A tissue sample run through a mass spectrometer yields one averaged chemical profile for millions of cells of many types. When a tumor's cells shift toward burning glucose into lactate, the bulk profile cannot say which cells did it, or whether the tumor simply contained more of the cells that always do. Two further problems make this worse. Most of the thousands of detected chemical signals have no matching reference spectrum and get thrown away. And signal intensities drift from one instrument run to the next, so when a large study spans several runs, these "batch effects" are easily mistaken for biology.
This proposal treats the three problems as one inverse problem with a known physical order. Cells mix first, and the instrument distorts the mixture afterwards. We fit a model in which each measured signal is a run-specific, monotone distortion of a proportion-weighted sum of per-cell-type abundances. The cell proportions come from paired transcriptomics. The model never needs to know what molecule a signal is, so unannotated features are unmixed alongside annotated ones. The output is a cell-type-resolved metabolome for each condition, and from it a list of which cell type's chemistry changed, including chemistry no database can name.
Core Claim
Current batch correctors for metabolomics align the distributions of whole samples across runs. The closest example is the encoder–decoder–classifier model of Joint deep learning for batch effect removal and classification toward MALDI MS based metabolomics, which applies a distribution-matching penalty (maximum mean discrepancy) to serum spectra. On tissue, this alignment wipes out real differences in cell composition whenever composition differs between runs, and in clinical cohorts it usually does. Deconvolution methods do resolve cell composition, but they are built for transcript counts. Count Bridges enable Modeling and Deconvolving Transcriptomic Data models integer counts with birth–death dynamics. POPPY aligns bulk RNA to single-cell atlases. Neither can take continuous ion intensities, and neither models instrument drift. Our claim is that putting cell mixing before the instrument distortion, and anchoring the distortion with the pooled quality-control (QC) samples that LC-MS labs already run in every batch, makes the two effects separable. Under this structure, batch correction and cell-type unmixing of untargeted metabolomes become one estimable problem. We further claim the unmixed profiles carry enough signal to attribute condition-driven metabolic shifts, such as a hypoxia-induced glycolytic switch, to the right cell type for unannotated features as well as annotated ones.
Model
For sample $s$ in run $b(s)$ and feature $f$, an untargeted LC-MS feature is defined only by its mass-to-charge ratio and retention time and needs no identity. We write
$$ y_{sf} = g_{b(s)f}\Big( m_s \sum_{k} p_{sk}\, c^{(t(s))}_{kf} \Big) + \varepsilon_{sf}. $$
Here $p_{sk}$ is the fraction of cell type $k$ in sample $s$, taken from paired bulk RNA-seq deconvolution or known by design. The term $c^{(t)}_{kf} \ge 0$ is the abundance of feature $f$ per unit of cell type $k$ under condition $t$, such as normoxia or hypoxia. The scalar $m_s$ absorbs sample-level loading and global matrix effects. The function $g_{bf}$ is a monotone increasing spline that captures run-specific response and drift for that feature. The error $\varepsilon_{sf}$ is heteroscedastic noise.
Pooled QC samples have identical composition in every run, so differences between their measurements estimate the $g_{bf}$ directly. Varying composition across study samples then identifies $c$ up to a per-feature scale. That scale is irrelevant for within-feature comparisons across cell types and conditions.
Fitting alternates between two steps. Nonnegative least squares, or a log-normal likelihood, updates $c$. Constrained spline regression updates $g$. A hierarchical prior shrinks $c^{(\text{hypoxia})}_{kf}$ toward $c^{(\text{normoxia})}_{kf}$, so that a "shift" must be supported by data. Uncertainty in $p$ is propagated by resampling the transcriptomic deconvolution.
Ion suppression that depends nonlinearly on the whole matrix is the most likely way the model fails. We test it directly instead of assuming it away (see the decisive checks).
Research Program
The work has three parts.
The first is the estimator just described, released with QC-anchored identifiability diagnostics. These diagnostics flag features whose run distortion cannot be estimated, which happens when the feature is missing from QC pools.
The second is a controlled ground-truth resource, because no public dataset in the file provides cell-type metabolomes under batch shift. We borrow the logic of the HYE124 known-ratio tri-proteome benchmark used by Dear-DIA XMBD. Three to four human cell lines are mixed at designed proportions, with one line also cultured under hypoxia. All mixtures, pure lines and QC pools are acquired across at least four LC-MS runs. Composition is deliberately confounded with run in some arms. The pure-line measurements are held out as ground truth.
The third is downstream interpretation. Features carrying a cell-type-specific shift but no library match are grouped by shared fragmentation motifs with mineMS2 and given candidate structures with GEMS. Annotated features are scored for pathway activity with PUMA, now applied per cell type rather than per bulk sample. Annotation is deliberately kept downstream. The unmixing does not depend on it, and that is why dark features survive.
The file's metabolic-reprogramming literature sets the expected biology. HIF-driven glycolysis and the PDHK block on pyruvate entry into the mitochondrial TCA cycle (Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development; Hypoxia and hypoxia-inducible factor signals regulate the development, metabolism, and function of B cells) predict which annotated features should shift, and in which line. That prediction is a sanity check on attribution.
