Time-Varying Co-Expression Networks to Stage Chronic Diet-Induced Liver Disease

Drugs for fatty-liver disease keep failing in the clinic partly because the animal models that discover them are the wrong shape: a chemical toxin destroys a mouse liver in days, while the human disease smolders for decades under a sustained nutritional overload. A genuine chronic model does exist — the Gubra Amylin NASH (GAN) diet mouse, fed high fat, fructose, and cholesterol for up to 72 weeks, spontaneously develops advanced fibrosis and liver cancer, and even reproduces the human paradox in which liver fat recedes as the disease worsens, with hepatic lipid area peaking around week 24 and declining by week 72 (Semaglutide reduces tumor burden in the GAN diet-induced obese and biopsy-confirmed mouse model of NASH-HCC with advanced fibrosis). But the analysis tools have not caught up with the model. The standard systems-biology method, weighted gene co-expression network analysis (WGCNA — grouping genes into modules of co-regulated expression and summarizing each module by its first principal component), pools all samples into one static network, treating a week-24 and a week-72 liver as interchangeable if their gross phenotype looks similar. On a biphasic disease trajectory, that is exactly the wrong assumption.

This proposal rebuilds co-expression analysis so the network itself is a function of time, applies it to the longitudinal GAN cohorts, and extracts a continuous "chronicity coordinate" that replaces the field's categorical acute-versus-chronic vocabulary. The payoff is concrete: in the 72-week GAN study, semaglutide cut tumor incidence from 88% (14/16 vehicle) to 40% (6/15) and tumor burden from 8.3 to 1.5 tumors per mouse, while lanifibranor — more effective at reducing established fibrosis — left tumor burden untouched at 10.3 tumors per mouse (Semaglutide reduces tumor burden in the GAN diet-induced obese and biopsy-confirmed mouse model of NASH-HCC with advanced fibrosis). Two drugs, both aimed at the same chronic injury, diverge on the outcome that matters. No current analysis method can say in network terms why.

Core Claim

The acute-to-chronic gap is usually described categorically — an animal model is "acute" or "chronic" — but it can be measured continuously as a property of the co-expression network's trajectory: the point at which module structure stops responding to removal of the inciting stress. Prior work supplies the ingredients but never combines them. A conserved 189-gene signature already works as a molecular ruler of biological time, quantifying de-differentiation in tumors (cancer gene expression), and principal-component projection of stem-cell-derived cardiomyocytes against fetal and adult tissue established a quantitative "maturation gap" (Ellen Poon et al.). Both are rulers of developmental time built from static snapshots. The closest temporal-and-network studies stop short of dynamics: an acute myocardial infarction study ran WGCNA and observed that the peripheral-blood transcriptome shifts from inflammatory to metabolic programs over six months, but the network itself was built statically on pooled cohorts and the longitudinal data used only post hoc to watch markers move (Potential biomarkers of acute myocardial infarction based on weighted gene co-expression network analysis); a cortical-development study correlated static module eigengenes with age but never modeled edge rewiring between stages (Temporal proteomic profiling of postnatal human cortical development). Neither touched a diet-induced chronic disease, and neither asked a therapeutic question. Our claim is that a time-varying, cell-mixture-corrected network, fit to the GAN cohorts, will (i) reproduce the biphasic burned-out trajectory without being told the histology, (ii) separate the modules governing fibrosis remodeling from those governing neoplastic transformation, and (iii) define a scalar staging coordinate whose saturation under lifestyle reversal quantifies irreversibility — the continuous version of the acute-to-chronic boundary the field currently lacks.

