UNKAI: A Protein Functional Identity Prediction Model Based on ESM-C Latent Representations and the Attention...

Area of Science:

• Bioinformatics
• Computational Biology
• Protein Science

Background:

• Vast protein sequence data exists, but many proteins lack functional characterization.
• Protein language models (pLMs) are emerging tools for predicting protein structure and function.
• Existing methods for predicting enzymatic activity have limitations.

Purpose of the Study:

• To develop a deep learning method for predicting whether two proteins catalyze the same enzymatic reaction.
• To leverage latent representations from state-of-the-art protein language models.
• To improve the accuracy and interpretability of protein function prediction.

Main Methods:

• Utilized latent representations from ESM Cambrian (ESM-C), a protein language model.
• Developed a neural network architecture incorporating an attention mechanism.
• Compared the new method against sequence similarity and AlphaFold-based structural models.

Main Results:

• The deep learning method demonstrated superior performance in predicting shared enzymatic reactions compared to existing approaches.
• The model outperformed methods based on full-length sequence similarity and previous AlphaFold-based models.
• Attention weights analysis revealed the model's ability to identify biologically significant residues (e.g., catalytic sites).

Conclusions:

• Integrating protein language models with attention mechanisms enhances protein function prediction accuracy and interpretability.
• This approach reduces the need for manual feature engineering in predicting enzymatic activity.
• The study showcases a powerful new tool for functional annotation of uncharacterized proteins.