Fig. 6: DIA transfer learning for discovery of modified peptides.

a, A custom deep learning model was trained for every experiment using the identifications from the DIA search engine. Evosep liquid chromatography illustration created with BioRender. b, Multiple properties were optimized, resulting in smaller and better matching spectral libraries. c, Observed and predicted retention times for dimethylated precursors before transfer learning. d, DIA transfer learning for the retention times of dimethylated peptides. During training by stochastic gradient descent, a 20% validation set of precursors was held out to mitigate overfitting and ensure generalization to the peptide space of interest. e, Retention times after transfer learning. f, Comparison of the number of unique peptides identified with the pretrained base model (default) to the transfer learned model after retention time and MS2 transfer learning. g, Distribution of absolute retention time errors for the pretrained base model (default), the nonlinear calibration within alphaDIA and after transfer learning. h, Comparison of spectral correlation before and after MS2 transfer learning. i, Number of unique observed modifications by type.