DeepDigest: Prediction of Protein Proteolytic Digestion with Deep Learning.

Proteolytic digestion of proteins by one or more proteases is a key step in shotgun proteomics, in which the proteolytic products, i.e., peptides, are taken as the surrogates of their parent proteins for further qualitative or quantitative analysis. The proteases generally cleave proteins at specific amino acid residue sites, but digestion is hardly complete (wide existence of missed cleavage sites). Therefore, it would be of great help to improve the prior experimental design and the posterior data analysis if the digestion behaviors of proteases can be accurately modeled and predicted. At present, systematic studies about the commonly used proteases in proteomics are insufficient, and there is a lack of easy-to-use tools to predict the cleavage sites of different proteases. Here, we propose a novel sequence-based deep learning algorithm-DeepDigest, which integrates convolutional neural networks and long short-term memory networks for protein digestion prediction. DeepDigest can predict the cleavage probability of each potential cleavage site on the protein sequences for eight popular proteases including trypsin, ArgC, chymotrypsin, GluC, LysC, AspN, LysN, and LysargiNase. We compared DeepDigest with three traditional machine learning algorithms, i.e., logistic regression, random forest, and support vector machine. On the eight training data sets, the 10-fold cross-validation accuracies (AUCs) of DeepDigest were 0.956-0.982, significantly higher than those of the three traditional algorithms. On the 11 independent test data sets, DeepDigest achieved AUCs between 0.849 and 0.978, outperforming the other traditional algorithms in most cases. Transfer learning then further improved the prediction accuracy. Besides, some interesting characteristics of different proteases were revealed and discussed. Ultimately, as an application, we used DeepDigest to predict the digestibilities of peptides and demonstrated that peptide digestibility is an informative new feature to discriminate between correct and incorrect peptide identifications.