Improved protein model quality assessment by integrating sequential and pairwise features using deep learning.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Accurate protein model quality estimation is crucial for evaluating, selecting, and refining protein structures in the absence of experimental data.
• Despite advancements in deep learning, current protein quality assessment (QA) methods, especially for local quality on challenging targets, remain suboptimal.

Purpose of the Study:

• To introduce ResNetQA, a new single-model-based method for both local and global protein model quality assessment.
• To enhance the accuracy of protein quality estimation using integrated sequential and pairwise features.

Main Methods:

• ResNetQA employs a deep neural network integrating 1D and 2D convolutional residual neural networks (ResNet).
• A 2D ResNet module processes pairwise features (e.g., distance maps, co-evolution data).
• A 1D ResNet predicts quality from sequential features and pooled pairwise information.

Main Results:

• ResNetQA demonstrated superior performance over state-of-the-art methods on CASP12 and CASP13 datasets.
• Ablation studies confirmed the significant contribution of the 2D ResNet module and pairwise features to improved QA accuracy.

Conclusions:

• ResNetQA offers a powerful new approach for protein model quality assessment.
• The integration of diverse features and a hybrid ResNet architecture is key to its enhanced performance.