From Fingerprint to Flux: Drug-Grounded Metabolic Prediction in Single Cells

When a small molecule enters a cell, it does not uniformly nudge twenty thousand genes; it binds particular enzymes and diverts carbon flow through specific metabolic pathways. Today's single-cell perturbation models ignore this. They compress a drug's structure into a molecular fingerprint — a fixed bit vector recording which chemical substructures a molecule contains — and then let a generative model smear that fingerprint's influence across the whole transcriptome through an attention mechanism. Meanwhile, metabolic-network models can convert a single cell's gene expression into physically consistent reaction fluxes — the rates at which metabolites flow through each step of metabolism — but they have never been given a drug as an input. Neither family of models can answer the question a pharmacologist actually asks: how will this molecule rewire metabolism, cell by cell?

The question matters because the cells that matter most are the ones that averaging hides. Bulk assays pool millions of cells into one measurement and thereby erase rare subpopulations whose metabolic rewiring drives therapy resistance — stem-like cells surviving tamoxifen 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, or lipid-rewired gemcitabine-resistant pancreatic cancer cells Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma — a masking problem that motivates single-cell measurement in the first place Genome 2011. We propose to build the first model that predicts a drug's metabolic effect at single-cell resolution, with the drug's fingerprint decoded into inhibition of named enzymes and the predicted response constrained to be chemically possible.

Core Claim

The mapping from (untreated cell, molecule) to treated cell should pass through the metabolic network, because that is how drugs actually work. Concretely, we will train a model on the Tahoe-100M perturbation atlas Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling that takes an unperturbed single-cell expression profile and a drug's Morgan fingerprint and outputs, for that cell, a post-treatment flux vector that satisfies mass balance by construction, together with the gene-expression changes needed for benchmark comparison. Three design choices distinguish this from everything in the literature. First, the fingerprint is grounded: a learned decoder assigns an inhibition weight to each metabolic reaction, mirroring how real compounds shut down gatekeeper enzymes such as pyruvate dehydrogenase kinase, which dams pyruvate out of the TCA cycle Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development or hexokinase 2, which traps glucose inside the cell Elevated expression of glycolytic genes as a prominent feature of early-onset preeclampsia: insights from integrative transcriptomic analysis. Second, the output is physically constrained: predicted fluxes live in the stoichiometric null space, so the model cannot hallucinate biochemically impossible states. Third, the model is allowed to change the cell's metabolic identity rather than pinning it to its baseline manifold — because stressed cells genuinely do shift between glycolytic and oxidative states Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development and pass through hybrid intermediate states Intermediate cell states in epithelial-to-mesenchymal transition.

The closest existing work makes the contrast sharp. scPPDM conditions a latent diffusion model on Morgan fingerprints to generate perturbed single-cell profiles, but operates in an unconstrained statistical latent space with no metabolic pathway structure, and explicitly forces retention of baseline cell identity through a classifier-free guidance knob scPPDM: A Diffusion Model for Single-Cell Drug-Response Prediction — the exact opposite of our third design choice. GEFMAP infers single-cell fluxes with stoichiometric null-space constraints, achieving Pearson correlations of 0.928 on synthetic networks and 0.841 on an E. coli network, but takes no chemical input of any kind and was validated only on unperturbed differentiation trajectories Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning. PerturbDiff models treated-cell distributions with a kernel-embedding diffusion objective that handles sparsity well, but treats the transcriptome as an unstructured space with no pathway topology PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling. We borrow one ingredient from each — fingerprint conditioning from scPPDM, the null-space parameterization from GEFMAP, and the distributional loss from PerturbDiff — and add what none of them has: a learned chemical-to-enzyme decoder sitting between the fingerprint and the metabolic network.

Research Program

The model has three coupled components. The reaction-level drug decoder $g_\phi$ maps a drug's fingerprint $m$ to a vector of inhibition weights $a$, one per metabolic reaction, so $a = g_\phi(m)$. The flux head predicts a per-cell flux vector $v$ in the null space of the stoichiometric matrix $S$ of a human metabolic network (Recon3D, as used by GEFMAP on human embryoid-body data Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning), with drug action entering as capacity restrictions:

$$v = N\,z,\qquad 0 \le v_i \le u_i\,(1 - a_i),\qquad z = f_\theta(x,\,a)$$

Here $N$ is an orthonormal basis of the null space of $S$, so mass balance $Sv = 0$ holds by construction; $x$ is the untreated cell's expression profile; $u_i$ is reaction $i$'s capacity, set by the cell's own enzyme abundance; and $f_\theta$ is the network predictor. The decoder $g_\phi$ is trained end-to-end but weakly supervised: for drugs in Tahoe-100M with known enzymatic targets, inhibition weights on the corresponding reactions are anchored to those targets, and for the rest the model must infer the target from the fingerprint alone — which is precisely the generalization we want to test. Finally, an expression head decodes $(x, v, a)$ into predicted perturbed expression, so the model can be scored on the same differentially-expressed-gene metrics as existing benchmarks. Training uses a distributional objective — energy distance in the spirit of the scPerturb E-distance statistic scPerturb: Harmonized Single-Cell Perturbation Data or the MMD-style loss of PerturbDiff — rather than mean squared error, because more than 90% of single-cell expression entries are zero and squared-error losses collapse onto shared zeros PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling.

