Metabolic Phenocopy: Attributing Drug-Induced Metabolic Rewiring to Off-Target Proteins

A drug almost never touches just one protein. It binds a primary target and a handful of secondary ones, and the cell answers by rewiring its metabolism — switching between burning sugar in mitochondria and fermenting it, rerouting carbon through collateral pathways, or resisting the drug entirely. The hard question is attribution: which of the secondary proteins actually caused the metabolic shift? Today this is answered with slow, per-target knockdown experiments — silence one gene at a time and see whether the drug's footprint reappears. The study that pioneered this "phenocopy" logic needed 621 microarrays to qualify seven compounds against a single target Phenocopy – A Strategy to Qualify Chemical Compounds during Hit-to-Lead and/or Lead Optimization.

I propose to answer the attribution question computationally, at proteome scale, by fusing three existing pieces that have never been connected: chemical-similarity prediction of a compound's target panel, genome-scale metabolic models in which any gene can be deleted for free, and enrichment analysis that includes the non-enzyme regulators of metabolism. The output, for each drug and cell line, is a ranked list of the off-target proteins whose perturbation best explains the drug's measured metabolic response — a target-validation signal built on causal consistency rather than expression correlation, which is precisely the trap that makes most prognostic biomarkers useless as drug targets Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes.

Core Claim

For a compound with a measured transcriptomic response in a given cell line, the metabolically causal off-targets can be identified by scoring each predicted binding partner on how well an in silico perturbation of that partner, projected through protein-interaction-expanded metabolic gene sets, reproduces the compound's own metabolic enrichment signature. Two tiers of evidence are combined. For any candidate target, we test whether the drug's differential response genes are overrepresented in the target's physical-interaction neighborhood — the same expansion that made network-based metabolic enrichment dramatically more sensitive Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer. For targets that are metabolic enzymes, we add a stronger check: delete the gene in a cell-line-specific metabolic model, record which reactions change flux, and require flux-level agreement with the observed response.

The attribution score for compound $c$ and candidate target $t$ is

$$ S(c,t)=\pi(t\mid c)\;\rho\!\left(\mathbf{e}_c,\mathbf{e}_{t}\right), $$

where $\pi(t\mid c)$ is the fishing prior that compound $c$ binds target $t$ (from chemical-similarity target prediction), $\rho$ is a rank correlation, $\mathbf{e}_c$ is the drug's enrichment vector over metabolic pathways computed with interaction-expanded gene sets, and $\mathbf{e}_{t}$ is the enrichment vector of the target's perturbation — knockout flux changes for enzymes, neighborhood membership for non-enzymes such as kinases. Targets ranking high on both evidence tiers are proposed as validated metabolic effector targets, and their predicted exchange fluxes (for example, lactate secretion) become directly testable metabolomics readouts.

This differs from the closest existing work in both directions it borrows from. MOMENTA introduced the interaction-expanded gene sets and validated them against mass spectrometry, but it maps proteomic states to pathways and never touches a compound, a binding prediction, or a target-ranking problem Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer. The phenocopy study attributes compound signatures to targets, but requires a measured knockdown signature per target and traces no metabolic network or pathway statistics Phenocopy – A Strategy to Qualify Chemical Compounds during Hit-to-Lead and/or Lead Optimization. A third line connects chemistry to response: canonical correlation between 3D molecular descriptors and genome-wide response profiles across 1,159 drugs Comprehensive data-driven analysis of the impact of chemoinformatic structure on the genome-wide biological response profiles of cancer cells to 1159 drugs — but that mapping is linear and stops at "this chemical feature predicts mitochondrial stress," never naming the protein responsible. My proposal closes that loop: from structure, to target panel, to in silico perturbation, to a ranked, flux-consistent causal claim.

Research Program

Three modules and a scoring layer are what actually gets built. The first module is target fishing. Following the benchmark that compared seven target-prediction tools on ChEMBL 34 with 100 held-out FDA drugs, I use ligand-centric prediction with Morgan fingerprints and Tanimoto similarity (MolTarPred), which outperformed target-centric and deep-learning predictors in precision, recall, and Matthews correlation coefficient A precise comparison of molecular target prediction methods. Crucially, training keeps lower-confidence annotations: that same benchmark found a "high-confidence paradox," where strict potency filtering degraded repurposing performance because the moderate-affinity off-targets that matter most for phenotype get discarded.

The second module is cell-line-specific metabolic models. The GEMbench study built 3,740 context-specific models from Recon3D across six omics datasets and four integration algorithms, and found that transcriptomic integration with GIMME and INIT yields markedly more consistent models than proteomic integration with iMAT or FASTCORE Exploring the Metabolic Heterogeneity of Cancers: A Benchmark Study of Context-Specific Models. I therefore build models for each profiling cell line with GIMME or INIT from RNA-seq, and use single-gene deletions to obtain flux-change profiles for every fished enzymatic target.

