Cytokine-Conditioned Deconvolution of Precancerous Bulk Tissue

Chronic inflammation does not merely accompany cancer — it rewires the cells that will become cancerous. Sustained exposure to cytokines like IL-6 and IL-1β progressively reprograms epithelial and stromal cells toward stem-like, apoptosis-resistant states, and recruits transient myeloid populations that act as ignition switches for malignant transformation. Yet nearly everything we know about this cascade in patients comes from either single-cell snapshots of fully developed tumors or aggregate bulk measurements of precancerous tissue. The bulk measurements are the clinically abundant resource — thousands of banked gastritis, colitis, and cirrhosis biopsies with long-term follow-up — but they pool all cells into one average, and the averaging is uniquely damaging here: it conflates "more inflammatory cells moved in" with "the same cells got reprogrammed," two mechanistically opposite explanations that demand opposite interventions.

Standard deconvolution — the family of methods that unmix a bulk transcriptome into constituent cell types — cannot resolve this ambiguity because almost every existing tool assumes that a cell type's transcriptional signature is stable across health and disease. This "stable state hypothesis" is stated explicitly as the foundation of Bulk2Space, which synthesizes single-cell profiles from bulk RNA-seq via a variational autoencoder and maps them into space. It is exactly the assumption that inflammation-driven oncogenesis violates: IL-6 continuously deacetylates FRA1 and expands cancer stem populations in colorectal tissue (Wang et al.), and stromal fibroblasts and dendritic cells — not epithelium — emerge as the dominant secretors of the master gastric carcinogenesis driver CHI3L1 (Li et al.). The proposal here is to build and validate a deconvolution model in which each cell type carries a continuous activation state that is an explicit, learned function of the local inflammatory milieu, and to apply it to longitudinal precancerous cohorts where the proportion-versus-reprogramming question is open and answerable.

Core Claim

The claim is that bulk RNA-seq of chronically inflamed precancerous tissue contains enough signal to jointly estimate, per sample: (i) the proportions of constituent cell types, (ii) a continuous per-cell-type inflammatory activation state, and (iii) the presence of transitional cell states absent from any single-cell reference atlas — provided the deconvolution model is conditioned on inflammation rather than assuming static signatures. Two lines of evidence motivate this. DestVI showed, in spatial spots of infected lymph node and fibrosarcoma, that modeling a continuous latent state per cell type (its $\gamma$ vector) recovers inflammation-driven activation programs — monocytes upregulating interferon-response genes only in a specific tissue compartment — that discrete-type methods like RCTD and SPOTlight miss entirely. And Count Bridges established a generative process that operates natively on integer transcript counts and outperformed CIBERSORTx and MuSiC at recovering cell-type proportions. What neither does, and what no paper in this corpus does, is deploy state-aware deconvolution on non-spatial clinical bulk cohorts across a precancerous cascade, where the state shift is the disease process itself.

The model — call it a cytokine-conditioned count deconvolver — treats a bulk sample's counts $y_g$ for gene $g$ as arising from cell types $k$ with proportions $\pi_k$ and expression profiles parameterized by a per-type latent state $\gamma_k(s)$, where $s$ is a sample-level inflammatory covariate (a composite IL-6/IL-1β/CHI3L1 signature score computed from the bulk data itself). The continuous dependence $\gamma_k(\cdot)$ is the point: it is learned as a smooth, low-dimensional function, so the model can extrapolate to inflammatory regimes where reference single-cell profiles were never collected, and it can flag cell types whose inferred state has moved off the reference manifold — the signature of a transitional population like the week-6 monocyte-like macrophages (Ly6C⁺ MHCII⁻) that Yang et al. showed peak in the AOM-DSS colitis model before any macroscopic tumor forms, or the CD68⁺CD206⁺CCL18⁺ macrophage subset that Song et al. tied to poor hepatocellular carcinoma survival ($P = 0.001$). The count layer follows the integer birth-death machinery of Count Bridges, with an EM procedure that treats single-cell composition as latent and enforces that inferred cells sum exactly to the observed bulk counts. Unlike the reference-atlas transfers of POPPY, whose Cell Ontology alignment presumes the relevant states exist in 3CA (an atlas of established tumors), the orphan-state component here is discovered from the bulk data itself.

Research Program

Three building blocks, then three validation campaigns. First, the generative core: extend the DestVI-style per-type latent state from spatial spots to non-spatial bulk counts, replacing its MAP inference with a proper posterior over $\pi_k$ and $\gamma_k$ so that rare transitional states carry calibrated uncertainty rather than point estimates. Second, the cytokine conditioning: $\gamma_k$ is regressed on $s$ through a monotone spline basis shared across samples within a cohort, so the model learns, e.g., that fibroblast CHI3L1 output rises with inflammatory load rather than merely that fibroblast abundance rises. Third, the orphan-state detector: a low-rank non-negative component that absorbs residual structure no reference type explains, flagged and characterized post hoc by gene-set enrichment against known transitional programs (the MLM signature of Yang et al., the M2-axis of Song et al.).

