Benchmarking Deep Networks for Predicting Residue-Specific Quality of Individual Protein Models in CASP11.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in structural biology

Background:

• Accurate protein model quality assessment is crucial for understanding protein function before experimental structures are available.
• Single-model methods predict protein model quality using only one model, enabling residue-specific quality evaluation.
• Existing methods require rigorous evaluation, especially in distinguishing reliable from unreliable protein model regions.

Purpose of the Study:

• To develop and evaluate novel single-model methods for protein quality assessment.
• To compare the performance of stacked denoising autoencoders (SdAs) and support vector machines (SVMs) in this task.
• To identify the most effective computational strategies for residue-specific quality prediction.

Main Methods:

• Development of four novel single-model methods: Wang_deep_1, Wang_deep_2, Wang_deep_3 (based on SdAs), and Wang_SVM (based on SVMs).
• Evaluation using Pearson's correlation coefficients and Receiver Operating Characteristic (ROC) analysis.
• Comparison against six other methods from the CASP11 competition at global and local levels.

Main Results:

• The developed methods demonstrated superior performance in residue-specific quality assessment compared to most CASP11 competitors, as shown by ROC analysis.
• The SdA-based method, Wang_deep_1, achieved the highest accuracy (0.77) among individual methods.
• Ensemble methods (Wang_deep_2, Wang_deep_3), integrating SdAs and SVMs, slightly outperformed Wang_deep_1 in ROC analysis, highlighting the benefit of combining deep networks with SVMs.

Conclusions:

• Novel computational methods based on SdAs and SVMs significantly enhance protein model quality assessment.
• Integrating deep learning networks with SVMs proves effective for improving residue-specific quality prediction accuracy.
• The developed methods offer a promising advancement for evaluating protein models in structural bioinformatics.