Ground-Truthing Bulk Deconvolution with Laser-Capture qRT-PCR

When a molecular assay measures a whole tissue biopsy, it reports an average over all the cells in the sample — tumor cells, stromal cells, immune infiltrate, endothelium. Computational deconvolution promises to undo this averaging: given a bulk gene-expression profile and a reference atlas of pure cell types, it estimates both the cell-type proportions in the sample and the expression profile of each component. The problem is that these methods are almost never tested against real, physically separated tissue. Benchmarks construct "ground truth" by summing single-cell RNA-sequencing reads into synthetic bulk profiles (pseudo-bulks), so the reference atlas and the evaluation target share the same biases — dissociation artifacts, dropout, and whatever cell states the atlas happens to contain. The field has, in effect, been grading deconvolution on its own homework.

We propose to close this loop physically. Laser capture microdissection (LCM) is a microscope-guided technique in which a laser cuts out a researcher-outlined group of cells from a tissue section — a tumor core, a stromal margin, an immune aggregate — while leaving the rest behind. Quantitative real-time PCR (qRT-PCR) then measures a handful of chosen transcripts in that dissectate by counting the amplification cycle at which fluorescence crosses a threshold, which is inversely proportional to the starting copy number. Together they form a cheap, orthogonal, low-input ruler for what a specific physical region of a tissue actually contains. We will use LCM-plus-qRT-PCR anchors to benchmark deconvolution against physical ground truth, audit whether cell-line-derived reference profiles inject cell-line-specific artifacts into clinical deconvolution, and purity-correct both physical and computational readouts for the unavoidable contamination of laser cuts.

Core Claim

Our claim is threefold. First, method rankings established on pseudo-bulk benchmarks (as in the large breast-cancer evaluation of nine deconvolution methods by Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures) will not fully transfer to physical ground truth, particularly at high tumor purity, where that study documented a systematic "purity trap" — deconvolution methods confusing normal epithelial cells for cancer cells because the two share lineage programs. Second, swapping patient-atlas references for cheaper cell-line-derived references introduces a measurable, quantifiable bias in deconvolved proportions, because cell lines carry culture-specific transcriptional states rather than in-tissue states; this is the expression-level analogue of the drug-sensitivity discordance documented between the GDSC and CCLE screens (Assessment of pharmacogenomic agreement). Third, a small number of laser cuts plus one qRT-PCR plate per tissue block is enough to calibrate a purity correction that improves both the physical assay (which suffers from laser-cut contamination) and the computational one (which suffers from the purity trap).

Research Program

The program has three components, all run on the same set of archived formalin-fixed paraffin-embedded (FFPE) breast-cancer blocks — the routine currency of clinical pathology, chosen deliberately because FFPE tissue cannot be dissociated into viable single-cell suspensions, which rules out the most obvious competing approach.

The first component builds the anchor panel. From each block, we cut serial sections. On one section, a pathologist outlines compartments: tumor core, tumor–stroma margin, stromal region, and immune aggregates. Automated laser microdissection of the kind used in Deep Visual Proteomics (Deep Visual Proteomics defines single-cell identity and heterogeneity) — which achieves roughly 200 nm cutting precision, about 1,250 excised contours per hour, and quantifies thousands of proteins from as few as 100 cells — demonstrates that archived FFPE material, even 17 years old, supports high-quality molecular readout after dissection. Each dissectate goes to a low-input qRT-PCR panel: the eight genes of the clinically validated PCR-GGI breast-grading assay (four cell-cycle targets, four reference genes) from Improvement of the clinical applicability of the Genomic Grade Index through a qRT-PCR test performed on frozen and formalin-fixed paraffin-embedded tissues, plus roughly ten lineage-marker genes (epithelial, fibroblast, macrophage, T-cell, endothelial) that let us estimate, per dissectate, both the target-cell signature and the residual contamination from off-target cells. That contamination is expected and documented: cutting a 3D tissue cylinder based on a 2D image inevitably admixes underlying cytosol or stroma, as the Deep Visual Proteomics authors themselves note (Deep Visual Proteomics defines single-cell identity and heterogeneity).

