When researchers tune a cell's environment — temperature, salt, nutrient load, or the stiffness of the gel the cells sit on — they usually turn one knob at a time and read out a single coarse number: how fast cells grow, how much product they make, or how far a tumor spreads. A largely separate community builds co-expression networks, in which genes that rise and fall together are clustered into modules and a few highly connected "hub" genes are nominated as drug targets. These two efforts almost never meet: environmental optimization treats the gene-regulatory network as an invisible black box, while network-based target discovery almost always runs under one standard laboratory condition. That separation is dangerous, because the microenvironment can flip which genes a cell actually depends on. Patient-derived colorectal tumor organoids grown on mechanically soft matrices turn on oxidative phosphorylation and become vulnerable to the inhibitor Gboxin and the ADORA2B antagonist ZM241385, which do nothing on standard matrices (A Soft Matrix Microenvironment Promotes Laterally Spreading Tumors via Oxidative Phosphorylation‐Dependent Cell Adhesion). Target lists derived under one condition may be quietly wrong under another.
The core idea of this proposal is to make the co-expression network itself the thing being optimized and the thing doing the nominating, across both environment and time. We will grow laterally spreading tumor organoids on hydrogels whose stiffness and extracellular-ligand density are tuned independently, sample their transcriptomes densely from 30 minutes to two weeks, and fit each gene module's activity — its eigengene, the first principal component of the module's expression matrix — as a smooth response surface over stiffness, ligand density, and time. Hubs are then nominated not by static connectivity but by how their position in the network consolidates or dissolves along that surface, and validated with drugs and knockdowns timed to the windows the network predicts.
Core Claim
Three gaps in the compiled literature define the claim. First, no study fits a response surface — the standard tool of environmental parameter optimization, in which a quadratic polynomial maps several input factors to an output — where the output is a network quantity such as a module eigengene or hub connectivity, rather than a macroscopic endpoint like cell density or degradation yield; existing optimization work (Preparation, biocontrol activity and growth promotion of biofertilizer containing Streptomyces aureoverticillatus HN6) stops at endpoints. Second, the closest existing work (A Soft Matrix Microenvironment...) connects environmental parameters, co-expression networks, and drug validation, but only at static endpoints after adaptation, with stiffness confounded by Matrigel dilution (which changes ligand density and diffusion too), leaving the early-versus-late dynamics of adaptation unmapped. Third, co-expression hubs have a credibility problem: a co-essentiality network built from CRISPR knockout screens across 769 cancer cell lines identified cancer driver genes far better than co-expression networks, with a cancer-related pathway enrichment odds ratio of 4.27 versus only 0.49 for co-expression (A Co-essentiality Network of Cancer Driver Genes Better Prioritizes Anticancer Drugs). We hypothesize that this gap shrinks when co-expression is measured as a function of a systematically varied environment and over time: perturbing the environment along a designed gradient should separate genes that merely ride along with the phenotype from genes whose network position actively reorganizes as the cell adapts. Temporal profiling already shows this reorganization happens — aflatoxin-exposed rat liver shifts from metabolic disruption to persistent cell-cycle arrest over seven days (New Insights Into the Persistent Effects of Acute Exposure to AFB1 on Rat Liver) — but nobody has yet used it to rank targets.
Research Program
The methodological centerpiece is a longitudinal, mixed-effects version of Weighted Gene Co-expression Network Analysis (WGCNA), the framework that clusters genes into modules using soft, continuously weighted correlations and then links module activity to traits (Session Introduction). Standard WGCNA assumes every sample is independent, which is false for repeated measurements of the same organoid line over time. Application of Weighted Gene Co-expression Network Analysis for Data from Paired Design fixed this for paired tumor/normal biopsies by putting the subject into a linear mixed-effects model as a random effect; we extend that formulation from two paired time points to full time series, modeling the module eigengene as
$$\mathrm{ME}_{lt} = \beta_0 + \beta_1 S_l + \beta_2 t + \beta_3 S_l t + u_l + \varepsilon_{lt}$$
where $\mathrm{ME}_{lt}$ is the eigengene of a module in organoid line $l$ at time $t$, $S_l$ is matrix stiffness, $u_l$ is a random intercept for the line, and the residuals $\varepsilon_{lt}$ follow an autocorrelated (AR(1)) process within each line. The stiffness-by-time interaction $\beta_3$ is the scientific quantity of interest: it measures whether a module's response to stiffness grows, fades, or reverses as adaptation proceeds. We will calibrate false-discovery rates for this model on simulated autocorrelated data, where naive pooled WGCNA is expected to overstate significance.
