Fig. 4. Model performance with transfer learning on 21 PTMs from Proteome Tools.

a The accuracy of MS2 prediction with different numbers of peptides for transfer learning for each PTM. Each PTM is tested separately. “80% seqs” refers to using 80% of the identified modified sequences for transfer learning. b Overall accuracy without unmodified peptides from (a). Boxplots for n = 21 PCC90 values of 21 PTM types are drawn with the interquartile range within the boxes and the median as a horizontal line. The whiskers extend to 1.5 times the size of the interquartile range. c Transfer learning dramatically improves the MS2 prediction of the example peptide “AGPNASIISLKSDK-Biotin@K11” (tuned by 50 other peptides). d Comparisons of RT prediction for each PTM on pre-trained and transfer learning (by 50% of all the identified peptides) models, as well as DeepLC models. e Overall R² distribution without unmodified peptides from (d). Boxplots for n = 21 R² values of 21 PTM types are drawn with the interquartile range within the boxes and the median as a horizontal line. The whiskers extend to 1.5 times the size of the interquartile range. RT retention time, MS2 intensities of fragment spectra, PCC Pearson correlation coefficient, PTM post-translational modification.