Baselines
The naive reference is per-run median scaling followed by attributing bulk condition differences to whichever cell type is most abundant. This is what a lab without the method effectively does, and any unmixing method must clear it.
The first serious baseline is the sequential pipeline. ComBat-style location–scale correction per feature is followed by nonnegative linear deconvolution with known proportions. It receives the same features, the same QC samples (used for its batch model) and the same proportions.
The second is the learned corrector of Joint deep learning for batch effect removal and classification toward MALDI MS based metabolomics, retrained on our runs and followed by the same linear deconvolution. That paper reports diagnostic accuracy gains of 5.1% to 7.9% on average over sequential pipelines, and up to 13.6% accuracy and 15.3% AUC on specific batches, on its private 598-subject serum dataset. Those numbers measure classification on homogeneous biofluid, not profile recovery. For cell-type deconvolution of untargeted metabolomes, no reported result was found on Lacuna, so all comparisons will be made on our own resource.
The strongest simple competitor is a per-feature linear regression of log intensity on cell proportions with run as a fixed effect, which is the obvious move a statistician would make. All learned baselines get the same hyperparameter search budget as our model: 50 configurations, chosen on a validation run that is excluded from testing.
Two ablations isolate the key ingredient. In the first, the composition terms are removed, so the model applies the same monotone warps but aligns runs without knowing proportions. In the second, the order is reversed, unmixing log-transformed intensities after a per-run correction.
Beating the baselines means two things on confounded-design runs. Per-feature correlation between recovered and held-out pure-line profiles must be significantly higher, with paired bootstrap over features, in both the annotated and the unannotated strata. Recall of cell-type-specific hypoxia shifts must also be higher at a fixed 10% false discovery rate.
Why not just…?
Why not just regress log intensity on cell proportions with a batch covariate? This is the real threat. If run effects are close to a per-feature multiplicative constant and composition varies little between samples, log-linear regression approximates our model well. Its weakness is structural: mixing is additive in concentration, not in log space. Log-regression therefore misattributes signal when one minor cell type dominates a feature, which is exactly the rare-subpopulation case raised by Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance. Our mixture design sweeps the minor-population fraction from 2% to 50%. If log-regression matches us across that sweep, we will report that and ship it as the recommended tool.
Why not just batch-correct first, then deconvolve? Global or distribution-level alignment, whether location–scale or MMD-based as in the MALDI model and in BERMUDA, cannot tell a composition difference between runs from a technical one. BERMUDA avoids merging populations only because it can cluster single cells, and bulk spectra offer nothing to cluster. The confounded-design arm tests this directly. The prediction is that correct-then-deconvolve shrinks true cell-type differences toward zero in proportion to the confounding.
Why not just infer cell-type metabolism from single-cell RNA, as GEFMAP does? GEFMAP predicts mass-balanced fluxes from enzyme transcripts. Its authors acknowledge that transcript levels miss post-translational control such as PDH phosphorylation, and the model cannot see any metabolite absent from Recon3D. We run GEFMAP on scRNA-seq of the same lines and check whether its predicted glycolytic shift agrees with our measured per-cell-type lactate and TCA-intermediate shifts. Agreement validates both methods. Disagreement on PDHK-regulated steps is the outcome we expect, and it would show what measured chemistry adds.
Why not just measure sorted cells directly? Pure-line measurement is our ground truth. For primary tissue, though, the file notes that direct single-cell mass spectrometry suffers sample loss, low ion yield and severe batch variability. The method is meant for the many existing bulk cohorts that cannot be re-collected.
Decisive Checks
We run three checks. The first is recovery on confounded runs: recovered versus held-out pure-line profiles, reported separately for annotated and unannotated features, against all baselines and ablations. A negligible margin over the reversed-order ablation would refute the core claim.
The second targets matrix effects. On a spiked-standard subset with known concentrations, we test whether the residuals of the mixing model grow with composition. If they do, an interaction term between $m_s$ and composition is added and its benefit reported.
The third is attribution, with two parts. The method must place the hypoxia-induced lactate rise in the hypoxic line alone. It must also recover unannotated hypoxia-shifted features at a rate comparable to annotated ones, which we can verify because the pure lines are measured. Before any wet-lab spending, the whole pipeline is rehearsed in silico. Simulated bulk profiles are built from pure-line spectra in the manner of the Kassandra sorted-cell simulations used by POPPY, and synthetic monotone run distortions are imposed on them.
References
- Joint deep learning for batch effect removal and classification toward MALDI MS based metabolomics
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Dear-DIA XMBD : Deep Autoencoder Enables Deconvolution of Data-Independent Acquisition Proteomics
- mineMS2: annotation of spectral libraries with exact fragmentation patterns
- GEMS: Molecular Structure Identification via Geodesic Navigation of the Isomer Manifold
- Pathway Activity Analysis and Metabolite Annotation for Untargeted Metabolomics using Probabilistic Modeling
- Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
- Hypoxia and hypoxia-inducible factor signals regulate the development, metabolism, and function of B cells
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes
- Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning