Somatic Alleles as a Built-In Index for Tumor RNA Deconvolution

Bulk RNA sequencing grinds a tumor into one tube and reports how much of each gene's message is present, averaged over every cell. Methods that unmix this average back into cell types work reasonably on immune cells. They fail in a specific, documented way on the cells that matter most: as a tumor gets purer, every tested method increasingly mistakes normal epithelial cells for cancer cells. They do this because cancer cells still express much of the program of the normal cell they came from. In the language of error-correcting codes, the "codewords" for malignant and normal epithelium sit too close together for any decoder to separate. A better decoder cannot fix that. The fix is to read a second, independent signal that the two cell types do not share.

We propose that the reads themselves already carry such a signal. Cancer cells carry somatic mutations and copy-number changes, and RNA reads that cover heterozygous or mutated positions are effectively tagged with their genome of origin. Coding theory for DNA storage has shown that embedding an index inside the payload is the optimal way to recover which molecule a read came from. We borrow that structure. An inner stage reads allele counts at informative sites to estimate the malignant fraction, and it declares an explicit "erasure" when too few informative reads exist. An outer stage performs ordinary expression deconvolution with that fraction pinned. The goal is to break the purity trap using only the RNA-seq data already collected. The theory says how many informative reads are needed before the anchor can be trusted.

Core Claim

The purity trap reported in Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures is a problem of confusable signatures, not of fitting. That benchmark simulated 38,000 mixtures from a 100,064-cell, 26-patient breast atlas. It found that all nine methods, including BayesPrism and DWLS, increasingly misassign normal epithelium as malignant as purity rises from 5% to 95%. It also found that subtype memory sends triple-negative samples toward luminal progenitors.

Our claim is that a genotype-anchored, two-stage decoder reduces this error, and that it does so in a predictable way. The amount of error removed should be governed by the number of informative allelic reads. When that number falls below a threshold the decoder should abstain rather than guess. The method should never be worse than expression-only deconvolution, because an erasure falls back to it.

Why Coding Theory Is the Right Lens

The method comes from concatenated coding: an inner code absorbs messy channel noise and hands the outer code clean erasures at known positions. In Multilayer Codes for Synchronization from Deletions, inner synchronization codes turn insertions and deletions into block erasures that a simple outer code then fills. We use this division of labor. The allelic stage handles the sparse, noisy genotype channel and either outputs a malignant-fraction estimate with a stated confidence or declares an erasure. The expression stage then solves an easier problem.

DNA-Based Storage: Models and Fundamental Limits treats sequencing as random sampling of reads from an unordered pool, where some molecules are never seen. Exact Error Exponents of Concatenated Codes for DNA Storage shows that, under exact multinomial sampling, failure is dominated by "outage" events where too few distinct molecules are observed, and that coded-index schemes achieve the best error decay. We take two lessons from these papers. The first is that the right quantity to track is the count of informative observations, not total depth. The second is that an index carried inside the payload is the efficient design.

Both papers concern synthetic storage, where the encoder is designed. Our encoder is evolution: the tumor's somatic genome writes the index for us, unevenly. That difference is why the outage analysis has to be redone for this setting.

Concatenated Codes for Multiple Reads of a DNA Sequence shows that combining per-read soft probabilities separately is nearly as good as exponentially expensive joint decoding. We use the same choice and pool soft allelic evidence across sites and haplotype blocks rather than decoding them jointly.

Error Exponents for DNA Storage Codes with a Variable Number of Reads suggests an optional extension. When the inner stage declares an erasure, the same library can be topped up with more sequencing, and the error should then fall faster than under fixed depth.

Research Program

The inner stage uses the following observation model. Take a heterozygous germline site in a region where the tumor has lost one allele. Let $\pi$ be the malignant cell fraction, and let $e_N$ and $e_T$ be the per-cell expression of that gene in the non-malignant and malignant compartments. The expected fraction of reads carrying the lost allele is

$$ b = \frac{(1-\pi)\, e_N / 2}{(1-\pi)\, e_N + \pi\, e_T}. $$

Analogous expressions hold for gains, for allele-specific imbalance, and for expressed somatic point mutations. The equation shows that the genotype channel and the expression channel are coupled through $e_N$ and $e_T$. Because of this coupling, the two stages exchange soft information iteratively, like inner and outer decoders, rather than running once in sequence.

Allele counts at each site are binomial given $b$. A Chernoff bound on the pooled counts gives a confidence statement on $\pi$ whose error decays exponentially in $K$, the number of reads covering informative sites. That bound sets the erasure threshold. Haplotype phasing across adjacent SNPs, inferred from population reference panels, lets many weak sites be pooled into blocks, which raises the effective $K$ without deeper sequencing.

The outer stage is a standard reference-based deconvolver; we will use BayesPrism and DWLS as interchangeable outer modules. It is run with the malignant fraction constrained to the inner estimate, or left free when the inner stage declares an erasure. Its malignant expression profile is fitted per sample, so a triple-negative tumor's luminal-progenitor-like program is absorbed into the malignant compartment instead of being assigned to normal luminal progenitors.

