Mass spectrometry reads the proteome one fragmentation pattern at a time — but for decades no one could predict what that pattern would look like. Today’s digest is about the deep-learning models that now can, and how that unlocks faster, deeper proteomes. Zhou et al.’s pDeep predicted peptide spectra ‘with >0.9 median Pearson correlation coefficients.’ Gessulat et al.’s Prosit produced predictions that ’exceed the quality of the experimental data,’ cutting false discovery rates ‘>10×.’ Tiwary et al.’s DeepMass:Prism reached accuracy ‘within the uncertainty of measurement.’ Yang et al.’s DeepDIA built in-silico spectral libraries ‘directly from protein sequence databases.’ Demichev et al.’s DIA-NN brought neural networks to high-throughput DIA. And Zeng et al.’s AlphaPeptDeep unified retention time, ion mobility and fragment prediction in one framework. Reading the proteome, faster.
