Subclone-Resolved Metabolic Fluxes from Bulk Tumors Predict Chemotherapy Escape

Most of what we know about why chemotherapy stops working comes from measurements that average a whole tumor into one number. Bulk sequencing mixes millions of cells into a single composite readout, and this averaging hides exactly the cells that matter: the rare, plastic subpopulations that survive treatment. A growing body of work shows these survivors often escape not by acquiring new driver mutations but by rewiring metabolism — switching fuel sources, slowing down energy consumption, or rebuilding lipid membranes to block iron-mediated cell death. Yet the tools we have sit in separate silos: subclonal deconvolution models recover genetic subpopulations from bulk data but ignore metabolism entirely, while metabolic flux models need clean single-cell inputs that clinical bulk samples never provide.

This proposal builds a single model that closes the loop. The idea is to deconvolve bulk tumor transcriptomes directly into subpopulation-specific metabolic fluxes — descriptions of how fast each latent subpopulation runs each chemical reaction — and to use those fluxes to predict chemotherapy response and to diagnose why resistance arose: clonal selection of a pre-existing genetic lineage, or non-mutational metabolic plasticity spread across cells. The question matters clinically, because the two failure modes call for different interventions: an expanding resistant clone demands a targeted attack on that lineage, while a plastic metabolic state demands combination therapy that closes the escape route, as already shown for quiescent EGFR-inhibitor escape in lung cancer, where adding a BH3 mimetic or glutamine deprivation eradicated the tolerant cells (Transcriptomic-metabolomic reprogramming in EGFR-mutant NSCLC early adaptive drug escape linking TGFβ2-bioenergetics-mitochondrial priming).

Core Claim

The central technical claim is that metabolic stoichiometry makes bulk deconvolution tractable in a way that raw gene-expression decompositions are not. Standard deconvolution asks a latent variable model to recover per-subpopulation gene counts — thousands of underdetermined dimensions from one bulk profile. Instead, we parameterize each latent subpopulation by its intracellular flux vector $v_k$ over a curated reaction network, decoded into expected transcript counts through an encoder trained on single-cell data, with the bulk profile explained as

$$x \;\approx\; \sum_k \pi_k \, g(v_k), \qquad S v_k = 0 \;\; \text{for each } k,$$

where $\pi_k$ is the fraction of subpopulation $k$, $g$ maps fluxes to expected expression, and $S v_k = 0$ enforces mass balance ($S$ is the stoichiometric matrix of the metabolic network, as in GEFMAP). The stoichiometric null space is far smaller than gene space, which regularizes an otherwise ill-posed decomposition, and it forces every inferred subpopulation to be biochemically realizable — no subpopulation can "run" a reaction set that violates conservation of mass.

The second ingredient is a two-channel decomposition that couples this metabolic channel to a genetic channel: variant-allele-frequency clustering of the kind used by SubNetDL (A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer) identifies genetic subclones from bulk DNA, and the model learns whether the resistant flux state co-segregates with an expanding genetic clone or is distributed across clones. This co-segregation signal is what turns a prediction tool into a diagnostic: clonal selection shows up as a flux state riding on a VAF-shifted lineage; metabolic plasticity shows up as a flux state appearing across clones with no VAF shift. SubNetDL cannot make this distinction because its subclones are defined purely by driver mutations propagated on a protein-interaction network — it structurally cannot see a resistant subpopulation defined by, say, lipid-remodeling flux, even though such enzymes are precisely the engines of drug escape. Conversely, the ACOT8 study proved gemcitabine resistance in pancreatic cancer is executed by a lipid-remodeling enzyme that depletes peroxidation-prone lipids and blocks ferroptotic death (Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma) — but it analyzed bulk tumors and bulk lipidomics, so it cannot say whether that resistance lives in a specific subpopulation or in every malignant cell. Our model answers exactly that question, and the ACOT8 and TGFβ2 papers supply the ground-truth mechanisms we will try to recover unprompted.

Research Program

Three components. First, the generative deconvolution core: we extend the discrete latent-variable machinery of Count Bridges, which models integer count data with birth–death dynamics and Expectation–Maximization and already outperforms CIBERSORTx and MuSiC at recovering cell-type proportions from bulk RNA (Count Bridges enable Modeling and Deconvolving Transcriptomic Data), so that its latent single-cell profiles are parameterized by flux vectors rather than gene counts. The flux decoder reuses GEFMAP's design — a graph neural network that infers each subpopulation's metabolic objective from enzyme-weighted reaction graphs, followed by a null-space-projected network that predicts fluxes satisfying $S v = 0$ — pre-trained on single-cell atlases such as the curated cancer cell atlas used by POPPY, which contains over 2.7 million cells across 15 cancer types (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level). Second, the genetic channel: SciClone-style VAF clustering per patient, with the model learning a joint assignment between flux states and genetic subclones. Third, downstream heads: a drug-response head trained on TCGA and CCLE/GDSC response labels, and a combination-suggestion head that proposes which metabolic pathway to co-inhibit given the inferred escape flux.

