Yannek Nowatzky

Non-targeted metabolomics holds great promise for advancing precision medicine and biomarker discovery. However, identifying compounds from tandem mass spectra remains a challenging task due to the incomplete nature of spectral reference libraries. Augmenting these libraries with simulated mass spectra can provide the necessary references to resolve unmatched spectra, but generating high-quality data is difficult. In this study, we present FIORA, an open-source graph neural network designed to simulate tandem mass spectra. Our main contribution lies in utilizing the molecular neighborhood of bonds to learn breaking patterns and derive fragment ion probabilities. FIORA not only surpasses state-of-the-art fragmentation algorithms, ICEBERG and CFM-ID, in prediction quality, but also facilitates the prediction of additional features, such as retention time and collision cross section. Utilizing GPU acceleration, FIORA enables rapid validation of putative compound annotations and large-scale expansion of spectral reference libraries with high-quality predictions.

Non-targeted metabolomics holds great promise for advancing precision medicine and facilitating the discovery of novel biomarkers. However, the identification of compounds from tandem mass spectra remains a non-trivial task due to the incomplete nature of spectral reference libraries. Augmenting these libraries with simulated mass spectra can provide the necessary reference to resolve unmatched mass spectra, but remains a difficult undertaking to this day. In this study, we introduce Fiora, an innovative open-source algorithm using graph neural networks to simulate tandem mass spectra in silico . Our objective is to improve fragment intensity prediction with an intricate graph model architecture that facilitates edge prediction, thereby modeling fragment ions as the result of singular bond breaks and their local molecular neighborhood. We evaluate the performance on test data from NIST (2017) and the curated MS-Dial spectral library, as well as compounds from the 2016 and 2022 CASMI challenges. Fiora not only surpasses state-of-the-art fragmentation algorithms, ICEBERG and CFM-ID, in terms of prediction quality, but also predicts additional features, such as retention time and collision cross section. In addition, Fiora demonstrates significant speed improvements through the use of GPUs. This enables rapid (re)scoring of putative compound identifications in non-targeted experiments and facilitates large-scale expansion of spectral reference libraries with accurate spectral predictions. Abstract Figure:

Motivation: Deep learning has moved to the forefront of tandem mass spectrometry-driven proteomics and authentic prediction for peptide fragmentation is more feasible than ever. Still, at this point spectral prediction is mainly used to validate database search results or for confined search spaces. Fully predicted spectral libraries have not yet been efficiently adapted to large search space problems that often occur in metaproteomics or proteogenomics. Results: In this study, we showcase a workflow that uses Prosit for spectral library predictions on two common metaproteomes and implement an indexing and search algorithm, Mistle, to efficiently identify experimental mass spectra within the library. Hence, the workflow emulates a classic protein sequence database search with protein digestion but builds a searchable index from spectral predictions as an in-between step. We compare Mistle to popular search engines, both on a spectral and database search level, and provide evidence that this approach is more accurate than a database search using MSFragger. Mistle outperforms other spectral library search engines in terms of run time and proves to be extremely memory efficient with a 4- to 22-fold decrease in RAM usage. This makes Mistle universally applicable to large search spaces, e.g. covering comprehensive sequence databases of diverse microbiomes. Availability and implementation: Mistle is freely available on GitHub at https://github.com/BAMeScience/Mistle.

Motivation Deep learning has moved to the forefront of tandem mass spectrometry-driven proteomics and authentic prediction for peptide fragmentation is more feasible than ever. Still, at this point spectral prediction is mainly used to validate database search results or used for confined search spaces. Fully predicted spectral libraries have not yet been efficiently adapted to large search space problems that often occur in metaproteomics or proteogenomics. Results In this study, we showcase a workflow that uses Prosit for spectral library predictions on two common metaproteomes and implement an indexing and search algorithm, Mistle, to efficiently identify experimental mass spectra within the library. Hence, the workflow emulates a classic protein sequence database search with protein digestion but builds a searchable index from spectral predictions as an in-between step. We compare Mistle to popular search engines, both on a spectral and database search level, and provide evidence that this approach is more accurate than a database search using MSFragger. Mistle outperforms other spectral library search engines in terms of run time and proves to be extremely memory efficient with an 8 to 22-fold decrease in RAM usage. This makes Mistle universally applicable to large search spaces, e.g. covering comprehensive sequence databases of diverse microbiomes. Availability Mistle is freely available on GitHub at https://github.com/BAMeScience/Mistle. Contact thilo.muth{at}bam.de