Cancer cells rewire their metabolism. They push glucose toward lactate, feed side pathways that build sugars for protein modification, and switch lipid use by subtype. Each rewiring suggests a drug target, yet many such targets fail. Often the enzyme was highly expressed but not required: the cell had another enzyme or another route that did the same job. The standard tool for interpreting gene lists, functional enrichment analysis (testing whether a list of genes contains more members of a curated pathway than chance predicts), cannot see this. It counts pathway members and ignores whether any one of them is a choke point.
We propose an enrichment statistic whose unit is not a pathway's member list. Its unit is a route through the metabolic network that a specific tumor's expression profile makes active. A gene counts toward that route only if the tumor has no expressed alternative for the step it catalyzes, either a second enzyme for the same reaction or a detour through other reactions. We call this route-cut enrichment. The hypothesis is testable on public data: route-cut irreplaceability computed from expression alone should predict which metabolic genes a particular cell line depends on in genome-wide CRISPR knockout screens. It should also predict how that dependence differs between cell lines, which is the property a therapeutic target needs and which gene-level priors cannot capture.
Core Claim
Knockout lethality of a metabolic enzyme depends on context. It is lethal when the enzyme is the only expressed way to carry flux along a route the cell actually uses. We make this operational by building on CaMeRe, which models HumanCyc (2,835 reactions, 3,543 enzymes) as a graph, weights edges with tumor expression, and ranks non-shortest metabolic routes by the expression of their weakest (bottleneck) enzyme and the stability of expression along the route. CaMeRe placed 7 of 8 canonical cancer pathways in its top three results. It is, however, a descriptive tool: it never asks whether cutting a route kills the cell.
We add two things. The first is a per-sample irreplaceability score for each gene. For a sample $s$ and gene $g$, it is
$$I_s(g) = \sum_{p \in P_s} w_s(p)\,\mathbf{1}[\text{removing } g \text{ leaves no expressed alternative for } p],$$
where $P_s$ is the set of active routes CaMeRe-style scoring retains for $s$, and $w_s(p)$ is the normalized route score. "Alternative" covers both isozymes for the same reaction and bypass routes between the same input and output compounds whose bottleneck expression stays above threshold. The second is an enrichment test over this score. Given a CRISPR hit list, we ask whether its members are over-represented among high-$I_s$ genes relative to metabolic genes matched on expression. This replaces "pathway X is enriched" with "these hits cut these active routes."
The idea differs from its nearest neighbors in specific ways. CaMeRe ranks routes but does not score genes against essentiality. DeepSyn replaces static enrichment with reasoning over a literature knowledge graph and can condition on a disease, but it has no notion of reaction alternatives. SynLethDB 2.0 offers enrichment over curated gene pairs that are fixed regardless of cellular state.
Research Program
The build has three steps. We first reimplement CaMeRe-style route scoring per sample on HumanCyc, using expression profiles for cancer cell lines from the Cancer Dependency Map (DepMap, genome-scale CRISPR knockout fitness scores plus baseline expression for hundreds of lines). We also run it on TCGA tumors and on GTEx normal tissues. Second, we compute $I_s(g)$ for all metabolic genes in each sample. Third, we implement the route-cut enrichment test and apply it to hit lists from BioGRID-ORCS and the Toronto KnockOut screens.
Target validation enters in two places. The first is the DepMap test itself, which checks the score against measured fitness. The second is a therapeutic-window score: a gene is flagged only if it is irreplaceable in a tumor lineage's active routes but replaceable in the matched GTEx normal tissue. This speaks directly to the failure described in the glioblastoma PDHK study. There, genetic knockout of PDHK1 or PDHK2 extended survival by about 10 days in orthotopic xenografts. Systemic inhibition with dichloroacetate impaired cells in vitro but shortened survival in mice. That study motivates the window score but also constrains our claims. Microenvironmental compensation of that kind lies outside a cell-autonomous network model, and we do not claim to predict it.
Hypoxia is the other open question. DepMap screens are run in normoxic culture. Larrieu et al. show PDHK1 concentrated in hypoxic tumor cores and PDHK2 at invasive margins. As an exploratory analysis only, we will compute route sets on IvyGAP core and margin expression and ask whether $I_s$ separates the two paralogs as their spatial pattern suggests.
A secondary extension scores gene pairs. A pair is predicted synthetic lethal (lethal together, tolerated singly) when the two genes jointly cut an active route that neither cuts alone. This is scored on metabolic pairs in SynLethDB under the cold-start split of the synthetic lethality benchmark. That benchmark reports that learned models memorize graph neighborhoods and degrade when neither gene was seen in training. Our score uses no labels, so it has no training set to memorize, and this is the setting where that property should matter.
Baselines
The naive reference is expression alone. We rank each line's metabolic genes by expression, and separately by mean essentiality across all other lines, the "common essential" prior. The second baseline is hard to beat on raw ranking and is the reason we report within-gene metrics.
The standard enrichment baseline is overrepresentation analysis with Fisher's exact test on KEGG and GO, run through funkea. Each gene is scored by the significance of the most enriched pathway it belongs to, among the line's top expressed genes. This is the method route-cut enrichment is meant to replace.
