit energy modeling offers a promising direction for advancing computational approaches to PPI prediction.