A Deep Learning Network Approach to ab initio Protein Secondary Structure Prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Bioinformatics

Background:

• Accurate protein secondary structure (SS) prediction is crucial for tertiary structure modeling, driven by rapid protein discovery.
• Current SS prediction methods have accuracy stagnating around 80%, prompting exploration of novel techniques.
• Neural networks have historically been vital for SS prediction, suggesting potential for deep learning advancements.

Purpose of the Study:

• To investigate the efficacy of deep learning techniques in improving protein secondary structure prediction accuracy.
• To develop and evaluate a novel deep learning-based SS predictor, termed DNSS.

Main Methods:

• Developed DNSS, a deep learning network architecture utilizing position-specific scoring matrices from PSI-BLAST.
• Employed graphical processing units (GPUs) and CUDA software for optimizing and training deep networks.
• Constructed a workflow of three separately trained deep networks for refined SS predictions.

Main Results:

• Achieved a Q3 accuracy of 80.7% and a Sov accuracy of 74.2% on an independent test dataset of 198 proteins.
• Demonstrated that the deep learning approach can achieve competitive and potentially superior SS prediction performance.

Conclusions:

• Deep learning offers a promising avenue to advance protein secondary structure prediction beyond current accuracy ceilings.
• The DNSS approach provides a robust framework for accurate SS prediction, contributing to tertiary structure modeling efforts.