When a tumor biopsy is sent for RNA sequencing, the assay grinds hundreds of thousands of cells — cancer cells, macrophages, T cells, fibroblasts — into a single averaged molecular profile. Computational tools can already unmix that average back into cell types: how many T cells, how many macrophages. But they treat every macrophage as the same macrophage. In a living tumor this is wrong in a way that matters clinically: whether a cell burns glucose, burns fat, or uses its mitochondria is rewritten by its local surroundings — how stiff the surrounding matrix is, how much oxygen reaches it, whether it is in physical contact with a tumor cell — and those metabolic choices determine whether immune cells get suppressed and whether drugs work. Because existing unmixing tools assume each cell type has one fixed molecular signature, a cell whose metabolism has been rewired by its neighborhood is misread as a different cell, or missed entirely.
We propose to fix the reference library rather than the algorithm alone. Patient-derived tumor organoids — miniature tumors grown in gel from a patient's own cells, which retain up to 96% of the original tumor's mutational spectrum (Li et al.) — can be co-cultured with the same patient's immune cells under controlled conditions: soft versus stiff matrix, normal versus low oxygen, direct contact versus separation. Sequencing single cells from these cultures yields a training set in which every cell carries both a known identity and a known imposed metabolic condition. From this we build an unmixing model that reads an archival bulk sample and reports, for each cell type, its metabolic state — and we test whether that recovered information predicts therapy response where existing bulk markers fail.
Core Claim
The claim is that metabolic states, not just cell lineages, can be recovered from bulk tumor transcriptomes — provided the reference atlas is built from perturbed microenvironments rather than from unperturbed single-cell atlases. Formally, for a bulk sample with integer gene counts $y_g$, we model
$$y_g \sim \mathrm{Poisson}\Big(\sum_{k,s} \pi_{ks}\,\mu_{g,ks}(c)\Big),$$
where $\pi_{ks}$ is the unknown fraction of cells of lineage $k$ in metabolic state $s$ (glycolytic, oxidative, fatty-acid-oxidizing), and $\mu_{g,ks}(c)$ is the expected expression of gene $g$ for that lineage–state pair under microenvironmental covariates $c$ such as matrix stiffness and oxygen tension. The unknowns $\pi_{ks}$ are what we want. The two ingredients are new: the joint lineage-by-state latent space, and the conditioning of reference profiles $\mu$ on experimentally controlled covariates $c$.
Prior Work and the Specific Hole
Three lines of work approach this intersection and each stops short. Benchmarks of tumor deconvolution, most comprehensively Aitken et al. with nine algorithms evaluated on 38,000 simulated bulk mixtures from a 100,064-cell breast cancer atlas, established that high tumor purity (>80%) causes systematic misclassification of normal cells as cancer — but the ground truth there is lineage only; no method was scored on whether deconvolved cells were glycolytic or oxidative. POPPY maps bulk clinical profiles into single-cell foundation-model spaces using Earth Mover's Distance over the Cell Ontology tree and achieves zero-shot composition recovery, but that ontology organizes cells by lineage and anatomy and has no nodes for transient metabolic states, and its references are static atlases. Count Bridges supplies the count machinery we reuse — a Poisson birth–death generative bridge over integer reads with Expectation–Maximization, which beat continuous and reference-free baselines on spatial spot deconvolution (root-mean-square error 0.110 versus 0.177 for STDeconvolve; Jensen–Shannon divergence 0.231 versus 0.288) — but it has never been applied to metabolic states or to tumor microenvironments.
On the biological side, three papers motivate the joint design by showing that microenvironmental conditions rewrite metabolism in ways bulk measurements hide. Gao et al. showed that soft extracellular matrix drives laterally spreading colorectal tumors toward oxidative phosphorylation through an ENTPD1–ADORA2B purinergic axis, and noted that their serially diluted Matrigel confounded stiffness with ligand density. Chen et al. showed that cholangiocarcinoma cells physically transfer the lipid arachidonic acid to macrophages through tunneling nanotubes, flipping them from anti-tumor to immunosuppressive. Fu et al. showed that suppressive myeloid cells in esophageal cancer depend on a PKN2–STAT3–CPT1B fatty-acid-oxidation program, and had to resort to physical cell sorting because their bulk microarray (GSE20347) could not tell them which compartment carried the signal. Finally, Bao et al. review the underlying tension: tumors contain co-existing glycolytic and oxidative subpopulations engaged in lactate exchange, yet bulk profiles wash this out, and no existing algorithm resolves it. No paper in any of these three clusters builds a perturbation-conditioned, lineage-and-metabolic-state reference from organoid co-cultures and uses it to unmix clinical bulk data. That is the hole.
