Flux-Conditioned Essentiality: Predicting Which Metabolic Genes Matter Only Under Rewiring

Every computational method that predicts which genes a cell needs treats the cell's environment as a fixed backdrop. Yet the biology says the backdrop is the story: when oxygen drops, tumors and stressed tissues switch from mitochondrial respiration to glycolysis, and the set of enzymes the cell cannot live without changes with it. This proposal argues that gene essentiality in metabolism is not a property of a gene but of a gene-environment pair, and that we can now build and test predictors of that pair. The plan is to extend flux-sampling graph neural networks — models that read a cell's metabolic network the way a language model reads a sentence — so they accept a continuous environmental context (oxygen level, nutrient mix, pH) as an input, train them on CRISPR knockout screens tagged with growth conditions, and then prospectively validate, in wet-lab assays, the top predicted genes that flip from dispensable to essential under hypoxia.

The motivation is concrete. The strongest existing flux-based essentiality model, FluxGAT, was trained and evaluated in exactly one growth condition and, per its own authors, was never tested across metabolic microenvironments such as hypoxia or nutrient shifts; the same single-condition limitation is stated explicitly for the earlier Mass Flow Graph classifier in E. coli, which was run only on aerobic glucose. Meanwhile the glioblastoma PDHK study gives a live warning of what the missing environmental dimension costs: dichloroacetate killed tumor cells in standard culture yet shortened survival in mice, a contradiction no current predictor could foresee. And on the validation side, every machine-learning framework for context-dependent dependency — LEAP, Virtual CRISPR, co-essentiality networks — is evaluated only retrospectively against held-out databases. The intersection of these three directions (experimental validation, context-dependent protein function, metabolic reprogramming) is therefore not a vague triangulation: it names a specific missing object, a condition-conditioned essentiality predictor with a prospective experimental test.

Core Claim

The claim is that a model which conditions on the metabolic state of the cell — represented not as a tissue label but as continuous environmental parameters that change the feasible flux distribution — will predict condition-specific gene essentiality substantially better than the same architecture trained on pooled, context-blind data, and that its most valuable output is a ranked list of "flip genes": genes whose knockout is tolerable in normoxia but lethal after hypoxic metabolic reprogramming, or vice versa. For each metabolic gene $i$ and context vector $c$ (oxygen tension, carbon source, extracellular pH), the model outputs $P(e_i = 1 \mid c)$, the probability that knocking out gene $i$ prevents growth in context $c$. The quantity of interest is the flip score

$$\Delta_i(c_{\text{hyp}}, c_{\text{norm}}) = P(e_i = 1 \mid c_{\text{hyp}}) - P(e_i = 1 \mid c_{\text{norm}}),$$

which is exactly what pooled models cannot produce, because they average over contexts and thereby erase the reversals that define context-dependent protein function — the phenomenon documented for INPP4B, which acts as a tumor suppressor in most settings but promotes survival in NPM1-mutated leukemia, and for S100A11, which flips oncogene versus suppressor roles across tissues.

Research Program

Three components. First, a context-conditioned extension of the flux-sampling graph architecture. FluxGAT samples steady-state flux distributions with an Artificial Centring Hit-and-Run sampler over a genome-scale metabolic model (3,972 metabolites and 6,236 reactions for Chinese hamster ovary cells), builds a weighted reaction graph from the samples, and trains a graph attention network against a CRISPR screen (1,980 essential of 15,028 targeted genes). We reuse that machinery but sample under many constraint regimes — varying oxygen uptake bounds, carbon sources, and pH-dependent exchange constraints — so that each reaction node carries a distribution of flux features across contexts rather than in one. A continuous context vector then modulates the network's intermediate features, using the same Feature-wise Linear Modulation trick that ProtQueSt uses to condition protein retrieval on textual queries; here the "query" is not text but the physical state of the cell. This is the deliberate contrast with CDBridge, which injects context as a static categorical tissue embedding and, by its own analysis, does not model dynamic metabolic states or hypoxia gradients: we treat context as continuous and physically grounded in the constraint set, so that unseen conditions can be interpolated rather than only looked up among known tissue labels.

