Teaching Cell-Line Drug Models to Expect Stroma

Cancer drug response models are trained on cell lines, because those are the only systems with hundreds of thousands of labeled drug responses. They are then applied to patient tumor biopsies, where they usually transfer poorly. A large and avoidable cause of the failure is a mismatch in what the measurement contains. A cell line is a pure culture of cancer cells. A bulk RNA sequencing profile of a biopsy, which reads out the average gene activity of every cell in the sample, mixes cancer cells with fibroblasts, blood-vessel cells and immune cells in proportions that differ from patient to patient. We propose a simple correction. Before training, artificially contaminate each cell-line profile with realistic amounts of stromal and immune signal, sampled from the compositions actually seen in patient tumors, and keep the cell line's drug-response label unchanged. The model then learns, by construction, that tumor composition is not evidence about drug sensitivity.

This reverses the direction everyone currently works in. Existing transfer methods either try to align whole patient profiles to whole cell-line profiles, or they would need to "deconvolve" (computationally unmix) a biopsy into its tumor-only part. Unmixing a tumor-intrinsic profile without a tumor reference is badly underdetermined. Mixing forward is well posed: it needs only cell-type reference profiles and patient composition estimates, both of which are routine. Our central claim is testable on public data. If composition mismatch is a major part of the transfer gap, then the errors of current methods should grow as tumor purity falls, and forward mixing should flatten that trend.

Core Claim and Positioning

Several published transfer methods motivate this work, and all of them treat the biopsy as one homogeneous profile:

  • TRANSACT aligns nonlinear subspaces of cell lines and tumors. It assumes that the alignment implicitly discards non-tumor signal.
  • TransDRP uses adversarial and prototype alignment across 1,186 labeled patients.
  • DeepSADR builds drug-fragment × pathway interaction graphs and fine-tunes a readout on patients.
  • CSG²A pretrains gene-gene attention on drug-perturbation profiles.

None of them represents stromal and immune content explicitly. Alignment-based methods therefore face a bind: they must either push patient immune variation onto cell lines that have no immune cells, or quietly absorb it. One finding in the file sharpens this tension. A model built on tumor-microenvironment features and drug fingerprints (Wang et al.) found that microenvironment features carry clinical signal, with 6-month progression-free survival AUC 0.793. So composition should be modeled, not merely removed.

Our design accommodates this. The augmented cell-line profiles teach the tumor-intrinsic predictor to be invariant to composition. A separate, small head receives the estimated composition as an explicit input and is fit on patient labels where they exist.

Single-cell work constrains the claim. Comparisons of in vitro and in vivo trajectories (chondrocyte scRNA-seq) found that cell states diverge between culture and tissue. Organoid-fibroblast co-culture (colorectal organoid model) was needed to recover in-vivo-like tumor programs. Together these show that stroma also changes cancer-cell state, not just the measured average. Additive mixing captures contamination but not that interaction. We therefore claim to close only the contamination part of the gap, and we design a measurement of what remains.

Research Program

For each cell line $c$ with expression $x_c$ (in linear, not log, abundance units), we generate patient-like profiles as follows:

$$\tilde{y} = (1-s)\,x_c + s\sum_k \pi_k r_k$$

The terms are:

  • $s$ is a non-tumor fraction drawn from the purity distribution of patients of the same cancer type.
  • $\pi_k$ are the proportions of non-tumor cell type $k$, drawn from that cancer type's TCGA deconvolution estimates. These come from established tools such as CIBERSORT-style regression or TIMER.
  • $r_k$ is a reference expression profile for cell type $k$, taken from public tumor single-cell data.

Each $\tilde{y}$ inherits the drug-response label of $c$. We then add sampled measurement noise, so that variance is realistic for bulk RNA sequencing. Mixing is followed by renormalization and log-transform, so that the predictor sees inputs on the same scale as patients.

The augmentation sits in front of any learner. We evaluate it with three learners:

  • an elastic net, as the simplest learner;
  • TRANSACT's consensus features, as the strongest unsupervised aligner;
  • DeepSADR's two-stage pipeline, as the strongest supervised transfer method.

This separates the contribution of augmentation from that of architecture. In the full model, the composition head adds $\pi$ and $s$ as inputs at patient fine-tuning time only.

Baselines

Naive reference. An elastic net trained on raw cell-line expression and applied unchanged to patients. This anchors how large the gap is.

ComBat batch correction followed by a deep regressor. This is the batch-correction view of the gap. TRANSACT reports that it is beaten, though the file gives no scalar (no reported result found on Lacuna for ComBat+DL on TCGA). We rerun it ourselves on identical splits.

TRANSACT. Run with its published kernel choices, retuned under the same hyperparameter budget as every other method. No reported scalar AUC was found on Lacuna for TRANSACT on TCGA, PDXE or HMF, so we reproduce it ourselves.

DeepSADR. The strongest reported competitor. On TCGA it achieved AUC 0.957 on sorafenib against 0.811 for the next baseline, and an average AUC of 0.856 across its five drugs (DeepSADR). We use its 555-patient, five-drug TCGA setup and its frozen-encoder fine-tuning protocol, both with and without our augmentation, with matched compute and tuning trials. TransDRP's reported advantage is qualitative in the file: it significantly beat unsupervised-transfer baselines on 6 of 9 drugs. We include it on its own nine drugs.

Ablations isolating the key ingredient. We run three:

  • mixing with uniformly random proportions instead of patient-matched ones;
  • mixing with shuffled reference profiles;
  • the purity-covariate and deconvolve-then-align alternatives described below.

What counts as success. Beating DeepSADR means either a higher mean AUC across its five drugs on identical splits, or an equal AUC together with a significantly flatter error-versus-purity slope.

Why not just…?

Why not just add purity and immune scores as covariates, or regress them out linearly? This is the most serious competitor, and it may succeed. Linear residualization on log expression cannot undo a mixture that is additive in linear space. It also has to learn the interaction between composition and genes from a few hundred labeled patients. Augmentation instead learns invariance from tens of thousands of cell-line labels. We run residualization and the covariate model as baselines on every drug. If either matches augmentation on both AUC and purity slope, we will report that the cheap fix suffices.

Why not just deconvolve each biopsy into a tumor-only profile and then run TRANSACT? Recovering a tumor-specific expression vector per sample requires a malignant reference that the patient's own cancer does not supply. The error of that reconstruction grows exactly in low-purity samples, which are the cases we care about. We implement this pipeline anyway and compare the two pipelines' error-versus-purity curves.

Why not just train on patient data directly? The file reports a random forest trained on TCGA whose 5-FU accuracy fell from 84.1% in cross-validation to 52.9% on an external histology (TCGA-trained models). With roughly 100 labeled patients per drug, preclinical labels remain necessary.

Proposed Experiments

We use GDSC and DepMap cell lines as source data. Targets are the following:

  • TCGA labeled patients, on the DeepSADR and TransDRP drug sets;
  • NIBR PDX pharmacology;
  • Hartwig metastatic biopsies, on TRANSACT's three drugs.

We measure per-drug AUC and responder/non-responder separation. CSG²A, for comparison, separated responders from non-responders for 5 of 13 drugs at p < 0.05.

The first decisive check stratifies patients by estimated purity and fits error against non-tumor fraction for each method. Composition mismatch predicts a positive slope for the baselines and a near-zero slope for the augmented models.

The second check is a natural control from PDX models. After mouse sequencing reads are filtered out, PDX profiles are nearly tumor-only, so augmentation should help much less there than on TCGA. A large gain on PDX would mean the method works for some other reason than the one we claim.

The third check estimates the interaction residual left by additive mixing. We train with references derived from co-culture versus monoculture where public data permit. Any remaining gap at high purity is attributed to cell-state change rather than contamination.