The experimental platform addresses the Matrigel confound head-on. Organoids from laterally spreading tumors will be cultured on hydrogels in which stiffness (roughly 0.25–4 kPa) and collagen ligand density are set independently, crossing the two factors in a two-factor response-surface design of the kind used in fermentation optimization, but with network readouts in place of yield:
$$\widehat{\mathrm{ME}}(S, C) = b_0 + b_1 S + b_2 C + b_{12} SC + b_{11} S^2 + b_{22} C^2$$
where $C$ is ligand density. Because stiffness and ligand density enter as separate factors, mechanosensitive modules (stiffness-driven) and ligand-driven modules can be separated statistically — something the Matrigel-dilution design of A Soft Matrix Microenvironment... structurally cannot do. Sampling will span early windows (30 minutes, 2, 6, 24 hours) that coarse Day-0/Day-7 designs miss, plus 3, 7, and 14 days, with bulk RNA-seq across the factorial grid and single-cell RNA-seq at the predicted transition points.
Target nomination then works on hub trajectories. Within stiffness-responsive modules, each gene's module membership (correlation with the module eigengene) is tracked over time, and hubs are classified as early-transient (connectivity spikes within hours, then decays) or late-committed (connectivity consolidates over days). The contrasting predictions are testable: an early-transient hub should be required to initiate adaptation, while a late-committed hub should be required to maintain the soft-matrix phenotype. This directly attacks the bystander problem — the observation that many deregulated proteins are passive consequences of cellular stress rather than active drivers (Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes) — by replacing "is it correlated and connected" with "does its network position reorganize when the environment does."
Baselines
The naive reference point is endpoint differential expression: a fold-change ranking of genes at day 14 on the softest versus stiffest matrices, ignoring time and network structure. It costs nothing, is what most labs would do first, and it is the floor any network method must clear; beating it means nominating, among the top-ranked genes, targets whose perturbation actually changes the phenotype, since differential expression alone is known to nominate reactive bystanders (Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes) and to discard coordinated sub-threshold shifts (Identifying Functional Gene Modules via Co-Expression Networks).
The strongest direct competitor is the existing soft-matrix pipeline itself: single-condition organoid culture, static WGCNA at one endpoint, and pharmacological rescue with Gboxin and ZM241385 (A Soft Matrix Microenvironment...). We will run this exact pipeline on our own data — same organoid lines, same sequencing, endpoint samples only, Matrigel dilution for the stiffness contrast, identical soft-thresholding and module-count tuning budget — so the only differences are the temporal dimension, the decoupled matrix, and the mixed-effects model. No reported result found on Lacuna for the quantitative performance of that pipeline's network analyses on its datasets (GSE161277 and GSE261388); its drug-rescue findings are reported qualitatively. We would count as beating it if (a) our temporal hub ranking diverges materially from its static ranking on the same samples, and (b) at least one target nominated only by the temporal method, and none nominated only by the static method, passes functional validation.
The strongest target-nomination method in the compiled literature is the co-essentiality network of A Co-essentiality Network of Cancer Driver Genes Better Prioritizes Anticancer Drugs, which reports a cancer-related pathway enrichment odds ratio of 4.27 (versus 0.49 for co-expression and 0.79 for protein-interaction networks) and higher PRISM IC50 correlation in 13 of 15 cancer types. We will apply its guilt-by-association propagation to our module gene sets using DepMap dependency scores, with the same input genes and no extra tuning. The test is pointed: does DepMap-derived essentiality — measured on cells grown on rigid plastic under ambient conditions — predict knockdown effect size on soft matrices? The soft-matrix paper already showed vulnerability flips with stiffness, so we expect partial failure, but this is the benchmark our claims must be measured against.
Finally, an ablation isolates the two key ingredients: the full pipeline run (i) without the mixed-effects model — samples pooled across time as in Identification of Potential Candidate Genes From Co-Expression Module Analysis During Preadipocyte Differentiation in Landrace Pig and Temporal proteomic profiling of postnatal human cortical development, which built WGCNA across time courses but treated time points as ordinary samples — and (ii) without time, at the endpoint only. If the full model's nominations and significance calibration do not differ from these ablations, the added machinery is not earning its cost.
Why not just…?
Why not just run differential expression at each time point? This is cheap and familiar, but it fails in two documented ways. Genes passing a hard fold-change cutoff are frequently passive stress responses rather than drivers (Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes), and thresholding erases the subtle, coordinated shifts that modules are designed to capture (Identifying Functional Gene Modules via Co-Expression Networks). More decisively, the proposal includes a direct comparison: the top differentially expressed genes will be carried into the same validation queue as the network-nominated hubs, so if fold-change ranking validates as well, the networks lose and we will say so.
