Decoupling Pharmacokinetics from Pharmacodynamics: Host CYP Inhibition as a Perturbation Axis in Xenograft Transcriptomics

When a drug is tested in a mouse carrying a human tumor, the measured outcome mixes two different things: how strongly the drug acts on the tumor cell, and how much of it actually survives the mouse's drug-clearance machinery. The liver's cytochrome P450 (CYP450) enzymes — a family of proteins that chemically break down most drugs — set that exposure, and inhibiting them can multiply or collapse systemic drug concentrations. Yet the flagship preclinical effort linking tumor gene expression to drug response, the Novartis Patient-Derived Xenograft (PDX) Encyclopedia analysis built on the Xeva platform, treats response purely as an intrinsic property of the tumor's mutations and baseline RNA, with no model of host drug metabolism at all. Meanwhile, an entire separate literature predicts CYP450 inhibition and drug-drug interactions computationally, but never touches a living animal or measures what inhibition does downstream. Nobody has used controlled host metabolic enzyme inhibition as an experimental dial — deliberately turning the knob on systemic drug exposure in vivo — and read out the consequences in the xenografted tumor's transcriptome with modern, topology-aware enrichment methods.

The core claim of this proposal is that deliberate hepatic CYP450 co-inhibition in xenograft-bearing mice, followed by bulk and single-cell RNA sequencing of the tumor, can separate exposure-driven from efficacy-driven transcriptomic responses, and that standard enrichment methods will misread this signal because metabolic pathways share multifunctional enzymes. We will build a perturbation design — drug alone versus drug plus CYP inhibitor (pharmacokinetic boosting) versus inhibitor alone — and an analysis pipeline that combines topological pathway inference with a new exposure-aware enrichment model, then test whether pathway-level features that incorporate host metabolic state predict in vivo response better than tumor-intrinsic features alone.

Background and the Gap

Three bodies of work approach this intersection but stop short. The pharmacogenomic line, represented by the Xeva analysis of the PDX Encyclopedia, associated baseline tumor RNA, copy number, and mutations with drug response across 1,075 PDXs spanning 5 tissue types and 62 anticancer drugs, identifying 87 pathways significantly associated with response to 51 drugs at FDR < 5% and predictive biomarkers with concordance index above 0.60. Its blind spot is explicit: no hepatic clearance, no CYP-mediated exposure variation, no drug-drug interactions — response is a property of the tumor, not of the tumor-host pharmacokinetic system. Notably, the same paper showed that mouse passaging itself perturbs specific pathways (mRNA 3'-end processing and elongation arrest pathways at FDR < 5%, median Pearson correlation 0.97 across passages), proving the host already leaks into tumor transcriptomes in uncontrolled ways.

The computational metabolism line — the tree-based consensus models for the "Big Five" CYP isozymes, the multimodal MEN deep network achieving 93.7% accuracy and 0.985 AUC on CYP inhibitor prediction, and the long-tail TFMD framework reaching F1 of 0.7995 on DDI-DB171 where the DDI-SCL baseline manages only 0.5860 — predicts inhibition from molecular structure but is uncoupled from any living system. These models cannot say what inhibition does to a tumor's gene expression, and the OpenDDI benchmark's finding that accuracy drops by more than 50% on unseen drugs suggests in silico prediction alone is not a substitute for in vivo measurement.

The enrichment line — pareg's regularized regression with inter-pathway dependency penalties, and XPathway's graph-topological pathway activity inference from RNA-seq — fixed the "bag-of-genes" fallacy, but neither has ever been applied to pharmacological enzyme inhibition in a mammalian in vivo context. XPathway was validated in a marine bryozoan symbiosis; pareg on synthetic data and TCGA breast cancer. Both are exactly the tools needed to read out a perturbation that acts on multifunctional metabolic enzymes, and neither has met that perturbation.

There is also a genuine tension to confront: proteomic drug-response models outperform transcriptomic ones (Pearson r of 0.5467 versus 0.5311 for ln(IC50) prediction), and median RNA-to-protein correlation is only 0.436–0.539. This proposal does not ignore that — it makes the tension a measured outcome, checking which transcriptomic pathway signals survive at the protein level.

Core Claim and Hypothesis

Deliberate, controlled inhibition of host hepatic CYP450 activity in xenograft-bearing mice creates a graded exposure axis along which the tumor transcriptome can be decomposed. Genes and pathways whose response scales with exposure are pharmacokinetic echoes; those that change only when the tumor actually responds (shrinks) are efficacy signals; and standard overrepresentation analysis conflates them because metabolic enzymes participate in many pathways at once. An enrichment model that (a) uses pathway graph topology rather than gene lists, and (b) explicitly conditions on measured systemic exposure, will recover this structure where classical Gene Ontology/KEGG overrepresentation analysis and even tumor-intrinsic PDX pharmacogenomics cannot.

