When researchers sequence a tumour biopsy in bulk, every cell's RNA is pooled into one measurement. Computational deconvolution then tries to recover what fraction of the sample came from each cell type. It does this by matching the pooled profile against a dictionary of marker genes, which are genes expressed mainly by one cell type. In breast cancer this dictionary fails in a specific, well-documented way. As the fraction of cancer cells in a biopsy rises toward clinical levels, all nine benchmarked methods start reporting healthy epithelial cells that are not there. The cause is that tumours keep the transcriptional "memory" of the healthy cell they came from (Tran et al.).
We argue this is a dictionary problem that sentiment-lexicon research has already met in another form. A word like hot is positive for pizza and negative for a laptop processor (Zhou et al.). In the same way, a gene can be a clean normal-epithelium marker on average across patients yet be switched on by the malignant cells of some patients or subtypes. The proposal has two parts. First, choose markers by their worst-case specificity across held-out patients rather than their average specificity. Second, condition that choice on tumour subtype. The resulting marker panels are handed to existing deconvolution tools, which are otherwise left unchanged. The claim is that this should remove most of the purity-trap false positives without the missed detections that simpler conservative filters cause.
Core Claim
The purity trap comes from what we will call polysemous markers: genes whose specificity for a cell type holds in the pooled atlas but reverses in a subset of patients, because those patients' malignant cells express them. The standard approach picks markers by differential expression on a pooled single-cell atlas. That rewards average contrast, so it keeps these genes. Replacing the average with a worst case over patients should expose them.
Let $\bar x_{g,c}$ be the mean expression of gene $g$ in cell type $c$ over training patients. Let $\bar x^{(p)}_{g,c'}$ be its mean in cell type $c'$ within patient $p$, where the malignant cells of patient $p$ count as their own patient-specific type. We score
$$ s_{g,c} = \min_{p \in P}\;\Big(\bar x_{g,c} - \max_{c' \neq c}\ \bar x^{(p)}_{g,c'}\Big), $$
where $P$ is the set of training patients within a subtype. A gene enters the panel for $c$ only if $s_{g,c}$ exceeds a margin $\delta$. The panel is computed separately for each breast cancer subtype (ER+, HER2+, TNBC). This is the "target" in Zhou et al.'s sense: the context that decides what a marker means.
Where the Lexicon Ideas Come From, and Which Ones We Drop
Three pieces of lexicon work shape the method, each in a different role.
Hamilton et al. induce domain-specific sentiment lexicons by propagating polarity from seed words. They attach bootstrap confidence intervals to every score and show that more than 5% of non-neutral words shift polarity across time or community. We reuse the resampling idea, but we resample patients rather than seed words: a marker's reliability is how its specificity survives the removal of patients.
Zhou et al. show that polarity must be scored per target, and they report 74.8% precision on context-dependent words. That paper motivates conditioning on subtype.
Dragut et al. check sentiment dictionaries for internal contradictions. Entries flagged as inconsistent agreed with human raters only 16% of the time, against 55% for consistent ones. We borrow the consistency principle, that an entry must agree with the dominant evidence behind it. We do not borrow the satisfiability solver. Our constraints are independent per gene, so a direct minimum over patients is exact and a solver adds nothing.
We also deliberately set aside signed spectral clustering. Sedoc et al. use it to keep antonyms apart, raising valence correlation from 0.659 to 0.768. Its biological analogue would be repulsive edges between opposing gene programs. That targets spillover among immune subsets, not the malignant-versus-normal confusion that defines the purity trap, and bundling it in would blur what any improvement is due to.
Relation to Deconvolution Work
The deconvolution literature enters as motivation and as the plug-in targets.
xCell curates 489 fixed signatures and removes genes overexpressed in cancer cell lines. That is a static, context-free version of our filter. SCDC guards against inconsistent references by weighting whole atlases, reaching Pearson 0.991 on a three-cell-line mixture. Our method audits individual genes across patients instead. POPPY and Count Bridges improve how bulk data is represented and modelled, but both treat the reference labels and gene dimensions as given. Our work is complementary: it changes which genes the reference offers.
Research Program
We will build marker panels from the Wu et al. breast cancer atlas, which covers 100,064 cells from 26 patients across ER+, HER2+ and TNBC. Panels will be built under three schemes: pooled differential expression, mean-margin selection, and worst-case selection, each with and without subtype conditioning.
Mixtures are simulated following the Tran et al. protocol across purities from 5% to 95%. The critical change is that mixtures are built only from patients held out of marker selection. Otherwise, worst-case selection would be rewarded for memorising its own patients. A second, harder split derives panels on Wu and simulates mixtures from the external Bassez et al. and Pal et al. atlases, testing whether the panels transfer across labs and protocols.
