Assigning secondary structure in proteins using AI.

Area of Science:

• * Structural Biology
• * Bioinformatics
• * Computational Chemistry

Background:

• * Protein structure assignment is vital for understanding protein function and for secondary structure prediction.
• * Traditional methods face challenges with missing atomic data in protein structures.
• * Existing approaches include hydrogen bond analysis, geometric methods, and machine learning.

Purpose of the Study:

• * To develop a robust method for protein secondary structure element (SSE) assignment, particularly addressing challenges posed by missing atoms.
• * To implement a deep learning approach using convolutional neural networks (CNNs) for accurate SSE prediction.
• * To utilize only Cα coordinates for efficient and reliable structure assignment.

Main Methods:

• * Development of DLFSA, a multi-class classifier program utilizing CNNs for SSE assignment.
• * Implementation of a GPU-based parallel procedure for efficient extraction of protein fragments.
• * Training and testing the CNN model on protein fragments, using only Cα coordinates.

Main Results:

• * The DLFSA model achieved 88.1% training accuracy and 82.5% testing accuracy on protein fragments.
• * The method demonstrated successful application to full-length proteins.
• * Fragment-based studies confirmed the feasibility of deep learning for structure assignment.

Conclusions:

• * Deep learning, specifically CNNs, offers a powerful solution for protein secondary structure assignment.
• * The DLFSA method provides accurate assignments using minimal Cα coordinate data, even with missing atoms.
• * This approach enhances the reliability of structural data for downstream applications in structural biology and bioinformatics.