Biphasic Dose-Time Network Surfaces for In Vitro–In Vivo Translation

Many chemicals do opposite things to a cell depending on how much of the chemical the cell actually sees. A small dose of a toxic metal, a drug, or reactive oxygen species can stress a cell into a stronger, defended state; a large dose of the same agent exhausts the cell and kills it. This non-monotonic behavior — called hormesis or biphasic dose-response — is documented everywhere from sunflower seedlings germinating better under trace arsenic (Arsenic Toxicity to Germination and Vegetative Growth of Sunflower) to dihydrotestosterone lengthening hair follicles at low concentration and shrinking them at high concentration through the same Wnt/β-catenin pathway (Dihydrotestosterone Regulates Hair Growth Through the Wnt/β-Catenin Pathway). Yet standard analysis pipelines assume monotonicity, and standard explanations for why a petri-dish result fails in an animal invoke biological complexity: missing immune cells, missing tissue structure. There is a simpler, testable suspect that the literature has noticed but never followed through on: the animal's tissue never experiences the same dose as the dish did, so it sits elsewhere on the same U-shaped curve. In Dose-Dependent Effects in Plasma Oncotherapy, cold plasma that was cytotoxic to tumor cells in vitro accelerated tumor growth in mice precisely because the tumor interstitial fluid received far lower reactive-species concentrations — a sub-lethal, growth-stimulating dose.

I propose to test the "dose-shift" explanation of the in vitro–in vivo translation gap at transcriptome scale, using the one public resource designed for it: Open TG-GATEs, which holds 20,118 GeneChip assays on 170 compounds tested in parallel in primary rat and human hepatocytes (in vitro) and rat liver and kidney (in vivo), across multiple doses and multiple time points. The plan is to replace static gene-expression comparison with a dose- and time-resolved co-expression network model, fit non-monotonic response surfaces on the in vitro arm, and ask whether the in vivo arm is predicted by evaluating those surfaces at a learned, compound-specific effective dose.

Core Claim

The claim has two parts, one methodological and one substantive. Methodologically: gene co-expression should be modeled as a smooth surface over dose and time rather than as a single static network, because a biphasic agent makes two genes correlate positively in the adaptive zone and uncouple or invert in the toxic zone, and a static network averages that opposition toward zero and erases it. Weighted gene co-expression network analysis — WGCNA, which groups genes that move together and summarizes each group by its module eigengene, the first principal component of the group's expression matrix (Session Introduction) — must therefore be made dose-time-aware before it can see hormesis at all. Substantively: for each compound, module eigengenes trace a conserved, non-monotonic surface over dose and time, and the in vivo state is the in vitro surface evaluated at a shifted dose. Formally, for compound $c$, administered dose $d$, and time $t$, I model each module eigengene as

$$\mathrm{ME}_m(c, d, t) = f_m\big(w_c(d),\, t\big),$$

where $f_m$ is a non-monotonic surface fitted on in vitro data and $w_c$ is a monotone increasing "dose warp" mapping the in vivo administered dose to an in vitro-equivalent effective dose. The warp absorbs pharmacokinetics, protein binding, and tissue metabolism; it is learned, not measured, because TG-GATEs has no matched pharmacokinetic data. If a single learned warp per compound predicts held-out in vivo eigengene trajectories, then a large fraction of the translation gap is not mysterious biology but a predictable slide along a biphasic manifold — and, crucially, the tipping point where adaptation collapses into toxicity becomes a mapped, forecastable region of dose-time space rather than an unstable surprise.

This is unambiguous about what would refute it: if the warps are unstable across cross-validation folds, or if in vivo eigengenes are predicted no better by warped in vitro surfaces than by unwarped ones, then the dose-shift story is quantitatively minor and the biology-only explanation wins. The proposal is built to let the data decide.

Research Program

The program has four components. First, I will construct dose-time-stratified co-expression networks on the TG-GATEs in vitro arm. Rather than one global network, each (compound, dose, time) condition gets its own covariance estimate, pooled hierarchically so modules are shared across conditions while their wiring varies. Full-transcriptome covariance per condition is computationally prohibitive, so I will reuse the amortized pairwise-regression surrogate from Cell-Level Virtual Screening, which the authors show reduces network prediction error by 12–47% over population-based estimators, and the consensus-module machinery validated cross-species in Cross-Species Gene Expression Analysis Reveals Gene Modules Implicated in Human Osteosarcoma, where conserved modules and a hub gene survived independent validation cohorts.