Method: Networks That Move

The core obstacle is that temporal profiling in vivo is destructive: animals are sacrificed at discrete intervals, so no single liver is tracked continuously. We borrow the solution from a supervised Bayesian method that used Gaussian Process (GP) regression — a flexible curve-fitter that reports predictive uncertainty — to re-annotate the timing of transcriptomic samples, predicting time points within a 30-minute margin of true time and keeping 97% of annotations within 0.5 hours even when 30% of genes were perturbed (A supervised Bayesian method for time (re)annotation of transcriptomics data). We fit a GP trajectory $\hat{g}_i(t)$ for every gene across the population of sacrificed animals, then replace WGCNA's single pooled adjacency with a sliding-window version computed on the fitted curves:

$$A_{ij}(t) = \left| \mathrm{cor}\left(\hat{g}_i(s), \hat{g}_j(s)\right)_{s \in [t-\Delta, t+\Delta]} \right|^{\beta}$$

where $\beta$ is the usual soft-thresholding power and $\Delta$ the window half-width, with GP posterior variance propagated into confidence intervals on every edge and module eigengene. Where static WGCNA computes one adjacency $A_{ij} = |\mathrm{cor}(x_i,x_j)|^\beta$ over all samples (Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease), this yields a family of networks indexed by disease time, from which we can read off when modules form, dissolve, and swap hub genes. The chronicity coordinate is then a simple projection:

$$\kappa(t) = \frac{\lVert e(t) - e_{\mathrm{early}}\rVert}{\lVert e_{\mathrm{chronic}} - e_{\mathrm{early}}\rVert}$$

where $e(t)$ is the vector of module eigengenes at time $t$ and $e_{\mathrm{early}}$, $e_{\mathrm{chronic}}$ are the eigengene states of early adaptive and terminal fibrotic reference groups.

One confound must be removed for $\kappa$ to mean anything. Bulk liver homogenates mix shifting cell populations with genuine intracellular regulation, and the RNF181 study explicitly lists this lack of cell-type specificity as a core WGCNA limitation (Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease). The concern is real, not hypothetical: sleeve gastrectomy raises macrophage and NK-cell fractions while dietary restriction depletes hepatocytes and stellate cells and disrupts periportal zonation (Multiomics analyses decipher intricate changes in the cellular and metabolic landscape of steatotic livers upon dietary restriction and sleeve gastrectomy), and $RAMP1^{-/-}$ mice on high-fat diet fail to mount compensatory monocyte-derived macrophage recruitment (RAMP1 signaling attenuates diet-induced steatotic liver disease in mice). We will deconvolve each bulk sample against single-nucleus references collected at anchor time points and report composition-corrected and uncorrected networks side by side, so that "the cells changed" is never silently sold as "the regulation changed."

Research Program

The first aim is retrospective and computational: fit the time-varying networks to the two public GAN transcriptomic cohorts — the 72-week vehicle/semaglutide/lanifibranor study deposited as GSE243976 and the 34-week-then-chow-reversal study deposited as GSE246328 (Semaglutide reduces tumor burden in the GAN diet-induced obese and biopsy-confirmed mouse model of NASH-HCC with advanced fibrosis; Dietary intervention reverses molecular markers of hepatocellular senescence in the GAN diet-induced obese and biopsy-confirmed mouse model of NASH). Both studies measured transcriptomes but analyzed them either as baseline/endpoint comparisons or through pre-selected senescence genes, never as genome-wide network trajectories. The second aim is dissociation: identify modules whose eigengenes shift under semaglutide but not lanifibranor (tumor-protective candidates) versus the reverse (fibrosis-remodeling candidates), and test whether held-out module states at the week-54 biopsy predict which mice later develop tumors. The third aim is irreversibility: track $\kappa(t)$ through the chow-reversal time series, where body weight normalized by 22–25% within eight weeks and p21-positive senescent cells — 40-fold elevated in NASH livers — returned to baseline, yet the clinical fibrosis stage did not significantly improve within 24 weeks (Dietary intervention reverses molecular markers of hepatocellular senescence in the GAN diet-induced obese and biopsy-confirmed mouse model of NASH). Our prediction is that $\kappa$ keeps rising or saturates while gross physiology recovers, giving the field its first continuous molecular readout of the point of no return — the quantitative staging metric the acute-to-chronic literature says is missing.

