ProCeSa: Contrast-Enhanced Structure-Aware Network for Thermostability Prediction with Protein Language Models

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Bioinformatics
• Artificial Intelligence in Life Sciences

Background:

• Protein thermostability is critical for biological function and traditionally requires extensive experimental measurement.
• Deep learning, especially protein language models (PLMs), has advanced protein thermostability prediction.
• Integrating structural insights from PLM embeddings without atomic data remains a significant challenge.

Purpose of the Study:

• To introduce a novel model, ProCeSa, for enhanced protein thermostability prediction.
• To effectively combine sequence and structural information derived from PLMs.
• To overcome the limitations of existing methods in leveraging structural context from embeddings.

Main Methods:

• Developed the Protein Contrast-enhanced Structure-Aware (ProCeSa) model.
• Utilized a contrastive learning scheme guided by amino acid residue categories.
• Extracted integrated sequence and structural information from PLM embeddings.
• Evaluated performance on publicly available datasets for classification and regression tasks.

Main Results:

• ProCeSa demonstrated superior performance compared to state-of-the-art methods.
• The model achieved high accuracy in both protein thermostability classification and regression tasks.
• The approach successfully integrated structural information from PLM embeddings without needing atomic structural data.

Conclusions:

• ProCeSa offers a powerful new approach for predicting protein thermostability.
• The model's ability to leverage PLM-derived structural context enhances prediction accuracy.
• This method advances computational approaches in protein engineering and functional prediction.