Each panel is fed to DWLS, BayesPrism, CIBERSORTx and MuSiC with no other changes. We will also analyse which genes the worst-case rule rejects. The prediction is that luminal progenitor genes will be rejected in TNBC and mature luminal genes in ER+, which is exactly the lineage-memory pattern Tran et al. describe. If the rejected genes look instead like housekeeping noise, the lexicon interpretation is wrong.
Baselines
The naive reference point is pooled Wilcoxon markers from the full atlas, run through each deconvolver with default settings. This is the practice the purity trap was measured under.
The strongest existing competitors are BayesPrism and DWLS. Tran et al. report that these two had the lowest combined false-positive and false-negative counts and the best accuracy on granular immune lineages. BayesPrism also explicitly models cell-type-specific expression shifts and patient-level tumour heterogeneity, which makes it the most direct threat to the premise of this proposal. The file gives no numeric error or false-positive rate for either tool on this benchmark (no reported result found on Lacuna for BayesPrism or DWLS normal-epithelial false-positive rate). We will therefore rerun both ourselves on identical held-out mixtures, with identical gene universes and the same hyperparameter search budget per method.
We also include hspe as the conservative-filter reference. Tran et al. found it had the lowest false-positive rate but a high false-negative rate, so it shows what "beating" must avoid.
Two further baselines represent the simple static filters a reviewer would reach for. One is a CCLE-style removal of cancer-expressed genes, as in xCell. The other is a filter that drops genes with high between-patient variance.
The key ablation contrasts mean-margin with worst-case selection, holding subtype conditioning fixed. A second ablation removes subtype conditioning while keeping the worst case. Together they separate the contribution of the worst-case rule from that of the context.
Beating a baseline means lowering the predicted normal-epithelial fraction in the 80–95% purity bins where the true normal fraction is zero. It also means not raising the false-negative rate when normal epithelium is genuinely present, and not raising RMSE on immune lineages.
Why not just…?
Why not just use BayesPrism, which already models tumour-specific expression? This is the most serious threat. BayesPrism still exhibited the trap in the Tran et al. benchmark, which suggests that modelling the expression shift alone is not sufficient. It remains possible that BayesPrism with good markers is already nearly optimal. If BayesPrism with worst-case panels shows no gain over BayesPrism with pooled panels, we will report that the marker lexicon is not the bottleneck for model-based methods and that our gains are confined to regression-based tools.
Why not just build the reference from subtype-matched patients? Subtype conditioning alone is a real and cheap alternative, and it is one of our ablations. Our expectation is that it helps but falls short. Lineage memory varies between patients within a subtype, so an average over matched patients can still keep genes that flip in individual tumours. If subtype matching alone closes the gap, the worst-case ingredient is unnecessary and we will say so.
Why not just filter genes expressed by cancer cell lines, as xCell does? Cell lines do not reproduce the lineage memory of individual patient tumours. Tran et al.'s results show that static signature approaches remain vulnerable at high purity. The CCLE-filter baseline tests this directly on the same held-out mixtures.
Why not just drop genes that vary across patients? High variance is not the same as flipped specificity. Many strong markers vary in level while keeping their rank order across cell types, so a variance filter should discard useful genes and increase false negatives, much like hspe. The variance-filter baseline measures this.
Decisive Checks
The central check is the 95% purity bin on held-out Wu patients and on the Bassez and Pal external atlases. We look for a substantial drop in phantom normal epithelium under DWLS and CIBERSORTx, some drop under BayesPrism, and flat false-negative rates.
A second check is whether the genes rejected by the worst-case rule are enriched for the luminal programs specific to each subtype.
On TCGA BRCA's 1,038 bulk tumours there is no ground truth for cell proportions. There we will only check that the inferred normal-epithelial fraction stops rising with inferred purity, and we treat this as a supporting sanity check rather than proof.
The main failure mode is that worst-case selection leaves too few markers for rare cell types. We will report panel sizes per cell type and relax $\delta$ only on patient splits that are not used for evaluation.
References
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Is "hot pizza" Positive or Negative? Mining Target-aware Sentiment Lexicons
- Inducing Domain-Specific Sentiment Lexicons from Unlabeled Corpora
- Polarity Consistency Checking for Sentiment Dictionaries
- Predicting Emotional Word Ratings using Distributional Representations and Signed Clustering
- xCell: digitally portraying the tissue cellular heterogeneity landscape
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data