Substrate Surface Energy as a Hidden Confounder in Translational Cell Culture

Most drug candidates that look potent in a laboratory dish fail in patients, and the standard explanations — missing immune cells, missing blood flow, missing three-dimensional architecture — are all about what the dish lacks. This proposal is about a variable the dish has but nobody measures: its own surface. Every culture substrate has a wettability, characterized by the contact angle — the angle a resting droplet makes with the surface — and by dosing the surface with several probe liquids one can decompose its surface free energy into a polar (hydrogen-bonding) part and a dispersive (non-polar) part. Cells sitting on that surface sense it: which cells in a mixed population attach, spread, and survive depends on it. A bulk assay then averages over whatever mixture the surface selected, so the readout can be a property of the dish rather than of the cells — and the surfaces cells occupy in a living tissue are different again.

The connection to masked heterogeneity is direct. Bulk assays already blur rare subpopulations, as Genome 2011 argued at the birth of single-cell genomics, and the consequences are documented: bulk profiling mistook the selective outgrowth of rare stem-like cells for a uniform epigenetic shift in tamoxifen-resistant breast cancer Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties, and conserved pluripotency programs in rare cancer subpopulations are invisible at population level A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers. My claim is that the culture substrate is one of the selectors of these mixtures, that its selection pressure is quantifiable with a century-old, fifty-cent technique, and that measuring it is a missing covariate in every computational pipeline that transfers in vitro drug screens to patients. The Lacuna index itself flags this hole: in vitro-to-in vivo transfer models "focus almost entirely on chemical structures and genomic expressions, ignoring how the physical culture substrate (wettability, surface energy, stiffness) alters cellular signaling, drug permeability, and gene network topology." No indexed work unites measured surface energy, single-cell resolution, and in vivo validation.

Core Claim

Substrate surface free energy — measured by multi-liquid contact angle goniometry and decomposed into polar and dispersive components via the Owens–Wendt–Rabel–Kaelble relation — is a first-order, cheaply measurable confounder of bulk in vitro readouts because it reweights heterogeneous cultures toward particular subpopulations. Explicitly conditioning on it should (i) unmask which subpopulations actually drive an observed drug response, and (ii) improve prediction of in vivo outcomes. As a corollary, the drift of the contact angle on a cell-conditioned interface over days — cells secrete matrix and remodel the liquid–solid boundary, so the angle moves — is a non-destructive, label-free early reporter of composition drift in a living culture. The OWRK decomposition we will use throughout solves, for each probe liquid,

$$(1+\cos\theta)\,\gamma_{lg} = 2\left(\sqrt{\gamma_s^d\gamma_l^d} + \sqrt{\gamma_s^p\gamma_l^p}\right),$$

where $\theta$ is the measured contact angle, $\gamma_{lg}$ the known liquid surface tension, and $\gamma^d$, $\gamma^p$ the dispersive and polar components of the solid and liquid surface energies. The point of the decomposition is that it separates surface chemistry from surface roughness, the conflation the file identifies as a core limitation of single-angle wettability metrics — the same distinction exploited in Water Damage Resistance of Tuff Asphalt Mixtures with Admixtures Based on Surface Energy Theory for asphalt adhesion.

Research Program

The closest existing work defines what must be exceeded, and both nearest neighbors stop one direction short. The taurine-grafted collagen corneal film Preparation and Characterization of a Photo-Crosslinked Methacryloyl-Collagen Composite Film to Promote Corneal Nerve Regeneration via Surface Grafting of Taurine Molecules is the only indexed study coupling rigorous contact angle characterization to cell response, but its readouts are population-averaged MTT viability assays on two cell lines, and it never leaves the dish — it resolves no heterogeneity and tests no in vivo translation. Conversely, the patient-derived organoid–fibroblast co-culture Mimicking tumor cell heterogeneity of colorectal cancer in a patient-derived organoid-fibroblast model resolves heterogeneity and improves on naked organoids, but grows everything in undefined basement membrane extract whose surface energy was never measured. This proposal occupies exactly the vacated corner.

The program has four parts. First, a substrate library: roughly ten surfaces spanning a grid of polar and dispersive surface energy, from plasma-treated polystyrene (standard dish) through collagen-grafted hydrophilic films (the ColMA-Tr chemistry) to hydrophobic silanized variants (the KH-570 chemistry from Surface Hydrophobic Modification of Biochar by Silane Coupling Agent KH-570), each fully goniometrically characterized with water, diiodomethane, and ethylene glycol, and each fabricated in roughness-matched and stiffness-matched pairs so that surface energy, not mechanics, carries the signal. Second, native-interface targets: the same multi-liquid goniometry applied to excised native tissue interfaces (corneal stroma; colorectal tumor extracellular matrix), so that "surface-energy-matched" culture has a defined physical target. Third, heterogeneity resolution on two systems with known hidden-subpopulation problems: the tamoxifen-resistant breast cancer models, whose resistant stem-like minority is the canonical masking failure, and the colorectal organoid–fibroblast co-culture with oxaliplatin response. Each system is cultured across the substrate grid; terminal single-cell RNA sequencing gives ground-truth composition, while non-destructive longitudinal single-cell estimates come from the conditional-diffusion approach of Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion, and bulk RNA-seq from each substrate is deconvolved with Count Bridges enable Modeling and Deconvolving Transcriptomic Data — whose EM formulation treats single-cell counts as latent integers — to quantify masking directly. Fourth, the translation test: extend the transfer architecture of DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction with a surface-energy conditioning input, and validate against an in vivo ground truth using the pooled screening platform of Library transgenesis in zebrafish through delayed site-specific mosaic integration for in vivo pooled screening of transgenes, which permits the same clones to be assayed in a living organism at throughput comparable to culture.

