BioStructNet: Structure-Based Network with Transfer Learning for Predicting Biocatalyst Functions.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Chemical Engineering

Background:

• Enzyme-substrate interactions are crucial for biological processes and industrial applications.
• Machine learning accelerates biocatalysis research but faces challenges with limited data for specific enzyme functions.
• Predicting enzyme activity, conversion efficiency, and stereoselectivity is vital for discovering novel biocatalysts.

Purpose of the Study:

• To develop BioStructNet, a structure-based deep learning network for predicting enzyme-substrate interactions.
• To integrate protein and ligand structural data for enhanced prediction accuracy.
• To address challenges posed by limited data in biocatalysis research using transfer learning.

Main Methods:

• Developed BioStructNet, a deep learning network integrating protein and ligand structural information.
• Implemented transfer learning by training a source model on a large dataset and fine-tuning on a specific dataset (CalB).
• Validated model performance using attention heat maps and molecular dynamics simulations.

Main Results:

• BioStructNet demonstrated enhanced predictive accuracy compared to other algorithms.
• Transfer learning significantly optimized prediction accuracy for small, function-specific datasets.
• Attention heat maps from BioStructNet aligned with molecular dynamics simulations, validating interaction predictions.

Conclusions:

• BioStructNet effectively captures enzyme-substrate interaction complexity using structural data.
• Transfer learning is a viable strategy to improve prediction accuracy with limited data.
• BioStructNet can accelerate the discovery of functional enzymes for industrial applications, especially with small datasets.