Drugs that work in a plastic dish routinely fail in patients, and one under-examined reason is metabolic: cells in culture run their chemistry differently than cells in a tissue, and cells in a mouse run it differently than cells in a human. Flux balance analysis (FBA) offers a way to reason about this precisely. It reconstructs an organism's full metabolic network as a stoichiometric matrix and asks: given mass balance and physical limits, which distributions of reaction rates (fluxes) are possible? The answer is a linear program, not a simulation of kinetics, which makes it cheap enough to run at genome scale. What FBA cannot currently do is say how the achievable flux space changes when the regulatory wiring around the metabolic genes differs between species, or between a culture flask and a tumor. Meanwhile, the comparative-genomics community has documented that this wiring differs enormously — only half to two-thirds of mouse enhancers and promoters have identifiable human counterparts, and only about half the human genome can even be aligned to mouse (A comparative encyclopedia of DNA elements in the mouse genome) — but it stops at mRNA and chromatin, never projecting divergence down to flux.
The proposal here is to make that projection explicit and testable. Both the in vitro–in vivo shift and cross-species regulatory divergence act on the same mathematical object in FBA: the reaction bounds. The environment (nutrient availability, oxygen, growth demand) sets uptake and capacity bounds; regulation determines whether the enzyme catalyzing a reaction is expressed at all, which is exactly how expression-gating methods like GIMME, iMAT, and E-FLUX constrain genome-scale models (Multi –omics and metabolic modelling pipelines: challenges and tools for systems microbiology). So instead of treating lab-to-clinic transfer as a shift in an unconstrained neural embedding, we model it as a structured shift in the bound vector of a genome-scale metabolic model — a shift that decomposes into an environment component and a regulatory-divergence component that can be measured separately.
Core Claim
The claim is threefold. First, a genome-scale metabolic model whose reaction bounds are modulated by a regulatory state including species-specific non-coding regulators can predict metabolic differences between contexts that expression-only models cannot, because mRNA abundance is known to decouple from flux through post-transcriptional and allosteric buffering (Multi –omics and metabolic modelling pipelines: challenges and tools for systems microbiology). Second, the in vitro–in vivo translation gap in drug response is partly a stoichiometrically predictable quantity: preclinical transfer models built on abstract pathway lists (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction) should be beaten, or at least matched, by a model whose internal representation is a physically consistent flux distribution. Third, and most distinctively, the portion of the gap attributable to species-specific regulation (regulators that exist in one species and not the other) is quantifiable, because such regulators enter the model only for the species that has them.
The mechanistic picture is one equation. A standard FBA solves $S v = 0$ (mass balance for flux vector $v$ given stoichiometric matrix $S$) within fixed reaction bounds; we replace fixed upper bounds with learned, context-dependent ones:
$$0 \le v_j \le v_j^{0} \cdot \sigma\!\left(\beta_j^{\top} r + \eta_j^{\top} e\right)$$
where $r$ is the regulatory state of a cell (transcription factor activities and non-coding regulator presence, including entries with no ortholog across species), $e$ is the environment (substrate, oxygen, growth demand), $\sigma$ is a squashing function, and the coefficient vectors $\beta_j, \eta_j$ are learned end-to-end from pretraining data. A regulator absent in mouse contributes to mouse predictions only through the learned residual bound $v_j^0$, never through its own row — divergence is encoded structurally, not discovered post hoc.
Research Program
Three things get built. The first is the divergence-aware bound modulator itself: human and mouse genome-scale metabolic reconstructions, with reaction bounds parameterized as above, trained so that predicted fluxes reproduce observed context phenotypes. Pretraining uses the DepMap screening corpus — 966 cancer cell lines, 20 drugs, 225,781 labeled response samples (DeepSADR) — with the learning signal being drug-response labels plus growth phenotypes, in the spirit of existing multi-omics FBA pipelines that already reach 75–95% accuracy against Biolog growth assays in microbes (Multi –omics and metabolic modelling pipelines). The regulatory input $r$ is assembled from two sources in the file: the Mouse ENCODE comparative atlas of DNase, ChIP, and expression data across 123 mouse cell types (A comparative encyclopedia of DNA elements in the mouse genome), and the scCompass-126M corpus of 126 million human and mouse single-cell transcriptomes that underlies GeneCompass (GeneCompass: deciphering universal gene regulatory mechanisms with a knowledge-informed cross-species foundation model). Crucially, where GeneCompass restricts itself to 17,465 one-to-one orthologs and therefore discards species-specific lncRNAs, we deliberately include them as bound-relevant inputs: the human-specific CYP1B1-AS1 circuit that Coxiella exploits, which is absent in mice and whose knockdown causes a 3-fold drop in mitochondrial membrane potential (CYP1B1-AS1 regulates CYP1B1 to promote Coxiella burnetii pathogenesis by inhibiting ROS and host cell death); the primate-specific Alu-dependent MIR205HG fibrosis circuit (Takashima et al.); and the topologically inverted p53/miR-138 loop between human and rodent cells (Species-specific mutual regulation of p53 and miR-138 between human, rat and mouse). These are the concrete cases where a mouse experiment cannot see the human regulatory gate on a metabolic enzyme.
