INSIGHT SERIES
Through a combination of matching MS2 fragmentation against reference spectral libraries as well as de novo spectrum to structure models [4, 5], Pyxis annotated more than 3,200 analytes in SRM 1950 from just two LC-MS acquisition methods. Determination of analyte concentrations from untargeted LC-MS produced values that agreed with NIST certified values and published values from independent laboratories using LC-MS and NMR analytical methods. The entire data analysis from receipt of raw files, through untargeted structural annotation, de novo identification of library-absent compounds, and concentration determination for more than 2,000 high-confidence analytes required approximately 15 minutes.
Background
Pyxis is a co-scientist trained to analyze multi-omic data for the purpose of expanding biological interpretation and connecting omic data to phenotypic outcomes. One of Pyxis’ enabling capabilities is access to an LSM foundation model for machine interpretation of biochemical omic data, the first of its kind [4, 5]. Like other foundation models, e.g., for text, image, and sequence data, the LSM allows direct machine ingestion of raw instrument data commonly encountered in biochemical omics. Pyxis’ biochemistry knowledge derives from a collection of ML models trained to interpret these data for tasks including annotating chemical structure from MS2 fragmentation data [4, 5], determining untargeted concentrations from MS1 peak feature data with and without matched internal standards [6, 7, 8, 9], and coupling of these data with advanced biology reasoning models.
Our group recently launched de novo chemotyping, which aims to unify key attributes of targeted and untargeted mass spectrometry into a single concept: (1) the broad biochemical coverage (often tens of thousands of small molecules and lipids) typical of untargeted acquisition; with (2) the concentration determination of targeted methods that enable those analyses to be compared across studies, cohorts, instruments, researchers, and other batch effects. De novo chemotyping is akin to next generation sequencing, and advances in machine intelligence have made it possible for us to accelerate the data analysis, reduce the expertise burden, and bring the cost and scalability of biochemical omics to parity with the sequence-based omics.
That said, it is important for us to challenge the coverage and the concentration accuracy of Pyxis’ underlying mass spectrometry models. This is part of an overall effort of continual improvement through training, validation, and model improvement. As part of this effort, we asked Pyxis to measure and analyze NIST SRM 1950 [1] reference human plasma and compare untargeted analytes and their concentrations against certified and generally accepted values. We also asked Pyxis to analyze linearity and reproducibility across a six-level dilution series of SRM 1950 and across a 12-level amino acid standard.
Result
Value
NIST-valued amino acids, SRM 1950 plasma
Median deviation from NIST, 10 certified amino acids
5.8%
Mean deviation from NIST, 10 certified amino acids
6.0%
Injections within ±20% of NIST values, all 13 NIST-valued amino acids
96%
Injections within ±30% of NIST values, all 13 NIST-valued amino acids
100%
Linearity and precision
Dilution linearity, 12 amino acids in plasma and 16 in a 17 amino acid standard
R² ≥ 0.98, slope 0.92 to 1.05
Replicate precision across dilution levels, plasma
2% to 6% CV
Biochemome annotation
Analytes annotated in SRM 1950 from two LC-MS methods with DDA triggering
more than 3,200
De novo identifications in SRM 1950
422
High-abundance plasma metabolites and lipids identified
88%
Table 1. Summary of results. Concentration figures use the 6x dilution, four replicate preparations; annotation figures combine HILIC and reversed-phase acquisitions.
5.8%
Median deviation of Pyxis from NIST certified values across the ten certified amino acids in SRM 1950 plasma.
Pyxis matches NIST certified amino acid values
For all 13 amino acids with NIST values, Pyxis concentrations in SRM 1950 plasma fall within ±20% of the NIST value in 50 of 52 injections and within ±30% in all 52 [1].
NIST assigns certified values to ten amino acids in SRM 1950 and reference (non-certified) values to arginine, phenylalanine, and threonine; together these are the 13 NIST-valued amino acids. Like any measurement, each NIST value carries an expanded uncertainty, shown alongside the Pyxis values below. All Pyxis concentrations in this report are taken from a single dilution, 6x, the most concentrated in the series, as the median of four replicate preparations.
