Deciphering Membrane Proteins Through Deep Learning Models by Revealing Their Locale Within the Cell.

Area of Science:

• Proteomics
• Computational Biology
• Molecular Biology

Background:

• Membrane proteins are vital biomolecules with critical roles in cellular processes.
• Accurate prediction of membrane protein subcellular localization is essential for functional studies.
• Current prediction models perform poorly on membrane proteins, especially deep learning models.

Purpose of the Study:

• To develop and evaluate deep learning models for accurate membrane protein subcellular localization.
• To differentiate membrane proteins into three specific locations: plasma membrane, internal membrane, and organelle membrane.
• To improve the performance of prediction tools for membrane proteins in proteomics.

Main Methods:

• Utilized deep learning algorithms, specifically Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM).
• Employed a curated dataset of 3000 proteins from the MemLoci approach, with reduced redundancy.
• Incorporated pseudo amino acid composition (PseAAC) to extract sequence-based protein features.

Main Results:

• The Long Short-Term Memory (LSTM) model achieved a prediction accuracy of 83.4%.
• The Recurrent Neural Network (RNN) model achieved a prediction accuracy of 80.5%.
• The LSTM model demonstrated superior performance compared to the RNN model for membrane protein localization.

Conclusions:

• Deep learning models, particularly LSTM, show significant promise for accurate membrane protein localization.
• The developed approach enhances the capability of computational tools in proteomics.
• This work contributes to a better understanding of membrane protein functions through improved localization prediction.