ProPept-MT: A Multi-Task Learning Model for Peptide Feature Prediction.

Area of Science:

• Proteomics
• Computational Biology
• Analytical Chemistry

Background:

• Data-independent acquisition (DIA) offers improved reproducibility in quantitative proteomics over data-dependent acquisition (DDA).
• Current DIA analysis relies on DDA-derived spectral libraries, limiting proteome coverage and requiring extensive time.
• There is a need for efficient methods to generate comprehensive spectral libraries for DIA proteomics.

Purpose of the Study:

• To develop a novel deep learning model, ProPept-MT, for predicting key peptide features.
• To enable the construction of high-quality 4D in silico spectral libraries for DIA proteomics.
• To enhance the efficiency and scope of DIA-based quantitative proteomics.

Main Methods:

• ProPept-MT, a multi-task deep learning model utilizing multi-head attention and BiLSTM for feature extraction.
• Nash-MTL for gradient coordination to optimize prediction performance.
• Integration of predicted retention time (RT), ion mobility (IM), and ion intensity for 4D library generation.

Main Results:

• ProPept-MT achieved high prediction accuracy: 99.9% Pearson correlation coefficient (PCC) for RT, 96.0% median dot product (DP) for fragment ion intensity, and 99.3% PCC for IM.
• The model demonstrated efficacy in predicting features for both unmodified and phosphorylated peptides.
• Successful application in constructing 4D DIA in silico libraries, validated on a benchmark dataset.

Conclusions:

• ProPept-MT significantly improves the prediction of peptide features essential for DIA proteomics.
• The developed 4D in silico libraries enhance proteome coverage and streamline DIA data analysis.
• ProPept-MT represents a valuable tool for advancing quantitative proteomics research using DIA methods.