DeepDetect: deep learning of peptide detectability enhanced by peptide digestibility

In tandem mass spectrometry-based proteomics, proteins are digested by specific protease(s) into peptides, but generally only a fraction of theoretical peptides can be detected. To explore the characteristics of proteotypic peptides, we have developed a series of methods for peptide digestibility and detectability prediction, and this is the latest one of them. Specifically, we propose here a bidirectional long short-term memory (BiLSTM)-based algorithm, named DeepDetect, for the prediction of peptide detectability enhanced by peptide digestibility. Compared with existing algorithms, DeepDetect is featured by its improved accuracy of prediction and wide applicability to commonly used proteases covering Trypsin, ArgC, Chymotrypsin, GluC, LysC, AspN, LysN, and LysargiNase. On eleven independent test data sets, DeepDetect achieved AUCs of 0.858~0.957, superior to the state-of-the-art algorithms, AP3 and PepFormer, which either used conventional machine learning methods or disregarded the digestibility feature. As an application, the predicted peptide detectability was utilized to re-rank and validate the peptide-spectrum matches, and showed the potential to promote the sensitivity of peptide identification.