Flexible protein-protein docking with a multitrack iterative transformer.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Conventional protein-protein docking is slow due to extensive sampling and reranking, limiting high-throughput applications like structure-based virtual screening.
• Existing deep learning methods for protein docking are faster but often achieve low success rates and assume rigid docking, ignoring crucial binding-induced conformational changes.

Purpose of the Study:

• To develop a novel deep learning method, GeoDock, for accurate and efficient protein-protein docking that accounts for conformational flexibility.
• To enable high-throughput complex structure prediction and facilitate applications like virtual screening and the study of allosteric mechanisms.

Main Methods:

• GeoDock utilizes a multitrack iterative transformer network, taking only protein sequences and structures as input.
• The model is designed to be flexible at the residue level, predicting conformational changes upon protein binding.
• Unlike other deep learning models, GeoDock does not require multiple sequence alignments.

Main Results:

• GeoDock achieved a 43% top-1 success rate on the DIPS test set, outperforming other tested methods.
• After addressing training set contamination, the success rate on DIPS was 31%.
• On DB5.5 and antibody-antigen datasets, GeoDock surpassed other deep learning models but lagged behind conventional methods and AlphaFold-Multimer.

Conclusions:

• GeoDock offers a significant speed improvement for protein-protein docking, with inference times under 1 second per GPU.
• The architecture provides a foundation for capturing backbone flexibility during binding, although binding-induced conformational changes remain a challenge.
• GeoDock's speed makes it suitable for large-scale structure screening and virtual screening applications.