The validation campaigns target the three open questions left dangling by the corpus. The gastric cascade (bulk GSE55696 and GSE66229 paired with single-cell GSE167297, from Li et al.) is the cleanest testbed for proportion-versus-reprogramming: Li and colleagues located CHI3L1 production in fibroblasts and dendritic cells with single-cell data but never ran deconvolution on their bulk gastritis cohorts, so whether rising bulk CHI3L1 across gastritis → atrophy → metaplasia → dysplasia reflects expanding secretor populations or per-cell activation is unmeasured. The HBV/HCV hepatocellular carcinoma path is the clinical-outcome test: Song et al. translated their Mφ_c4 macrophage signature into TCGA-LIHC only as an aggregate bulk score; we will re-run that translation with state-aware deconvolution and ask whether the continuous macrophage state estimate — not just the proportion — carries the prognostic signal ($P = 0.001$ for overall survival in their single-cell analysis). The AOM-DSS colitis model is the temporal test: pseudo-bulk the published single-cell time course at weeks 0–10 and ask whether the model recovers the week-6 MLM peak that Yang et al. found by flow cytometry, and whether the antibiotic perturbation (which cut tumor load and CCL2 nearly 10-fold) is visible as a state reversion rather than a proportion change.

Proposed Experiments

The methodological evaluation runs on the benchmark of Tran et al.: 38,000 pseudo-bulk mixtures simulated from the Wu et al. breast cancer single-cell atlas (100,064 cells, 26 patients) across 5%–95% tumor purities, on which nine deconvolution tools were benchmarked. We add a second, biologically motivated stress test: the same mixtures but with reference profiles perturbed by simulated cytokine exposure (IL-6-style stemness induction on epithelial cells, following the FRA1/NANOG program documented by Wang et al.), so that the "stable state hypothesis" is violated in a controlled, dose-graduated way. Metrics are proportion-recovery RMSE and correlation, false-positive/false-negative rates per lineage, and — uniquely enabled by the design — state-recovery error: the distance between inferred $\gamma_k$ and the true applied perturbation. Tran et al. report qualitatively that BayesPrism and DWLS achieved the lowest combined false positives and false negatives and best granular immune-lineage performance, with BayesPrism, Scaden, and MuSiC most stable across purities; no reported result found on Lacuna for exact per-method RMSE values in that benchmark, so we will reproduce all baselines ourselves under matched conditions rather than compare to quoted numbers.

For the precancerous campaigns, the decisive metrics are: in gastric tissue, whether the model's decomposition of bulk CHI3L1 rise into secretor-proportion versus per-cell-secretion terms matches the single-cell ground truth from GSE167297 at each cascade stage, and whether the early-cancer separation approaches the aggregate-classifier ceiling that Li et al. obtained with an SVM on bulk features (AUC 0.98 for identifying early gastric cancer) while additionally yielding the cell-type attribution that their machine-learning classifiers lack. In TCGA-LIHC, whether a fitted Cox model on continuous macrophage state beats the same model on aggregate Mφ_c4 signature score and on macrophage proportion alone, using the survival endpoints Song et al. established ($P = 0.001$ for the subset, $P = 0.031$ for the favorable XCL1⁺ CD8⁺ T-cell counterpart). In the colitis model, whether the inferred orphan component's temporal profile peaks at week 6 and reverts under antibiotic treatment, matching the flow-cytometry ground truth. A final negative control comes from SubNetDL, whose subclone-level network model achieved median AUROC 0.77 on both MSK NSCLC immunotherapy cohorts (242 and 240 patients), significantly beating aggregate tumor mutational burden ($P = 0.02$; $P = 6.0\times10^{-3}$) — establishing that finer-grained decomposition pays off clinically, and simultaneously serving as a DNA-side comparator: we will test whether coupling our inferred inflammatory niche states to SubNetDL-style subclonal structure improves on either alone, since SubNetDL is blind to microenvironmental cell states by construction.

Baselines

The do-nothing reference point is the aggregate bulk gene-signature score, the practice Song et al. actually used when translating their single-cell HCC findings into TCGA-LIHC: score each bulk sample with the mean expression of a marker set (their $CD68^+CD206^+CCL18^+$ macrophage genes, or an IL-6/CHI3L1 inflammatory signature) and correlate with outcome or stage. This is what the field does today, it requires no model, and it already carries real signal — the gastric SVM of Li et al. reached AUC 0.98 with bulk features alone — so it is the bar any new method must clear while adding something the score cannot say: which cells the signal comes from and whether their abundance or their internal state changed. It will be run on identical cohorts and identical endpoints, and beaten only if the decomposition changes any clinical conclusion (e.g., shows a signature attributed to epithelium actually originates in fibroblasts, as the CHI3L1 case suggests it might).

