Deep learning-derived optimal annotation strategies to power the systematic mapping of peptide space

Rapid and reliable peptide identification techniques are essential for proteomics. High-resolution tandem mass spectrometry acquires a large amount of data through data-dependent acquisition (DDA) and data-independent acquisition (DIA), but traditional parsing methods are difficult to process efficiently. Improvement of peptide identification methods by deep learning provides new ideas for peptide sequence characterization: (i) combining MS/MS spectra prediction tools Prosit and pDeep with database searching to improve identification efficiency and accuracy; (ii) using deep neural networks for mass spectral feature extraction and MS/MS spectra clustering to achieve high-throughput protein coverage; (iii) combining graph theory with convolutional neural networks and other deep learning models to achieve powerful learning discrimination of MS data; (iv) computational tools such as TagGraph, Open-pFind and DBReducer based on sequence tags to improve the recall and precision of peptide identification. This study provides a more efficient and accurate solution for peptide identification and promotes the development of proteomics.