Amino acid
NIST status
NIST value ± U (µmol/L)
Pyxis mean ± 1σ, µmol/L
Deviation of mean
Pyxis median (IQR), µmol/L
Deviation of median
Injections within ±20%
Methionine
Certified
22.3 ± 1.8
22.0 ± 1.4
1.4%
22.0 (21.2 to 22.8)
1.2%
4 of 4
Leucine
Certified
100.4 ± 6.3
99.4 ± 6.9
1.0%
99.1 (96.5 to 102.1)
1.3%
4 of 4
Proline
Certified
177 ± 9
175.4 ± 11.5
0.9%
174.3 (169.4 to 180.3)
1.5%
4 of 4
Lysine
Certified
140 ± 14
144.4 ± 9.9
3.1%
144.1 (139.3 to 149.1)
2.9%
4 of 4
Tyrosine
Certified
57.3 ± 3.0
59.6 ± 5.2
4.0%
59.7 (57.3 to 62.0)
4.2%
4 of 4
Alanine
Certified
300 ± 26
279.9 ± 20.3
6.7%
277.6 (269.1 to 288.5)
7.5%
4 of 4
Isoleucine
Certified
55.5 ± 3.4
60.6 ± 4.8
9.1%
59.6 (58.1 to 62.1)
7.5%
3 of 4
Serine
Certified
95.9 ± 4.3
87.8 ± 6.2
8.5%
87.0 (84.5 to 90.3)
9.3%
4 of 4
Glycine
Certified
245 ± 16
222.0 ± 14.0
9.4%
220.5 (213.3 to 229.2)
10.0%
4 of 4
Valine
Certified
182.2 ± 10.4
206.7 ± 27.5
13.4%
209.9 (194.4 to 222.1)
15.2%
3 of 4
Threonine
Reference
119.5 ± 6.1
121.6 ± 10.3
1.8%
120.7 (116.9 to 125.4)
1.0%
4 of 4
Phenylalanine
Reference
51 ± 7
54.7 ± 4.4
7.2%
53.4 (52.3 to 55.7)
4.7%
4 of 4
Arginine
Reference
81.4 ± 2.3
73.8 ± 4.9
9.4%
73.2 (71.7 to 75.3)
10.0%
4 of 4
Table 2. Pyxis concentrations against NIST SRM 1950 values. Pyxis median and interquartile range across four replicate preparations at 6x, and Pyxis mean and standard deviation (1σ) across the same preparations. U is NIST’s expanded uncertainty [1]. Deviation is the absolute difference between the Pyxis median, or mean, and the NIST value, relative to the NIST value.
Concentration linearity with dilution
Pyxis concentrations follow the applied dilution with slopes of 0.92 to 1.05 and R² of 0.98 to 1.000, for 12 amino acids in plasma over a 16-fold range and for 16 of the 17 amino acids in a pure standard over a 64-fold range.
Figure 1. Dilution slopes. Plasma: NIST SRM 1950 at 6x to 96x. Standard: the upper seven of 12 levels (0.39 to 25 µM) of a 17 amino acid standard spanning 0.012 to 25 µM; valine not shown. Slope of log measured against log expected concentration; shaded band, slope 0.90 to 1.10. September 2026.
Replicate precision across dilution levels is equally tight: coefficients of variation run 2% to 6% in plasma and 1% to 7% in the standard. Across 0.39 to 25 µM, 82% of standard injections for the same 16 amino acids fall within ±20% of nominal. The linearity fit uses the upper seven of the standard’s 12 levels (0.39 to 25 µM); the full series extends to 0.012 µM.
Pyxis agrees with independent laboratories
On 22 plasma metabolites measured by at least two independent platforms [2], Pyxis concentrations fall within the range those platforms report, or within 20% of it.
Figure 2. Pyxis concentrations against independent laboratory values. Independent values compiled by the Wishart group [2] and others: Wishart, Wishart-Prime, NMR, Biocrates MxP500, and reference values. Pyxis values from the 6x dilution, four replicate preparations, September 2026. Both shown relative to the median of the independent methods.
