Improving protein and protein interactions using pseudo-dimers derived from monomeric proteins

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Predicting protein-protein interactions (PPIs) in dimeric complexes is crucial but challenging.
• Current models like AlphaFold-Multimer and AlphaFold3 have limitations due to scarce dimer structural data.
• Acquiring protein dimer structures is expensive and time-consuming.

Purpose of the Study:

• Introduce a novel pre-training method, split and merge proxy (SMP), to leverage abundant monomeric protein data.
• Enhance the accuracy and generalization of protein complex prediction models.
• Develop a cost-effective strategy for large-scale pre-training of PPI prediction models.

Main Methods:

• SMP constructs pseudo-dimers by splitting monomer data into pseudo-receptors and pseudo-ligands.
• Models are pre-trained on a proxy task of merging pseudo-dimers and predicting pseudo-interactions.
• Pre-trained models are fine-tuned on real protein dimer datasets.

Main Results:

• SMP pre-training significantly improves accuracy and generalization across multiple benchmarks.
• Models fine-tuned with SMP surpass existing strong baselines in PPI prediction.
• SMP-enhanced models provide more accurate structure predictions than AlphaFold-Multimer and AlphaFold3 on CASP15 dimer targets.

Conclusions:

• SMP offers a scalable strategy for utilizing monomeric data to advance protein complex modeling.
• The method provides insights into the relationship between monomers and multimers.
• SMP represents a cost-effective approach to improving PPI prediction accuracy.