Radiotherapy and most chemotherapy drugs kill tumor cells by drowning them in reactive oxygen species and lipid peroxidation. Tumors survive because they maintain a layered antioxidant shield—enzymes like catalase and peroxidoxins, the lipid peroxidation brake GPX4, and the master regulator NRF2—and because they rewire metabolism to feed that shield with NADPH and the right lipids. The central practical problem is redundancy: knocking out any single enzyme often does nothing, because parallel buffers compensate, while drugs that hit a whole enzyme family at once poison normal tissue that depends on the same chemistry. This proposal asks a question that, according to the assembled literature, no study has addressed: can we use human genetics to decide, before building an inhibitor, which member of a redundant antioxidant-metabolic enzyme family actually causes therapy resistance in people?
The tool is Mendelian randomization—using inherited genetic variants, randomized at conception, as instruments that mimic a lifelong clinical trial, immune to reverse causation and most confounding. Its multivariable form (MVMR) is designed exactly for correlated exposures: it fits the causal effect of each family member while conditioning on the others, isolating the one true driver from confounded bystanders. The proposal builds a genetic triage pipeline over the enzyme families where experimental dissection has already been done, backtests it against those known answers, and then runs it forward to pick—and experimentally test—new targets in pancreatic cancer models.
Core Claim
Germline cis-expression quantitative trait loci (cis-eQTLs, inherited variants that dial a gene's expression up or down) can be used as instruments in an MVMR framework to rank the members of redundant antioxidant and metabolic enzyme families by their causal effect on cancer outcomes. The claim is falsifiable in two directions: the pipeline should recover the experimentally verified drivers (catalase, not peroxiredoxins, in leukemia; GOT1, not GLUD1, in pancreatic cancer stem cells; PDHK1, not PDHK2, in glioblastoma), and it should make predictions about untested family members that hold up under knockdown-plus-drug experiments.
Research Program
Three stages. First, the instrument panel. For each enzyme family implicated in antioxidant-mediated resistance, I will collect tissue-specific cis-eQTLs from GTEx v10, following the protocol of the metformin drug-target MR study, which used GTEx v10 cis-eQTLs for eleven putative drug targets with instrument $F$-statistics above 10 and found that genetically proxied PRKAG1 expression causally reduces colorectal cancer risk (OR 0.74, 95% CI 0.63–0.87, adjusted $P = 0.001$) independent of glycemic control Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes: A Drug-Target Mendelian Randomization Study. That paper supplies the method but only examined cancer incidence; we apply the same machinery to correlated enzyme families and to progression and survival endpoints. The families are: catalase versus PRDX1/2/6 versus GPX4; GOT1 versus GLUD1; PDHK1 versus PDHK2 (and the remaining isoforms); ACOT8 versus related acyl-CoA thioesterases; RRM1 versus RRM2; and NRF2-axis regulators including USP35, which stabilizes NRF2 protein post-translationally and confers cisplatin resistance in esophageal squamous cell carcinoma The deubiquitinating enzyme USP35 regulates the stability of NRF2 protein. As a second exposure layer, circulating metabolites from the Shin et al. (2014) genome-wide association study of 452 blood metabolites—already shown to carry causal signal for multiple myeloma, including oxidized bilirubin (OR 0.83) mapping onto the heme oxygenase-1 redox pathway Causal effects of genetically determined blood metabolites on multiple myeloma: a Mendelian randomization study—will be tested jointly with enzyme expression, so that a metabolite signal is not mistaken for an enzyme signal.
Second, the causal estimation. For each family, the outcome SNP effects from GWAS Catalog, UK Biobank, and FinnGen summary statistics are regressed jointly on each instrument's effect on every family member's expression:
$$\hat{\beta}_g = \sum_{j=1}^{k} \hat{\theta}_j\, \hat{\gamma}_{gj}$$
where $\hat{\beta}_g$ is SNP $g$'s association with the outcome, $\hat{\gamma}_{gj}$ is its association with expression of enzyme $j$, and $\hat{\theta}_j$ is the direct causal effect of enzyme $j$ holding the other family members fixed. Sensitivity follows the established pipeline—inverse-variance-weighted estimates as the primary result, MR-Egger intercepts to detect directional pleiotropy, weighted median and mode estimators, and leave-one-out analysis—as used in the cathepsin study, which is the direct methodological precedent: among nine correlated proteases, multivariable MR isolated Cathepsin L2 as the sole independent driver of asthma (multivariable OR 1.060, $P = 0.008$) while reverse MR ($P = 0.550$) showed the other cathepsins were consequences, not causes Causal association between cathepsins and asthma: a Mendelian randomization study. That paper did this for proteases and asthma; we do it for redox-metabolic enzyme families and cancer outcomes, which no paper in the corpus attempts. Multivariable MR is mathematically required here because correlated exposures violate the exclusion restriction in univariable designs Multivariable Mendelian Randomization for Correlated Exposures.