Research Program
The program has three parts. First, we build the perturbed atlas. Patient-derived organoids from colorectal and esophageal tumors will be co-cultured with autologous peripheral blood immune cells, following the autologous co-culture designs of Fu et al. and Yu et al., in synthetic hydrogel matrices whose stiffness and ligand density are tuned independently — directly addressing the Matrigel confound Gao et al. identified. Each culture arm imposes a defined condition: soft versus stiff matrix, normoxia versus hypoxia, direct contact versus transwell separation (with tunneling-nanotube blockade by Cytochalasin B in one arm, per Chen et al.). Single-cell RNA-seq of these cultures gives every cell a known imposed condition, which we convert into metabolic-state labels using pathway signatures plus Seahorse flux measurements on matched cultures, so that the state label reflects actual energetics rather than only transcript correlations.
Second, we build the model. We extend the Count Bridges Poisson birth–death bridge so the latent classes are lineage–state pairs rather than lineages, and we learn a small conditioning function that outputs reference profiles $\mu_{g,ks}(c)$ as a function of the imposed covariates, so that deconvolving a hypoxic or soft-matrix tumor uses appropriately rewired profiles instead of a normoxic atlas. Inference over $\pi_{ks}$ stays within the existing Expectation–Maximization machinery, keeping the method reference-based and applicable to any bulk cohort without retraining.
Third, we validate clinically. The esophageal cohort of Fu et al. offers an unusually clean clinical test: independent flow cytometry from 40 patients gives measured fractions of PKN2-expressing suppressive myeloid cells, which correlated with advanced stage ($p=0.022$) — a ground-truth target that bulk analysis alone could not recover. We will ask whether our deconvolved fatty-acid-oxidation-active myeloid fraction reproduces those measurements. We acknowledge the program inherits known organoid limits catalogued by Li et al. — absent vasculature and mature stroma, and batch variability — which is why the conditioning function matters: it must interpolate between culture conditions, and a dedicated experiment tests where it fails.
Baselines
The naive reference point is a sample-level metabolic pathway score: sum or average the expression of glycolysis, oxidative-phosphorylation, and fatty-acid-oxidation gene sets directly on the bulk profile, with no unmixing at all. This is what most retrospective metabolic analyses of archival cohorts actually do, and it is the null hypothesis: any per-lineage, per-state estimate must add predictive power beyond it, for instance by separating laterally spreading from protruding colorectal tumors on Gao et al.'s cohorts (GSE161277, GSE261388) where a bulk oxidative-phosphorylation score is confounded by how many epithelial cells the sample happens to contain.
The strongest existing competitor is BayesPrism, which Aitken et al. identified as the most accurate and purity-stable deconvolver across 38,000 pseudo-bulk mixtures; the exact RMSE values for BayesPrism and DWLS in that benchmark are not numerically reported — no reported result found on Lacuna for those figures — so we will reproduce it ourselves on our mixtures rather than compare to a published number. We will run BayesPrism twice: once with its standard lineage references and post-hoc metabolic pathway scoring of its output (the strongest simple version of the competing approach), and once in a reference-swap configuration where we hand it our perturbed organoid atlas. All methods get the same input mixtures, the same gene universe, and the same tuning budget. Beating it means lower error on state proportions in ground-truth pseudo-bulk mixtures — especially at tumor purity above 80%, where the purity trap bites and where rare fatty-acid-oxidizing suppressive myeloid cells must still be recovered — and correlation with the 40-patient flow-cytometry fractions in the esophageal cohort where post-hoc scoring does not correlate.
Two further comparisons bracket the design space. POPPY will be run zero-shot on the same mixtures to test whether ontology-aware cross-modal transfer alone recovers metabolic structure; it reported up to 73% silhouette-score improvement and a 714% increase in composition-similarity correlation on its lineage tasks, so it is a serious contender for composition but is architecturally blind to states. And a continuous flow-matching deconvolver built on the TorchCFM library will be trained on identical data, testing whether the discrete count machinery matters here as it did in Count Bridges' benchmarks, where the count-based model beat STDeconvolve with RMSE 0.110 versus 0.177 and Jensen–Shannon divergence 0.231 versus 0.288. Finally, one ablation isolates the proposal's key ingredient: the same count-based model, run with lineage-only references from unperturbed organoid cultures. If the perturbed, joint-state atlas is doing the work, this ablation should collapse specifically on state recovery while remaining adequate on lineage recovery.
Why not just…?
