|
Abstract INTRODUCTION:
Liquid chromatography–mass spectrometry remains a foundational technology across the life sciences, pharma R&D, and environmental research. Rapid acquisition and conversion of mass spectra into actionable data is highly desirable in both discovery and in routine workflows. Traditionally, researchers have had to choose between onerous targeted analyses—measuring the absolute concentration of a small list of known compounds— and untargeted, broad coverage analyses to identify, but not quantify, as many known and unknown compounds as possible. In 2022 we released PyxisTM, a deep learning-based tool capable of inferring the analyte concentrations directly from raw mass spectra with a standardized HILIC method. In this work, we expand Pyxis’ capabilities to a broader spectrum of small molecules and metabolic pathways, partially facilitated by the extension of the model to incorporate data acquired using a newly developed Reversed Phase liquid chromatography method.
To enable direct machine ingestion of raw instrument data, we developed Large Spectral Models (LSMs) via self-supervised training on massive quantities of unlabeled raw data aggregated across many downstream applications, instrument platforms, and configurations. We adapt our LSMs to specific prediction problems by fine-tuning on relatively small high-quality application-specific datasets.
METHODS:
To benchmark Pyxis’ latest model, we selected the NIST Standard Reference Material (SRM) 1950 (Metabolites in Frozen Human Plasma), and the Reference Material (RM) 8231 (Frozen Human Plasma Suite for Metabolomics), representing a total of four different phenotypes. Metabolites were extracted from these biological samples by precipitating proteins with an 80% organic solution. For each sample type, multiple different sample to solvent ratios were used to create a dilution series. The initial analysis of SRM 1950 specifically was prepared by two different operators, thus generating two separate batches. The extracts were mixed with StandardCandles™, Pyxis’ analyte and matrix independent calibrators. Four microliters from each sample extract were analyzed using a Transcend LX-2 multichannel system coupled to an Orbitrap Exploris 120 mass spectrometer (both Thermo Fisher Scientific™). Metabolites were extracted from biological matrices by precipitating proteins with an organic solution with StandardCandlesTM, Pyxis’ analyte and matrix independent calibrators. HILIC-based chromatographic separation was performed with the Waters™ Atlantis Premier BEH Z-HILIC column (2.5 mm, 2.1 x 50 mm), and a mobile phase consisting of 20 mM ammonium carbonate with 0.25% (v/v) ammonium hydroxide (pH=9.6) (solvent A), and acetonitrile (solvent B). High resolution spectra were acquired for 6.7 minutes in full scan and polarity switching mode. For RP analyses, we recorded MS1 data for 7 minutes in polarity switching mode and used the ACQUITY HSS T3 UPLC analytical (1.8 mm, 2.1 x 50 mm) and guard columns (Waters). The RP mobile phase consisted of 0.2% formic acid in water (solvent A) and 0.1% formic acid in methanol (solvent B). The raw instrument files were analyzed using Pyxis to identify and infer the concentration of metabolites in the biological samples.
RESULTS:
Sample extracts from distinct phenotypes were analyzed using Pyxis. Briefly, the SRM 1950 consists of pooled frozen human plasma from healthy donors, and the RM 8231 is a suite of pooled plasma from three distinct phenotypes, diabetic, hypertriglyceridemic, and young African Americans (20-25 years of age).
The predicted concentrations in the widely used NIST Human Plasma Reference Material (SRM 1950), were benchmarked against recently published data.The reference dataset generated by the Wishart lab provides an extensive and quantitative characterization of the human reference plasma, with over a thousand high-confidence values for the scientific community to easily benchmark their targeted and untargeted metabolomics and lipidomics analyses. The median R-squared of the 50 most abundant analytes was 0.95 and the slope was 1.24, which demonstrates Pyxis’ ability to infer analyte concentrations with accuracy. Importantly, Pyxis does not require the use of stable isotope labeled internal standards, calibration curves, and the associated extensive sample preparation steps.
Upon validating Pyxis-inferred concentrations, we sought to make a comparison between different sub-population cohorts in the sample set: healthy population, young African-American, diabetic, and hyperglyceridemic. Using both HILIC and RP column chemistries, we quantified over 1000 different analytes across multiple classes—namely, endogenous metabolites, natural products, products of the gut microbiome metabolism, and drugs—to identify phenotype-specific markers. The metabolic profiles inferred by our model enabled the accurate segregation of plasma samples from the different phenotypes.
CONCLUSION:
This work showcases the expanded applicability of Pyxis to biological samples using both RP and HILIC LC-MS methods. We demonstrate Pyxis’ ability to discern changes in biological samples and were able to replicate consensus concentrations in standard reference materials. Finally, we achieved quantitative separation of phenotypes with, not just analytical accuracy, but also with significant efficiency when compared to traditional workflows. |
= Discovery stage. (53.14%, 2025)
= Translation stage. (22.33%, 2025)
= Clinically available. (24.53%, 2025)
