ABSTRACT: To create a dataset of PSMs that does not exhibit tryptic bias, we selected PSMs with a uniform distribution of amino acids at the C-terminal peptide positions from two datasets: MassIVE-KB [Wang2018] and PROSPECT [Shouman2022]. The MassIVE-KB dataset contains 30 million PSMs and consists entirely of data generated using trypsin; hence, only a small proportion of the MassIVE-KB peptides do not end in K or R, corresponding to those that appear at the C-terminus of the corresponding protein. The PROSPECT dataset is a collection of 61 million PSMs generated from synthetic peptides. To avoid overrepresentation of some peptides in this dataset, we randomly selected at most 100 PSMs per unique peptide, similar to the processing that was done during the creation of the MassIVE-KB dataset. This pre-selection step reduced the size of the PROSPECT dataset to 12.6 million PSMs. Finally, to create a non-enzymatic dataset containing 1 million peptides, we selected 50,000 PSMs for each canonical amino acid. These PSMs were selected at random from MassIVE-KB, then supplemented as necessary from PROSPECT to obtain the desired count (see Yilmaz et al. [Yilmaz2023] Supplementary Table S1). This dataset contained PSMs from 247,859 unique peptides, which were randomly split into training, validation and test sets in the ratio 80/10/10. FTP directory contains the mgf files corresponding to train, validation and test splits.
[Wang2018] M. Wang, J. Wang, J. Carver, B. Pullman, S. Won Cha, N. Bandeira, "Assembling the Community-Scale Discoverable Human Proteome", Cell Systems, Volume 7, Issue 4, 2018.
[Shouman2022] O. Shouman, W. Gabriel, V. Giurcoiu, V. Sternlicht, M. Wilhelm, "PROSPECT: Labeled Tandem Mass Spectrometry Dataset for Machine Learning in Proteomics", NeurIPS Datasets and Benchmarks, 2022
[Yilmaz2023] M. Yilmaz*, W. Fondrie*, W. Bittremieux, R. Nelson, V. Ananth, S. Oh, and W. Noble,"Sequence-to-sequence translation from mass spectra to peptides with a transformer model", bioRxiv, 2023
William Noble 