Research Program

The program has three components: an in vivo perturbation experiment, an analysis method, and a predictive test.

The in vivo experiment uses subcutaneous PDX cohorts (the ectopic site keeps the design affordable and compatible with the replicate-based and 1×1×1 designs that Xeva already handles) in immunodeficient mice, treated with an anticancer drug that is a known CYP3A4 substrate, under four arms: vehicle, drug alone, drug plus a CYP inhibitor selected to boost exposure (the pharmacokinetic-boosting logic established clinically with ritonavir), and inhibitor alone. The inhibitor-alone arm is essential — it isolates any direct tumor or host effect of the inhibitor itself, which is exactly the confound the current literature cannot rule out. Systemic exposure is measured by drug plasma levels at multiple time points per arm, so each mouse has a measured exposure value rather than an assumed one. Tumors are harvested for bulk RNA-seq and a subset for single-cell RNA-seq, targeting the rare resistant subclonal states that bulk averaging masks and that SCAD showed govern sensitivity (average AUC 0.831 for pre-treatment resistance prediction across seven drugs).

The analysis method is a new enrichment model we call exposure-conditioned pathway regression. It extends pareg's regularized regression — which regresses per-gene differential expression p-values on a gene-pathway membership matrix with LASSO and network-fusion penalties — in one specific way: pathway enrichment scores are conditioned on the measured exposure covariate and its interaction with pathway membership, so the model estimates, per pathway, both a baseline drug effect and an exposure-scaling coefficient. Pathway definitions use KEGG graph structure (compounds as nodes, enzymes as edges) following XPathway, so that inhibition-driven shifts in shared multifunctional enzymes are attributed across connected pathways rather than double-counted. This is the proposal's key technical ingredient: existing methods estimate whether a pathway is enriched; ours estimates whether a pathway's response is driven by how much drug the host let through.

The predictive test asks whether adding host-metabolic features improves response prediction over the tumor-intrinsic baseline. Using the Xeva analysis of the PDX Encyclopedia as the reference — 87 drug-pathway associations, biomarkers with concordance index above 0.60 — we test whether exposure-conditioned pathway features from our perturbation cohort achieve higher concordance between predicted and actual mRECIST response than baseline-expression-only features, in our own held-out replicate mice.

Baselines

The naive reference point is vehicle-only tumors: no drug, no inhibitor, expression measured at the same time points. Any pathway called "enriched" in treated arms must be compared against this to exclude handling and time effects. The strongest existing method on this question is the tumor-intrinsic pharmacogenomic pipeline of the Xeva platform, which identified the 87 drug-pathway associations and concordance-index-above-0.60 biomarkers on 1,075 PDXs; we will run its association machinery on our own cohort, feeding it the same tumors, the same RNA-seq depth, and the same replicate structure, but only baseline tumor expression — no exposure covariate, no inhibitor arm. The comparison is fair because the only difference is the presence of the host-metabolic axis, which is precisely the question. Beating it means a higher concordance index between predicted and observed response in held-out mice, with the improvement attributable to exposure-conditioned pathway features (a statistically significant increase in concordance, and pathway coefficients whose signs reproduce across independent PDX models of the same tumor type). Two ablations isolate the key ingredient: first, the same exposure-conditioned regression but with flat gene-set membership (no KEGG graph topology), to test whether topology matters or just the exposure covariate; second, the full model run on the drug-alone arm only, with exposure imputed rather than measured, to test whether measured inhibitor-arm exposure is doing real work.

Why not just…?

Why not just test drug combinations in 2D cell culture? This is the cheapest alternative, and it genuinely fails here for a structural reason: hepatic CYP clearance happens in the liver and intestine while the tumor grows elsewhere, so no plastic-plate assay contains the variable under study. As the xenograft literature in the file notes, drugs that work on plates frequently fail in vivo because of systemic pharmacokinetics — which is the whole point of this proposal. A 2D co-culture cannot have a plasma concentration. The inhibitor-alone arm in our design is the one piece a culture system could approximate (direct inhibitor effect on tumor cells), and we include it precisely to preempt that objection.

Why not just predict the interactions computationally with the existing CYP and DDI models? Models like MEN (93.7% accuracy on CYP inhibition) and TFMD (F1 0.7995 on long-tail DDI events) predict whether two molecules interact, but not what a changed exposure does to a tumor's gene expression or adaptive pathway rewiring. More damning, the OpenDDI benchmark reports that performance on unseen drugs drops by more than 50%, and clinical external validation of the consensus CYP models reached only ~79% accuracy on 132 real positive DDIs — these are not accurate enough to substitute for measurement in a preclinical efficacy context where the output variable (tumor transcriptome and response) does not exist in any training set. We will, however, use these models for drug and inhibitor selection, which is what they are good at: screening candidate booster-inhibitor pairs before committing animals.

