Using deep learning and large protein language models to predict protein-membrane interfaces of peripheral membrane...

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Biology

Background:

• Protein-membrane interactions are vital, with abnormalities linked to diseases.
• Identifying membrane-binding domains in peripheral membrane proteins (PMPs) is challenging.
• Current methods for analyzing protein-membrane interfaces can be time-consuming and require structural data.

Purpose of the Study:

• To assess the efficacy of natural language processing (NLP) and protein language models (pLMs) in predicting membrane-interacting amino acids for PMPs.
• To establish a faster and potentially more accessible method for protein-membrane interface analysis.

Main Methods:

• Utilized experimental data and protein embeddings from two pLMs (ProtTrans and ESM).
• Trained classifier models using these embeddings.
• Evaluated the performance of NLP and pLM-based prediction against existing methods.

Main Results:

• Demonstrated the first proof-of-concept for using deep learning and pLMs to predict protein-membrane interfaces.
• Achieved comparable accuracy to existing tools.
• Significantly reduced prediction time and eliminated the need for 3D structural data.

Conclusions:

• NLP and pLMs show significant promise for predicting protein-membrane interactions.
• This approach offers a faster and more efficient alternative for studying PMPs.
• The developed method can advance research into diseases associated with abnormal protein-membrane attachment.