When a tumor biopsy shows high succinate or kynurenine, the measurement is an average over cancer cells, macrophages, fibroblasts and blood vessels. Standard pathway enrichment then reports "TCA cycle altered" without saying which cells altered it. That gap matters most when the metabolite is also a signal: lactate, succinate and kynurenine are secreted by one cell type and sensed by receptors on another. We propose to recover the cellular origin of bulk metabolite changes with one physical constraint. In a tissue at steady state, the extracellular pool of each metabolite must balance the production and consumption of every cell type in it, weighted by how many cells of that type are present. Coupling cell-type deconvolution of bulk RNA to per-cell-type metabolic flux models through this shared-pool constraint, and fitting it to measured bulk metabolomics, should split a bulk metabolite shift into cell-type contributions. Those contributions can then drive cell-type-resolved pathway enrichment and metabolite-mediated communication maps.
The core claim is narrow and testable. Treating the extracellular pool as a shared mass-balance constraint across deconvolved cell types assigns metabolite changes to their source cells more accurately than the obvious alternative, which is regressing each measured metabolite on estimated cell fractions across patients. The advantage should be largest where that regression is weakest: small cohorts, closely related cell types, and metabolites whose bulk level reflects consumption rather than production. Ligand-receptor analysis becomes a downstream use, not the core contribution. Pathway enrichment becomes the readout.
Where the Pieces Stand
Four lines of work each hold part of this problem, and none connects them. For cell-cell communication, the methods score signaling from ligand and receptor mRNA in sender and receiver groups. CellChat uses mass-action scoring over curated multi-subunit complexes. stMLnet adds distance-dependent signaling on spatial data. SigXTalk uses hypergraphs to trace downstream crosstalk from receptor to transcription factor to target gene. All three treat ligands as genetically encoded proteins, so none can represent a metabolite whose abundance depends on fluxes in other cells. These papers motivate the downstream communication module. They do not supply its method.
On the metabolic side, GEFMAP predicts reaction fluxes per single cell. It uses a graph neural network on the reaction-metabolite network and a null-space layer that guarantees mass balance ($Sv=0$, where $S$ is the reaction stoichiometry matrix and $v$ the flux vector). It models each cell as an isolated system with no exchange. UnitedMet imputes bulk metabolite levels from bulk RNA. It uses a rank-based Plackett–Luce model, a probability model over orderings, trained on paired RNA-metabolomics data from clear-cell renal carcinoma. It works only at the bulk level and discards absolute scale. GEFMAP supplies our per-cell-type flux layer. UnitedMet supplies our observation model. The file's own analysis notes a tension between UnitedMet's ranks and GEFMAP's cardinal stoichiometry. We resolve it by keeping stoichiometry inside cells and using ranks only at the point where model predictions meet mass-spectrometry data.
For deconvolution, the job of splitting a mixed RNA profile into cell-type fractions and profiles, there are several options. Count Bridges produces integer single-cell count profiles that sum exactly to bulk totals. POPPY attributes bulk phenotypes to cell types through ontology-guided alignment. A 50-dataset spatial benchmark ranks CARD, Cell2location and Tangram as top methods and warns that closely related subtypes stay hard to separate. All of these stop at proportions or unmixed expression. We reuse them as the front end, unmodified. Finally, MOMENTA expands enzyme-only gene sets with protein-interaction neighbors. It reports a 25-fold increase in significant metabolic pathways in glucose-starved breast cancer cells relative to base gene sets, but it works on bulk populations. We use its expanded sets as the gene-set layer for cell-type-resolved enrichment.
