ContrastQA: A label-guided graph contrastive learning-based approach for protein complex structure quality assessment

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Protein complex accuracy estimation lags behind monomer estimation.
• Integrating interface-specific and global structural data is crucial for accurate protein complex quality assessment.
• Existing methods struggle to effectively combine local and global structural information.

Purpose of the Study:

• To introduce ContrastQA, the first Estimation of Model Accuracy (EMA) framework for protein complexes.
• To develop a novel approach integrating label-guided graph contrastive learning and geometric graph neural networks.
• To improve the accuracy of assessing protein complex model quality.

Main Methods:

• Developed ContrastQA, a framework utilizing label-guided graph contrastive learning based on interface quality.
• Integrated a geometric graph neural network to model global structural features.
• Evaluated ContrastQA on the CASP16 dataset using TMscore and GDT-TS metrics.

Main Results:

• ContrastQA achieved superior performance in protein complex EMA.
• Ranking losses on TMscore and GDT-TS were 0.123 and 0.116, respectively.
• Outperformed the second-best method by 10.9% (TMscore) and 8.7% (GDT-TS).

Conclusions:

• ContrastQA effectively integrates local and global structural information for accurate model quality estimation.
• The label-guided graph contrastive learning module is particularly effective in selecting high-quality models.
• The graph contrastive learning framework shows promise as a pre-training strategy for protein structure representations.