The second component treats those anchors as the evaluation the field is missing. All major deconvolution benchmarking to date validates against synthetic aggregations of single-cell data: the foundational linear mixing model $Y = A \cdot X$ of In silico microdissection of microarray data from heterogeneous cell populations, the nine-method breast-cancer benchmark (Performance of tumour microenvironment deconvolution methods...), the Poisson count-bridge deconvolver of Count Bridges enable Modeling and Deconvolving Transcriptomic Data, and the ontology-aligned cross-modal transfer of POPPY (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level) all stop short of physical validation, and POPPY's authors explicitly note that states missing from the reference atlas cannot be recovered at all. We will run the same nine methods, under their original settings, on bulk RNA-seq from the section adjacent to each anchor, and score their predicted cell-type proportions against the LCM–qRT-PCR physical proportions from the same block.

The third component is the cell-line reference audit, which imports the fourth direction — cell line specificity — as a controlled variable. Deconvolution needs a reference profile for each cell type; patient-derived single-cell atlases are expensive, while cell-line expression profiles are abundant and standardized. But cell lines are known to drift into culture-specific states: the tamoxifen-resistance breast-cancer line models of Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance... show that rare stem-like states within lines dominate resistance phenotypes in ways bulk line profiles misrepresent, and the pharmacogenomic re-analysis of GDSC and CCLE (Assessment of pharmacogenomic agreement) shows that even genomic baselines are stable while line-level functional measurements are not (only 1 of 15 shared drugs met a $\rho \ge 0.8$ reproducibility bar; internal replicates correlated at just 0.39–0.57). We will deconvolve the same patient bulk profiles twice — once with patient-atlas references and once with cell-line-derived references — and measure the shift in estimated proportions and in the predicted cell-type-specific expression of the panel genes.

Finally, we purity-correct the physical readout itself, applying the linear unmixing idea of In silico microdissection... not to bulk tissue but to the laser dissectates. If $y$ is the measured panel in a dissectate, $\pi$ the fraction of on-target cells, and $x_\text{off}$ the panel measured in the adjacent stromal dissectate, we take

$$\hat{x}_\text{target} = \frac{y - (1-\hat{\pi})\, x_\text{off}}{\hat{\pi}},$$

with $\hat{\pi}$ estimated from the lineage-marker genes in the same plate. This inverts the usual logic: computational unmixing applied to physical dissection, correcting the Z-axis contamination that no downstream analytic currently addresses.

Proposed Experiments

Experiment 1 establishes feasibility on roughly 25 FFPE blocks (150–200 dissectates). The key risks are FFPE RNA degradation and low-input qPCR failure; the reference point is the PCR-GGI study, where 14% of bulk FFPE samples were excluded for degraded RNA and correlation with the microarray signature was $r = 0.89$ on FFPE versus $r = 0.95$ on frozen tissue (Improvement of the clinical applicability of the Genomic Grade Index...). Success means a dissectate-level failure rate below that 14% bulk figure and technical-replicate correlation approaching the $r = 0.96$ reproducibility that Deep Visual Proteomics achieves for low-input measurement of ~100-cell samples (Deep Visual Proteomics defines single-cell identity and heterogeneity).

Experiment 2 is the benchmark transfer test. Metric: per-block Spearman correlation between deconvolved and physical proportions, stratified by pathologist-estimated purity. Decisive comparisons: (a) does the pseudo-bulk ranking (BayesPrism and DWLS best on granular immune lineages; BayesPrism, Scaden, MuSiC most purity-stable, per Performance of tumour microenvironment deconvolution methods...) survive physical ground truth; (b) does the purity trap appear at high purity even for the most stable methods. We will state up front that if BayesPrism matches the physical anchors across the purity range, the central premise weakens — and that this negative result would itself be the first physically validated calibration of deconvolution and worth reporting.

Experiment 3 quantifies cell-line reference bias as the per-block difference in deconvolved proportions between the atlas-referenced and line-referenced runs, benchmarked against qPCR technical noise. We predict the bias will be non-uniform, pushing toward line-like lineage states, mirroring the subtype memory errors (TNBC mixtures over-predicting luminal progenitors) already seen with atlas references (Performance of tumour microenvironment deconvolution methods...).

Experiment 4 tests whether purity-corrected, LCM-anchored PCR-GGI scoring improves on uncorrected bulk PCR-GGI, whose FFPE performance ceiling is the published $r = 0.89$ concordance and $\text{HR} = 1.89$ for distant recurrence (Improvement of the clinical applicability of the Genomic Grade Index...). The specific target is the intermediate Grade 2 population, where stromal dilution plausibly distorts the proliferation signal.