Why not just run single-cell or spatial sequencing on everything? For prospective studies this is partially right, and we depend on it: single-cell RNA-seq of our organoid cultures is how the atlas gets built. But it cannot answer the archival question. Hundreds of thousands of stored clinical trial samples, including the 330 melanoma samples across six trials used by POPPY and the 1,038 TCGA breast samples used by Aitken et al., exist only as bulk profiles and cannot be re-sequenced cell by cell. Dissociation also destroys exactly the structures this field cares about — Chen et al. note that tunneling nanotubes are sheared during tissue processing — and common spatial platforms still aggregate 10 to 50 cells per spot, requiring deconvolution anyway. The threat is real for new trials, and Experiment 5, which checks our deconvolved states against organoid-level single-cell ground truth, is what quantifies the accuracy we give up by going computational.
Why not just run BayesPrism and score metabolic gene sets on its cell-type fractions? This is the alternative that could sink the proposal, and we treat it as the primary baseline rather than arguing around it. It falls short for a reason the literature documents: pathway scores computed on a deconvolved fraction conflate state with residual lineage signal, because the reference profile a metabolically rewired cell is matched against comes from an unperturbed atlas, so the rewiring leaks into the lineage assignment itself. Aitken et al. showed how badly lineage assignment degrades at high purity, and Fu et al. showed that bulk-level metabolic signals cannot be attributed to a compartment without physical sorting. Our reference-swap experiment is the honest test: if BayesPrism plus post-hoc scoring matches joint-state estimates on ground-truth mixtures at all purity levels, the simpler route wins and we will report that.
Why not just build the reference from standard 2D co-cultures or conditioned media? Because the states we most want to capture are the ones 2D systems cannot produce. Chen et al. showed the lipid transfer that reprograms macrophages requires direct physical contact, impossible across a transwell; Gao et al. showed matrix stiffness itself is the metabolic stimulus, and tissue-culture plastic is orders of magnitude stiffer than any tumor matrix; and Li et al. showed 2D cell lines lose the patient's clonal architecture that shapes patient-specific metabolic rewiring. A 2D-derived atlas would be cheaper and more standardized, but it would systematically miss precisely the stiffness-dependent and contact-dependent metabolic states that motivated the work, and our stiffness- and contact-arm ablations will show directly how much each perturbation arm contributes to downstream recovery.
Proposed Experiments and Decisive Checks
Experiment 1 establishes methodological footing by reproducing the Count Bridges pipeline on its published benchmarks — the Onek1k peripheral-blood dataset and MERFISH/10X Visium spatial sections — confirming we can match the reported spot-deconvolution performance before extending the latent space. Experiment 2 is the core ground-truth test: pseudo-bulk mixtures simulated from our perturbed organoid single-cell data, swept across tumor purity from 5% to 95% following the SMOTE design of Aitken et al., with recovery error measured separately for lineage proportions and for state proportions. The decisive outcome is state-proportion error that stays controlled above 80% purity where the baseline methods' lineage errors cascade into state errors.
Experiment 3 tests the conditioning function: deconvolving the laterally-spreading versus protruding-adenoma cohorts of Gao et al., where the correct answer — an oxidative-phosphorylation shift in the epithelial compartment of soft-matrix tumors — is known in advance. Experiment 4 is the clinical anchor: recovering the fatty-acid-oxidation-active suppressive myeloid fraction from the esophageal bulk cohort and testing correlation with the 40-patient flow-cytometry measurements of Fu et al.. Experiment 5 tests clinical utility: whether deconvolved lineage-by-metabolic-state features predict checkpoint-inhibitor response in the 330-sample melanoma cohorts better than bulk metabolic scores and lineage-only features, with SubNetDL's median AUROC of 0.74 on TCGA and 0.77 on MSKCC immunotherapy cohorts as the reference point for what masked-heterogeneity features have achieved for response prediction; no reported result found on Lacuna for a directly comparable metabolic-state-based predictor, which is itself the gap. A prospective addendum tests mechanism: whether the deconvolved oxidative-phosphorylation fraction in individual organoid lines predicts their growth arrest under the OXPHOS inhibitor Gboxin, the response Gao et al. demonstrated. If the joint-state model wins only on lineage and not on state, or only under low purity, the perturbed-atlas hypothesis is wrong and the simpler post-hoc approach should stand.
References
- The advancements in organoids: Potential and challenges in researching the esophagus and esophageal squamous cell carcinoma
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- A Soft Matrix Microenvironment Promotes Laterally Spreading Tumors via Oxidative Phosphorylation‐Dependent Cell Adhesion
- Role of Tunneling Nanotubes in Arachidonic Acid Transfer and Macrophage Function Reprogramming in Intrahepatic Cholangiocarcinoma
- PKN2 enhances the immunosuppressive activity of polymorphonuclear myeloid-derived suppressor cells in esophageal carcinoma by mediating fatty acid oxidation
- Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development
- D-xylose suppresses hepatocellular carcinoma progression by regulating dihydrodiol dehydrogenase and remodeling the immune microenvironment
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer