Most tumor RNA sequencing measures a blended sample: one test captures cancer cells, the healthy cells around them, and immune cells together. Splitting that blend back into its parts is called deconvolution. Deconvolution tools fail in one case that matters a great deal for biomarker discovery. When a sample is mostly tumor, which is the usual case in surgical breast cancer specimens, the tools start labelling normal epithelial cells as cancer cells. Cancer cells keep much of the gene-expression "memory" of the normal tissue they came from, so in expression space the two look nearly the same. A biomarker search run on top of these estimates cannot tell a gene the tumor truly switches on apart from a gene that ordinary breast epithelium expresses. The second kind of gene makes a poor diagnostic marker, because healthy tissue expresses it too.
We propose to break this tie with a signal that only cancer cells carry: their DNA copy number, meaning how many copies of each stretch of chromosome a cell holds. In an amplified region, a malignant cell's expression tends to rise roughly in proportion to the extra copies. A diploid normal epithelial cell shows no such pattern. If we tie the malignant component's expected expression to the patient's own copy-number profile, deconvolution gains a direction along the genome that normal epithelium cannot mimic. The core claim is that this anchor recovers correct malignant and normal-epithelial proportions at high purity, where current tools fail. It also yields per-patient estimates of expression that belongs to the tumor cells themselves. Prognostic biomarkers built from those estimates should hold up against independent single-cell data in a way that bulk-derived candidates do not.
Where this sits
The problem comes straight from Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. That study built 38,000 artificial bulk samples ("pseudo-bulks") by mixing cells from the 100,064-cell Wu et al. breast cancer atlas at tumor purities from 5% to 95%. It then benchmarked nine tools: BayesPrism, Scaden, CIBERSORTx, MuSiC, DWLS, hspe, CPM, Bisque and EPIC. All of them lost accuracy as purity rose, and the authors traced the errors to normal epithelium being called malignant because of lineage memory. That paper diagnoses the problem and offers no fix. Our proposal is the fix, and we evaluate it on their benchmark design.
Other work shapes the design without solving the problem. TIMER, reviewed in Digitally deconvolving the tumor microenvironment, filters genes by their correlation with purity and resolves only six immune lineages. The same review documents "molecular mimicry," where cancer cells express immune genes such as CTLA-4. That is the same failure in a different setting: a reference profile cannot express what a malignant cell actually does. Statistical inference for cell type deconvolution (MEAD) gives us two things. The first is a likelihood that models gene-specific scaling between platforms and weights genes by noise. The second is a negative result: picking marker genes by hand does not reliably help. That rules out the cheap fix of dropping lineage-memory genes. SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References shows that weighting several single-cell references together reduces reference bias. Count Bridges enable Modeling and Deconvolving Transcriptomic Data argues that RNA counts should be modelled as integers rather than log-transformed. We adopt a count likelihood for the same reason, although we do not need its generative bridge machinery.
None of these methods uses a malignancy-specific covariate inside the mixing model. The closest neighbour in spirit is classical two-component tumor/normal unmixing, which takes a DNA-based purity estimate as a fixed number. That line of work is not covered in our research file, so we treat it as a baseline to beat rather than claiming to be the first to use DNA. The new ingredient here is per-gene, per-patient copy number as a shaping term for the malignant profile, not as one scalar purity value. It is used within a deconvolution over many cell types.
Core Claim and Model
For gene $g$ in bulk sample $j$, we model the observed count $y_{gj}$ as negative binomial. A negative binomial is a count distribution that allows more spread than a Poisson. Its mean is
$$ \mathbb{E}[y_{gj}] = s_j \Big( \sum_{k \neq m} p_{jk}\,\mu_{gk} \;+\; p_{jm}\,\mu_{gm}\,\big(\tfrac{c_{gj}}{2}\big)^{\beta_g} \Big), $$
where the symbols are as follows. $s_j$ is the sequencing depth of the sample. $p_{jk}$ is the fraction of cell type $k$, and $m$ indexes the malignant type. $\mu_{gk}$ is the reference expression of gene $g$ in type $k$, taken from single-cell data. $c_{gj}$ is the copy number of the region containing gene $g$ in patient $j$. $\beta_g \ge 0$ is a gene's dosage sensitivity, which measures how strongly its expression follows copy number.
We learn $\beta_g$ once, from malignant cells in held-out single-cell patients whose copy number has been called from expression. Genes with dosage compensation get $\beta_g$ near zero and contribute no anchor. We borrow MEAD's gene-specific platform scaling and noise weighting. We fit the model by maximum likelihood over proportions while holding $\beta$ fixed.
The identifiability argument is simple. Suppose $\mu_{\cdot m}$ and the normal epithelial profile are close to collinear. The term $(c_{gj}/2)^{\beta_g}$ then adds a gene-by-gene modulation that is specific to the patient and appears only in the malignant column. The more of a tumor's genome is altered, the better separated the two columns become. The model therefore makes a testable prediction: the gain over baselines should grow with the fraction of the genome that carries copy-number changes.
A residual step produces the biomarker output. After fitting, we attribute the remaining expression to the malignant component, scaled by the expected dosage. This gives a per-patient profile of tumor-intrinsic expression. Survival and subtype analyses run on these profiles instead of on raw bulk counts.
Research Program
The first stage rebuilds the purity-sweep benchmark on the Wu atlas with one change: the reference cells and the cells used to build the mixtures come from disjoint patients. This removes the optimistic leakage that comes from reusing the same patients. Each patient's malignant cells carry copy-number calls from expression, which gives $c_{gj}$ for the pseudo-bulks.
