DDA-BERT: end-to-end training for data-dependent acquisition mass spectrometry-based proteomics.

Area of Science:

• Proteomics
• Bioinformatics
• Artificial Intelligence

Background:

• Peptide-spectrum match (PSM) rescoring is vital for accurate peptide identification in data-dependent acquisition (DDA) proteomics.
• Current methods often use shallow classifiers, limiting the potential of deep learning features for PSM ranking and confidence estimation.

Purpose of the Study:

• To introduce DDA-BERT, a transformer-based deep learning model for end-to-end PSM rescoring in DDA proteomics.
• To evaluate DDA-BERT's performance against existing tools across various species and datasets.

Main Methods:

• Developed DDA-BERT, a transformer-based deep learning model.
• Trained the model on approximately 271 million PSMs from 11 species.
• Benchmarked DDA-BERT against existing rescoring tools on human, yeast, Drosophila, and Arabidopsis datasets, including HLA immunopeptidomics data.

Main Results:

• DDA-BERT demonstrated consistent performance improvements across species-specific benchmarks.
• Achieved significant increases in peptide identifications, ranging from 2.24% to 269.35% depending on the dataset.
• Showcased high sensitivity in trace-level proteomics and enhanced identifications in HLA immunopeptidomics.

Conclusions:

• DDA-BERT represents a novel, high-performing approach to DDA proteomics PSM rescoring.
• The model offers a scalable, AI-driven foundation for peptide identification, though it requires GPU computing and extensive training data.
• This work advances the field by integrating end-to-end deep learning for more accurate and sensitive peptide identification.