Protein language models can capture protein quaternary state

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Deep Learning in Proteomics

Background:

• Determining a protein's quaternary state (number of monomers in a functional unit) is vital for understanding protein function.
• Experimental methods for quaternary state determination are often challenging and labor-intensive.
• Existing computational tools typically rely on experimentally validated structural information.

Purpose of the Study:

• To investigate if protein sequence information alone, via deep learning embeddings, can predict protein quaternary states.
• To develop and evaluate a novel deep learning model for quaternary state prediction based solely on protein sequences.

Main Methods:

• Generated ESM-2 embeddings for a large dataset of proteins with known quaternary states (QSbio dataset).
• Trained a deep learning model, named QUEEN (QUaternary state prediction using dEEp learNing), for quaternary state classification.
• Assessed QUEEN's performance on a distinct set of protein folds (ECOD family level) to ensure generalization.

Main Results:

• QUEEN successfully distinguishes between monomers and multimers using only sequence-derived embeddings.
• The model achieves moderate success in predicting specific quaternary states, outperforming simple sequence similarity transfer.
• Performance is lower than structure-based methods but demonstrates the presence of quaternary state information within protein sequences.

Conclusions:

• QUEEN is the first model to leverage protein language model embeddings for quaternary state prediction, highlighting sequence-based approach limitations and strengths.
• The model offers a fast, structure-free alternative for quaternary state prediction, valuable for large-scale sequence analysis and specific protein investigations.
• A Colab implementation is available for wider accessibility and application.