Proposed Experiments

The first experiment runs entirely on public data. Applying the dynamic pipeline to GSE243976 and GSE246328, the decisive check is whether a steatosis module's eigengene trajectory, fit blind to histology, reproduces the known biphasic lipid curve — peaking near week 24 and declining toward week 72 — which a pooled static network cannot in principle recover, since it discards sample order. Success also requires that predicted hub genes in the semaglutide-shifted modules be validated against the tumor-incidence outcome, using held-out mice: train on module states at biopsy, predict tumor development under vehicle, and report area under the ROC curve with GP-derived uncertainty. The second experiment fills the sampling gap the retrospective data cannot: a staggered-sacrifice GAN cohort (weeks 12, 24, 36, 54, 72; n = 4–6 per point) with single-nucleus RNA-seq at three anchor points to build the deconvolution reference, plus targeted validation of two predicted hub genes by spatial or proteomic assay. We acknowledge the cost — this is the same 34-to-72-week maintenance burden the GAN studies already accept — and note that RNA–protein concordance decays over chronic intervals, with paired cortical profiling showing average RNA–protein correlation decreasing with age ($r = -0.56$, $p = 6.4 \times 10^{-5}$) (Temporal proteomic profiling of postnatal human cortical development); every hub-gene claim must therefore be checked against protein-level or histological endpoints, not transcript abundance alone. The third experiment tests generalization across species and interventions: applying cell-type-resolved dynamic networks to the rat sleeve-gastrectomy versus dietary-restriction cohort, whose single-cell data exist but only at one terminal endpoint (Multiomics analyses decipher intricate changes in the cellular and metabolic landscape of steatotic livers upon dietary restriction and sleeve gastrectomy), asks whether dietary and surgical resolution of the same steatotic baseline route through different module trajectories or converge on one.

Baselines

The "do nothing" reference point is the field's default: pooled static WGCNA with scale-free topology selection of $\beta$, exactly as deployed in the RNF181 coronary-artery study, which identified 7 modules comprising 8,525 genes from cross-sectional left-ventricular data (Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease). It is the strongest competitor for the network question because it is what every published WGCNA analysis — including both temporal-cohort studies — actually did; no reported result found on Lacuna for static WGCNA applied to the GAN cohorts, so we will run it ourselves on the same data, the same soft-thresholding budget, and the same module-count conventions. It will be beaten if it fails to order samples along the disease trajectory and cannot separate semaglutide- from lanifibranor-responsive modules, while our dynamic pipeline does. The strongest existing longitudinal competitor is Time-Course Gene Set Analysis, which fits linear mixed-effects models with splines to predefined gene sets and, in the DALIA-1 HIV vaccine trial (19 patients, 14 time points), identified 69 significant gene sets where standard univariate differential expression and standard time-series GSEA detected zero after multiple-testing correction (Time-Course Gene Set Analysis for Longitudinal Gene Expression Data). No reported result found on Lacuna for TcGSA applied to NASH or diet-induced cohorts, so we will run it on the same cohorts with its published machinery. Beating it means two things it cannot do by construction: discovering de novo modules absent from MSigDB/KEGG that carry disease-specific hubs predicting held-out tumor incidence, and reporting edge rewiring rather than aggregate set-level trajectory. Pairwise differential expression between successive time points (limma/DESeq2-style testing, runnable in the InMoose environment with its empirical-Bayes batch correction across platforms) serves as the naive floor InMoose. Finally, two ablations isolate the proposal's key ingredients: the dynamic method without cell-mixture deconvolution, and per-timepoint separate WGCNA without GP smoothing (which collapses to tiny-n networks at each sacrifice point) — both on identical data and tuning budgets.

Why not just…?

