Using AI to Enable Rapid Decision Making by Transforming Raw LC-MS Data to Identified and Quantified Molecules

Determining the identity and abundance of molecules present in a biological system is the ultimate promise of mass spectrometry. While improvements in high resolution MS have led to the detection of an ever-increasing number of “features”, the identities are elusive, and concentrations are typically unknown. Given how AI has successfully addressed other long standing scientific problems, notably in protein folding, why not also employ AI advances to model the critical issue in MS of transforming raw data directly into ID and absolute concentration?
In conventional LC-MS experiments, obtaining absolute concentration necessitates both expertise and the purchasing of commercial standards. We have developed a technology, Pyxis, founded on a powerful deep learning model which enables the transformation of raw data to ID and concentrations of detected analytes. For Pyxis to work sample preparation must incorporate the proprietary universal calibrators, (StandardCandles™) and data acquisition must use a prescribed LC-MS method (HILIC or RP) and be acquired using full scan MS data in both positive and negative modes. We have internally generated exemplars to verify model performance in a series of biological critical matrices – including cell lysate, urine, plasma, and DBSs. When compared to the traditional method of obtaining quantitative data using isotopologues, the current model (Pyxis 1.4 – July 2024) has a median absolute concentration error of ~25% (identical for all matrices) on over 50000 data points for a broad set of cellular metabolites. We are currently extending our models to generalize MS1-based quantitation across molecular structure – meaning that a user can suggest an analyte of interest based on biological hypotheses and be provided with an absolute concentration which is within a factor of two of its “ground truth”. The performance attributes of the model will be discussed along with some critical demonstration of applications.