Second, a condition-stratified training corpus. The retrospective data are CRISPR screens harmonized in BioGRID-ORCS and DepMap, filtered to entries with recorded culture conditions, plus the existing single-condition benchmarks for direct comparison: the CHO cell screen behind FluxGAT and the 255-reaction E. coli knockout growth data behind the Mass Flow Graph classifier. An E. coli pilot on the iML1515 model is the cheap first test, because matched knockout growth assays exist and alternative media conditions are trivially imposed in the sampler; the harder mammalian track trains on condition-tagged screens and is evaluated on held-out conditions, not just held-out genes.

Third, prospective wet-lab validation — the component the entire field is missing. We will select the top predicted flip genes (and matched controls predicted to be condition-invariant) in a DepMap-profiled human cancer cell line, run competitive CRISPR knockout growth screens under 21% versus 1% oxygen, and confirm the metabolic mechanism of a subset with Seahorse extracellular flux assays showing that the knockout's lethality coincides with the predicted failure to sustain glycolytic flux. The YesU glycan study is the cautionary template here: structural modeling correctly identified the broad functional class but misidentified the specific target, resolved only by glycan array experiments. We take that lesson as a design constraint — the pipeline's claims are calibrated against physical assays, not database labels.

Proposed Experiments

The first experiment is retrospective and condition-stratified. Splitting by condition rather than only by gene, the model must predict essentiality in contexts unseen during training. Metrics are PR-AUC and sensitivity per condition, plus a dedicated evaluation set of literature-documented conditional essentials (genes known to become essential under hypoxia or nutrient restriction). The reference numbers to beat come from the file: FluxGAT on the CHO screen reported PR-AUC of 0.765, accuracy of 0.871 ± 0.012, and sensitivity of 0.576 against flux balance analysis's 0.304, and the E. coli Mass Flow Graph Random Forest reported 76% accuracy, 82.5% precision, and 86.8% recall under aerobic glucose (deletion-FBA reached 84.3% accuracy there). No reported result found on Lacuna for any method's accuracy on condition-shifted essentiality — that is the number this project creates.

The second experiment is the E. coli pilot on iML1515, predicting knockout growth under aerobic versus anaerobic glucose and a carbon-source swap, benchmarked against the single-condition Mass Flow Graph classifier retrained on the pooled data. This isolates whether conditioning helps even where data are sparse and the organism is simple.

The third experiment is the prospective screen. Prediction success is measured as the fraction of top-ranked flip genes whose knockout shows a statistically significant oxygen-condition-dependent fitness defect, compared against matched condition-invariant controls and a random-gene reference set. A secondary mechanistic check asks whether the observed fitness defect tracks the predicted flux failure, using Seahorse ECAR/OCR readouts.

The fourth experiment is the ablation suite: the identical architecture with the conditioning removed (context-blind, pooled training), with continuous context replaced by categorical condition labels, and with flux sampling replaced by stoichiometry-only graph features. This tells us whether the gain comes from conditioning, from continuity of the context, or from flux information itself.

Baselines

The naive reference point is the pooled prior: predict every gene's essentiality at its base rate in the training corpus, ignoring condition entirely. On flip genes this predicts a flip score of exactly zero for every gene, so any real signal beats it trivially, but it calibrates the class imbalance and defines the floor for PR-AUC.

The strongest existing method is FluxGAT, the graph attention network over ACHR-sampled flux distributions, which on its CHO benchmark reported sensitivity 0.576 versus FBA's 0.304, specificity 0.977, PR-AUC 0.765, and accuracy 0.871 ± 0.012. It is the strongest competitor because it is the only prior model that already learns essentiality from flux states without an observer-biased objective function. For a fair comparison we retrain FluxGAT from its released procedure on our condition-stratified corpus with the same compute, the same number of flux samples per condition, and the same hyperparameter search budget as our model — the only difference being the absence of the context-conditioning pathway. Beating it means a higher condition-stratified PR-AUC overall and, decisively, nonzero versus near-zero retrieval on the flip-gene evaluation set, where a context-blind model should collapse toward chance because flip genes are essential in some training conditions and not others.