Third, the backtest and the prospective test, described under Proposed Experiments. One honest limitation shapes the design: these are germline instruments in mostly normal tissues, and the corpus itself flags that circulating or tissue-level genetic effects may not reflect intratumoral, microenvironment-localized enzyme activity. No reported result found on Lacuna for a germline GWAS of gemcitabine or radiotherapy response specifically, so the MR layer speaks to cancer incidence and survival endpoints, and therapy-specificity must come from the experimental layer. Tissue-specific (pancreas, brain, blood) rather than pan-tissue eQTLs partially mitigate the localization problem, and the wet-lab layer is the decisive arbiter.
Baselines
The naive reference point is correlation-based target ranking: order the candidate enzymes by the strength of the observational association between their tumor expression and patient survival in TCGA, which several cited papers effectively do—the ACOT8 study found high ACOT8 correlated with poor overall survival in PDAC, and the RRM1 study confirmed RRM1 overexpression in breast, colon, and lung tumors Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma; Knockdown of RRM1 in tumor cells promotes radio-/chemotherapy induced ferroptosis by regulating p53 ubiquitination and p21-GPX4 signaling axis. This baseline gets the same candidate list, the same outcome data, and the same wet-lab validation slots. Beating it means a higher fraction of top-ranked targets producing a significant drug-sensitization shift, and correct demotion of bystanders that expression correlation would wrongly promote.
The strongest existing competitor is univariable drug-target MR, run exactly as in the metformin study Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes: A Drug-Target Mendelian Randomization Study: one enzyme at a time, same GTEx instruments, same GWAS outcomes, same sensitivity analyses, same compute. It is the strongest competitor because it already works—it produced the PRKAG1 finding—and it is the simplest defensible upgrade over correlation. Beating it is specifically about the correlated-family problem: on the backtest, univariable MR should attribute causal effects to multiple members of each family because shared cis-regulatory structure leaks across the single-exposure regression, while MVMR should isolate the verified single driver. If univariable MR already cleanly separates catalase from PRDXs, the multivariable ingredient adds nothing and the proposal's core claim fails; this ablation is the experiment that isolates it.
A third baseline is ECLIPSE's VULNCAUSAL module, which used invariant risk minimization across cancer cell line lineages in DepMap to find causal vulnerabilities, reporting an 80× enrichment over chance ($P < 10^{-5}$) and a 3.7× higher validation rate than correlation-based models ECLIPSE: A Composable Pipeline for Predicting ecDNA Formation, Evolution, and Therapeutic Vulnerabilities in Cancer. It is the strongest published causal-prioritization pipeline in the corpus, but it works on somatic cell-line data rather than human germline instruments and never modeled redox pathways. We will run its IRM procedure on our candidate enzyme set in DepMap under the same conditions and compare its wet-lab validation rate against ours.
Why not just…?
Why not just run observational multi-omics regressions on patient cohorts? Because the corpus documents exactly how this fails in this setting. Observational asthma studies repeatedly nominated Cathepsin B as a culprit; bidirectional MR showed it was a downstream consequence, with Cathepsin L2 the real driver Causal association between cathepsins and asthma: a Mendelian randomization study. In the antioxidant context the failure mode is worse: catalase and the peroxiredoxins are co-upregulated in resistant leukemia cells, so expression-survival correlation promotes all of them, yet knockdown showed only catalase matters Targeting Catalase but Not Peroxiredoxins Enhances Arsenic Trioxide-Induced Apoptosis in K562 Cells. The backtest in this proposal directly measures how often correlation ranking is wrong where the genetic answer is known.
Why not just use deep learning drug-response prediction like DeepCDR? DeepCDR integrates genomic, epigenomic, and transcriptomic features into a graph convolutional network and reaches a Pearson correlation of 0.923 and AUC of 0.841 on the GDSC benchmark DeepCDR: a hybrid graph convolutional network for predicting cancer drug response. It predicts; it does not explain which enzyme to inhibit and cannot tell a causal driver from a passenger. ECLIPSE's authors showed standard predictive models suffer circular reasoning and that causal filtering raised validation rates 3.7-fold ECLIPSE: A Composable Pipeline for Predicting ecDNA Formation, Evolution, and Therapeutic Vulnerabilities in Cancer. Our question is target selection under redundancy, not response prediction; DeepCDR's numbers are a benchmark for a different task.
Why not just do a CRISPR or shRNA screen in cell lines or organoids? This is the most serious alternative, and it is how the field currently works—catalase, GOT1, PDHK1, and ACOT8 were all found this way. But the corpus documents its failure mode: in glioblastoma, genetic PDHK knockout extended mouse survival, yet the small-molecule inhibitor dichloroacetate shortened it, showing that in-vitro efficacy does not translate when inhibition is systemic and unselective Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development. Screens also cannot tell you whether the target's lifelong, whole-organism perturbation is tolerable in humans, which germline instruments approximate for free. The honest resolution: screens remain the validation layer here, and the wet-lab experiments in this proposal are exactly those screens—but run only on the genetically triaged shortlist, rather than genome-wide.