The second stage tests transfer to independent single-cell cohorts, Bassez et al. and Pal et al. These are listed alongside the Wu atlas in the benchmark paper.
The third stage moves to the 1,038 TCGA breast cancer bulk profiles described in the same paper. Measured copy number from TCGA replaces the inferred calls at this stage. This is where biomarkers are discovered.
Baselines
The naive reference point is CIBERSORTx run with a standard single-cell signature, together with plain differential expression on bulk TCGA counts for the biomarker task. That combination is how most bulk-derived tissue biomarker studies proceed, including SLC22A8: An indicator for tumor immune microenvironment and prognosis of ccRCC from a comprehensive analysis of bioinformatics.
The strongest existing competitor is BayesPrism. Among the nine tools benchmarked in Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures, it alone updates the malignant profile for each sample. It therefore targets exactly the mismatch we address, but without a DNA anchor. The research file reports no per-method accuracy numbers from that benchmark, so the record reads: no reported result found on Lacuna for BayesPrism on the Wu pseudo-bulks. The only quantitative deconvolution figure in the file is SCDC's Pearson correlation of 0.991 with ground truth on a 6:3:1 mixture of MDA-MB-468, MCF-7 and fibroblast cells. That figure comes from SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References. It describes cell lines that are easy to separate, not tumor against its own normal tissue, so we cite it only to show how easy settings flatter these methods. We will also run MEAD, because our likelihood extends it, and SCDC with Wu, Bassez and Pal as its multiple references.
Every method gets the same reference cells, the same held-out mixtures, the same gene universe, and an equal hyperparameter search budget. Each method is tuned on a validation split of patients that is separate from the test patients.
Two ablations isolate the key ingredient. The first sets $\beta_g = 0$ for all genes, which leaves MEAD-style count deconvolution with nothing else changed. The second feeds each sample another patient's copy-number profile. If the shuffled version performs as well as the real one, the gain comes from extra flexibility and not from the anchor.
Beating BayesPrism means three things. At purity of 70% or higher, the fraction of true normal epithelium assigned to malignant must drop significantly across held-out patients, using a paired test over mixtures. Error on the other cell types must not get worse. The gain must shrink in tumors with few copy-number changes, as the model predicts.
Why not just…?
Why not just give existing tools a DNA-based purity estimate? This is the most serious threat. Fixing the total malignant fraction could repair much of the proportion error. A single number, however, cannot tell the model which epithelial genes belong to the tumor, so the tumor-intrinsic expression profiles should stay contaminated. We will add a "purity-constrained" variant of our own model and of BayesPrism, which fixes $p_{jm}$ and sets $\beta=0$. That variant gets the true purity in pseudo-bulks and the TCGA purity estimate in real data. If it matches our method on both the proportion metrics and the recovery of malignant expression, the per-gene anchor is unnecessary, and we will report that result.
Why not just drop the lineage-memory genes? MEAD found that selecting marker genes does not reliably improve accuracy and can make it worse (Statistical inference for cell type deconvolution). Removing shared epithelial genes also throws away the very genes whose tumor-versus-normal status a biomarker study needs to know. We will still run a version of CIBERSORTx and BayesPrism with those genes filtered out, for completeness.
Why not just use richer references or a learned cross-modal encoder? SCDC-style weighting of several references and the contrastive bulk-to-single-cell alignment in Cross-modal transfer learning for mapping bulk transcriptomes at cellular level both still rely on reference cell states. That paper itself notes that states missing from the atlas cannot be resolved. A patient's own malignant profile is never fully present in a reference, so more references should narrow the gap but not close it. SCDC with three references is included as a baseline to measure how much of the gap remains.
Proposed Experiments
On held-out Wu pseudo-bulks across the 5–95% purity sweep, we report three measurements. The first is absolute proportion error for each cell type. The second is the misassignment rate of normal epithelium against purity. The third is the correlation between recovered tumor-intrinsic profiles and the true malignant pseudo-bulk of each patient. All results are stratified by the fraction of the genome with copy-number changes, which tests the predicted dose-response.
The same models are then applied unchanged to Bassez and Pal mixtures to test transfer across cohorts.
The decisive biomarker check runs on TCGA. Each pipeline nominates genes whose tumor-intrinsic expression carries independent prognostic value in a Cox survival model, adjusted for subtype and purity. We then measure precision: the fraction of nominated genes that the independent Bassez and Pal single-cell data confirm are expressed mainly in malignant cells and not in normal epithelium. A pipeline whose candidates are mostly normal-epithelium genes has reproduced the purity trap at the biomarker level. We predict that bulk differential expression and BayesPrism-based nomination will show lower precision than copy-number-anchored nomination. This is the measurement that matters most for later liquid-biopsy work. A circulating marker that healthy epithelium also sheds would face the background-signal problem described in A multimodal cell-free RNA language model for liquid biopsy applications. Moving these candidates into plasma assays is left for later.
Two outcomes would falsify the claim. The first is the shuffled-copy-number ablation matching the real anchor. The second is the purity-constrained baseline matching us on malignant-profile recovery and biomarker precision.
References
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Digitally deconvolving the tumor microenvironment
- Statistical inference for cell type deconvolution
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- SLC22A8: An indicator for tumor immune microenvironment and prognosis of ccRCC from a comprehensive analysis of bioinformatics
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- A multimodal cell-free RNA language model for liquid biopsy applications