In silico spectral libraries by deep learning facilitate data-independent acquisition proteomics

Area of Science:

• Proteomics
• Computational Biology
• Biotechnology

Background:

• Data-independent acquisition (DIA) is a powerful quantitative proteomics technique for large sample cohorts.
• Current DIA workflows rely on time-consuming data-dependent acquisition (DDA) to build spectral libraries, limiting identification scope.
• This dependence on DDA restricts DIA's ability to identify and quantify peptides beyond those detected in DDA experiments.

Purpose of the Study:

• To develop a deep learning-based method, DeepDIA, for generating in silico spectral libraries for DIA.
• To enable direct spectral library generation from protein sequence databases, bypassing DDA.
• To enhance the capabilities of DIA proteomics by overcoming DDA-dependent limitations.

Main Methods:

• DeepDIA utilizes deep learning models trained on proteomic data to predict peptide spectra.
• Instrument-specific models were developed for improved prediction accuracy.
• Peptide detectability prediction was integrated to build libraries directly from sequence databases.

Main Results:

• DeepDIA's in silico spectral libraries demonstrate quality comparable to experimental libraries.
• Instrument-specific DeepDIA models outperform global models in library generation.
• DeepDIA enhanced peptide and protein detection in human serum samples compared to DDA-based libraries.

Conclusions:

• DeepDIA offers a robust, AI-driven alternative to DDA for generating spectral libraries in DIA proteomics.
• This approach expands the scope of peptide and protein identification and quantification in DIA.
• DeepDIA represents a significant advancement in the DIA proteomics toolbox, enabling more comprehensive analyses.