When Protein Structure Embedding Meets Large Language Models.

Area of Science:

• Bioinformatics and Structural Biology
• Computational Biology
• Machine Learning in Life Sciences

Background:

• Protein structure analysis is crucial for drug discovery, disease diagnosis, and evolutionary studies.
• Current protein classification methods often rely on sequence-based embeddings, neglecting vital 3D structural information.
• Existing approaches lack a unified strategy combining structural and sequence features for efficient protein analysis.

Purpose of the Study:

• To develop a novel method for creating numerical protein embeddings that integrate 3D structural information with sequence data.
• To enhance the performance of protein classification and function prediction by synergistically combining diverse feature sets.
• To address the limitations of sequence-only embeddings and Euclidean space assumptions in representing complex protein data.

Main Methods:

• Leveraging 3D protein structure information through contact maps to design Euclidean space embeddings.
• Integrating structure-based embeddings with features from large language models (LLMs) and traditional feature engineering.
• Utilizing benchmark datasets such as PDB Bind and STCRDAB for experimental validation.

Main Results:

• The proposed method demonstrates superior performance in supervised protein analysis and function prediction compared to existing approaches.
• Combined embeddings effectively capture both structural and sequential characteristics of proteins.
• Experimental results validate the efficacy of the novel embedding strategy on diverse protein datasets.

Conclusions:

• The novel embedding approach significantly advances protein classification and function prediction by incorporating 3D structural data.
• This method offers a more comprehensive representation of proteins, improving accuracy in bioinformatics applications.
• The findings pave the way for more sophisticated machine learning models in structural biology and drug discovery.