A second workstream quantifies what bulk averaging destroys. For each drug and cell line we will pseudo-bulk the predicted and measured single-cell flux states, then measure how much within-line metabolic divergence between subpopulations disappears in the aggregate — connecting our predictions to the heterogeneity-masking literature, which shows that bulk profiles cannot separate whole-population shifts from rare resistant clones 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, and that deconvolution methods recover coarse cell types but not intra-lineage metabolic states Cross-modal transfer learning for mapping bulk transcriptomes at cellular level Count Bridges enable Modeling and Deconvolving Transcriptomic Data.

Baselines

The naive reference point is the no-change baseline: predict that treated cells are simply a fresh sample of the untreated cells. This is the standard floor on perturbation benchmarks — it scores zero predicted effect, so any method must beat it on effect-size metrics like E-distance before its more complex machinery is worth discussing; we will report it on every metric.

The strongest existing competitor is scPPDM scPPDM: A Diffusion Model for Single-Cell Perturbation Prediction, a latent diffusion model that conditions on Morgan fingerprints via guided decomposable attention with factorized classifier-free guidance, evaluated on Tahoe-100M under Unseen Covariates and Unseen Drugs splits. It reports a +36.11% improvement in DEG Spearman correlation and +34.21% in DEG Pearson correlation over its nearest competitor under Unseen Drugs, plus a +28.57% lead in top-1000 DEG accuracy; no absolute values for these metrics were reported in the material we searched, so we will re-run it ourselves. To make the comparison fair, scPPDM will be trained and tuned on exactly our data — the Tahoe-100M subset of 50 cancer cell lines and 379 drugs it originally used — under the same compute budget and the same hyperparameter search count as our model, and scored with the same evaluation code. Because scPPDM outputs expression profiles rather than fluxes, we will apply our expression-to-flux readout (or GEFMAP itself) to its generated profiles so both methods are compared in flux space as well as gene space. Beating it means matching or exceeding its DEG Spearman and Pearson under Unseen Drugs while additionally producing mass-balanced fluxes that track measured pathway shifts — something scPPDM cannot represent by design. We will also run PerturbDiff, which reports 96% to 100% win rates against its strongest baselines on unseen treatments across Tahoe-100M and PBMC PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling, as a second generative baseline under the same protocol.

The most important baseline is the ablation that isolates our key ingredient: a pipeline that chains scPPDM (generation) with GEFMAP (flux inference) post hoc, with no end-to-end training. If that pipeline matches our model, the end-to-end stoichiometric conditioning earns nothing and the proposal's central claim fails. We will additionally run two cheaper ablations of our own model: one with the null-space head removed and metabolism reduced to mean expression of glycolytic versus oxidative-phosphorylation enzyme genes (a pathway-score model), and one with the drug decoder replaced by scPPDM-style ungrounded fingerprint conditioning into the attention layers. For reference on flux-inference quality, GEFMAP's null-space network reports Pearson correlations of 0.928 ± 0.016 on synthetic networks and 0.841 ± 0.014 on the E. coli network Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning; no reported result found on Lacuna for GEFMAP's flux accuracy on chemically perturbed human cells, which is exactly the regime our model targets.

Why not just…?

Why not just chain scPPDM and GEFMAP as a pipeline? This is the most serious objection, and we treat it as a first-class baseline rather than arguing around it. There are concrete reasons to expect the chain to fall short: GEFMAP was trained and validated on unperturbed developmental trajectories, and generated perturbed profiles may sit off its input manifold; no gradient flows from the fingerprint to the flux, so the pipeline cannot learn that two fingerprint-similar drugs hit different enzymes; and errors compound across stages. But the honest answer is that the decisive experiment settles it — if the chained pipeline matches our model on flux-ratio prediction and on the activity-cliff stress test below, we will report the pipeline as the practical contribution and the end-to-end claim as falsified.

Why not just score pathways from enzyme mRNA? Averaging the expression of canonical pathway genes is cheap and interpretable, and a reviewer will rightly ask why we need a stoichiometric network at all. The evidence against it is that transcript abundance is a poor proxy for reaction rate — post-transcriptional regulation, allostery, and compartmentalization break the correspondence Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning — and that GEFMAP's data-driven objective discovery revealed condition-specific programs (such as ketolysis during neuronal differentiation) that both biomass-maximizing flux balance analysis and naive pathway summaries obscured. Still, we include the pathway-score model as an explicit ablation; if it predicts drug-induced glycolysis/OXPHOS ratios as well as the null-space head, the stoichiometric machinery is unjustified for this task and we will say so.

