Adaptive gradient scaling: integrating Adam and landscape modification for protein structure prediction.

Area of Science:

• Computational biology
• Structural biology
• Machine learning

Background:

• Protein structure prediction is a critical scientific challenge with significant applications in drug discovery and biotechnology.
• Experimental structure determination is costly and time-consuming, making computational methods essential.
• Machine learning has advanced protein structure prediction but struggles with optimizing complex energy landscapes.

Purpose of the Study:

• To develop novel optimization methods for protein structure prediction.
• To address limitations of current machine learning approaches in navigating complex energy landscapes.
• To improve the robustness and performance of protein folding prediction algorithms.

Main Methods:

• Integration of the Landscape Modification (LM) method with the Adam optimizer for OpenFold.
• Introduction of a gradient scaling mechanism based on energy landscape transformations.
• Development of LM SA, incorporating simulated annealing for enhanced convergence and exploration.

Main Results:

• LM and LM SA demonstrated superior performance compared to standard Adam across multiple evaluation metrics.
• The novel methods exhibited faster convergence and better generalization capabilities.
• Performance improvements were particularly noted on proteins outside the training dataset.

Conclusions:

• Integrating landscape-aware gradient scaling into optimizers enhances computational optimization.
• The developed LM and LM SA methods offer improved prediction performance for complex protein folding problems.
• This research advances the field of computational structural biology and optimization algorithms.