The scientific payoff is not just better AUROC. It is the ability to take any bulk-profiled tumor and report: the resistant metabolic state exists or does not, its flux character (glycolytic, lipogenic, glutamine-dependent, quiescent), and whether it is being carried by an expanding clone or induced across cells. That directly addresses the unresolved conflict in the metabolism literature — whether treated tumors fail through elevated glycolysis or through metabolically quiescent survival (Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development, Transcriptomic-metabolomic reprogramming...) — which bulk profiling cannot separate because symbiotic and divergent states average into one ambiguous net signal.

Proposed Experiments

Experiment 1 (pseudo-bulk validation). We build pseudo-bulk mixtures from the scDrugMap single-cell collections — 326,751 cells across 36 datasets and 23 studies, plus 18,856 validation cells (scDrugMap: benchmarking large foundation models for drug response prediction) — so ground-truth per-cell-type flux states are known by running the GEFMAP decoder directly on the single-cell profiles. Success is measured as Pearson correlation between deconvolved and true fluxes, with the direct single-cell prediction as the ceiling. GEFMAP reports flux prediction PCC of 0.928 ± 0.016 on synthetic networks and 0.841 ± 0.014 on E. coli networks (GEFMAP); we target deconvolved flux recovery above roughly 0.7 on the human network from pseudo-bulk mixtures at realistic admixtures, and we measure composition error against CIBERSORTx, MuSiC, and Count Bridges run on identical mixtures.

Experiment 2 (response prediction). We replicate SubNetDL's evaluation exactly: 507 patients across 10 cancer–drug pairs in TCGA, six cancer–drug pairs in CCLE/GDSC, and the MSK NSCLC immunotherapy cohorts (MSK_2020, n=242; MSK_2018, n=240). The numbers to beat are median AUROC 0.74 on TCGA, 0.81 on CCLE/GDSC, and 0.77 on MSK, all reported by SubNetDL. We additionally run SubNetDL's own perturbation control on our model — randomly shuffling subpopulation assignments should degrade performance, confirming the subpopulation-resolved signal is real.

Experiment 3 (mechanism recovery, gemcitabine in pancreatic cancer). On the Fudan cohort of 28 gemcitabine-sensitive and resistant patients and the PAAD_GSE148673 cohort used by the ACOT8 study, we test whether the model attributes resistance to antiferroptotic lipid flux without any ACOT8 hint. The falsifiable outcome: resistant patients show either a lipogenic resistant subpopulation (selection-like) or uniform lipogenic flux across all malignant subpopulations (plasticity-like). The ACOT8 study reports that knockdown drops MIA PaCa-2 gemcitabine IC50 from roughly 60 nM to 10–15 nM, and orlistat co-treatment restores sensitivity in organoids and xenografts; our combination head should independently rank gemcitabine-plus-lipase-inhibition above non-metabolic combinations for flux-typed resistant lines.

Experiment 4 (plasticity versus selection). On the EGFR-mutant NSCLC perturbation data — HCC827 and H1975 cells treated with erlotinib or CL-387,785 — the model should detect a suppressed-glycolysis, suppressed-TCA, glutamine-dependent quiescent flux state within days of drug exposure, co-occurring across genetic subclones with no VAF shift, matching the mechanism established by the TGFβ2 study. The tamoxifen-resistant breast cancer models, where resistance tracks a rare stem-like subpopulation (Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance...), serve as the selection-positive control where the resistant flux state should co-segregate with one lineage.

Experiment 5 (combination prediction). On the O'Neil synergy dataset of 23,052 drug combinations across 39 cell lines, we test whether adding inferred flux-state features to the multimodal graph convolutional network improves synergy ranking; that model reports mean AUC 0.84 and mean accuracy 0.91 (Deep graph embedding for prioritizing synergistic anticancer drug combinations). No reported result found on Lacuna for flux-feature-augmented synergy prediction, so this comparison is genuinely open.

Baselines

The naive reference point is aggregate bulk biomarkers: tumor mutational burden and bulk-expression classifiers (elastic-net logistic regression on the same TCGA profiles). This is what most clinical pipelines use today, and SubNetDL already quantified its weakness: it significantly outperformed TMB on both MSK cohorts ($P = 0.02$ for MSK_2020; $P = 6.0\times10^{-3}$ for MSK_2018) and reclassified dozens of TMB-indicated false positives (SubNetDL). We will run these baselines on identical patient splits with the same tuning budget; beating them means higher median AUROC on the 10 TCGA pairs with confidence intervals excluding overlap.

The strongest existing competitor is SubNetDL itself, because it is the only published model that both unmasks subclonal structure from bulk clinical data and predicts therapy response on these exact benchmarks, reporting median AUROC 0.74 (TCGA), 0.81 (CCLE/GDSC), and 0.77 (MSK). We will retrain or use its published setup on the same 507 TCGA patients, same CCLE/GDSC pairs, and same MSK cohorts, with equal compute and hyperparameter search, and we count success as exceeding its median AUROC on at least two of the three benchmark families while matching it on the third. Its driver-gene focus is also its blind spot — it prunes its network to 253 canonical cancer drivers, which excludes metabolic enzymes like ACOT8 and PDHK1 by construction — so the mechanistic experiments (3 and 4) provide a second, orthogonal win condition it cannot contest.