The strongest existing competitor is BayesPrism, which in the head-to-head benchmark of Tran et al. achieved the lowest combined false positives and false negatives and the best performance on granular immune lineages, with the highest stability across purity levels of any tested method; CIBERSORTx and MuSiC, the most widely deployed tools clinically, are weaker on this benchmark and additionally treat counts as continuous, which Count Bridges showed costs accuracy at the count level. No reported result found on Lacuna for numeric BayesPrism RMSE, so it will be rerun on our two mixture benchmarks (original Tran-style and cytokine-perturbed) with the same single-cell references, the same gene universe, and the same tuning budget as our model; beating it means lower proportion RMSE across purity regimes and, on the perturbed benchmark specifically, non-degenerate state recovery where BayesPrism's static reference forces it to misattribute reprogrammed cells to the nearest unchanged type. Bulk2Space (Liao et al.) is the conceptual closest competitor and will be run under the same protocol; the prediction is that it degrades specifically under the perturbation benchmark because its stable-state assumption is what the perturbation breaks.

The key ablation is our own model with the cytokine conditioning removed — a static per-type $\gamma_k$ identical for all samples, which reduces the method to a count-native DestVI analogue. This isolates the proposal's single claimed ingredient: that conditioning cell state on the inflammatory covariate, rather than merely allowing free per-sample states, is what enables extrapolation to regimes lacking references and what separates proportion change from reprogramming. If the static-state ablation matches the conditioned model on the gastric and LIHC campaigns, the central claim is wrong and the proposal collapses to an engineering improvement on DestVI — a result we would report.

Why not just…?

Why not just use routine blood inflammation indices (dNLR, SIRI, MLR)? These are essentially free and already prognostic — Qin et al. showed SIRI, ALI, and MLR identify distant metastasis risk in breast cancer, and Jiang et al. showed dNLR partially mediates outcomes in laryngeal carcinoma. But they compress the whole tissue into one non-specific number: Song et al. demonstrated that within the very same HCC patients, an immunosuppressive macrophage subset tracks death ($P = 0.001$) while an activated CD8⁺ T-cell subset tracks survival ($P = 0.031$) — opposing forces a blood ratio cannot separate. Our TCGA-LIHC experiment will make this concrete: if the aggregate-style index and the deconvolved states carry the same prognostic information, the skeptic wins; the design tests it rather than assumes it.

Why not just run CIBERSORTx or MuSiC on the precancerous cohorts? Beyond Tran et al.'s purity-trap findings (systematic misclassification of normal epithelium as malignant above 80–90% purity, with obvious relevance to dysplastic epithelium that shares lineage programs with cancer), the deeper failure for this application is that all of these tools force every observed profile onto reference types that were cataloged in established disease. The transitional MLM state that Yang et al. found peaking at week 6 does not exist in any atlas — 3CA, the reference behind POPPY, is built from 15 established cancer types. A skeptic can fairly object that retraining these tools with expanded references might suffice; that is precisely what the orphan-state detector answers empirically, because the week-6 colitis recovery is a state no reference we supply contains.

Why not just sequence everything at single-cell resolution? This is the real threat, not a strawman: single-cell profiling is how every biological insight in this proposal was discovered. The honest answer is part cost, part coverage, part physics. Cost: the clinically valuable cohorts — TCGA's hundreds of samples, longitudinal gastritis and cirrhosis biopsies with outcome data — are bulk, and will remain bulk; reprofiling them single-cell is not going to happen at scale, as Venkat et al. and Tran et al. both argue. Physics: even spatial platforms aggregate — 10x Visium spots hold dozens of cells, and MERFISH panels cover only hundreds of genes (Lopez et al.) — so unmixing is needed regardless. But the proposal concedes the point where it bites: our validation strategy deliberately uses single-cell data as ground truth at every step (GSE167297, the HCC immune atlas, the colitis time course), so if bulk signal proves insufficient, the experiments will say so cleanly rather than burying the failure.

Why not just retrain Bulk2Space or POPPY on better atlases? Retraining addresses reference completeness, not the stable-state assumption. The contradiction documented by the corpus is structural: Wang et al. showed the same cells changing state under IL-6 within a patient across time, which no fixed reference profile — however large — can represent. The cytokine-perturbation benchmark is designed to settle this directly: Bulk2Space run on identical data with identical references, with performance degradation attributable only to its state-stability assumption.

Decisive Checks

Three outcomes decide the proposal. First, on the cytokine-perturbed mixtures, the conditioned model must beat BayesPrism and the static-$\gamma$ ablation on state recovery, not merely on proportions — proportion-only parity would mean the conditioning adds nothing. Second, the gastric decomposition must be non-trivially checkable against GSE167297: if the model attributes the bulk CHI3L1 rise across the cascade to per-cell fibroblast activation where the single-cell data show proportion expansion (or vice versa), it fails its one mechanistic prediction. Third, the colitis time course must recover the week-6 MLM peak and its antibiotic reversion without any MLM profile in the reference — the strictest test of the orphan-state mechanism, and the one no existing tool in this corpus could pass. If all three hold, the deliverable is a validated method that turns the thousands of banked precancerous bulk biopsies into a cell-resolved, inflammation-resolved record of early oncogenesis — the record the field currently lacks.