The third module is the enrichment engine: MOMENTA-style "Neighborhood" gene sets — metabolic pathway annotations expanded with first-degree physical interactors — applied to both the drug response and the in silico perturbation. The scoring layer then computes $S(c,t)$ for every drug–target–cell-line triple and, for enzymatic targets, additionally checks that predicted exchange-flux changes point in the same direction as any measured metabolite shifts.

A calibration step anchors the method on ground truth we already trust: for drugs in the profiling compendium whose primary target is well established, the primary target should rank first. If it does not, the mapping from flux change to enrichment vector is mis-calibrated and gets fixed before any off-target claims are made.

Proposed Experiments

The primary data source is the compendium of 1,159 drugs profiled across more than 11,000 genes in three cancer cell lines (MCF7, PC3, HL60), coupled to 76 three-dimensional chemical descriptors Comprehensive data-driven analysis of the impact of chemoinformatic structure on the genome-wide biological response profiles of cancer cells to 1159 drugs. This is the only resource in the literature I have that joins compound structures to genome-wide cellular responses at scale, and it is exactly what the attribution task needs. Evaluation proceeds in four experiments.

The first experiment re-runs the enrichment power comparison on drug data: base enzyme-only KEGG sets versus interaction-expanded neighborhood sets, counting pathways significant at adjusted p < 0.05. The reference point is MOMENTA's own result — in glucose-starved MDA-MB-231 cells, neighborhood sets yielded a 25-fold increase in significant metabolic pathways relative to traditional GSEA on base sets, and expanded sets carried up to five times more unique genes per pathway Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer. If a comparable gap does not appear on drug-perturbation transcriptomes, the premise of pathway-level off-target analysis weakens and the proposal should be scaled back.

The second experiment is the attribution benchmark proper. Ground-truth off-targets are assembled from ChEMBL annotations and literature-flagged secondary targets for the 1,159 compounds; metrics are recall at rank 20 and area under the receiver operating characteristic curve for recovering true secondary targets, plus rank correlation between predicted and curated perturbation signatures. The third experiment is the head-to-head on the TGF-β receptor 1 phenocopy cohort: seven compounds (five indolinones, two pyridopyrimidinones) with siRNA-matched signatures across 621 microarrays, where the promiscuous compound BI1 was measured to inhibit 75 kinases while the cleaner BI2 inhibited 21 Phenocopy – A Strategy to Qualify Chemical Compounds during Hit-to-Lead and/or Lead Optimization. The test is whether the pipeline, given no siRNA data, predicts BI1's broad off-target burden and BI2's cleanliness, and whether the predicted off-target neighborhoods explain the measured caspase-3 and interleukin-6 phenotypes that the phenocopy authors linked to scaffold-clustered off-target signatures.

The fourth experiment probes a genuine tension in the source literature. GEMbench found proteomic integration inferior; MOMENTA found proteomic data highly informative once interaction networks are included Exploring the Metabolic Heterogeneity of Cancers: A Benchmark Study of Context-Specific Models, Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer. Drug responses offer an independent perturbation readout: I test whether GIMME/INIT-derived or iMAT/FASTCORE-derived knockouts better match measured responses to metabolic drugs, using agreement as an external arbiter between integration algorithms. Where profiling lines overlap the 928-line, 225-metabolite mass-spectrometry compendium, predicted exchange-flux changes are checked against measured metabolite shifts; if the overlap turns out to be empty, this sub-experiment is reported as untestable on current data rather than forced. All methods run on a single CPU-scale budget, with equal hyperparameter search allocated to every competitor.

Baselines

The naive reference point is ChEMBL lookup: take a compound's annotated targets directly from ChEMBL 34, with no prediction and no metabolic modeling, and ask how often annotated secondary targets alone explain the metabolic enrichment signature A precise comparison of molecular target prediction methods. This measures how far curation alone gets; beating it means recovering metabolically active off-targets that are absent from annotations, at a precision the lookup cannot reach by construction.

The strongest existing competitor is the phenocopy method itself: matching compound-induced transcriptomes against measured siRNA knockdown signatures, which demonstrably separates on-target from off-target signals and predicts downstream toxicity phenotypes Phenocopy – A Strategy to Qualify Chemical Compounds during Hit-to-Lead and/or Lead Optimization. To make the comparison fair, I run both on the identical TGF-βR1 cohort — same seven compounds, same 621 microarrays, same ground-truth kinase panels — building the HaCaT metabolic model from that study's own expression data so no extra information favors my pipeline. The honest expectation is that measured knockdowns are more faithful where they exist; the win condition is coverage and agreement: recovering a comparable off-target panel for BI1 and BI2 without a single siRNA experiment, and matching or exceeding the phenocopy method's recovery of the measured 75-versus-21 kinase asymmetry.