Baselines

The naive reference point is doing nothing: run the eight-gene PCR-GGI assay on unseparated FFPE sections exactly as published (Improvement of the clinical applicability of the Genomic Grade Index...), on the same blocks, with no dissection and no deconvolution. This is the strongest "do nothing" baseline because it is a clinically validated assay with reported FFPE concordance of $r = 0.89$ to the microarray Genomic Grade Index and a distant-recurrence hazard ratio of 1.89 ($p < 0.001$) in the primary FFPE validation cohort. Beating it means improving intermediate-Grade-2 risk reclassification at equal assay cost, with improvement measured on held-out blocks.

The strongest existing competitor is BayesPrism, identified by Performance of tumour microenvironment deconvolution methods... as the top performer for granular immune lineages and among the most stable across tumor purity on their 38,000-mixture pseudo-bulk benchmark built from the Wu et al. breast atlas (100,064 cells, 26 patients). For fairness, we run BayesPrism (and DWLS, Scaden, and MuSiC as secondary comparisons) on adjacent-section bulk RNA-seq from our blocks, with the same Wu et al. atlas reference and the authors' original hyperparameters, so the only changed variable is the ground truth: physical anchors instead of pseudo-bulks. No reported result found on Lacuna for a numeric accuracy value of BayesPrism on that benchmark — the file records only qualitative rankings — which is precisely why a physical anchor with a hard number matters. BayesPrism is beaten if its per-block Spearman correlation with physical proportions falls materially below the LCM–qPCR panel's own reproducibility ceiling, or if it fails where the anchor pipeline does not at high purity.

Two ablations isolate our key ingredients. Replacing the physical anchors with pseudo-bulk ground truth (same methods, same data) isolates the contribution of physical validation — if method rankings are identical under both, the anchor adds cost without information, and we will say so. Swapping the atlas reference for cell-line-derived references while holding everything else fixed isolates the cell-line bias ingredient and directly quantifies the direction we care about.

Why not just…?

Why not just run single-cell RNA sequencing on the tissue? For fresh surgical tissue this is a real threat, and we do not claim superiority there. But three things limit it for our setting: dissociation destroys the spatial and morphological context that motivates compartment-level measurement (Deep Visual Proteomics defines single-cell identity and heterogeneity); clinical archives consist overwhelmingly of FFPE blocks that cannot be dissociated into usable suspensions, while laser dissection recovers molecular profiles even from 17-year-old FFPE; and the clinical-deployment argument that carried the PCR-GGI into routine use — cheap, fast, platform-agnostic targeted assays — applies to qRT-PCR, not to droplet sequencing (Improvement of the clinical applicability of the Genomic Grade Index...). We will, however, include one fresh-frozen comparison set so the scRNA-seq alternative is answered empirically rather than by assertion.

Why not just use computational deconvolution alone and skip the laser entirely? This is the alternative our Experiment 2 is designed to settle rather than argue around. The evidence for shortfall is real but indirect: the purity trap at high tumor purity and the subtype memory errors documented across nine methods (Performance of tumour microenvironment deconvolution methods...), and the atlas dependency that makes unlisted cell states unrecoverable (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level). None of that was measured against physically separated tissue. If deconvolution matches our anchors everywhere, deconvolution-only wins and the honest conclusion is that physical anchoring is unnecessary for proportion estimation — while still leaving our cell-line reference audit and purity-corrected readout as the contribution.

Why not just use the existing bulk PCR-GGI panel without dissection? It is validated, cheap, and already works on FFPE ($r = 0.89$; tamoxifen response 74.5% in low-risk versus 53% in high-risk patients, per Improvement of the clinical applicability of the Genomic Grade Index...). But it assays whatever mixture the section happens to contain, its authors did not test ultra-low-input microdissected samples or correct for stromal dilution, and its proliferation-only design limits it in ER-negative and HER2-positive disease where proliferation is uniformly high. Experiment 4 measures exactly how much purity correction buys in the intermediate-grade population where dilution bites; if the gain is negligible, we will report that the bulk panel is robust to composition — a useful negative result for the assay's users either way.

Decisive Checks

The proposal succeeds or fails on three numbers: the transfer (or flip) of the pseudo-bulk method ranking under physical ground truth, stratified by purity; the magnitude of the cell-line reference bias relative to qPCR technical noise; and whether purity-corrected PCR-GGI improves intermediate-grade reclassification over the published $r = 0.89$ FFPE ceiling. Each is measurable with a few hundred dissectates, one sequencing run per block, and standard clinical-lab equipment — small experiments, but the first ones in this literature where the truth comes from the tissue instead of from the model's own inputs.