Comprehensive Summary
This article by Wen et al. discusses the training of DeepMVP, deep learning-based post-translational modification and variant-induced alteration prediction. It is a machine learning framework for discovering new post-translational modification (PTM) sites. To train the model, 241 mass spectrometry datasets were used, encompassing 397,524 PTM sites in 6 selected PTM types: phosphorylation, acetylation, ubiquitination, sumoylation, methylation and N-glycosylation. This large combined dataset was used to create PTMAtlas, a more comprehensive and “high-quality” database for PTM sites. To test the efficacy of their model, the researchers selected 2 cancer proteogenomic datasets with experimental evidence and DeepMVP did well in identifying the expected PTM-altering variants. DeepMVP is now publicly available. After inputting a protein sequence of 31 to 61 residues in length and selecting one of six possible PTM types, it will predict the location of a possible PTM site. DeepMVP was compared to existing tools and outperformed them in all of the six selected PTM types.
Outcomes and Implications
Genetic missense variants contribute to many diseases by altering PTM sites. The current knowledge on these sites is very limited, with most existing predictions centering around well-characterized kinases. DeepMVP is a significant improvement over past methods as it targets six main types of PTM sites rather than only phosphorylation mediated PTM sites found within known kinases. The researchers mention that DeepMVP can identify 7,713 PTM-altering pathogenic germline variants and 230,092 somatic mutations in cancer and has implications for development of therapeutics.