Advances in artificial intelligence (AI) algorithms and models have been driven largely by access to large, high quality datasets. The wealth of publicly available MS/MS mass spectra makes peptide mass spectra prediction a potent application for deep neural networks, which can recognize very complex patterns in peptide fragmentation. For example, the NIST tandem MS Human libraries alone contain 2M+ spectra from 951K unique peptide ions, many of which are high quality consensus spectra. The mass spectra of peptides are an essential means of identification of proteins and their modifications in complex biological samples, such as those found in the field of proteomics. AI spectrum prediction may greatly expand the ability of spectrum matching software to positively identify peptides and proteins in these samples. We establish best practices to produce a useful model, including neural network architecture, input embedding, training and validation protocol, and practical use for peptide identification. Our model takes as input an embedding of peptide sequence, charge, modifications, collision energy, and then outputs a predicted spectrum. The model is composed of multiple independently trained convolutional neural networks, which yield mean spectra with calculated uncertainties. Each network is trained on peptide reference spectra from the NIST human spectral library and then validated and tested on spectra from the NIST Chinese hamster spectral library, and other sources. Predicted spectra for the test and validation sets show high similarity to experimentally determined spectra. We also analyze the out-of-distribution behavior of our model’s predictions and find that it follows basic principles of the gas phase fragmentation of peptides.

AIomics: Comprehensive MS/MS peptide mass spectra prediction with uncertainty estimates using ensembles of neural networks

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.