Second, I will fit non-monotonic dose-time surfaces $f_m$ for each module eigengene using low-rank spline or Gaussian-process regression in two dimensions (dose, time). This is where the biphasic literature enters as a design constraint: the fitter must allow an interior extremum, unlike the Hill or four-parameter logistic models that pharmacology uses by default and that mathematically cannot produce a stimulatory low-dose zone, as Predicting Dose-Response Curves with Deep Neural Networks demonstrates. The temporal axis matters equally: dose and time are coupled, since a low dose held long enough can become cumulatively toxic and an initial shock can resolve into over-compensation, a coupling seen in the evolving toxic footprint of Jiahui Yan et al. and in the four-phase (alarm, remodel, integrate, rewire) progression of the low-dose radiation profiling study. That paper also supplies a hard methodological lesson: dose dominated phenotypic variance and standard cross-validation leaked temporal information through culture drift, which they fixed with a two-stage residualization and leave-one-week-out protocol. I will port that protocol — dose-only baseline, residual temporal modeling, leave-one-time-point-out validation — to TG-GATEs, whose own documented hazards (10-year batch variability across contract research organizations, cell-death artifacts at cytotoxic doses) demand it.

Third, I will learn the dose warps $w_c$ by minimizing discrepancy between predicted and observed in vivo eigengene trajectories, and validate under leave-one-compound-out and leave-one-dose-out splits. The rat hepatocyte-to-rat-liver pair is the cleanest test (same species); the human hepatocyte arm then adds a cross-species dimension on top. Fourth, I will decompose the residual in vivo variance: whatever the warped surfaces fail to predict should be enriched for genuinely in vivo biology — immune, stromal, vascular processes invisible to hepatocyte monoculture — giving a quantitative split between the dose-shift explanation and the complexity explanation, in the spirit of the CAP oncotherapy finding that low interstitial doses engaged PD-1/PD-L1 checkpoints no dish could contain.

Baselines

The naive reference point is nominal-dose carryover: predict the in vivo state at dose $d$ and time $t$ by reading the in vitro state at the same nominal dose and time, plus a do-nothing variant that predicts no change from vehicle control. This baseline is not a strawman — it is what the field implicitly does whenever an in vitro concentration is assumed to equal the in vivo concentration — and the entire warp hypothesis lives or dies on whether it beats this reference on held-out compounds.

The field-standard competitor is classic static WGCNA exactly as specified in the canonical pipeline: one global network over all in vitro samples regardless of dose or time, module eigengenes regressed onto dose and time with linear or logistic terms, then applied to the in vivo arm. It is the strongest standard competitor because it uses the identical data and the same module-eigengene representation, differing only in the ingredient under test — dose-time stratification. It doubles as the decisive ablation: if hormetic inversion is real, static WGCNA should average it to zero and lose specifically on modules that invert.

The strongest existing method is CellVS-Net's amortized network prediction (Cell-Level Virtual Screening), which is the closest prior art: it predicts context-specific network rewiring across dosages and chemicals and showed that network representations carry more therapeutic signal than even oracle measured expression snapshots, retrieving clinical indications at a Hits@1 of 0.1250 versus 0.0179 for random and 0.0357 for molecular fingerprints on its DDR-Bench. It has never been run on in vivo data or with a non-monotonic dose-time surface, so I will port its surrogate-loss architecture to TG-GATEs, train it on the in vitro arm, and apply it to the in vivo arm at matched nominal dose. To keep the comparison fair, all methods get the same gene sets, the same cross-validation splits, the same soft-threshold and module-count tuning budgets, and the same compute. Beating it means lower held-out error on in vivo module eigengene trajectories (RMSE and per-module correlation) under leave-one-compound-out, with the gap attributed to the two ingredients it lacks — the biphasic surface and the warp — via the ablations below.

