Leakage-Aware Spatial Granger Tests for Directed Cell–Cell Signaling

Spatial transcriptomics measures which genes are switched on at known positions in a slice of tissue. Biologists want to use it to answer a directional question: does a signal sent by one cell change what its neighbor does? Granger causality offers a natural test. Signal A "Granger-causes" B if A's past improves the prediction of B's future beyond what B's own past already provides. Applying it to tissue runs into two obstacles. The first is that each tissue slice is a single killed snapshot, so there is no "past" to condition on. The second is that transcripts physically bleed between neighboring capture spots or cells. That bleed is the spatial counterpart of volume conduction in brain recordings, where electrical signal spreads instantly across sensors, and in that setting it is known to fabricate most of the apparent directed connectivity. Lalla et al. attributed roughly 90% of cortical–striatal theta interdependence to a volume-conducted hippocampal source.

We propose a Granger test that runs inside one tissue slice. It handles the first obstacle by taking the "past" from within each cell's own RNA. Unspliced pre-mRNA reports transcription happening now. Spliced mRNA reports expression built up earlier. Our test asks whether a neighbor's spliced ligand predicts a receiver's unspliced target transcripts beyond what the receiver's own spliced state predicts. It handles the second obstacle by modelling transcript bleed explicitly. The receiver is conditioned on the sender's own copy of the target gene, with a bleed kernel estimated from cell-type-exclusive markers. Direction is then scored as the difference between the forward and reversed tests, which cancels bleed because bleed is symmetric. The deliverable is a directed sender→receiver graph with a calibrated false-positive rate under realistic bleed. No current spatial method provides this.

Core Claim

Our central claim concerns false directed edges. In sequencing-based spatial data, most edges that a naive directed-signaling analysis calls are bleed artifacts. A conditional, splicing-lagged spatial Granger test with an antisymmetric direction score removes them while keeping real paracrine edges. Real edges are the ones that disappear when the receiver's receptor is knocked out.

The claim has two separable ingredients:

  1. Bleed conditioning plus the antisymmetric score. This is the primary contribution.
  2. The splicing-defined lag. We treat this as a hypothesis to be tested by ablation, not assumed.

If the lag adds nothing over lag-free conditional regression, the contribution reduces to the bleed-control machinery, and we will report it that way.

Relation to Prior Work

Three methods motivate the gap by stopping short of directed, lagged edges between physical neighbors.

GEASS. GEASS runs a transfer-entropy feature selector over a spatial diffusion operator on MERFISH data. It returns a sparse set of causally relevant genes, but no sender→receiver edges and no lags.

Celcomen. Celcomen separates within-cell from between-cell gene interactions with identifiability guarantees. However, it explicitly cannot orient A→B versus B→A, and it assumes static equilibrium.

spVelo. spVelo does model spliced and unspliced counts over a spatial neighbor graph. Its communication analysis, though, correlates instantaneous velocity with ligand–receptor expression. It runs no predictive test and does not model bleed.

A second group supplies the method we reuse and adapt.

GrID-Net. GrID-Net generalized Granger tests to pseudotime-ordered graphs using full-versus-reduced models and an F-test. It reports up to 71% higher eQTL agreement than correlation. It works only on dissociated cells ordered along a developmental axis, never on physical neighbors.

TimeTalk. TimeTalk applied VAR-style Granger testing to ligand–receptor "talkers" and transcription-factor "listeners" along embryonic pseudotime, again without spatial coordinates.

GVAR. The self-explaining nonlinear model of GVAR gives us a nonlinear variant that also reports the sign of each effect.

A third group constrains the design.

Lalla et al. Their volume-conduction result is the reason we do not trust raw directed scores between neighbors.

FlowSig. FlowSig orients inflow → gene module → outflow chains through conditional independence on static snapshots. It shows that direction can be imposed by structural priors. It also shows the limit of that approach: without a lag, the orientation comes from the prior, not from the data.

COMMOT. COMMOT models competition among ligand–receptor pairs. It warns that isolated pair scores overestimate signaling, so our sender term is built from COMMOT-style competition-adjusted ligand availability, not raw products.

Research Program

Take a receiver cell or spot $i$ and a candidate target gene $g$. Let $u_{i,g}$ be its unspliced count, $s_i$ its spliced expression vector, and $w_i$ its deconvolved cell-type composition (from stereoscope on Visium-like data). The full model is a negative-binomial regression:

$$ \log \mathbb{E}[u_{i,g}] = f(s_i, w_i) + \beta \sum_j K(d_{ij})\, a_{j\ell}\, s_{i,r} + \sum_j L(d_{ij})\,\big(\gamma_u u_{j,g} + \gamma_s s_{j,g}\big) $$

The terms are as follows:

  • $f$ is the receiver's own-state predictor. It is linear in the main version and a GVAR-style network in the nonlinear version.
  • $a_{j\ell}$ is the competition-adjusted availability of ligand $\ell$ from neighbor $j$.
  • $s_{i,r}$ is the receiver's spliced receptor $r$.
  • $K$ is a signaling-range kernel over distance $d_{ij}$.
  • $L$ is a bleed kernel.
  • The final sum conditions on the neighbor's own copy of the target gene. This is the spatial analogue of conditioning on the volume-conducted source.

The reduced model drops the $\beta$ term, and the Granger statistic is the likelihood ratio between the two.

We estimate the bleed kernel $L$ separately rather than letting it absorb signal. We pick genes that are exclusive to one cell type, for example epithelial markers in immune-rich regions. Their detection in neighbors of the wrong type, as a function of distance, gives an empirical bleed profile for each platform and slice.