The second deliverable is a decomposition analysis. For any pair of contexts (cell line vs. patient tumor; human vs. mouse homologous tissue), we re-solve the model with $e$ shifted but $r$ frozen, then $r$ shifted but $e$ frozen, then both, and report what fraction of the predicted reaction-level flux change — and, downstream, of predicted drug-response changes — each component explains. This gives the field something it currently lacks: a number for how much of the culture-shock phenomenon, in which cell lines universally upregulate proliferative programs while solid tumors sit at a Spearman fidelity as low as 0.50 (Comprehensive transcriptomic analysis of cell lines as models of primary tumors across 22 tumor types), is microenvironment versus regulatory wiring. A useful internal sanity check comes from CONTRABASS, which showed in Plasmodium that the set of essential reactions expands and network flexibility contracts as the growth-rate demand rises (CONTRABASS: exploiting flux constraints in genome-scale models for the detection of vulnerabilities); our environment term should reproduce that scaling without being told it.
The third deliverable is the transfer head: a lightweight predictor from flux distributions (under a candidate drug's bound perturbation) to response, pretrained on DepMap and evaluated on the TCGA fine-tuning cohort of 555 patients across 5 drugs (DeepSADR). We deliberately keep this head simple, because the scientific claim lives in the stoichiometric representation, not in a bigger decoder.
Baselines
The do-nothing reference on the retrieval side is the random baseline on DDR-Bench, drug-disease retrieval at Hits@1 of 0.0179, with molecular fingerprints barely better at 0.0357, as reported by Cell-Level Virtual Screening; anything we build must clear 0.1250, the Hits@1 that CellVS-Net achieves on the same benchmark by predicting drug impact as rewiring of gene-gene dependency networks rather than expression snapshots. CellVS-Net is therefore the strongest non-clinical competitor, and we will retrain it on identical inputs with an equal compute and tuning budget, comparing on DDR-Bench Hits@1 and on its perturbation-prediction MSE setting. On the clinical side, the strongest existing method is DeepSADR, which reports a Sorafenib response AUC of 0.957 against a next-best baseline of 0.811, and an average AUC of 0.856 across drugs that collapses to 0.662 when its adaptive readout is replaced with simple pooling (DeepSADR). We will run DeepSADR on the same DepMap-to-TCGA split, same 555-patient cohort, same tuning budget, and count ourselves as beating it only if mean AUC exceeds 0.856 with non-overlapping confidence intervals across drugs; matching it while additionally producing flux-level explanations would already be informative, and we will say so honestly. Two ablations isolate our ingredients: the full model with the GEM replaced by DeepSADR-style static pathway annotations (isolating the stoichiometric constraint), and the full model with uniform bounds across contexts (isolating the divergence-aware modulation). A classical FBA baseline — re-solving with maximum-biomass after each perturbation, using context-free bounds — is included because it is what a metabolic modeler would reach for first, though the essentiality literature warns its re-optimization assumption fails in altered cells (Prediction of gene essentiality using machine learning and genome-scale metabolic models).
Why not just…?
Why not just batch-correct the cell line and tumor data? This was tested directly: Seurat, Harmony, and LIGER were applied to the paired Kinker cell-line (53,512 cells, 198 lines) and Kim patient (208,506 cells, 58 patients) single-cell datasets and failed to integrate them, because the differences are genuine biology — absence of immune selection, altered mechanical stress — not technical batch noise (Identifying commonalities between cell lines and tumors at the single cell level using Sobolev Alignment of deep generative models). Sobolev alignment itself is the stronger version of this alternative and a real influence on our design; but it stops at transcriptomic manifolds and never maps its conserved-versus-context-specific factors into a metabolic model, so it cannot say whether a context-specific factor is stoichiometrically consequential. Our experiment 2 below is exactly the head-to-head: conserved/contextual factors from Sobolev alignment versus our bound decomposition, predicting the same flux outcomes.
Why not just use expression snapshots or differential expression? CellVS-Net outperformed models given oracle ground-truth expression data on disease retrieval (Cell-Level Virtual Screening), and mRNA is documented to decouple from flux (Multi –omics and metabolic modelling pipelines). The sharpest test of this alternative is built into our cross-species experiment: mouse ENCODE's neighborhood analysis found core metabolic pathways conserved in co-expression while immune and extracellular programs diverged (A comparative encyclopedia of DNA elements in the mouse genome); if our model predicts flux divergence in some expression-conserved metabolic pathways — gated by species-specific non-coding regulators like CYP1B1-AS1 — while an expression-similarity baseline predicts conservation there, that is a direct, falsifiable separation.