Two further ablations isolate the proposal's key ingredients: one removes the warp only (non-monotonic surface, nominal dose), and one removes non-monotonicity only (monotone four-parameter-logistic surface in dose per module, with warp). For the binary in vivo pathology endpoints, I will additionally retrain a DeepSADR-style static transfer model — fragment-pathway graphs with adaptive readout, which achieved a clinical Sorafenib AUC of 0.957 versus 0.811 for the next best baseline in DeepSADR — on TG-GATEs. That reported number is from a different task and corpus (DepMap to TCGA clinical response); no reported result found on Lacuna for DeepSADR, CellVS-Net, ARCANet, or any static transfer method applied to in vivo prediction on Open TG-GATEs, which is precisely the open seat this proposal occupies.

Why not just…?

Why not just run differential expression at each dose and time point? The evidence says single-gene analysis throws away exactly the signal at stake. CellVS-Net found that predicted network rewiring beat even measured expression snapshots at retrieving clinical indications (Hits@1 of 0.1250 versus far lower baselines, in Cell-Level Virtual Screening), and the HGSK versus MSigDB comparison showed data-driven co-expression modules outperform curated gene sets for drug response in 40–60% of cases. Beyond predictive power, per-condition differential expression yields no continuous object in dose and time to warp — there is no surface to slide along. Still, the DEG analysis will be run and reported as a descriptive baseline so the reader can see the difference rather than take it on faith.

Why not just fit a Hill or four-parameter logistic curve per gene and read off the dose shift? These models are monotone by construction and so cannot represent the stimulatory zone at all, as ARCANet argues when replacing them with neural curve prediction. The honest counterpoint is that ARCANet's own authors found that below roughly 2,000 data points simple parametric models are more robust than high-capacity networks — a real threat here, since per-gene, per-compound fits in TG-GATEs would be that sparse. The design answers it by pooling: surfaces are fitted per shared module across compounds and conditions, giving thousands of observations per surface, and the monotone-surface-plus-warp ablation means that if monotonicity suffices, the ablation will say so and the biphasic claim will be appropriately weakened.

Why not just align global gene expression between the in vitro and in vivo domains (standard domain adaptation)? DeepSADR showed global expression alignment fails for two reasons — drugs act through substructures on specific pathways, not whole-molecule on whole-genome, and the mechanism itself shifts between contexts — and its own ablation found that replacing the adaptive readout with standard pooling collapsed average AUC from 0.856 to 0.662 (DeepSADR). Direct alignment additionally bakes in the wrong assumption: that the administered dose is the effective dose, which is exactly what the warp is supposed to correct. This alternative is a genuine threat for the binary pathology endpoint, where a strong static classifier might match my model, and I will settle it empirically by running the DeepSADR-style baseline on that endpoint; if it matches on pathology AUROC while failing on continuous module trajectories, that would itself be an informative result about where dose-awareness matters and where it does not.

Proposed Experiments and Decisive Checks

The primary evaluation is held-out prediction of in vivo module eigengene trajectories for unseen compounds: RMSE and Spearman correlation per module, with leave-one-compound-out as the headline split and leave-one-dose-out and leave-one-time-point-out splits (the last to prevent the temporal leakage documented in the radiation study). Secondary endpoints are AUROC for the in vivo pathology and clinical-chemistry labels that TG-GATEs carries alongside its 25 TB of digitized histology, and a tipping-point audit: the dose-time coordinates where adaptive modules (antioxidant and stress-defense eigengenes) invert should agree with the 36 mechanism-specific biomarker sets already defined on this database and with the cytotoxic-dose artifacts the TG-GATEs authors flag, which I will handle by excluding or separately flagging cytotoxic conditions rather than letting cell-death artifacts masquerade as network rewiring.

Three checks make the outcome decisive rather than anecdotal. First, warp stability: the learned $w_c$ must be recoverable across independent folds, and its implied effective-dose discount should be compound-plausible (steady, monotone, mostly sub-linear). Second, attribution: the gap between warped-surface prediction and nominal-dose carryover quantifies the dose-shift share of the translation gap, while the warped-surface residual, tested for enrichment of immune and stromal signatures, quantifies the biology-only share — the same contrast the CAP study dramatized but could not decompose transcriptome-wide with its targeted qRT-PCR panel. Third, negative controls: warps fitted on one compound must degrade prediction when applied to another, confirming the warps carry compound-specific pharmacokinetics rather than a generic fudge factor. If all three hold, the deliverable is a practical one — a fitted, transferable dose-time surface library that tells a toxicologist or pharmacologist, before any animal is dosed, where on the biphasic curve the target tissue is likely to land, and where the adaptive zone gives way to collapse.