Why not just run pairwise differential expression across time points? The evidence against it is direct: in the DALIA-1 longitudinal trial, per-gene univariate analysis detected zero significant temporal patterns after multiple-testing correction while the systems-level method found 69 (Time-Course Gene Set Analysis for Longitudinal Gene Expression Data), and in single-cell perturbation benchmarks differential expression failed to capture known mechanisms in 11 of 12 gene knockdowns where network-constrained dynamic models succeeded (A scalable gene network model of regulatory dynamics in single cells). Chronic low-grade dietary stress produces precisely the coordinated, low-amplitude, multi-gene shifts that per-gene testing is worst at; we include it as the naive floor anyway, and if it matched the dynamic pipeline's held-out tumor prediction we would consider the network machinery unjustified.

Why not just use longitudinal GSEA or TcGSA on predefined pathways? This is the most serious simple alternative, and we treat it as the strongest baseline rather than argue around it. Its limits are structural: predefined static sets cannot surface diet-specific modules that no database contains (the metabolic-myeloid rewiring in steatotic livers is a case in point, Multiomics analyses decipher intricate changes in the cellular and metabolic landscape of steatotic livers upon dietary restriction and sleeve gastrectomy), they yield no hub genes, no continuous eigengenes correlatable with tumor counts or respirometry, and no notion of topology change. The head-to-head on the GAN cohorts will settle it empirically.

Why not just apply an ODE network model like FLeCS? FLeCS embeds gene regulatory networks in learnable coupled differential equations and reproduced 72-hour drug-response trajectories in vitro (A scalable gene network model of regulatory dynamics in single cells; code at FLeCS). It was validated on short, dense, in vitro single-cell time series; a 72-week destructive cross-sectional bulk-tissue cohort is a different data regime, and whether its message-passing ODE machinery can even be fit there is unknown. Our sliding-window GP approach is deliberately the cheaper, more robust middle ground for sparse destructive sampling; if FLeCS can be adapted to this regime and outperforms it, that is a finding worth reporting, and we will attempt the fit as an exploratory arm.

Why not just use acute toxin models (CCl₄ or methionine-choline-deficient diets) instead of 34–72 weeks of feeding? Acute chemical shock reproduces none of the phenomena under study: the burned-out biphasic lipid trajectory, the spontaneous senescence plasticity, and the prophylaxis-paradox dynamics all emerge only under sustained dietary attrition (Semaglutide reduces tumor burden in the GAN diet-induced obese and biopsy-confirmed mouse model of NASH-HCC with advanced fibrosis; Dietary intervention reverses molecular markers of hepatocellular senescence in the GAN diet-induced obese and biopsy-confirmed mouse model of NASH). Indeed, the RNF181 study's own validation relied on acute doxorubicin/isoproterenol cardiotoxic shock rather than chronic diet — exemplifying the very gap we target. The retrospective first experiment runs on existing public data precisely so the expensive chronic design is only committed after the computational claims hold.

Why not just hard-threshold the correlations (r > 0.8) and track the resulting graphs over time? Hard thresholding treats a correlation of 0.79 as zero, erasing the graded, coordinated signals characteristic of chronic low-grade stress, and forfeits the scale-free topology on which hub identification depends; a benchmark of 13 structural-inference methods across 11 graph topologies further showed inference accuracy to be highly sensitive to such graph-property choices (Benchmarking Structural Inference Methods for Interacting Dynamical Systems with Synthetic Data). A hard-threshold variant of our own dynamic pipeline costs little and will be reported as an ablation.

Decisive Checks

The proposal stands or falls on three falsifiable outcomes: the steatosis-module eigengene trajectory must reproduce the biphasic lipid curve (peak near week 24, decline by 72 weeks) without histological supervision; module states at biopsy must predict held-out tumor incidence better than the static and gene-set baselines; and $\kappa$ must saturate under chow reversal while senescence markers clear and the fibrosis stage does not improve — the continuous signature of the point where adaptive response ends and chronic, drug-resistant remodeling begins.