The analysis deliverable is a two-term decomposition of deconvolution error. One term is outage error, driven by too few informative reads, and it should vanish exponentially as $K$ grows. The other is confusability error, driven by how close the reference signatures are. The decomposition will give practitioners a rule for when RNA-only purity anchoring can be trusted for a given tumor's copy-number burden and sequencing depth.

Baselines

The naive reference point is the expression-only deconvolver run unchanged. This is BayesPrism, identified as the strongest method for granular lineages in Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures, together with DWLS from the same benchmark. These are the strongest competitors because they are exactly what the purity trap was measured on. We will run them with the same reference atlas, the same gene sets and the same mixtures as our method. Our method adds only the allelic stage on top of the same outer module, so any gain comes from the anchor.

That paper's summary gives no per-tool error values: no reported result found on Lacuna for BayesPrism or DWLS RMSE on the breast pseudo-bulk mixtures. We will therefore re-run both and report our own numbers. Beating them means three things. Normal-epithelium misattribution must drop significantly in the 70–95% purity bins. Immune-fraction accuracy must not get worse. On erasure samples, results must be at least as good as the baseline, since those samples fall back to it.

The second baseline is BayesPrism given an oracle malignant fraction, taken from simulation ground truth or, on real data, from DNA-based purity. This ceiling tells us how much of the purity trap is fixable by anchoring at all. A related baseline uses an expression-only copy-number proxy: a smoothed gene-expression profile along each chromosome feeds the same inner stage, but without allele counts. This ablation isolates the key ingredient, allele-tagged reads, from the general idea of using copy-number structure.

Count Bridges, from Count Bridges enable Modeling and Deconvolving Transcriptomic Data, models integer counts and unmixes bulk samples into constrained single-cell profiles. It reportedly beats CIBERSORTx and MuSiC, though the file gives no numbers. We include it as the strongest generative deconvolver, run on the same mixtures. It sees only expression, so it should fall into the same trap.

POPPY, from Cross-modal transfer learning for mapping bulk transcriptomes at cellular level, maps bulk samples into single-cell embedding space using Cell Ontology structure. Its reported gains, up to 73% in silhouette and up to 714% in composition correlation, are relative cohort-integration gains on a different task. They are not comparable targets. We treat it as a secondary comparison only.

Why not just…?

Why not just plug in a DNA-based purity estimate? Where matched tumor DNA exists, this is the right thing to do, and our oracle-anchor baseline is exactly that. Most RNA-seq cohorts and archival clinical samples, however, have no matched DNA. Our contribution is recovering most of that anchor from RNA alone and knowing when the recovery is reliable. If the RNA-only anchor reaches the oracle's gain only on high copy-number tumors, the outage analysis should predict exactly that pattern. The pattern itself is the deliverable.

Why not just improve the reference, for example by adding tumor-matched normal epithelium or subtype-specific malignant profiles? The benchmark already shows the confusion persists across all nine methods, including those that model malignant cells per sample. That points to closeness of the signatures rather than a missing reference entry. We will test this directly by giving the baseline a subtype-matched malignant reference from held-out patients. If that closes the gap, the anchor is unnecessary, and we will say so.

Why not just infer copy number from expression, as is common for single cells? Expression-based copy-number inference reads the same channel that is confused. A luminal-like program shifts smoothed expression in ways that look like gains and losses. Allele counts are, to first order, independent of how strongly a gene is expressed. The expression-only copy-number ablation tests this, and it is a real threat on tumors with large arm-level changes. We expect allele reads to matter most on tumors with focal or copy-neutral loss-of-heterozygosity events.

Proposed Experiments

The first experiment uses the breast atlas of Wu et al. from the deconvolution benchmark. We will regenerate pseudo-bulk mixtures across the same 5–95% purity range and attach simulated allele counts. Malignant cells receive patient-specific copy-number and loss-of-heterozygosity profiles; non-malignant cells receive balanced germline heterozygosity. Reads are sampled multinomially at depths spanning typical bulk studies. The decisive check is that normal-epithelium misattribution under our method stays flat as purity rises, while the baselines degrade. The improvement should also track the predicted exponential dependence on informative reads $K$. If gains do not scale with $K$, the coding interpretation is wrong.

The second experiment uses held-out validation atlases (Bassez et al., Pal et al.), mixed with the breast atlas reference, to test transfer across studies.

The third experiment uses the 1,038 TCGA breast samples. There we compare RNA-only malignant fractions against DNA-derived purity, and we check whether erasure flags concentrate on tumors with few copy-number changes. Erasures should be rare in tumors with many copy-number changes and frequent in quiet genomes.

As an optional final check, we will simulate top-up sequencing only on erasure samples. This tests whether the variable-read error-exponent prediction carries over, which would determine whether adaptive resequencing is worth recommending.