Per-condition flux balance analysis with growth maximization is the classical alternative, and we run it with each context's constraints set to its measured exchange rates. The Mass Flow Graph study found deletion-FBA at 84.3% accuracy versus the classifier's 76% under aerobic glucose, so it is not weak; but FluxGAT showed its sensitivity ceiling (0.304 on CHO), and the growth-maximization assumption is known to fail precisely when knockouts force the cell onto a different objective, which is what reprogramming does. Beating FBA on the flip-gene set while matching its specificity is the target.

Finally, the key ablation is our own architecture with conditioning removed, described above; it isolates the one ingredient this proposal adds.

Why not just…?

Why not just run flux balance analysis under each condition's constraints? This is the most serious alternative, because FBA natively encodes environmental constraint changes. The evidence against it is two-fold: on the CHO benchmark, FBA detected only 0.304 of essential genes (FluxGAT reached 0.576), and the E. coli study showed that deletion-FBA silently assumes knockout cells still maximize wild-type biomass yield — an assumption that breaks exactly in the compensatory rewiring regime that hypoxia induces, and that the metabolic-reprogramming literature documents as rampant (OXPHOS upregulation when glycolysis is inhibited). But we do not argue this away; the head-to-head per-condition FBA baseline in the first experiment is designed to settle it, and if FBA matches our model on flip genes, the proposal's added machinery is not justified by that result.

Why not just prompt a frontier LLM, or extend the Virtual CRISPR embedding classifier with a condition string? Direct prompting already failed at the easier task: GPT-4o with chain-of-thought reached F1 0.35 on the out-of-distribution CRISPR benchmark, versus 0.84 F1 and 0.89 AUROC for the text-embedding MLP. But the embedding classifier is a real competitor, so we run it as an additional baseline, appending the culture condition to its input text on the same splits. Our expectation, grounded in that paper's own stated limitation of evaluating only baseline culture conditions, is that text embeddings cannot represent the continuous physics of flux rewiring; if it matches us, we will report that honestly and the project's contribution shifts entirely to the prospective wet-lab validation, which no embedding model has undergone.

Why not just add condition labels to a transcriptomic predictor like LEAP? LEAP predicts gene dependency across unseen cell lines from baseline RNA-seq using masked-autoencoder ensembles plus perturbation-specific LASSO, reporting a 16% average Spearman improvement over OmicsRPZ while being 30–150× faster than deep alternatives. It is fast, strong, and already handles per-cell-line context — so the skeptic's question is fair. The answer is architectural, not rhetorical: LEAP's inputs are static expression profiles, which cannot express what happens when the environment changes without a transcriptional signature, and its authors note it does not model metabolic fluxes or rewiring under environmental shifts. Still, we will train a LEAP-style baseline with condition covariates on our splits; if it recovers flip genes as well as the flux-conditioned model, that result would itself be informative and would reframe the field's priorities.

A note on scope: ProtQueSt-style query-conditioned retrieval could in principle take "under hypoxia" as a query, but its retrieval corpus ceiling — it can only re-annotate what Swiss-Prot already contains — makes it structurally unable to predict novel condition-specific dependencies, and the CAFA evaluation's finding that nearly 38% of annotations use generic terms like "protein binding" means the corpus itself is poorest exactly where context specificity is needed. That is why the conditioning here comes from physics (flux constraints) rather than text.

Decisive Checks

The project succeeds if three numbers come out favorably: a condition-stratified PR-AUC above the retrained context-blind FluxGAT baseline, with the gap concentrated on flip genes; a prospective validation confirming a materially enriched fraction of oxygen-dependent fitness defects among top-ranked flip genes versus matched controls; and the ablation showing that removing continuous conditioning, not merely adding parameters, causes the drop. The single most decisive outcome would be a confirmed flip gene that no pooled model — FBA, FluxGAT, LEAP, or the embedding classifier — flagged, validated by knockout screen under 1% oxygen and mechanistically tied to the predicted flux failure. That would be the first prospective, physically validated prediction of context-dependent essentiality under metabolic reprogramming, and the template the DCA paradox shows the field currently lacks.