The strongest simple competitor is a redundancy count: the number of expressed isozymes or paralogs for the gene's reactions. The key ablation isolates route-level alternatives. We run $I_s$ with only same-reaction isozymes counted as alternatives, with bypass routes also counted, and with route weights $w_s$ set uniform. If the bypass component adds nothing over the isozyme-only version, the network part of the method is unnecessary.
The strongest learned competitor is the text-embedding classifier from Virtual CRISPR, a text-embedding-3-large embedding with a 5-layer MLP. It reports F1 0.84 and AUROC 0.89 on its post-2024 "Difficult" benchmark. That task is different from ours, and for per-line DepMap metabolic dependency no reported result was found on Lacuna. We will retrain it on the same DepMap training lines, with cell-line context added to the text.
For pairs, the reference is SLMGAE, reported at F1 of about 0.73 in the cold-start setting by the synthetic lethality benchmark, with GRSMF as the stable matrix-factorization alternative.
The mechanistic competitor is expression-constrained flux balance analysis (FBA). FBA solves a linear program for steady-state reaction rates that maximize growth, then simulates each knockout. It is not in the research file, and no reported result was found on Lacuna for it on DepMap. It is the classical tool for exactly this question, so we will run it on the same lines.
All methods get the same lines, the same gene universe, and the same train/test splits. Learned baselines get an equal hyperparameter search budget. Our route-score thresholds are tuned only on training lines. Beating the field means a significantly higher within-gene across-line Spearman correlation with CRISPR fitness than the redundancy count, the embedding MLP, and FBA, with a bootstrap over genes. It also means higher precision at the top of the ranking for selective dependencies.
Why not just…?
Why not just run standard enrichment on the CRISPR hits? Overrepresentation analysis reports that a pathway's members are over-represented. It cannot say which member is a choke point in this cell, and the same pathway is returned for every line whose hits overlap it. DeepSyn's authors show static enrichment can return nothing at all for disease-critical glioblastoma clusters (DeepSyn). The within-gene evaluation settles the question directly, because a line-independent pathway label yields no across-line signal.
Why not just count expressed paralogs? This is the most serious threat, because much metabolic buffering may be isozyme-level. The ablation above tests it directly. If isozyme-only irreplaceability matches the full score, we will report that route-level detours do not matter for knockout fitness. The project would then shrink to a context-aware paralog enrichment test.
Why not just use expression-constrained FBA? FBA encodes mass balance, which our graph scoring does not. It may predict single knockouts better. Its weaknesses are its reliance on an assumed growth objective and culture-medium definition, and the fact that it does not natively produce an enrichment test over hit lists. If FBA wins on single-gene prediction, we will build route-cut enrichment on FBA-derived active reactions instead. We will report the comparison either way rather than defend the graph method.
Why not just train an embedding model or a graph neural network on DepMap? Learned models capture which genes are generally essential. The cold-start collapse in the synthetic lethality benchmark suggests they memorize neighborhoods rather than mechanism. Our held-out-lineage and held-out-gene splits test whether that also happens for single-gene context dependence.
Proposed Experiments
The primary experiment predicts per-line CRISPR fitness for all HumanCyc-mapped metabolic genes in DepMap. We use three splits: held-out lines, held-out lineages, and held-out genes, the last matching the cold-start setting. The headline metric is within-gene across-line Spearman correlation, computed only over genes with non-trivial fitness variance. It is reported alongside precision in the top 1% of genes for selective dependencies.
The decisive check is whether full $I_s$ beats the isozyme-only ablation and FBA on the lineage-held-out split. A positive result would be a gene like one of the PDHK paralogs, or the hexosamine route enzyme GFAT1 that CaMeRe highlights in bladder cancer, scoring irreplaceable in some lines and replaceable in others, with fitness tracking that split. A negative result, no gain over the paralog count, would be informative and will be reported.
The second experiment applies route-cut enrichment to Toronto KnockOut and BioGRID-ORCS hit lists. It measures how often the routes it names are cut by hits that were held out from the enrichment input, compared with overrepresentation analysis at matched false discovery rate.
The third experiment scores metabolic synthetic lethal pairs in SynLethDB under the cold-start split. It is compared with SLMGAE's reported F1 of about 0.73 and with retrained SLMGAE and GRSMF on the identical split.
Finally, the therapeutic-window filter, which compares tumor lineages with matched GTEx normal tissue, is checked for whether it removes known pan-essential metabolic genes while retaining lineage-selective ones. We will check this using the ACOX-1 and FABP4 subtype associations in Kim et al. as qualitative test cases for whether a prognostic marker is also a route-cutting dependency in matched breast cancer lines.
References
- CaMeRe: A Novel Tool for Inference of Cancer Metabolic Reprogramming
- Deep functional synthesis: a machine learning approach to gene functional enrichment
- SynLethDB 2.0: a web-based knowledge graph database on synthetic lethality for novel anticancer drug discovery
- Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
- Benchmarking machine learning methods for synthetic lethality prediction in cancer
- funkea: Functional Enrichment Analysis in Python
- Virtual CRISPR: Can LLMs Predict CRISPR Screen Results?
- Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes