Drug-response predictors are trained on cancer cell lines grown in plastic dishes and are then asked to rank drugs for patients. Two things go wrong. First, models trained on hundreds of cell lines often learn to tell tissue types apart rather than learning why a drug kills a cell. Second, the biology that governs a cell's response can change when it leaves the dish. Metabolism may be the most exposed part of that biology. Cultured cells sit in glucose-rich, oxygenated medium, while tumors are hypoxic and short of nutrients, so the metabolic routes a cell line depends on may not be the routes the tumor depends on.
This note proposes a specific test of that idea. For each cancer lineage, we build a metabolic reaction network twice: once weighted by tumor expression and once weighted by expression in matched cell lines. In each version we locate the "bottleneck" enzymes, meaning enzymes that many metabolic routes must pass through. We then keep only the bottlenecks that hold the same position in both versions. The bet is that drug-response features built from these conserved bottlenecks will transfer to unseen lineages and to patients better than whole-genome features or plain metabolic gene lists. The gain should be concentrated in drugs that act on metabolism, such as the nucleotide antimetabolites (gemcitabine, 5-fluorouracil) and PI3K/mTOR inhibitors. These are exactly the classes where current methods do worst.
Core Claim
The claim rests on one idea: an enzyme's position in the metabolic network, and not only its expression level, carries information about drug response. Positions that are stable between cell lines and tumors are the portion of that information that transfers. The proposal is new in three ways. The topology is metabolic mass flow rather than protein binding. It is recomputed for each context rather than fixed. And it is used as a filter against the dish-to-tumor shift, not merely as a prior for prediction. If the gain disappears when the network is randomly rewired, or when topology is replaced by simple membership in a metabolic gene list, the claim is false, and the experiments are built to find that out.
What Exists and What Is Missing
Three lines of work motivate the proposal.
NetBiTE biases tree ensembles toward drug targets and their neighbors on the STRING protein-interaction network. It improved correlation over random forests by 0.12 for receptor tyrosine kinase inhibitors and by 0.14 for EGFR inhibitors. However, it performed worse than an unbiased random forest for PI3K/mTOR inhibitors. The authors attribute this to target expression failing to reflect pathway engagement. That failure is the most direct motivation here. We replace the physical-interaction graph with a metabolic flux graph, where a gatekeeper enzyme's role is defined by the chemistry routed through it.
CellHit restricts features to mechanism-of-action gene sets mined by a language model. It aligns cell lines to TCGA tumors with Celligner, an unsupervised method that maps both into a shared expression space, and it reports a Pearson correlation of 0.89 on GDSC. It shows that curated feature sets beat all-gene models and that its models recover true targets. It does not use metabolic structure. Its alignment also corrects the whole transcriptome uniformly rather than asking which features actually carry over.
The benchmark study of Sharifi-Noghabi and colleagues constrains the design in two ways. It documents the tissue-type trap, and it shows cross-dataset correlation falling from about 0.41 to about 0.33 for random forests. It also establishes that area above the dose-response curve (AAC) is more stable than the half-maximal inhibitory concentration (IC50), so AAC is our label.
The metabolic tools supply the method. CaMeRe builds an expression-weighted reaction graph from HumanCyc and ranks rewired pathways by bottleneck enzymes. It recovered 7 of 8 known cancer pathways in its top three results, but it never predicts drug response. FluxGAT builds mass-flow graphs from objective-free flux sampling. It nearly doubled essential-gene sensitivity over flux balance analysis (0.576 versus 0.304), but only in hamster cells and only for knockout essentiality.
The lineage-specific studies show that the relevant enzymes differ by tissue. These include SQLE in sarcoma, glycolytic and glycogen-catabolic subtypes in pancreatic cancer, and PDH kinases as gatekeepers in glioblastoma. This is why we compute centrality per lineage rather than pan-cancer.
Two cautions shape the design. Jang and colleagues found, across more than 110,000 models, that Elastic Net and Ridge regression on expression alone are among the strongest methods. Any topology must therefore beat them, not just beat random forests. Kong and colleagues saw a 5-FU model fall from 84.1% cross-validation accuracy to 52.9% on stomach adenocarcinoma. This is the kind of transfer failure we aim to reduce, and it tells us that patient-level evaluation must stay within small, honest cohorts.
Research Program
Building the networks. We build a human reaction graph from HumanCyc following CaMeRe's construction, with reactions and metabolites as nodes. We remove currency metabolites such as ATP, NADH and water, because otherwise shortest paths would route trivially through them. For a context $k$, meaning a lineage in tumors or in cell lines, the cost of traversing a reaction is set inversely to the median expression of its catalyzing enzymes in that context. Shortest paths therefore favor routes the context actually expresses.
Tumor expression comes from TCGA. Where stromal contamination is high, we deconvolve tumor-intrinsic expression following the DeMixT approach used in the pancreatic subtype study. Cell-line expression for the same lineage comes from CCLE. As a sensitivity analysis, we replace shortest-path structure with FluxGAT-style flux sampling constrained by expression, to check that the results are not an artifact of the weighting rule.
Measuring conservation. For each enzyme $g$ we compute weighted betweenness (how often it lies on shortest routes between metabolite pairs) and weighted degree, separately in the tumor graph and the cell-line graph of lineage $\ell$. We define a conservation score
$$c_{g,\ell} = 1 - \frac{\left| r^{T}_{g,\ell} - r^{C}_{g,\ell} \right|}{N},$$
where $r^{T}_{g,\ell}$ and $r^{C}_{g,\ell}$ are the betweenness ranks of $g$ in the tumor and cell-line graphs of lineage $\ell$, and $N$ is the number of enzymes. An enzyme is a conserved bottleneck if it ranks in the top betweenness decile in both graphs and has high $c_{g,\ell}$. It is a culture-shifted bottleneck if it is central in one graph but not the other.
Turning this into predictors. Features for a given cell line are the expression values of conserved bottlenecks. Each value is scaled by the enzyme's centrality in that cell line's own expression-weighted graph, so topology enters the model directly and does not act only as a gene filter. The model is Elastic Net on AAC, trained per drug.
A biological sanity check comes before any prediction. Culture-shifted bottlenecks should be enriched for hypoxia- and glucose-responsive enzymes such as HK2 and the PDH kinases. If they are not, the hypothesized mechanism behind the dish-to-tumor shift is not what the data show.
Baselines
The naive reference is a per-drug mean predictor together with a lineage-only model that predicts AAC from tissue labels. The lineage-only model measures how much of any apparent success is the tissue-type trap. Anything that fails to clearly beat it under leave-one-lineage-out evaluation has learned nothing about the drug.
The strongest data-driven competitor is Elastic Net on all expressed genes, following Jang and colleagues. That paper reports it among the top performers, but in the file it gives only a significance result for regression over classification, not a per-drug correlation. It is therefore rerun here on identical splits, with the same nested cross-validation budget over the regularization path.
The strongest published curated-feature method is CellHit, which uses mechanism-of-action gene sets, Celligner alignment and XGBoost. It reports a GDSC Pearson correlation of 0.89. That number comes from its own within-GDSC protocol, so we rerun its feature sets and alignment on our splits rather than compare against the headline figure.
The strongest topology-prior method is NetBiTE, whose reported gains over random forests are 0.12 and 0.14 for the kinase classes. It is also rerun on our splits. For cross-screen transfer, the reference point is the drop from roughly 0.41 to 0.33 reported by Sharifi-Noghabi and colleagues. For leave-one-lineage-out evaluation, no reported result was found on Lacuna, so we establish that number ourselves.
Three ablations isolate the key ingredient. The first uses the same enzymes selected by membership in a metabolic gene set, with no centrality. The second uses centralities from a degree-preserving randomly rewired reaction graph. The third uses bottlenecks chosen from the cell-line graph alone, without the conservation filter.
All models receive the same features before selection, the same AAC labels, the same splits, and an equal hyperparameter search budget. Beating the field means two things. First, a leave-one-lineage-out Spearman correlation for metabolic-acting drugs that exceeds both all-gene Elastic Net and the gene-set ablation, with bootstrap confidence intervals over cell lines that exclude zero. Second, no significant loss on receptor tyrosine kinase inhibitors.
Why not just…?
Why not just use Elastic Net on all genes? This is a serious competitor, because the largest benchmark in the file finds it hard to beat. The counter-evidence is that curated feature spaces beat all-gene models in CellHit and network priors helped kinase inhibitors in NetBiTE. Neither result, however, concerns transfer across lineages or to patients. Our experiment settles the question directly: if all-gene Elastic Net matches the conserved-bottleneck model under leave-one-lineage-out evaluation for antimetabolites, the topology adds nothing and we report that.