Core Mechanism
For patient $j$, a deconvolution method gives cell-type fractions $\pi_{jk}$ and cell-type expression profiles $x_{jk}$ for cell types $k=1,\dots,K$. A shared GEFMAP-style network maps each $x_{jk}$ to internal fluxes and exchange fluxes $e_{jk}$, which are net secretion (positive) or uptake (negative) per cell for each extracellular metabolite. Internal mass balance holds by construction. The new constraint is the tissue-level balance
$$\sum_{k=1}^{K} \pi_{jk}\, e_{jk} + u_j = 0,$$
where $u_j$ is the net exchange of the tissue with blood. We penalize $u_j$ toward sparsity so that the cell types must explain most of the pool. Measured metabolites enter through a rank likelihood. A predicted tissue metabolite score $\hat m_j = \sum_k \pi_{jk}\, g(x_{jk}, e_{jk})$ is learned through a monotone link $g$. The cross-patient ordering of $\hat m_j$ must match the observed ordering under a Plackett–Luce model, as in UnitedMet. Training maximizes this rank likelihood subject to the pool constraint, with the flux network shared across cell types and patients.
The output is a per-patient, per-cell-type contribution $\pi_{jk}\, g(x_{jk}, e_{jk})$ to each metabolite. Three readouts follow. Differential metabolites between patient groups are decomposed into cell-type shares. Pathway enrichment is run per cell type on attributed flux changes, using both base and MOMENTA-expanded gene sets. A metabolite communication score is computed as producer secretion $\pi_{jk} e_{jk}$ times receiver sensor expression. It uses a small hand-curated table of metabolite-sensor pairs, such as succinate and its receptor and lactate and its receptor, and we restrict it to pairs with literature support.
Research Program
Work starts by building the coupled model on Recon3D, as used by GEFMAP. The model is restricted to central carbon, amino-acid and tryptophan metabolism, so that the exchange-flux layer stays identifiable. Count Bridges and Cell2location are interchangeable front ends, which lets us test whether attribution depends on the deconvolver. Next we build two attribution benchmarks with known ground truth. These benchmarks are the field's main missing resource, because no dataset in the file has cell-type-resolved metabolite measurements from real tissue. Last, we apply the model to the clear-cell renal carcinoma cohorts and ask whether attribution changes biological conclusions.
Baselines
The naive reference point is standard bulk metabolite pathway enrichment: hypergeometric over-representation on differential metabolites mapped to KEGG, with no cell attribution. It defines what an analyst gets today. It cannot attribute anything, so it serves only to show which pathway calls change once attribution is added.
The strongest competitor for attribution is fraction regression. For each metabolite we regress the measured level (or its rank-normal transform) on deconvolved cell fractions across patients, with non-negative coefficients. Each coefficient times the fraction is read as that cell type's contribution. This is the obvious statistical answer and it needs no metabolic model. We run it on identical fractions from the same deconvolver, with the same train/test patient splits and the same hyperparameter budget, using ridge penalty selection by inner cross-validation. To beat it, our method must show significantly higher per-metabolite correlation between inferred and true cell-type contributions on the mixture benchmarks. That comparison is paired across metabolites with a Wilcoxon test, on cohorts of 50 or fewer samples or with at least two transcriptionally similar cell types. If fraction regression ties us at large sample sizes, we will report that.
For held-out metabolite prediction, the strongest existing method is UnitedMet. It reports successful imputation of 48% to 67% of metabolites, measured as significant held-out Spearman correlation across four cross-validation datasets (UnitedMet). We rerun it on the same CPTAC and MSKCC splits. Our model is not expected to beat it at bulk prediction. It must stay within a few percentage points of the fraction of well-predicted metabolites, because an attribution model that predicts bulk data badly should not be trusted. For the flux layer, GEFMAP reports Pearson correlation of 0.928 on its synthetic set and 0.841 on augmented E. coli (GEFMAP). We use its architecture rather than competing with it.
Two further baselines are expression-only. The first is deconvolution followed by per-cell-type enzyme gene-set enrichment, using base and MOMENTA neighborhood sets. The second is a CellChat-style metabolite communication score in which secretion is proxied by the expression of the final biosynthetic enzyme. Both are run on the same deconvolved profiles.