The most important baseline is the key-ingredient ablation: the chained pipeline. We run Count Bridges to deconvolve bulk profiles into latent single-cell expression profiles, feed each through GEFMAP to get fluxes, then train the same response head on those flux features. This isolates the joint stoichiometric coupling — the single model in which decomposition and flux inference constrain each other — from the naive composition of existing parts. If the chained pipeline matches the joint model on Experiments 1–3, the joint design is not earning its complexity, and we will say so and report it; our prior is that it will not, because the chained decomposition operates in unconstrained gene space where the ill-posedness is worst and errors compound through each stage. A fourth reference is the single-cell upper bound: scDrugMap reports that scFoundation achieves mean F1 of 0.971 and 0.947 under pooled-data layer freezing and scGPT reaches mean F1 of 0.858 in tumor tissue under cross-data evaluation (scDrugMap); we report our pseudo-bulk results against these numbers as an honest ceiling, since our model works from bulk mixtures where that information has already been destroyed.

Why not just…?

Why not just chain the existing tools? This is the most serious objection, and the honest answer is that the chained pipeline is not a strawman — it is a real alternative, which is why it appears as the key ablation rather than a dismissal. Three concrete reasons predict it falls short. First, the chained decomposition must recover thousands of latent gene-count dimensions from one bulk profile with no biochemical constraint, while the joint model decomposes directly in the much smaller stoichiometric null space. Second, therapy-induced flux states are often absent from reference atlases, a limitation POPPY itself identifies as a failure mode of attribution when key transient states are missing from the reference (Cross-modal transfer learning...); a joint model trained end-to-end for response can carve latent subpopulations that the reference-guided chain never proposes. Third, the chain has no mechanism for the VAF–flux co-segregation analysis at all, so it cannot run Experiment 4. But if the chain matches the joint model empirically, the experiments settle it against us, and we will report that plainly.

Why not just use bulk biomarkers or bulk expression signatures? Because the averaging that produces them provably destroys the predictive signal. SubNetDL showed that randomly shuffling mutations across inferred subclones significantly degrades prediction, meaning the subpopulation-resolved distribution itself carries signal that any bulk aggregate discards, and it beat TMB with $P$-values of 0.02 and $6.0\times10^{-3}$ on two independent clinical cohorts (SubNetDL). The same argument applies to metabolic states: a tumor that is half glycolytic and half quiescent averages into a signal indistinguishable from a uniformly intermediate tumor.

Why not just run standard deconvolution and read out metabolic gene expression? Because expression is a poor proxy for flux. Enzyme transcript levels do not map linearly to in vivo flux — post-translational control like the succinylation that stabilizes glutaminase against degradation changes flux with no transcript change at all (BAG3 promotes autophagy and glutaminolysis via stabilizing glutaminase) — which is the central discordance GEFMAP was built to address (GEFMAP). Standard tools also stop at cell-type proportions, and even the strongest of them are already outperformed at that task by Count Bridges (Count Bridges...); none of them output fluxes, so the resistant-state characterization we need simply does not exist in their output.

Why not just run flux balance analysis with a biomass objective on the deconvolved profiles? Because biomass maximization assumes the cell is trying to grow, and the most dangerous chemotherapy-tolerant cells are precisely the ones that have stopped growing: the EGFR-inhibitor escape state is bioenergetically quiescent with suppressed glycolysis and TCA activity (Transcriptomic-metabolomic reprogramming...). GEFMAP showed that inferring per-cell objectives rather than imposing biomass reveals these non-growth states, and its objective-inference network outperformed biomass-only FBA on synthetic and bacterial networks. Our proposal inherits that per-subpopulation objective inference rather than fixing one objective for all cells.

Decisive Checks

Four results would establish the claim. First, exceeding SubNetDL's median AUROC of 0.74 on the TCGA pairs and 0.81 on CCLE/GDSC under matched conditions. Second, recovering the ACOT8 antiferroptotic lipid mechanism in gemcitabine-resistant pancreatic tumors without supervision, with the co-segregation analysis correctly reporting whether it is clonal or plastic. Third, correctly labeling the EGFR-inhibitor escape as non-mutational plasticity and the tamoxifen case as stem-like selection, using flux–VAF co-segregation as the discriminator. Fourth, the ablation showing the joint stoichiometric model beats the chained pipeline, confirming that the coupling — not just the added metabolic features — is what carries the improvement. If the first and fourth checks fail while the second and third succeed, the project still delivers something valuable: the first instrument for attributing bulk-level drug resistance to specific metabolic escape routes, even if its predictive edge over simpler pipelines turns out to be modest.