Physical-aware model accuracy estimation for protein complex using deep learning method.

Area of Science:

• Structural Biology
• Computational Biology
• Deep Learning

Background:

• The focus of protein structure prediction has shifted from monomers to complexes following AlphaFold2's success.
• Accurate quality estimation for protein complex models, independent of prediction methods, is crucial.
• Existing methods require improvement for evaluating the accuracy of predicted protein complex structures.

Purpose of the Study:

• To develop a novel physical-aware deep learning method for evaluating the residue-wise quality of protein complex models.
• To improve the accuracy of quality assessment for protein complex structures, particularly for flexible protein interactions.

Main Methods:

• Developed DeepUMQA-PA, a physical-aware deep learning approach for residue-wise quality evaluation of protein complex models.
• Constructed residue-based contact area and orientation features using Voronoi tessellation to represent physical interactions.
• Integrated geometry-based features, protein language model embeddings, and knowledge-based potentials into a fused network (graph neural network and ResNet).

Main Results:

• DeepUMQA-PA outperformed the state-of-the-art DeepUMQA3 on the CASP15 test set by 3.69% (Pearson) and 3.49% (Spearman).
• Achieved significant improvements for nanobody-antigen evaluation: 16.8% (Pearson) and 15.5% (Spearman).
• Demonstrated superior Mean Absolute Error (MAE) scores compared to AlphaFold-Multimer and AlphaFold3 self-assessment on a majority of targets.

Conclusions:

• Physical-aware features, incorporating contact area and orientation, effectively capture sequence-structure-quality relationships in proteins.
• DeepUMQA-PA shows particular promise for evaluating the quality of flexible protein complexes.
• The developed method provides a valuable tool for assessing the accuracy of predicted protein complex structures.