Enhancing Protein Structure Learning Using a Size-Guided Conditional Mixture-of-Experts.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in biology

Background:

• Deep learning on protein structures is advancing rapidly, with structure-based methods showing high performance.
• Existing models often overlook protein size, leading to size-specific biases and reduced accuracy.
• Protein structure dictates protein function, making accurate structure-based prediction essential.

Purpose of the Study:

• To develop a deep learning approach that effectively incorporates protein size information.
• To mitigate biases in protein property prediction models that arise from neglecting protein size.
• To enhance the accuracy and generalizability of deep learning models for protein structure analysis.

Main Methods:

• Proposed a protein size-guided conditional mixture-of-experts model.
• The model adaptively activates sub-networks based on protein size and network features.
• Developed a framework for protein property prediction integrating this novel approach.

Main Results:

• The proposed method improved performance across eight benchmark tasks.
• Achieved better results with different protein representations and dataset splits.
• Demonstrated seamless integration with existing deep learning models.

Conclusions:

• Protein size is a vital, often overlooked, source of prior knowledge in deep learning for protein structures.
• The size-guided mixture-of-experts approach effectively captures universal and size-specific protein characteristics.
• This method offers a flexible and powerful way to enhance protein property prediction accuracy.