Why not just run standard FBA harder — better thermodynamics, tighter bounds? Standard FBA with context-free bounds is our classical baseline, and it will fail on the clinical benchmark for a structural reason: it has no input slot for species-specific regulators at all, since a human lncRNA is not a gene in a mouse metabolic model, and no learning signal to fit bounds from 225,781 drug-response labels. Thermodynamic curation demonstrably matters — Gibbs-energy filtering took the iBsu1103 model of B. subtilis from 89.5% to 93.2% accuracy across 1,500 conditions (iBsu1103: a new genome-scale metabolic model of Bacillus subtilisbased on SEED annotations) — so we adopt it inside $v^0$; but it addresses a different axis (directionality, not context) and cannot by itself close a translation gap. A more real threat is extending GeneCompass-style cross-species models, since GeneCompass already transfers cell-type labels across species with a 7.5% accuracy boost and cuts perturbation-prediction MSE by 15.4% (GeneCompass); we address this head-on by keeping its orthologous core as the backbone of $r$ while showing, on held-out circuits, that the ortholog-only variant cannot reproduce the flux effects of the non-orthologous regulators — the same structural blind spot its authors acknowledge in restricting to 17,465 homologs.
Proposed Experiments
Experiment one is the clinical transfer: pretrain on the DepMap corpus, fine-tune and evaluate on the TCGA cohort of 555 patients across 5 drugs, reporting AUC per drug with DeepSADR's 0.856 average and 0.957 Sorafenib result as the bar (DeepSADR). Experiment two is retrieval and decomposition: evaluate on DDR-Bench against the 0.1250 CellVS-Net reference (Cell-Level Virtual Screening), and use the 666-line CCLE versus ~9,000-profile TCGA pan-cancer comparison to test whether our bound decomposition explains the fidelity gradient — liquid tumors near Spearman 0.75 versus cholangiocarcinoma at 0.50 (Comprehensive transcriptomic analysis) — predicting that the environment component dominates for liquid tumors while the regulatory component grows for solid, microenvironmentally embedded ones. Experiment three tests context transfer at the perturbation level on the Replogle Perturb-seq panel — 2,009 knockouts across K562, Jurkat, RPE1, and HepG2, where cross-cell-type response overlap is only about 19% (Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation) — reporting Wasserstein-2 distance against PT-RAG's 633.7, the no-retrieval baseline's 646.1, and vanilla retrieval's 1189.5; this is where we learn whether bound modulation generalizes across contexts better than retrieval does, and we will report it whichever way it lands. Experiment four is the cross-species prediction described above: which metabolic pathways diverge in flux between homologous human and mouse tissues, checked against the Mouse ENCODE conservation map, with the expression-conserved-but-flux-divergent stratum as the decisive novelty evidence. A fifth, aspirational validation against the FLIM imaging of organoid versus xenograft redox heterogeneity (Quantitative Spatial Analysis of Metabolic Heterogeneity Across in vivo and in vitro Tumor Models) is proposed as a new evaluation we would have to construct ourselves — no reported result found on Lacuna for a quantitative benchmark linking FLIM redox measurements to predicted flux distributions — so we flag it as a curatorial deliverable rather than an existing score to beat.
References
- A comparative encyclopedia of DNA elements in the mouse genome
- Multi -omics and metabolic modelling pipelines: Challenges and tools for systems microbiology
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- GeneCompass: deciphering universal gene regulatory mechanisms with a knowledge-informed cross-species foundation model
- CYP1B1-AS1 regulates CYP1B1 to promote Coxiella burnetii pathogenesis by inhibiting ROS and host cell death
- Involvement of lncRNA MIR205HG in idiopathic pulmonary fibrosis and IL-33 regulation via Alu elements
- Species-specific mutual regulation of p53 and miR-138 between human, rat and mouse
- Comprehensive transcriptomic analysis of cell lines as models of primary tumors across 22 tumor types
- CONTRABASS: exploiting flux constraints in genome-scale models for the detection of vulnerabilities
- Cell-Level Virtual Screening
- Prediction of gene essentiality using machine learning and genome-scale metabolic models
- Identifying commonalities between cell lines and tumors at the single cell level using Sobolev Alignment of deep generative models
- iBsu1103: a new genome-scale metabolic model of Bacillus subtilisbased on SEED annotations
- Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation
- Quantitative Spatial Analysis of Metabolic Heterogeneity Across in vivo and in vitro Tumor Models