pLMMoRF: A Web Server That Accurately Predicts Membrane-interacting Molecular Recognition Features by Employing a...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Protein-lipid interactions are vital for cellular functions.
• Intrinsically disordered regions (IDRs) often mediate these interactions, sometimes gaining structure upon binding.
• Experimentally identified lipid-interacting IDRs are termed membrane molecular recognition features (MemMoRFs).

Purpose of the Study:

• To develop and validate an accurate, fast, sequence-based predictor for MemMoRFs, named pLMMoRF.
• To support the identification of MemMoRFs, reducing reliance on time-consuming experimental methods.
• To analyze MemMoRF predictions across the human proteome for enhanced understanding of membrane-coupled proteins.

Main Methods:

• Collected and curated a dataset of experimentally annotated MemMoRFs.
• Utilized protein language models (pLMs) and deep convolutional neural networks for prediction.
• Applied feature selection to optimize pLM outputs for a compact neural network, selecting the Ankh-based model.
• Evaluated pLMMoRF against existing predictors on low-similarity datasets.

Main Results:

• pLMMoRF demonstrated higher accuracy than the current state-of-the-art predictor, CoMemMoRFPred.
• The predictor exhibits a relatively small computational footprint due to its compact network size and efficient processing.
• MemMoRF predictions for the entire human proteome were generated and made publicly available.

Conclusions:

• pLMMoRF is an accurate and efficient tool for predicting MemMoRFs from protein sequences.
• The selection of key embedding features is crucial for enhancing predictive performance and reducing computational costs.
• The publicly available predictions facilitate deeper insights into the roles of membrane-associated proteins.