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

We have used a transformer neural network for de novo peptide sequencing, formulating this as a translation task from peaks in mass spectra to sequences of amino acids. Our model, called Casanovo, achieves state-of-the-art performance across a series of tasks, including analysis of immunopeptidomics data, metaproteomics data, and an exploration of the dark human proteome. Additionally, the scoring function learned by Casanovo can be used to accurately match between mass spectra and peptide sequences, significantly outperforming traditional sequence database searching approaches.