DeepSSInter: Protein-protein contact prediction with a structure-aware protein language model

Area of Science:

• Computational biology
• Structural biology
• Bioinformatics

Background:

• Predicting protein-protein interactions is crucial for understanding protein complex structure and function.
• Deep learning has advanced protein contact prediction, but current methods using Multiple Sequence Alignments (MSA) face limitations in accuracy, speed, and efficiency.

Purpose of the Study:

• To develop a novel deep learning method, DeepSSInter, for predicting inter-protein residue-residue contacts.
• To overcome the limitations of MSA-based methods by utilizing single-sequence and structure-aware protein language models (PLMs).

Main Methods:

• Employed a transformer-powered deep learning architecture (DeepSSInter).
• Integrated intra-protein distance and graph representations with ESM2 and SaProt PLMs to generate structure-aware features.
• Utilized ResNet Inception and Triangle-aware modules for contact map prediction.

Main Results:

• DeepSSInter demonstrated significant improvements in both accuracy and speed for predicting inter-protein contacts compared to state-of-the-art methods.
• The model showed strong performance on both homo- and hetero-dimeric protein complexes.
• Incorporating DeepSSInter's predicted contacts enhanced protein docking performance.

Conclusions:

• DeepSSInter offers a more accurate, faster, and computationally efficient approach for predicting protein-protein contacts.
• The method advances the study of protein complex structure and function by providing reliable contact predictions.
• The DeepSSInter model is publicly available for further research and application.