Application of AI to Bottom-up Proteomics: Highlights From ASMS 2018

At ASMS 2018, I was impressed with the capability demonstrated by three posters from a team at the University of Munich.

One described the improved database called ProteomicsDB, where “DB” is large—consisting of 600 million recorded spectra plus another 280 million predicted spectra supported by more than 10 billion fragment ions (Schmidt, T.; Gessulat, S. et al. “ProteomicsDB: A Source for High-Quality Spectral Libraries and an Auxiliary Tool for the Development of Targeted Assays,” Poster MP 356, ASMS 2018). It covers 142 human cell lines, 13 body fluids, and 63 tissues. DB can be interrogated to search experimental, synthetic, and predicted. Predicted ions (m/z and intensity) are based on deep learning neural network and include predicted retention time of tryptic peptides as well as mass spectra.

The second poster described the deep learning artificial intelligence (AI) for predicting peptide MS/MS with high accuracy. First, more than 0.5 million increased the database for the MS synthetic peptides. This improved the accuracy of the prediction significantly. Adding a collision energy calibration, the experimental spectra matched the predicted y and b ion m/z value with greatly reduced deviations in intensity. This allowed a reduction in the false detection rate (FDR) of 0.1%, which is about a tenfold reduction over other MS protein prediction/identification programs (Gessulat, S.; Schmidt, T.K. et al. “PROSIT: Deep Learning Enables Proteome Wide Prediction of Peptide Tandem Mass Spectra With High Accuracy,” Poster TP 354, ASMS 2018).

The information above is a base for proteome tools, which use a bottom-up scheme based on the mass spectra of more than a million synthetic peptides, including 21 different post-translational modifications (PTMs) to explore the structure of the human proteome. Synthetic peptides were 7–20 amino acids long. Spectra were compared pairwise with PTMs—modified and unmodified. Mass spectra were obtained using 11 different fragmentation modes. Since the retention behavior of the unmodified peptides was well-documented, the retention variations were attributed to the particular PTM. The posters showed good correlation between PTMs. Those studied included hydroxyproline, succinimde, ubiquitin, acetate, propionate, butyrate, phosphate, monomethyl, dimethyl, trimethyl, and more. The poster presented data showing how PTMs can change spectral similarity.

Another panel compared high-energy collision dissociation (HCD) and electron transfer dissociation (ETD) fragmentation. In the future, the team expects to offer a library of 1.3 million endogenous peptides (Zolg, D.P.; Wilhelm, M. et al. “Systematic Characterization of the Chromatographic and Mass Spectrometric Properties of 21 Post-Translational Modifications Using Synthetic Peptides,” Poster WP 619, ASMS 2018).

Robert L. Stevenson, Ph.D., is Editor Emeritus, American Laboratory/Labcompare; e-mail: [email protected]

Related Products

Comments