Why not just run standard Fisher's exact test or GSEA on the differentially expressed genes? This is a real threat at the analysis stage, and the decisive evidence is already in the literature: pareg's synthetic benchmarks showed Fisher's exact test, topGO, MGSA, and blitzGSEA suffer from high false-positive rates under pathway redundancy and gene overlap, and XPathway showed classical enrichment mishandles multifunctional enzymes that participate in several pathways — exactly the enzymes this perturbation targets. Classical ORA also requires hard fold-change cutoffs that would discard the coordinated low-magnitude shifts across metabolic cascades that graded exposure produces. Rather than argue, our design includes a head-to-head: the same RNA-seq data analyzed by standard ORA, GSEA, pareg, and our exposure-conditioned model, scored on whether each recovers the exposure-scaling structure (dose-response of pathway scores against measured plasma levels) that the design builds in. If standard ORA recovers it equally well, the enrichment contribution of this proposal fails, and we will report that.

Why not just measure proteins directly instead of RNA? The proteomics-versus-transcriptomics paper shows RNA is a noisy proxy (median correlation 0.436–0.539) and proteomic models predict drug response slightly better (0.5467 vs. 0.5311 Pearson on ln(IC50)). But mass spectrometry carries severe missing-value problems, and no PDX-scale proteomic pharmacogenomic resource comparable to the PDX Encyclopedia exists, while standardized transcriptomic PDX pipelines do. Our answer is pragmatic: RNA-seq is the discovery screen, and we will validate a targeted panel of the top exposure-conditioned pathway enzymes at the protein level (Western blot or targeted mass spectrometry) on the same tumors. This turns the RNA-protein gap from an objection into a measured outcome: the fraction of transcriptomic pathway signals that are buffered post-transcriptionally is itself a result, and one that no study in the file has produced under controlled metabolic inhibition.

Proposed Experiments and Decisive Checks

Experiment 1 is the four-arm PDX perturbation, run in at least two PDX models of the same tumor type (to check model-to-model reproducibility of pathway coefficients) with replicate-based design as Xeva supports. Primary endpoints: measured plasma exposure ratios between boosted and unboosted arms (confirming the pharmacokinetic dial actually turned), tumor volume trajectories per arm (mRECIST), and bulk RNA-seq per tumor. The decisive check for the whole proposal is the exposure-scaling structure: per-pathway regression of pathway activity scores against measured individual-mouse plasma exposure should yield significant scaling coefficients (FDR < 5%) for a substantial set of metabolic pathways, and the scaling coefficients should be near zero for the inhibitor-alone arm.

Experiment 2 is the enrichment head-to-head on the Experiment 1 data: standard ORA, GSEA, pareg, and the exposure-conditioned model, evaluated on recovery of known exposure structure and on false-positive calls in the inhibitor-alone arm, where the inhibitor's own metabolic footprint (star-fruit-style multi-isoform inhibition patterns suggest simultaneous CYP3A4 and off-isoform effects) should generate confounds that only the exposure-aware model correctly attributes. XPathway's validation precedent — a 6.88-fold qPCR-confirmed change in fadA in the pathway it called — sets the bar: our top-called pathway enzymes get the same targeted qPCR/protein confirmation.

Experiment 3 is the predictive test: does adding exposure-conditioned pathway features beat tumor-intrinsic Xeva-style features on held-out mice? The number to beat is the tumor-intrinsic baseline from the PDX Encyclopedia analysis, where predictive biomarkers achieved concordance index above 0.60 at FDR < 0.05; our own tumor-intrinsic baseline run (identical data, identical compute, identical tuning budget) provides the exact within-cohort comparator. No reported result found on Lacuna for concordance of exposure-aware pharmacogenomic models in xenografts, which is precisely the empty cell this proposal fills.

Experiment 4 is the single-cell subclone arm: single-cell RNA-seq on a subset of tumors from the drug-alone versus boosted arms, testing whether increased exposure selectively eliminates or expands resistant subclonal states, using SCAD-style transfer of GDSC-derived resistance signatures (70 sensitive and 43 resistant biomarkers for cetuximab as the template) to annotate single-cell states, and deconvolving the bulk samples to check whether bulk-seq hides exposure-dependent subclone shifts that explain response outliers.

The program succeeds if the exposure axis is measurable, the enrichment model separates pharmacokinetic from efficacy signals where classical methods cannot, and response prediction improves over the tumor-intrinsic baseline. It fails informatively if pathway responses do not scale with exposure — which would itself argue that xenograft transcriptomic pharmacogenomics, as currently practiced, is not confounded by host metabolism and can safely ignore it. Either answer closes a gap that every paper in this literature currently leaves open.