Two ablations isolate the key ingredient. The decisive one removes the tissue pool constraint, which leaves independent per-cell-type flux models fitted to the same rank likelihood. The second replaces the rank likelihood with a Gaussian likelihood on log abundances. It tests whether the rank layer is what makes the model survive cross-cohort batch effects.
Why not just…?
Why not just regress each metabolite on cell fractions? This is the real threat. With many samples and well-separated cell types it may match us. It has two specific weaknesses. It needs fractions to vary independently across patients, which fails when cell types co-vary, and the deconvolution benchmark already shows that related subtypes are where fractions are least reliable (benchmark). It also cannot represent a cell type that lowers a metabolite by consuming it, except as a negative coefficient that the non-negative version forbids and the unconstrained version makes noisy. The mixture experiments sweep sample size and inter-type similarity, so they show exactly where, if anywhere, the constraint pays off. If no such regime exists, the proposal's core claim is wrong.
Why not just run per-cell-type gene-set enrichment on deconvolved expression? This asks whether a cell type expresses the enzymes, not whether it moves the metabolite. MOMENTA's 25-fold gain from network expansion (MOMENTA) shows how much enzyme-only sets miss even in pure populations. UnitedMet's ceiling of 48% to 67% predictable metabolites shows how loosely expression tracks metabolite levels. We include both gene-set variants as baselines. If they match our attribution accuracy on the cell-line mixtures, the metabolic model is unnecessary.
Why not just add metabolite-sensor pairs to CellChat with enzyme expression as the ligand proxy? That approach never sees a measured metabolite, so it cannot distinguish a cell that expresses a biosynthetic enzyme from one that actually exports the product. CellChat's own 10–14% false-positive reduction from modeling complexes (CellChat) shows how sensitive such scores are to proxy choice. The exchange simulation below gives ground truth for producer identity, and we test the proxy against it directly.
Proposed Experiments
The first benchmark is a cell-line mixture. Using Cancer Cell Line Encyclopedia lines with both RNA and metabolomics (MOMENTA describes 225 metabolites across 928 lines), we form in silico mixtures of three to six lines with Dirichlet-sampled fractions. Bulk RNA and metabolite abundance are summed in proportion. True per-line contributions are therefore known. We vary the number of mixtures from 20 to 500 and create hard settings by mixing lines of the same lineage. Metrics are per-metabolite correlation between true and inferred contributions and top-source accuracy. This benchmark has a weakness: linear mixing contains no exchange between cells, so it favors fraction regression.
The second benchmark addresses that weakness. It is a semi-synthetic exchange simulation in which coupled flux balance problems on Recon3D are solved jointly for line-specific models. One cell type secretes lactate or succinate and another takes it up, which yields bulk pools shaped by consumption. Here the coupling ablation is decisive. If removing the pool constraint does not hurt attribution of consumed metabolites, the mechanism adds nothing.
On real tissue, we run held-out metabolite prediction against UnitedMet on the CPTAC and MSKCC renal cohorts. Where the in vivo isotope tracing data in those cohorts allow, we check whether predicted tumor-cell flux directions agree with labeling patterns. The clinical check uses IMmotion151 (n=823). UnitedMet reports that imputed succinate and malate associate with shorter progression-free survival under sunitinib at $P<10^{-4}$. We ask whether that association concentrates in one attributed compartment, tumor or stroma, or spreads evenly. This is falsifiable but is not ground truth, so it is reported as a consistency test, not as validation.
The proposal succeeds if three results hold together. The pool-constrained model beats fraction regression in the small-sample and collinear mixture regimes. The coupling ablation degrades attribution of consumed metabolites in the exchange simulation. Held-out bulk prediction stays near UnitedMet. If only the third holds, we have rebuilt bulk imputation with extra parameters, and we will say so.
References
- Inference and analysis of cell-cell communication using CellChat
- Modeling and inference of spatial intercellular communications and multilayer signaling regulations using stMLnet
- Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data
- Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning
- UnitedMet harnesses RNA–metabolite covariation to impute metabolite levels in clinical samples
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
- Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer