Sequence-to-sequence translation from mass spectra to peptides with a transformer model

Area of Science:

• Proteomics
• Bioinformatics
• Machine Learning

Background:

• Accurate peptide identification from tandem mass spectra is crucial for proteomics.
• Database-dependent methods fail to identify novel or unexpected peptides.
• De novo peptide sequencing is essential for antibody sequencing, immunopeptidomics, and metaproteomics.

Purpose of the Study:

• To develop a novel machine learning model for de novo peptide sequencing.
• To address the challenge of modeling irregular tandem mass spectral data.
• To improve the analysis of complex proteomic datasets.

Main Methods:

• Developed Casanovo, a machine learning model utilizing a transformer neural network architecture.
• Trained Casanovo on a large dataset of 30 million labeled spectra.
• Fine-tuned a version of Casanovo for non-enzymatic peptides.

Main Results:

• Casanovo accurately translates tandem mass spectra peak sequences into amino acid sequences.
• The model significantly outperforms state-of-the-art methods on a cross-species benchmark dataset.
• Demonstrated improved performance in immunopeptidomics and metaproteomics analyses.

Conclusions:

• Casanovo offers a powerful solution for de novo peptide sequencing.
• The model enhances the exploration of the 'dark proteome' by identifying previously uncharacterized peptides.
• Casanovo is a valuable tool for advancing various fields within proteomics research.