The second strong competitor is the canonical-correlation approach, reimplemented on the same 1,159-drug, three-line compendium with the same descriptor set and an equal tuning budget Comprehensive data-driven analysis of the impact of chemoinformatic structure on the genome-wide biological response profiles of cancer cells to 1159 drugs. That work reports ten structure–response components, with its Component 3 linking hydrogen-bonding descriptors to mitochondrial and metabolic stress — but it names components, not proteins, and no target-attribution accuracy exists for it; no reported result found on Lacuna for a target-level attribution metric from that method. Beating it means demonstrating target-level resolution: ranking the correct off-target protein for a given metabolic signature, which a component-level correlation cannot produce by design.

Two ablations isolate the proposal's key ingredients. Running the full pipeline with enzyme-only base gene sets instead of neighborhood sets tests whether the interaction expansion, shown to matter for proteomic enrichment in MOMENTA's glucose-starvation data, also carries the attribution signal Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer. Running target fishing alone, without the perturbation-consistency layers, tests whether fishing priors by themselves already solve attribution — if they do, the metabolic modeling adds nothing and the proposal collapses to a known method.

Why not just…?

Why not just develop a selective inhibitor of one master metabolic enzyme? The evidence in the file says single-node blockade is systematically evaded: lung adenocarcinomas shunt glucose into the hexosamine pathway as an alternative outlet Elevated glucose metabolism via the hexosamine biosynthesis pathway: A metabolic signature of high-fluorodeoxyglucose-uptake lung adenocarcinoma, branched-chain amino acid reprogramming drives resistance to EGFR inhibitors Branched-Chain Amino Acid Metabolic Reprogramming Orchestrates Drug Resistance to EGFR Tyrosine Kinase Inhibitors, and ACOT8 blockade is bypassed through lipid and anti-ferroptotic routes Acyl-CoA thioesterase 8 induces gemcitabine resistance via regulation of lipid metabolism and antiferroptotic activity in pancreatic ductal adenocarcinoma. Synthetic lethality exists precisely because single-gene loss is tolerated Benchmarking machine learning methods for synthetic lethality prediction in cancer. But I should be honest: this proposal does not oppose selective inhibitors — it is a diagnostic that can nominate them. If a single target's in silico knockout alone flips the metabolic signature, the pipeline will rank it first, and the attribution experiment itself settles whether single or paired targets dominate.

Why not just run standard GSEA or KEGG overrepresentation on the drug-induced gene list? Direct measurement says it under-samples the biology: in glucose-starved breast cancer cells, expanding metabolic gene sets with first-degree physical interactors produced a 25-fold increase in significant metabolic pathways, and enzyme-only sets miss up to 80% of functionally relevant interactors, correlating poorly with measured metabolomics Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer. More fundamentally, flat gene sets return pathway names, not causes — they cannot say which protein a drug should avoid or hit next. The base-set ablation quantifies exactly this shortfall on drug data rather than asserting it.

Why not just match drug signatures against a large measured siRNA knockdown library? This is the real threat, because the phenocopy logic is proven Phenocopy – A Strategy to Qualify Chemical Compounds during Hit-to-Lead and/or Lead Optimization. Its shortfall is throughput and coverage: the original study needed 621 arrays to cover one target, and the approach's cost scales linearly with the number of candidate targets, while ChEMBL's sparseness leaves most of the proteome's bioactivity dark A precise comparison of molecular target prediction methods. My pipeline replaces measured knockdowns with free in silico ones and covers non-enzymes through neighborhood enrichment. The head-to-head on the TGF-βR1 cohort is designed to settle this: if, where knockdown data exist, the in silico version fails to reproduce the phenocopy findings, the claim shrinks to "useful only where no knockdown data exist" — a weaker but still honest contribution.

Decisive Checks

Four results would make or break the proposal. The neighborhood-versus-base ablation must reproduce a MOMENTA-scale sensitivity gap on drug-perturbation data, not just proteomic data. Attribution must beat both the ChEMBL-lookup reference and the reimplemented canonical-correlation method on recall at rank 20 and AUROC, and the fishing-only ablation must fall short — otherwise the metabolic modeling is decorative. On the phenocopy cohort, the pipeline must predict BI1's measured promiscuity (75 kinases) against BI2's cleanliness (21) without siRNA input. And for at least a handful of metabolic drugs, flux-consistent target predictions must agree in direction with measured metabolite shifts where line overlap permits the test. If all four hold, the field gains something it currently lacks: a way to turn any profiled compound into a ranked, causally-grounded list of the off-target proteins rewiring its metabolism — the input that polypharmacological design methods, which today pick target pairs like MEK1/mTOR by hand Geometric Pocket-Centric Protein Encoding for Polypharmacology-Guided Multi-Target Drug Design, Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design, have never had.