Combining pairwise structural similarity and deep learning interface contact prediction to estimate protein complex...

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning applications in protein structure prediction

Background:

• Accurate estimation of quaternary structural models (EMA) is vital for understanding protein function and interactions.
• Existing pairwise similarity methods for tertiary models are less effective for quaternary models, especially with low-quality, similar structures.
• A gap exists in reliable methods for assessing the quality of protein complex and assembly models.

Approach:

• Developed MULTICOM_qa, a hybrid method integrating a pairwise similarity score (PSS) with an interface contact probability score (ICPS).
• Utilized deep learning for inter-chain contact prediction to enhance the accuracy of the ICPS component.
• Blindly validated MULTICOM_qa in the Critical Assessment of Techniques for Protein Structure Prediction (CASP15), evaluating 24 predictors.

Key Points:

• MULTICOM_qa achieved first place in estimating global accuracy for assembly models at CASP15.
• Demonstrated strong performance with an average correlation coefficient of 0.66 between predicted and true model quality scores.
• Identified key factors influencing EMA, including target difficulty and model quality distribution.

Conclusions:

• The hybrid approach combining multi-model (PSS) and single-model (ICPS) methods offers a promising strategy for estimating protein complex accuracy.
• MULTICOM_qa effectively selects high-quality models for most targets, improving the reliability of structural predictions.
• The study highlights the potential of deep learning in advancing the field of protein structure quality assessment.