Baselines

The naive reference point is the field default: a bulk endpoint assay (IC50, viability) run on uncharacterized standard tissue-culture plastic — the design used, for instance, in the microRNA apoptosis studies on SW620 cells Effects of miR-135a-5p and miR-141 on proliferation, invasion and apoptosis of colorectal cancer cells. Beating it means the substrate-conditioned model predicts the in vivo readout with measurably lower error than the same assay ignoring the dish.

The strongest experimental competitor is the organoid–fibroblast co-culture of Mimicking tumor cell heterogeneity of colorectal cancer in a patient-derived organoid-fibroblast model, which restored patient tumor architecture and cellular diversity in minimal medium and identified 1,742 deregulated genes. It is the strongest competitor because it already attacks both heterogeneity masking and the culture-fidelity problem; we will run the same co-cultures, with the same patient material and media, but on goniometrically characterized substrates in place of undefined Matrigel — same samples, same sequencing depth, same compute — so the only manipulated variable is measured surface energy. Beating it means higher concordance with the zebrafish in vivo readout. No reported result found on Lacuna for its oxaliplatin prediction accuracy, so the comparison number will be reproduced under matched conditions rather than quoted.

The strongest computational competitor is DeepSADR DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction, which pre-trains on 966 cancer cell lines from DepMap (225,781 labeled samples) and transfers to 555 TCGA patients by freezing the chemical encoder and fine-tuning the adaptive readout. It is the strongest because it is the only indexed model that structurally addresses the in vitro–in vivo mechanism shift. We will re-run it on identical data, splits, and tuning budget, with and without surface-energy features. No reported result found on Lacuna for DeepSADR's transfer accuracy, so both arms are reproduced under equal compute and the paired difference is the outcome. Two ablations isolate the key ingredient: surface energy entered as a single static water contact angle versus the full polar–dispersive decomposition versus the dynamic conditioning trajectory; and the roughness/stiffness-matched substrate pairs, which test whether the effect is surface chemistry or mechanics.

Why not just…?

Why not just use organoids or animals and skip the surface physics? Organoids still sit on a matrix with unmeasured, batch-varying surface energy — the co-culture paper's own omission — so they inherit the confounder while adding cost; and whole-animal screens cannot cover the throughput of a 96-well plate. But this is a real threat: if matched in vivo assays were cheap, this proposal would be unnecessary. The decisive check uses the zebrafish pooled screen precisely because it makes the in vivo answer available at scale; if surface-energy-matched culture does not predict it better than standard culture, the idea fails on the evidence.

Why not just measure stiffness, since mechanotransduction is the established story? Stiffness and wettability are physically entangled in hydrogels, which is exactly why the library includes stiffness- and roughness-matched pairs that vary surface energy alone. The OWRK decomposition gives a chemical signature that stiffness cannot. If the matched-pair experiments show the stiffness covariate absorbs the effect, that result will be reported as a negative finding and the confounder claim narrowed accordingly — the factorial design exists to settle this, not to be argued around.

Why not just do single-cell sequencing at the endpoint? Because sequencing destroys the sample, so it cannot watch one culture drift, and because deconvolution depends on reference atlases that miss uncharacterized rare states — the documented failure mode of the tamoxifen study and the stated limitation of atlas-dependent methods. The contact-angle trajectory is what gives a live, non-destructive, per-culture signal; single-cell sequencing remains the ground truth it is validated against, not something the proposal replaces.

Why not just add substrate metadata as features to DeepSADR? That is literally part of this program — the difference is that no surface-energy-annotated drug response data exists to feed it. The wet-lab substrate grid and native-interface measurements are the rate-limiting contribution; without them the conditioning input has nothing to condition on.

Proposed Experiments

Three decisive checks, each with a pre-registered success criterion. First, masking mechanism: across the substrate grid, decompose the between-substrate variance in bulk IC50 (tamoxifen, oxaliplatin) into the part explained by deconvolved subpopulation composition shifts versus uniform response shifts. Success is composition explaining more than half the between-substrate variance in at least one system, with deconvolution proportion error against terminal scRNA-seq reported as the accuracy bound. Second, early warning: the contact-angle trajectory on the conditioned interface must predict the eventual composition drift and IC50 shift before bulk viability shows any change; success is a statistically significant lead time measured in days, evaluated against the diffusion-based non-destructive single-cell estimates as the running reference. Third, translation: Spearman correlation between in vitro response measured under each substrate condition and the zebrafish in vivo pooled-screen outcome for the same clones, with surface-energy-matched (to native tissue interfaces) versus standard-TCPS versus mismatched conditions; success is a paired improvement over both reproduced baselines — the naive TCPS assay and the DeepSADR transfer model without surface-energy features — exceeding the pre-registered margin. If the roughness-matched ablation transfers the effect to mechanics, or if TCPS-standard culture predicts the in vivo readout as well as matched culture, the hidden-confounder claim is refuted and the honest conclusion is that surface energy is a second-order variable in these systems — itself a result worth publishing, given that the entire translational pipeline currently assumes it away.