Why not just run standard univariable MR per enzyme? Because univariable MR assumes each instrument acts on the outcome only through its one exposure, which is false within a correlated enzyme family sharing cis-regulation Multivariable Mendelian Randomization for Correlated Exposures; it will spread causal credit across catalase and the peroxiredoxins alike. This alternative is fully embedded as the key ablation baseline, so if it performs as well as MVMR on the backtest, the proposal's central ingredient is refuted rather than argued around.
Proposed Experiments and Decisive Checks
The backtest comes first. Three enzyme dissections exist with quantitative readouts in the corpus, and the pipeline must reproduce all three before anything new is claimed. In arsenic-trioxide-resistant K562 leukemia cells, catalase knockdown raised apoptosis from roughly 20% to 80% while PRDX 1, 2, and 6 knockdown did nothing Targeting Catalase but Not Peroxiredoxins Enhances Arsenic Trioxide-Induced Apoptosis in K562 Cells; in pancreatic cancer stem cells, GOT1 knockdown raised post-irradiation (6 Gy) apoptosis from about 10% to 50% while GLUD1 knockdown did not Inhibition of glutamine metabolism counteracts pancreatic cancer stem cell features and sensitizes cells to radiotherapy; in orthotopic glioblastoma models, PDHK1 knockout extended median survival by roughly 10 days Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development. The pre-registered prediction: MVMR direct-effect estimates rank catalase, GOT1, and PDHK1 as sole significant drivers within their families, with null MR-Egger intercepts and stable leave-one-out estimates, while univariable MR and TCGA correlation ranking do not separate family members. Getting any family wrong counts against the core claim.
The prospective test then takes the pipeline's top-ranked and bottom-ranked untested family members into PANC-1 and MIA PaCa-2 cells, patient-derived organoids, and gemcitabine-plus-radiation challenge, following the experimental logic of the ACOT8 study, where ACOT8 depletion dropped the gemcitabine IC50 in MIA PaCa-2 cells from about 60 nM to 10–15 nM and suppressed pro-ferroptotic PUFA-PE lipid species Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma. Success means a prioritized target shifts the gemcitabine IC50 by a comparable order of magnitude with a matching lipid-peroxidation increase on the scale of the roughly two-fold change seen with RRM1 knockdown plus radiation Knockdown of RRM1 in tumor cells promotes radio-/chemotherapy induced ferroptosis by regulating p53 ubiquitination and p21-GPX4 signaling axis, while deprioritized bystanders produce null shifts. Both outcomes are informative: null effects in deprioritized enzymes confirm the triage, and a hit in a deprioritized enzyme refutes it.
Finally, a computational cross-check on DepMap and GDSC asks whether MVMR-prioritized genes show isoform-selective dependency signatures, benchmarked against ECLIPSE's reported 3.7-fold validation-rate advantage over correlation-based ranking ECLIPSE: A Composable Pipeline for Predicting ecDNA Formation, Evolution, and Therapeutic Vulnerabilities in Cancer. A longer-term extension, if the static pipeline works, is to apply ti-scMR—trajectory-inference-based dynamic single-cell MR, so far demonstrated only on healthy B-cell differentiation and multiple sclerosis brain tissue ti-scMR: trajectory-inference-based dynamic single-cell Mendelian randomization identifies causal genes underlying phenotypic differences—to tumor single-cell data along a drug-treatment pseudotime, tracking how the causal driver changes as cells transition from sensitive to resistant. That is a separate proposal, kept deliberately out of scope here so the static triage claim can be tested cleanly.
References
- Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes: A Drug‐Target Mendelian Randomization Study
- The deubiquitinating enzyme USP35 regulates the stability of NRF2 protein
- Causal effects of genetically determined blood metabolites on multiple myeloma: a Mendelian randomization study
- Causal association between cathepsins and asthma: a Mendelian randomization study
- The impact of metabolic syndrome on hepatocellular carcinoma: a mendelian randomization study
- Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma
- Knockdown of RRM1 in tumor cells promotes radio-/chemotherapy induced ferroptosis by regulating p53 ubiquitination and p21-GPX4 signaling axis
- ECLIPSE: A Composable Pipeline for Predicting ecDNA Formation, Evolution, and Therapeutic Vulnerabilities in Cancer
- Targeting Catalase but Not Peroxiredoxins Enhances Arsenic Trioxide-Induced Apoptosis in K562 Cells
- DeepCDR: a hybrid graph convolutional network for predicting cancer drug response
- Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
- Inhibition of glutamine metabolism counteracts pancreatic cancer stem cell features and sensitizes cells to radiotherapy
- ti-scMR: trajectory-inference-based dynamic single-cell Mendelian randomization identifies causal genes underlying phenotypic differences