The agreement extends past the amino acids to hippuric acid and octanoylcarnitine, and to clinical markers: creatinine matches its NIST certified value [1] within ±20% at all six dilutions.
Pyxis identifies the metabolites present in SRM 1950
From two LC-MS methods, Pyxis annotates more than 3,200 analytes in SRM 1950 and recovers 88% of the high-abundance metabolites and polar lipids the Wishart group reports in the same plasma [2].
SRM 1950, HILIC and RP
Pyxis
Analytes annotated
3,257
With an MS/MS spectral match
2,114
De novo annotations
422
Wishart-reported metabolites and polar lipids above 10 µM that Pyxis identifies
70 of 80
Table 3. Annotation in SRM 1950. Unique analytes across HILIC and reversed-phase acquisitions, seven replicate injections per polarity, isotope-labeled standards excluded.
>3,200
Analytes annotated in SRM 1950 from two LC-MS methods.
For scale, NIST’s own profiling of SRM 1950 identified 353 metabolites across GC-MS, LC-MS, and NMR [3], and the Wishart group quantified 728 using eight targeted assays on four platforms [2].
Established methods set the baseline for agreement
On the ten NIST-certified amino acids, Pyxis deviates from NIST by a mean of 6.0%, against 5.6% for 700 MHz NMR and 3.7% for targeted, isotope-dilution LC-MS/MS, and those two established methods differ from each other by a mean of 5.6% on the same analytes [2].
Reference methods
The Wishart group characterized SRM 1950 with several orthogonal platforms [2]:
• 700 MHz ¹H NMR. Bruker Avance III with a 5 mm HCN cryoprobe. Plasma was deproteinized by 3 kDa ultrafiltration, buffered with D₂O/phosphate containing DSS as internal standard, and fitted automatically with MagMet against a custom library, with manual review of every spectrum. Triplicate preparations.
• TMIC PRIME. Targeted direct-injection plus reversed-phase LC-MS/MS on a Sciex 4000 QTRAP, with phenyl isothiocyanate derivatization for amines and 3-nitrophenylhydrazine derivatization for organic acids, isotope-labeled internal standards, and seven-point external calibration; up to 142 metabolites.
• TMIC MEGA. The same chemistry scaled to up to 726 metabolites on a Sciex 5500 QTRAP with an Agilent 1290 UHPLC, using 48 isotope-labeled internal standards for amines and up to 102 for organic acids.
• Biocrates MxP 500. A commercial targeted kit on a Sciex 5500 QTRAP, covering up to 630 metabolites with kit-integrated internal standards and calibrators.
• Dedicated targeted LC-MS/MS assays for bile acids, catecholamines, and water- and fat-soluble vitamins, and ICP-MS for metal ions.
Every one of these is a targeted assay: each analyte has a predefined transition or spectral signature, and most have a matched isotope-labeled internal standard and a calibration curve. The NIST certified values [1] are themselves consensus values from multiple methods.
Agreement with NIST, method by method
Amino acid
NIST status
NIST (µmol/L)
NMR
Targeted LC-MS/MS
Pyxis
Alanine
Certified
300
299
295
277.6
Glycine
Certified
245
238
250.8
220.5
Isoleucine
Certified
55.5
52.4
58.3
59.6
Leucine
Certified
100.4
102.7
100.0
99.1
Lysine
Certified
140
157.7
137
144.1
Methionine
Certified
22.3
20.4
20.2
22.0
Proline
Certified
177
157.8
172.3
174.3
Serine
Certified
95.9
86.1
91.7
87.0
Tyrosine
Certified
57.3
57.1
55.2
59.7
Valine
Certified
182.2
178.1
173
209.9
Arginine
Reference
81.4
77.1
82.9
73.2
Phenylalanine
Reference
51.0
46.6
53.9
53.4
Threonine
Reference
119.5
117.0
118.4
120.7
Table 4. Amino acid concentrations in SRM 1950 by method. NMR and targeted LC-MS/MS values from [2], Table S2; NIST values from [1]. Pyxis values are the 6x medians from Table 2.