Why not just use shallow models on fingerprints? Random forests and support vector machines on fixed descriptors repeatedly match or beat deep models on scalar molecular properties Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models, and a study of 62,820 trained models found deep representations only pull ahead beyond roughly 10,000 to 100,000 samples A systematic study of key elements underlying molecular property prediction. Two responses: the output here is not a scalar but a roughly 20,000-dimensional distribution of perturbed expression plus a flux vector, which shallow models do not produce; and Tahoe-100M's 95.6 million quality-controlled cells sit far above the sample threshold where deep models are known to be handicapped. The legitimate scalar version of this objection — a random forest predicting per-condition pathway scores from fingerprints — is folded into our pathway-score baseline and will be run honestly.

Proposed Experiments

All experiments use Tahoe-100M: 100 million single-cell profiles across 1,100 small molecules and 50 cancer cell lines, 95.6 million cells passing quality control, with a reported biological-replicate Pearson correlation of 0.975 Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling. For head-to-head comparability we use the 379-drug subset evaluated by scPPDM; enzyme-coding genes are mapped onto the Recon3D human metabolic network following GEFMAP's procedure. Inclusion of a drug-cell-line condition requires at least 300 treated cells, per the subsampling guidelines established by scPerturb, which also recommend an average of 1,000 UMI counts per cell scPerturb: Harmonized Single-Cell Perturbation Data.

The first experiment reproduces the standard benchmark: predicting differentially expressed genes under Unseen Drugs and Unseen Covariates splits, scored by DEG Spearman and Pearson correlation and top-1000 DEG accuracy, directly against scPPDM and PerturbDiff under matched budgets. The second experiment is the one no existing method can attempt: predicting flux-level metabolic response, scored by correlation between predicted and measured changes in glycolysis-to-OXPHOS flux ratios and by energy distance between predicted and measured treated-cell distributions. The third experiment is mechanism validation on drugs with known metabolic targets: a pyruvate-dehydrogenase-kinase inhibitor should shift pyruvate carbon toward lactate and away from mitochondrial entry, exactly the dam described in the glioblastoma literature Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development, and drugs elevating glucose flux should surface hexosamine-biosynthetic signatures like those marking high-FDG-uptake lung adenocarcinoma Elevated glucose metabolism via the hexosamine biosynthesis pathway: A metabolic signature of high-fluorodeoxyglucose-uptake lung adenocarcinoma. The fourth experiment is the activity-cliff stress test: drug pairs with high fingerprint (Tanimoto) similarity but different enzymatic targets, where fingerprint-based property models are known to smooth over abrupt potency changes A systematic study of key elements underlying molecular property prediction. A grounded decoder should predict divergent metabolic responses for such pairs; an ungrounded attention token should not.

The closing checks are decisive rather than incremental. The pipeline-versus-end-to-end comparison decides whether stoichiometric conditioning must be inside the model. The pathway-score ablation decides whether fluxes earn their place over gene averages. Finally, we will sweep the strength of identity preservation during training and measure whether strong identity anchoring suppresses real drug-induced metabolic transitions, as the plasticity literature predicts it must Hybrid Stem Cell States: Insights Into the Relationship Between Mammary Development and Breast Cancer Using Single-Cell Transcriptomics — and we will quantify, via the pseudo-bulk analysis, exactly how much of this subpopulation-level metabolic divergence a bulk assay would have masked.

References

  1. 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
  2. Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma
  3. As we come to the end of 2011, several members of the Genome Biology Editorial Board give their views on the state of play in genomics
  4. Tahoe-100M : A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling
  5. Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
  6. Elevated expression of glycolytic genes as a prominent feature of early-onset preeclampsia: insights from integrative transcriptomic analysis
  7. Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development
  8. Intermediate cell states in epithelial-to-mesenchymal transition
  9. scPPDM: A Diffusion Model for Single-Cell Drug–Response Prediction
  10. Inferring Metabolic States from Single Cell Transcriptomic Data via Geometric Deep Learning
  11. PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling
  12. scPerturb: Harmonized Single-Cell Perturbation Data
  13. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
  14. Count Bridges enable Modeling and Deconvolving Transcriptomic Data
  15. Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models
  16. A systematic study of key elements underlying molecular property prediction
  17. Elevated glucose metabolism via the hexosamine biosynthesis pathway: A metabolic signature of high-fluorodeoxyglucose-uptake lung adenocarcinoma
  18. Hybrid Stem Cell States: Insights Into the Relationship Between Mammary Development and Breast Cancer Using Single-Cell Transcriptomics