AF2Complex predicts direct physical interactions in multimeric proteins with deep learning.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Understanding protein-protein interactions is crucial for deciphering biological systems.
• AlphaFold2 (AF2) has achieved remarkable accuracy in predicting individual protein structures.

Purpose of the Study:

• To adapt existing AlphaFold2 neural network models for predicting the structures of multimeric protein complexes without retraining.
• To develop a novel method, AF2Complex, for predicting protein-protein interactions and complex structures.

Main Methods:

• Utilized AlphaFold2 neural network models without retraining for multimeric complex structure prediction.
• Developed AF2Complex, a method that does not require paired multiple sequence alignments.
• Introduced new metrics for predicting direct protein-protein interactions.

Main Results:

• AF2Complex achieves higher accuracy than some protein-protein docking strategies.
• Demonstrated significant improvement over AF-Multimer for multimeric protein structures.
• Validated AF2Complex on benchmark datasets and the E. coli proteome, presenting high-confidence models for specific biological systems.

Conclusions:

• The same AlphaFold2 models can be effectively repurposed for predicting protein complex structures.
• AF2Complex offers a more accurate and efficient approach to modeling protein-protein interactions compared to existing methods.
• The method provides valuable insights into the assembly of complex biological systems.