Niche or Clone: Attributing the Sources of Gene Silencing in Spatial Tumor Maps

When a protective gene — a tumor suppressor, a metabolic enzyme, a regulatory RNA — switches off in a tumor, we usually see the aftermath (uncontrolled growth, immune evasion, metastasis) but not the initial off-switch. There are two candidate causes that demand completely different interventions: the cell's own internal changes, such as an activating mutation in a clonal lineage, or the cell's physical neighbors, such as inflammatory or hypoxic tissue niches that send repressive signals. No current method can tell these apart. Bulk sequencing averages the whole tissue into one tube and cannot even say which cell type silenced the gene; spatial transcriptomics — technologies that measure RNA while keeping each cell's original position in the tissue — can show that silencing co-occurs with a bad neighborhood, but co-occurrence is not cause; and mechanistic studies test silencing by artificially knocking genes down in flat 2D cultures, the very setting that removed the neighbor signals in the first place. The result is a literature full of silenced guardians whose consequences are meticulously mapped and whose triggers are unknown.

This proposal builds a computational method that splits every observed gene-silencing event in spatial tumor data into a clone-intrinsic part and a neighbor-triggered part, and then uses that split to predict which silencing events will reproduce in standard 2D culture. The in vitro–in vivo translation gap thus stops being a vague complaint about models and becomes a measurable, per-event quantity: a silencing event that is niche-attributed should fail in monoculture and recover in co-culture, and this is directly testable. The clearest motivating cases are two guardian genes whose triggers remain open questions in the literature: malic enzyme 3 (ME3), downregulated in gastric cancer across 442 patients ($P < 0.001$) with the initiating trigger explicitly unknown despite extensive 2D knockdown work Downregulation of malic enzyme 3 facilitates progression of gastric carcinoma..., and RACK1, progressively silenced during pancreatic metaplasia with repression by KRAS/TGF-α established but the microenvironmental mechanism unresolved RACK1 attenuates pancreatic tumorigenesis....

Core Claim

Per-gene, per-location attribution of silencing to niche versus clone is identifiable from data that already exists: deconvolved spatial transcriptomics plus a two-branch generative model with an explicit counterfactual. For a guardian gene $g$ in cell type $k$ at spot $s$, let $x_s$ be intrinsic features (cell type, expression-inferred clonal state and copy number) and $n_s$ the Dirichlet-encoded composition of neighboring cells. Train one branch to predict expression from $x_s$ alone and a second from $(x_s, n_s)$; the niche-attributable silencing is the counterfactual gap

$$A_g(s) \;=\; \pi_g(s)\,\big[\hat{y}^{\text{int}}_g(s) - \hat{y}^{\text{niche}}_g(s)\big],$$

where $\pi_g(s) \in [0,1]$ is a learned gate that shrinks the attribution to zero whenever neighbor composition carries no predictive information for that gene. This gate is the direct answer to a documented failure mode: NicheVI, the closest existing niche-conditioned model, assumes a cell's latent state is coupled to its neighbors and cannot represent cells whose regulation is niche-independent NICHEVI. Our claim is falsifiable in the simplest way: if the gate learns $\pi \approx 0$ everywhere, or if the intrinsic branch alone predicts everything, then niche attribution adds nothing and the proposal fails on its own terms.

Research Program

Three components. First, silencing quantification on mixed spots: each spatial spot in sequencing-based platforms pools 10–50 cells, so we deconvolve per cell type before asking who silenced what. We use the top reference-based tools (Cell2location, Tangram) identified by the 18-method, 50-dataset benchmark — which reports performance rankings but no absolute number on Lacuna, so we re-derive accuracy on our own simulated spots against Starfysh's reference-free accuracy of RMSE ≈ 0.09 on simulated mixtures as a sanity anchor A comprehensive benchmarking...; Starfysh. Second, the attribution model itself: the two branches above, with the niche branch consuming both neighbor composition (NicheVI-style) and ligand–receptor transport fields computed with collective optimal transport COMMOT, so the model can propose which neighbor signal, not just that neighbors matter. Third, the translation module: each attributed event is matched against in vitro perturbation signatures from LINCS L1000-style data using the dependency-network rewiring model of CellVS-Net, which reduced network prediction error by 12–47% across perturbation types Cell-Level Virtual Screening, producing a per-event probability of reproducing in 2D monoculture.

The design deliberately inverts the usual roles of the four research directions. Cellular heterogeneity masking is the obstacle the deconvolution step must clear; spatial transcriptomics is both the data substrate and the causal lens (neighbors are only observable with coordinates kept); upstream regulatory silencing supplies the target phenomenon and the case studies; and the in vitro–in vivo gap becomes the output — a predicted, per-event reproducibility score — rather than an unmodeled nuisance. This last move is what separates the proposal from existing transfer models: DeepSADR pre-trains on 966 DepMap cell lines and fine-tunes on 555 bulk TCGA tumors, reaching a Sorafenib response AUC of 0.957 versus 0.811 for the next best baseline, but its clinical targets are bulk averages that erase the spatially sequestered subclones where resistance actually arises DeepSADR; A network-based deep learning model integrating subclonal architecture.... We do not compete with DeepSADR on drug response; we supply the per-event spatial decomposition its bulk targets cannot contain.

