DEEPCON: protein contact prediction using dilated convolutional neural networks with dropout

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Protein contact prediction is crucial for understanding protein structure and function.
• Advancements in neural networks and high-throughput sequencing offer new avenues for solving this problem.

Purpose of the Study:

• To investigate the optimal design of deep convolutional neural networks (ConvNets) for protein contact prediction.
• To evaluate the performance of novel ConvNet architectures against state-of-the-art methods.

Main Methods:

• Designed and trained various ConvNet architectures using publicly available datasets.
• Employed deep learning techniques such as wide residual networks, dropouts, and dilated convolutions.
• Compared performance metrics with existing state-of-the-art protein contact prediction methods.

Main Results:

• Proposed ConvNet architectures achieve significantly higher precision in predicting protein contacts.
• Demonstrated up to 15% higher precision in medium and long-range contact prediction on benchmark datasets.
• Outperformed existing methods, including the ensembled DNCON2 method, by up to 4.8% in precision.

Conclusions:

• Deep convolutional neural networks offer a powerful approach to protein contact prediction.
• The developed DEEPCON method provides a significant improvement in prediction accuracy.
• The DEEPCON tool is publicly available for further research and application.