Deviation from NIST, ten certified amino acids
NMR
Targeted LC-MS/MS
Pyxis
Median
4.2%
3.2%
5.8%
Mean
5.6%
3.7%
6.0%
Largest
12.6% (lysine)
9.4% (methionine)
15.2% (valine)
Table 5. Deviation from NIST by method. Per amino acid absolute deviation of each method value from the NIST certified value.
NMR and targeted LC-MS/MS differ from each other by a median of 4.3% on the certified amino acids and 6.3% across all 19 amino acids both measured, reaching 43% for aspartate. Across the 13 NIST-valued amino acids, Pyxis lands closer to NIST than NMR does for seven (leucine, lysine, methionine, proline, serine, phenylalanine, and threonine) and closer than targeted LC-MS/MS for three (methionine, proline, and phenylalanine). Pyxis values are the same 6x medians used throughout this report; NMR, LC-MS/MS, and NIST values are per-method means [2].
Beyond the amino acids, established assays disagree far more
For analytes without a certified value, the spread among established targeted assays is the realistic baseline. Table 6 compares three published targeted assays, TMIC MEGA, TMIC PRIME, and Biocrates MxP 500, on the same SRM 1950 material [2]; these figures describe the published assays, not Pyxis.
Analyte class
Analytes compared
Median spread among published assays (highest ÷ lowest)
Largest spread among published assays
Acylcarnitines
7
2.3×
4.6× (C6:1)
Organic acids and others
6
2.7×
7.0× (succinic acid)
Cholesteryl esters (with two interlaboratory studies)
14
1.9×
47× (CE 20:0)
Table 6. Spread among published targeted assays in SRM 1950 (not Pyxis). From [2], Tables S4 and S5.
This spread among established assays is the appropriate yardstick for the independent-laboratory comparison in Figure 2.
Materials and methods
All samples were prepared with a single protein-precipitation workflow and acquired on Orbitrap Exploris 240 high-resolution accurate mass LC-MS/MS systems.
Samples and sample preparation
NIST SRM 1950 Metabolites in Frozen Human Plasma [1] was thawed on ice and centrifuged at 12,000 × g for 10 min at 4 °C. Proteins were precipitated with ice-cold methanol/acetonitrile/water (50:30:20, v/v/v), incubated at −80 °C for 1 h, mixed for 15 min at 4 °C, and centrifuged at 12,000 × g for 10 min at 4 °C. Supernatants were diluted 1:2 with extraction solution containing the augmented universal calibrants, giving final dilutions of 6x, 12x, 24x, 36x, 48x, and 96x relative to plasma, each prepared in four replicates. Extraction blanks, prepared with water in place of plasma, were processed alongside the samples.
A 17 amino acid standard was serially diluted two-fold to 12 levels spanning 0.012 to 25 µM (cystine at half these concentrations) and diluted 1:2 with the same calibrant-containing extraction solution.
Liquid chromatography
Hydrophilic interaction chromatography (HILIC) was performed on Thermo Scientific Vanquish UHPLC systems with a Waters Atlantis Premier BEH Z-HILIC VanGuard FIT column (2.1 × 50 mm, 2.5 µm) held at 25 °C. Mobile phase A was 20 mM ammonium carbonate in water with 0.25% (v/v) ammonium hydroxide (pH 9.55); mobile phase B was acetonitrile. The injection volume was 4 µL. Reversed-phase identification runs used C18 chromatography on the same LC platform.
Mass spectrometry
Data were acquired on Thermo Scientific Orbitrap Exploris 240 mass spectrometers with Ion Max NG heated electrospray sources. Quantitation runs used full-scan acquisition with positive and negative polarity switching in a single injection. Identification runs used data-dependent MS/MS with AcquireX deep-scan acquisition, in separate positive and negative injections. Each sequence began and ended with solvent blanks, extraction blanks, and quality control injections.