Baselines

The naive reference point is a single bulk-level scalar: the tissue-average expression of the guardian gene itself, alongside tumor mutational burden (TMB) as the classic global biomarker. This is what most clinical association studies effectively use, and it is known to be beatable — SubNetDL achieved a median AUROC of 0.77 on the MSK NSCLC immunotherapy cohorts ($n=242$ and $n=240$), significantly outperforming TMB with $P = 0.02$ and $P = 6.0 \times 10^{-3}$ SubNetDL. We will compute both scalars on the same patient cohorts; beating them means predicting which guardian genes are silenced where, and which patients progress, at higher AUROC than any tissue-average scalar.

The strongest existing competitor on the niche question is NicheVI, a variational autoencoder that conditions each cell's embedding on the Dirichlet-distributed composition of its spatial neighbors NICHEVI. It is the closest method because it already learns neighbor-conditioned cell states — but it produces phenotypic embeddings, not per-event causal attributions, and it breaks exactly where our gate is designed to work. To make the comparison fair, we run NicheVI on the same deconvolved spots, the same train/test gene splits, and the same tuning budget, adding a differential-expression head on top of its embedding so it can make the same predictions our model makes. For the spatial expression-prediction subtask, the strongest reported number on Lacuna is SpaOTsc's held-out spatial expression prediction: average AUC of 0.95 in zebrafish embryo (6hpf) and 0.88 in Drosophila embryo SpaOTsc. Those tissues are not ours, so we re-implement SpaOTsc's transport-based prediction on our data under matched compute; beating it means exceeding its per-tissue AUC on the same held-out guardian genes. Finally, the key ablation is the intrinsic-only branch — the full model with the niche input severed. If this ablation matches full performance on silencing prediction and on in vitro reproducibility, the central claim is dead, and we will report it as such.

Why not just…?

Why not just deconvolve bulk samples with CIBERSORTx or MuSiC? Because bulk unmixing returns only global fractions $w_k$ and cannot say whether the cell silencing a guardian gene sits next to the cells that might be triggering it. The masking problem is precisely the loss of physical co-localization Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography, and continuous-regression deconvolution additionally mishandles the integer geometry of counts, which Count Bridges demonstrates against exactly these methods Count Bridges. Niche attribution without coordinates is not a weaker version of this proposal; it is a category error.

Why not just profile more dissociated single-cell RNA-seq? Dissociated scRNA-seq can find which cell type silenced a gene but destroys the neighbor relationships that could explain why — the paracrine ranges and contacts are gone the moment the tissue is enzymatically digested SpaOTsc. This alternative is a real threat in one honest sense: if clonal state alone predicts silencing, dissociated data plus covariate regression would suffice. Our intrinsic-only ablation is exactly this method, run inside our own pipeline, so Experiment 2 settles it empirically rather than by argument.

Why not just run SpaGTL and COMMOT on the spatial data and correlate silencing with niche labels? This is the alternative I take most seriously. SpaGTL infers spatially aligned gene regulatory networks but models positive transcriptional regulons, not repressive cascades or epigenetic off-switches, and has not been evaluated against perturbation datasets SpaGTL; COMMOT maps downstream signaling directionality from mRNA proxies but cannot say why a sender stopped expressing a protective ligand COMMOT. A simple pipeline — niche labels from COMMOT fields, then per-niche differential expression — is a genuine competitor, so we run it as an explicit baseline in Experiment 2 with the same inputs. If its correlational attribution predicts in vitro reproducibility as well as our gated counterfactual, the counterfactual machinery is unnecessary, and the simpler pipeline should win the paper.

Proposed Experiments

Experiment 1 (calibration on ground truth we control): extend the synthetic spot-mixture construction from the 50-dataset deconvolution benchmark A comprehensive benchmarking... to inject silencing events whose probability depends either on a clonal genotype or on neighbor composition. We measure attribution accuracy (AUROC of source classification) and deconvolution quality, checking against Starfysh's RMSE ≈ 0.09 reference point. Experiment 2 (real tissues): apply the full pipeline to the 10x Xenium human breast cancer data and MERFISH mouse brain data used by NicheVI NICHEVI, predicting held-out guardian-gene silencing per cell type per location, against NicheVI-plus-DE-head, SpaOTsc re-run, and the COMMOT-niche-DE correlation baseline, all under matched data and tuning. Decisive check: does the gate $\pi$ concentrate on specific gene–niche pairs rather than staying near zero or saturating at one? Experiment 3 (the translation test): predict which attributed events reproduce in 2D monoculture, using CellVS-Net-style perturbation matching Cell-Level Virtual Screening; no reported result exists on Lacuna for predicting per-event in vitro reproducibility of endogenous silencing, so we construct the evaluation from documented cases — ME3, whose knockdown consequences were reproduced in AGS and MGC803 monocultures even though the trigger is unknown, and RACK1, whose repression is tied to inflammatory microenvironmental signaling ME3 study; RACK1 study. Experiment 4 (prospective closure of the loop): for one niche-attributed event, test reproduction in 2D monoculture versus 3D acinar explant co-culture with the predicted sender cell type — the RACK1 system already has the 3D explant and caerulein-pancreatitis infrastructure in place, and low RACK1 correlates with poorer survival at $p = 0.037$, so the clinical readout exists RACK1 study. Success across all four experiments means the field gains, for the first time, a per-event answer to "why did this guardian gene turn off" — and a principled rule for when a 2D culture was never going to show us the answer.