Direction is scored antisymmetrically. We compute the forward statistic for "$j$'s ligand → $i$'s target" and the reversed statistic with sender and receiver roles swapped for the same gene pair. Bleed inflates both by the same amount; paracrine signaling inflates only one. We call an edge only when the forward-minus-reverse difference exceeds a null built from spatially rotated coordinate permutations, which keep each slice's autocorrelation.

We will build the following:

  1. A tested package for the procedure above.
  2. A simulator of tissue sheets. Ground-truth directed signaling drives transcription with a known delay. Spliced and unspliced kinetics are explicit. Bleed is tunable, and so are dropouts.
  3. Analyses on real data with interventional ground truth.

Baselines

Do-nothing reference. Spatial ligand × receptor product scoring between neighbors within a distance threshold. Direction is assigned by which side expresses the ligand. This is what most practitioners do. It shows how much of any directed graph comes for free from expression patterns alone.

COMMOT. COMMOT is the strongest competition-aware spatial signaling method among our inputs. It is run with its default distance limits on the same gene panel and slices.

spVelo. spVelo uses the same spliced/unspliced information we do, so it is the closest competitor. If its velocity–ligand correlation matches us, our test adds nothing. It is run on identical count matrices, with its edge scores thresholded to the same number of calls as ours.

Celcomen. Celcomen is the strongest method with an interventional benchmark. On the Perturb-map data it reports a correlation between predicted and actual knockout effects of $p=0.0079$. It reports no edge-level precision or direction accuracy; no reported result on Lacuna for directional accuracy of Celcomen.

GrID-Net adapted. A GrID-Net-style pseudotime Granger test that ignores coordinates. On synthetic tissue with known delays and bleed, there is no reported result on Lacuna for any of these methods. The GEASS F1 of 0.91 is from a non-spatial time-series benchmark and is not comparable.

Shared conditions. All methods get the same genes, the same neighbor definitions where applicable, and an equal hyperparameter search budget (20 configurations each) on a held-out simulated tissue.

Ablations. Two ablations isolate our ingredients:

  1. Removing the bleed term and the antisymmetric score tests the primary claim.
  2. Replacing the unspliced target with spliced expression removes the lag and tests whether the lag matters.

What counts as beating the baselines.

  • On simulation: a higher area under the precision–recall curve for directed edges, and a false-positive rate on bleed-only edges held near the nominal level as bleed grows. The naive baselines' false-positive rate should climb with bleed.
  • On Perturb-map: a significant loss of called edges into receptor-knockout lesions with no loss out of them, at a stronger significance level than Celcomen's reported correlation.

Why not just…?

Why not just use ligand–receptor co-localization scores or COMMOT? These measure whether signaling is possible, not whether it changed the receiver. FlowSig and spVelo both show that co-expression does not establish downstream activation. More importantly here, bleed inflates co-localization directly. The bleed-sweep simulation will show whether their false-positive rate grows with bleed while ours stays flat. If it does not grow, bleed is not the problem we claim and the proposal weakens.

Why not just regress the receiver's target on neighbor ligand while conditioning on the receiver's state, with no splicing lag? This is the most serious threat. It is close to what Celcomen-style conditional models already do, and it may capture most of the benefit. We test it directly with the spliced-only ablation. If the lag adds nothing on Perturb-map direction accuracy, we drop the "Granger" framing and present the method as bleed-controlled conditional spatial regression with an antisymmetric direction score. The bleed modelling is the part we expect to survive either way.

Why not just run Granger tests along pseudotime, GrID-Net or TimeTalk style? Pseudotime orders a cell's developmental progress, not its neighbors. STORIES showed that dropping spatial constraints yields anatomically impossible transitions, and a non-spatial test cannot tell paracrine from cell-intrinsic drivers. The adapted GrID-Net baseline will show this concretely. On Perturb-map it should fail to localize edge loss to knockout lesions.

Why not just correlate spVelo velocities with neighbor ligand expression? That correlation shares the bleed confound: a neighbor's leaked unspliced target transcripts inflate the receiver's velocity. It also has no reduced model to compare against. If spVelo matches us after we hand it our bleed covariates, the contribution is the bleed correction, not the test. That is still a useful finding, and we will report it as such.

Proposed Experiments

Simulation. We generate 2D tissue sheets of 2,000–10,000 cells with five cell types and 20 directed ligand→target programs. Signaling acts with a delay longer than splicing time. We include feedback loops, since Celcomen's equilibrium assumption breaks there, and niche-restricted programs.

We sweep three parameters: bleed from none to strong, unspliced capture rate, and dropout. We report:

  • directed precision–recall;
  • false-positive rate on bleed-only gene pairs;
  • direction accuracy on bidirectional loops.

Perturb-map. Perturb-map mouse lung cancer is the decisive real-data test. Lesions carrying Jak2 or Tgfbr2 knockouts in tumor cells should lose incoming interferon-γ and TGF-β edges respectively. Edges sent from those lesions should be preserved. A method that is merely detecting co-expression or bleed will show symmetric or no change.

The main risk is that Visium-style unspliced counts are too sparse. We will measure intronic read fraction before committing. If it is insufficient, the lagged variant is evaluated only on Stereo-seq.

Stereo-seq atlases. On the Stereo-seq mouse organogenesis atlas used by STORIES, we run a cross-check that needs no knockout. We align consecutive time points with fused Gromov-Wasserstein transport, a method that matches cells across slices while preserving their spatial layout. Edges inferred within a single slice from splicing should then predict the direction of expression change between that slice and the next aligned slice. Agreement substantially above rotated-coordinate nulls would show that the within-slice lag carries real temporal information. Agreement at null would show that it does not, and would settle the lag question independently of Perturb-map.