Data analysis
Pyxis, the Matterworks co-scientist for omic data, reads raw files directly and draws on the Large Spectral Model, a foundation model trained on more than ten billion raw spectra. Each sample is interpreted into a de novo biochemome: an annotated inventory of the small molecules and lipids present, with concentrations attached, produced without per-dataset feature engineering and without a curated target list.
Raw files were uploaded to Pyxis [4, 5, 6, 7, 8, 9], which returned analyte identifications and concentrations. Plasma concentrations were obtained by multiplying the reported concentration by the dilution factor. Agreement with NIST was scored per injection as relative deviation from the certificate value [1]; dilution linearity as the slope and R² of log measured against log expected concentration; and identification coverage by matching Pyxis annotations to the compounds reported in SRM 1950 by the Wishart group [2].
Limitations
All results come from one PyxisLabs laboratory on a single lot of SRM 1950. Comparison against NIST values covers the 13 amino acids NIST certifies or reports; agreement for other analyte classes rests on the independent laboratory values compiled in [2].
For the New Insight Series, we asked Pyxis to measure NIST SRM 1950, the reference human plasma for metabolomics, and to set its identifications and concentrations against NIST certified values and independent laboratories. Pyxis annotated more than 3,200 analytes and matched the ten certified amino acids to a median deviation of 5.8%.
Data: PyxisLabs acquisitions of NIST SRM 1950 performed under the standard protocol with augmented universal calibrants. Quantitation: dilution series at 6x, 12x, 24x, 36x, 48x, and 96x, four replicates per level, HILIC, September 2026. Amino acid standard: 17 amino acids at 12 two-fold levels from 0.012 to 25 µM, six to nine replicates per level, HILIC, September 2026. Identification: SRM 1950 at 10x, seven replicate injections per polarity, HILIC and reversed-phase. All figures derive from the Pyxis analysis of those raw files. Questions about this analysis: info@matterworks.ai.
References
1. National Institute of Standards and Technology. Certificate of Analysis, Standard Reference Material 1950: Metabolites in Frozen Human Plasma. Gaithersburg, MD (revised 12 October 2023). tsapps.nist.gov/srmext/certificates/1950.pdf
2. Mandal, R. et al. Comprehensive, quantitative analysis of SRM 1950: the NIST human plasma reference material. Anal. Chem. 97, 667-675 (2025). doi:10.1021/acs.analchem.4c05018
3. Simón-Manso, Y. et al. Metabolite profiling of a NIST Standard Reference Material for human plasma (SRM 1950): GC-MS, LC-MS, NMR, and clinical laboratory analyses, libraries, and web-based resources. Anal. Chem. 85, 11725-11731 (2013).
4. Asher, G. et al. LSM-MS2: a foundation model bridging spectral identification and biological interpretation. Preprint at arXiv https://arxiv.org/abs/2510.26715 (2025).
5. Asher, G., Cadosch Delmar, M., Campbell, J. M., Geremia, J. & Kassis, T. Semantic AI interpretation of MS2 data, including fast spectral matching, property prediction and de novo generation. Poster, 72nd ASMS Conference on Mass Spectrometry and Allied Topics (June 2024).
6. Ferro, L. S. et al. Scalable absolute quantitation for untargeted metabolomics. Poster, 72nd ASMS Conference on Mass Spectrometry and Allied Topics (June 2024).
7. Ferro, L. S. et al. Expanded coverage and validation of generalized AI models for direct metabolomic annotation and quantitation of raw LC-MS data. Poster, 73rd ASMS Conference on Mass Spectrometry and Allied Topics (June 2025).
8. Costa, A. S. H., Asher, G., Shah, D., Kassis, T. & Campbell, J. M. Using large spectral models for simultaneous annotation and quantitation of complex metabolomic samples. Poster, 73rd ASMS Conference on Mass Spectrometry and Allied Topics (June 2025).
9. Costa, A. S. H. et al. Expanding the analyte coverage of Pyxis, an AI-powered platform for rapid, accurate identification and quantification of metabolites in biological samples. Poster, MANA 2025 (September 2025).