Large-scale proteomics analyses based on various mass spectrometers and laboratories have generated a tremendous amount of publicly available datasets, creating opportunities for developing novel peptide identification methods using AI-based spectrum prediction. However, multiple sources of variability in the spectral data pose considerable challenges to machine learning techniques that rely heavily on peptide fragmentation patterns to train neural networks for accurate prediction models. In particular,  inconsistencies in collision energies (CE) during peptide fragmentation were found to be an important factor in the variability of mass spectrometry data, resulting in suboptimal AI models of peptide fragmentation. To improve these AI models, we develop mass spectrometer-specific CE models to systematically investigate the relationship between charge-to-mass ratio (m/z) and applied CE of peptide spectra from 218 datasets consisting of 4731 raw data files collected in the NIST Human Peptide Spectral Library. Peptide spectrum similarity was used to compare collision energies between similar spectra from different mass spectrometers. Preliminary results indicate that around 5% to 20% of peptide ions show notably different fragmentation patterns at instruments' same nominal energy settings. We were able to deduce that CE inconsistencies come from factors including incorrect or undetermined ion charge states, the use of different CE conversion functions between instruments, and different fragmentation techniques. To further correct the CE of these spectra to train AI models of fragmentation, we created functions that recalibrate the stated CE of the spectra.   This study shows that mass spectrometry data quality is currently affected by non-negligible CE inconsistencies, significantly impacting the performance of AI models. It is a first step in understanding the effects of CE inconsistencies observed in proteomics data and can serve as a template for calibrating similar data. The approach may lead to effective calibration methods to establish large-scale, diverse, consistent MS/MS spectral training datasets across different instruments and laboratories.

In-depth analysis of collision energy inconsistencies in large-scale proteomics datasets for improving the spectrum prediction accuracy of deep learning models

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.