MTPrompt-PTM: A Multi-Task Method for Post-Translational Modification Prediction Using Prompt Tuning on a...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Post-translational modifications (PTMs) are crucial for protein function and cellular processes.
• Current computational models for PTM site prediction often rely on limited local sequence information.
• Existing protein language models (PLMs) for PTM prediction typically lack structural context and are trained as single-task models.

Purpose of the Study:

• To develop an advanced computational framework for accurate multi-type post-translational modification site prediction.
• To leverage structure-aware protein language models (S-PLMs) and multi-task learning for enhanced PTM prediction.
• To improve the efficiency and generalizability of PTM prediction models.

Main Methods:

• Developed MTPrompt-PTM, a multi-task PTM prediction framework using prompt tuning on a structure-aware protein language model (S-PLM).
• Implemented a multi-task learning strategy to predict multiple PTM types simultaneously using shared features.
• Incorporated a knowledge distillation strategy to optimize multi-task training.

Main Results:

• MTPrompt-PTM demonstrated superior performance over state-of-the-art tools for 13 different types of PTM sites.
• The framework effectively integrates structural information and multi-task learning for improved PTM site prediction.
• Results highlight the benefits of shared feature extraction and task-specific heads in multi-task learning.

Conclusions:

• MTPrompt-PTM represents a significant advancement in computational PTM site prediction.
• The integration of structural context and multi-task learning is key to improving prediction accuracy.
• This framework offers a more efficient and generalizable approach for predicting diverse PTM sites.