PreMemMoRF: A pretraining-fine-tuning framework for predicting membrane molecular recognition features

Area of Science:

• Biochemistry
• Computational Biology
• Molecular Biology

Background:

• Membrane molecular recognition features (MemMoRFs) are intrinsically disordered regions (IDRs) crucial for membrane dynamics.
• Dysregulation of MemMoRFs is linked to neurodegenerative diseases and viral infections.
• Limited annotations hinder accurate computational prediction of MemMoRFs.

Purpose of the Study:

• To develop a robust computational framework for large-scale MemMoRF identification.
• To address the challenge of data scarcity in MemMoRF annotation.
• To improve the accuracy of predicting MemMoRFs in proteins.

Main Methods:

• Developed PreMemMoRF, a deep learning framework utilizing transfer learning.
• Pre-trained the model on linear interacting peptides (LIPs) with similar conformational transitions.
• Fine-tuned the model on MemMoRF datasets to capture generalizable binding-related sequence features.

Main Results:

• PreMemMoRF significantly outperforms existing predictors across multiple metrics.
• Demonstrated robust performance on both transmembrane and membrane-associated proteins.
• Showed consistent performance in short linear motif prediction, indicating cross-task generalizability.
• Proteome-wide analysis revealed systematic differences in predicted scores across transmembrane regions, consistent with physicochemical constraints.

Conclusions:

• PreMemMoRF is a reliable computational framework for large-scale MemMoRF identification.
• The study highlights the effectiveness of transfer learning in addressing data scarcity for biological sequence prediction.
• Findings provide insights into the distribution and properties of MemMoRFs within membrane protein structures.