ISSEC: inferring contacts among protein secondary structure elements using deep object detection.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Protein folding relies on contacts between secondary structure elements (SSEs), crucial for tertiary structure determination.
• Existing methods for inferring inter-SSE contacts suffer from noise in predicted inter-residue contacts or inaccuracies in SSE boundary prediction.
• Accurate prediction of inter-SSE contacts is vital for advancing protein structure prediction.

Purpose of the Study:

• To develop an accurate computational approach for inferring contacts between protein secondary structure elements (SSEs).
• To overcome limitations of existing methods by leveraging deep learning for robust feature extraction and pattern recognition.

Main Methods:

• Utilized deep object detection techniques to identify characteristic rectangular patterns formed by contacting SSEs in inter-residue contact maps.
• Employed deep convolution to extract high-level features directly from inter-residue contact data, avoiding reliance on pre-defined SSEs.
• Developed a confidence scoring system for candidate regions and a greedy strategy to select non-overlapping regions for inferring inter-SSE contacts.

Main Results:

• The developed approach, ISSEC, accurately infers inter-SSE contacts by detecting specific patterns in predicted inter-residue contact maps.
• ISSEC effectively extracts relevant features using deep convolution, bypassing the need for explicit SSE boundary definition.
• The method demonstrated superior performance compared to state-of-the-art approaches in predicting inter-SSE contacts.

Conclusions:

• ISSEC significantly outperforms existing methods for predicting inter-SSE contacts.
• The successful application of ISSEC enhances the accuracy of both inter-residue contact prediction and overall protein tertiary structure prediction.