Why not just use a list of metabolic genes, without any network? This is the most likely way the idea fails. Much of any gain could come simply from restricting features to metabolism, which is a milder version of CellHit's curated-set effect. For that reason the gene-set ablation uses exactly the same enzymes as the full model. The rewired-graph control then tests whether real network position matters. The proposal survives only if real centrality beats both. We expect the difference to appear mainly for low-expression gatekeepers of the kind Dong and colleagues describe, which bridge modules without large expression changes.
Why not just align cell lines to tumors with Celligner and keep the existing model? Alignment corrects the whole transcriptome toward tumors, but it cannot restore a dependency that culture has erased. If glycolytic gatekeepers are peripheral in normoxic cells, shifting their expression values does not make the cell-line drug labels reflect tumor biology. The test is to combine alignment with each feature set. If Celligner alone closes the cell-to-patient gap as well as the conservation filter does, the filter is redundant. If the two are complementary, the gains should add.
Why not just use flux balance analysis to predict essential enzymes? FluxGAT shows that flux balance analysis, which assumes cells maximize growth, has poor sensitivity for essential genes (0.304). That assumption is least appropriate for stressed tumor cells. We include flux-balance essentiality scores as an additional feature baseline, and flux sampling as a robustness variant, rather than rely on either one.
Proposed Experiments and Decisive Checks
Data. Training uses GDSCv2 AAC with CCLE expression across lineages that have at least about 20 cell lines and a matching TCGA cohort. Drugs are grouped by class using target annotations. The metabolic-acting group comprises antimetabolites, PI3K/mTOR inhibitors and direct metabolic-enzyme inhibitors. Receptor tyrosine kinase inhibitors serve as the negative control group.
Leave-one-lineage-out evaluation. Each lineage's conservation scores come only from its own TCGA and CCLE data, so no labels leak across the split.
Cross-screen evaluation. Using the 11 drugs shared across CTRPv2, GDSCv1, GDSCv2 and gCSI, as in Sharifi-Noghabi and colleagues, this experiment checks that the features do not amplify assay noise. It is a robustness test, not the main claim, because the same cell lines appear in every screen and conservation between dish and tumor should not matter there.
Patient-level check. We apply cell-line-trained gemcitabine and 5-FU models to the TCGA response cohorts used by Kong and colleagues. Their tumor-trained models reached 85.7% validation accuracy for gemcitabine and 52.9% for 5-FU in stomach adenocarcinoma. These serve as reference points rather than direct competitors, because those models were trained on patients.
Pancreatic subtype check. We ask whether predicted gemcitabine sensitivity separates the M1 and M2 pancreatic subtypes in the direction implied by the COMPASS progression rates (48% versus 20.4%), as reported in the pancreatic subtype study.
What counts as evidence. The decisive pattern has four parts. Gains must appear for metabolic-acting drugs but not for receptor tyrosine kinase inhibitors. They must disappear on rewired graphs. They must exceed the matched gene-set ablation. And they must shrink when the conservation filter is removed, with the loss largest in cell-to-patient transfer.
Any single failure is informative. If gains appear equally for kinase inhibitors, we are seeing generic feature selection and not metabolism. If rewired graphs perform as well, topology is irrelevant. If the unfiltered cell-line bottlenecks transfer as well as the conserved ones, the dish-to-tumor shift in metabolism is not what limits generalization. That last outcome would itself answer an open question raised by the cross-domain drop in the existing benchmarks.
References
- Network-based Biased Tree Ensembles (NetBiTE) for Drug Sensitivity Prediction and Drug Sensitivity Biomarker Identification in Cancer
- Learning and actioning general principles of cancer cell drug sensitivity
- Drug sensitivity prediction from cell line-based pharmacogenomics data: guidelines for developing machine learning models
- CaMeRe: A Novel Tool for Inference of Cancer Metabolic Reprogramming
- FluxGAT: Integrating Flux Sampling with Graph Neural Networks for Unbiased Gene Essentiality Classification
- Transcriptome Sequencing Unveils a Molecular-Stratification-Predicting Prognosis of Sarcoma Associated with Lipid Metabolism
- Clinically impactful metabolic subtypes of pancreatic ductal adenocarcinoma (PDAC)
- Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
- Systematic Assessment of Analytical Methods for Drug Sensitivity Prediction from Cancer Cell Line Data
- Leveraging TCGA gene expression data to build predictive models for cancer drug response