DeepLoc 2.1: multi-label membrane protein type prediction using protein language models

Area of Science:

• Bioinformatics
• Computational Biology
• Proteomics

Background:

• Protein subcellular localization is crucial for cellular function.
• Accurate prediction of protein localization aids in understanding cellular mechanisms.
• Existing tools may lack comprehensive classification of membrane protein types.

Purpose of the Study:

• To introduce DeepLoc 2.1, an upgraded web server for protein subcellular localization and sorting signal prediction.
• To extend DeepLoc 2.1's functionality to classify proteins into specific membrane protein types: Transmembrane, Peripheral, Lipid-anchored, and Soluble.
• To evaluate the performance of DeepLoc 2.1 against established tools using a rigorously curated dataset.

Main Methods:

• Utilizing pre-trained transformer-based protein language models.
• Implementing a three-stage architecture for sequence-based, multi-label predictions.
• Conducting comparative evaluations on a large, homology-partitioned test set of eukaryotic protein sequences.

Main Results:

• DeepLoc 2.1 demonstrates state-of-the-art performance in predicting protein subcellular localization and membrane protein types.
• The server accurately classifies proteins into Transmembrane, Peripheral, Lipid-anchored, and Soluble categories.
• Comparative evaluations show DeepLoc 2.1 outperforms existing prediction tools.

Conclusions:

• DeepLoc 2.1 represents a significant advancement in predicting protein subcellular localization and classifying membrane protein subtypes.
• The use of transformer-based models and a multi-stage architecture contributes to its superior performance.
• The updated web server provides a valuable resource for researchers in proteomics and cell biology.