High-Coverage Four-Dimensional Data-Independent Acquisition Proteomics and Phosphoproteomics Enabled by Deep Learning-Driven Multidimensional Predictions

Four-dimensional (4D) data-independentacquisition (DIA)-basedproteomics is a promising technology. However, its full performanceis restricted by the time-consuming building and limited coverageof a project-specific experimental library. Herein, we developed aversatile multifunctional deep learning model Deep4D based on self-attentionthat could predict the collisional cross section, retention time,fragment ion intensity, and charge state with high accuracies forboth the unmodified and phosphorylated peptides and thus establishedthe complete workflows for high-coverage 4D DIA proteomics and phosphoproteomicsbased on multidimensional predictions. A 4D predicted library containing similar to 2 million peptides was established that could realize experimentallibrary-free DIA analysis, and 33% more proteins were identified thanusing an experimental library of single-shot measurement in the exampleof HeLa cells. These results show the great values of the convenienthigh-coverage 4D DIA proteomics methods.

Automated summary

4D DIA-based proteomics is a promising technology, but it is limited by the time-consuming building of a project-specific experimental library. In this study, a versatile deep learning model called Deep4D was developed to predict multiple parameters for both unmodified and phosphorylated peptides, enabling high-coverage 4D DIA proteomics and phosphoproteomics workflows based on multidimensional predictions. A 4D predicted library containing millions of peptides was established, allowing experimental library-free DIA analysis and identifying more proteins compared to a single-shot measurement with an experimental library in the example of HeLa cells.