Instrument-agnostic Machine Learning Framework for Accurate Prediction of Functional Groups from Tandem Mass Spectrometry

Structural elucidation of unknown compounds using tandem mass spectrometry (MS/MS) is an ongoing challenge. Expert chemists will often use common product ion peaks and neutral losses or peak matching software to predict structural aspects from MS/MS spectra, a process which is time consuming and limited by the molecular entries in the database. We introduce an instrument-agnostic machine learning (ML) framework that predicts functional groups directly from MS/MS spectra, eliminating the need for conventional database searches. The ML models are trained on data from different sources, the Mass Bank of North America (MoNA), a public database, and a custom MS/MS data-base acquired using a single high throughput desorption electrospray ionization (DESI) MS/MS platform to showcase transferability of model predictions independent of instrument acquisition. Additionally, a novel spectral representa-tion is developed for ML use by standardizing the spectra to the precursor ion m/z. We utilized molecular metrics, specifically, average molecular F1 score of 87% (MoNA) and 76% (DESI) and average molecular accuracy of 94% (MoNA) and 87% (DESI), indicating similar model performance when trained on different databases to validate the ML framework. To further illustrate the robustness and transferability of our ML framework, models trained on DESI-generated data successfully predicted functional groups, using spectral data from blind test set of molecules that were generated by two different company laboratories using distinct mass spectrometry instruments (Orbitrap and TOF) operated in positive-ion mode electrospray ionization. We believe that such an approach, where models trained on data from one instrument can reliably make accurate predictions on spectra acquired using different instruments in a database-free manner, will significantly expand the practical applications of ML to mass spectrometry.