Why not just simulate the perturbations with a virtual cell model instead of doing the experiments? The PerturbCausal benchmark evaluated GEARS, CPA, Geneformer, and STATE and found that no deep virtual cell model significantly outperformed a sparse linear comparator at recovering causal regulatory networks, while inflating predicted effect sizes by approximately 23-fold relative to real data (Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery). Simulated knockdowns therefore cannot substitute for the validation experiments here. They can, however, serve as an honest triage layer: we will run the sparse linear comparator as an in-silico baseline for predicting our knockdown readouts, since it is the model class that benchmark found hardest to beat.
Why not just do the static soft-matrix WGCNA that has already been published? This is the real threat, and the ablation above is designed to settle it rather than argue around it. There is genuine reason to expect divergence: adaptation programs are known to reorganize over time, from the day-scale drift of the aflatoxin footprint (New Insights Into the Persistent Effects of Acute Exposure to AFB1 on Rat Liver) to the dynamic hub behavior seen in differentiating preadipocytes, where the central hub ANGPTL4 is defined precisely by its temporal downregulation that a single endpoint would misread (Identification of Potential Candidate Genes From Co-Expression Module Analysis During Preadipocyte Differentiation in Landrace Pig). But it is possible that static and temporal rankings largely agree; if the overlap in validated-relevant nominations exceeds roughly ninety percent, we will report that the temporal machinery is unnecessary in this system, which would itself be a useful negative result for the field.
Why not just use co-essentiality networks for target nomination and skip co-expression entirely? Given the reported odds-ratio gap (4.27 versus 0.49), a reviewer would be right to push this. The counterargument is contextual: co-essentiality is measured under standard culture conditions, and the soft-matrix study demonstrated that drug sensitivity changes when the matrix changes (A Soft Matrix Microenvironment...); PINNACLE likewise succeeded by adding biological context that global models lacked, while explicitly leaving environmental conditions unmodeled (Contextual AI for Single-cell Therapeutic Target Discovery). The experiment answers the question empirically: co-essentiality rankings and temporal-network rankings will be compared head-to-head against measured knockdown effects on soft matrices.
Proposed Experiments and Decisive Checks
The program runs in three phases. Phase one is computational and methodological: implement the mixed-effects WGCNA, verify it reproduces the published paired-design result on the GSE45238 dataset (the turquoise module-trait association with t-value −18.68, p < 1.97 × 10⁻²⁰, and the miR-let-7c hub connected to 140 other miRNAs, from Application of Weighted Gene Co-expression Network Analysis for Data from Paired Design), and demonstrate on simulated autocorrelated time series that the naive pooled analysis inflates module-trait significance while the mixed-effects version controls false discovery. Phase two is the main experiment: the stiffness-by-ligand-density factorial organoid culture with dense time-course RNA-seq, module discovery, eigengene response surfaces, and hub-trajectory classification; modules discovered in vitro will be checked for preservation in the public patient lesion datasets GSE161277 and GSE261388 used by the soft-matrix study. Phase three is validation under the environmental optimum: CRISPR interference knockdown of one early-transient and one late-committed hub per major module, Gboxin and ZM241385 dosing, and — replacing the published 2D surface-area readout — 3D volumetric invasion tracking, since lateral spread is fundamentally a 3D process.
The decisive checks are concrete. First, the timing prediction: if late-committed hubs maintain the phenotype, then windowed drug or knockdown intervention applied during the network-predicted commitment window (days, not hours) should suppress invasion more than the same total intervention applied earlier or continuously; constant-dosing equivalence would falsify the temporal claim. Second, the initiation prediction: blocking an early-transient hub during the first hours should prevent the adaptation module from ever consolidating, measurable as absent late eigengene rise. Third, the context prediction: knockdown effect sizes on soft matrices should correlate worse with DepMap-derived essentiality than with our temporal hub scores — if DepMap predicts them equally well, environmental context adds nothing to target nomination and the simpler route wins. Fourth, the rank-divergence check against the static ablation, with the ninety-percent overlap threshold stated in advance. Each check has a named outcome that would make us abandon or sharply narrow the central claim, which is the standard this proposal should be held to.
References
- A Soft Matrix Microenvironment Promotes Laterally Spreading Tumors via Oxidative Phosphorylation‐Dependent Cell Adhesion
- Preparation, biocontrol activity and growth promotion of biofertilizer containing Streptomyces aureoverticillatus HN6
- A Co-essentiality Network of Cancer Driver Genes Better Prioritizes Anticancer Drugs
- New Insights Into the Persistent Effects of Acute Exposure to AFB1 on Rat Liver
- Session Introduction
- Application of Weighted Gene Co-expression Network Analysis for Data from Paired Design
- Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes
- Identification of Potential Candidate Genes From Co-Expression Module Analysis During Preadipocyte Differentiation in Landrace Pig
- Temporal proteomic profiling of postnatal human cortical development
- Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery
- Contextual AI for Single-cell Therapeutic Target Discovery