Plant biologists can now reliably find genes that make a crop tolerate salt. What they usually cannot say is how those genes do it. Silencing an annexin gene in cotton makes leaves wilt under salt, and overexpressing a chalcone-isomerase-like gene in soybean roots switches on sodium-transport and antioxidant genes. The step-by-step path from the gene to the protected tissue remains a black box. A key unanswered question is spatial. Does the gene act only inside the cells that express it, or does it change distant cells, for example the inner vascular core, through a mobile signal?
We propose to answer this with spatial transcriptomics, which measures gene expression at tens of thousands of positions across a tissue section. The model is a selective state-space model, a sequence model in the Mamba family that summarizes everything it has seen so far in a running memory and costs time linear in the number of positions. The model reads a root section along the root's own biological axes: tip to base, which roughly follows cell age, and outside to inside, which follows tissue layers. It does not use arbitrary pixel rasters. We then use the trained model to predict what happens when a candidate gene is switched off, and we test those predictions against real, spatially patchy gene silencing.
Core Claim
Two claims are testable, and the whole proposal stands or falls on them.
The first claim is about architecture. Read along the root's own coordinates, a linear-time model with a whole-section view predicts how silencing a stress gene changes expression in cells far from the silenced cells. It does so better than models that only see locally: a gated convolutional network, or a graph network that passes messages a few neighbors away. Any advantage should grow with distance from the perturbation.
The second claim is biological. The model's "virtual silencing" separates candidate downstream targets that respond inside the silenced cells from targets that respond in distant tissue layers. That gives each target a concrete mechanistic hypothesis: act-in-place or mobile-signal.
If a local model matches the global one at every distance, the correct conclusion is that salt signaling in these sections is local, and the state-space machinery is unnecessary. We treat that as a legitimate outcome, not a failure to explain away.
Why This Intersection, and What Already Exists
The motivating evidence comes from single-gene plant stress studies. Each identifies an effector and then stops at the mechanism.
In cotton, GhANN4 and GhANN11 were validated by virus-induced gene silencing (VIGS), which uses an engineered plant virus to knock down a chosen gene. Silencing GhANN4 compromised salt and drought tolerance. The authors say explicitly that the interacting partners and signaling cascade are unknown.
PgGRAS24 in pomegranate improves salt tolerance when overexpressed in Arabidopsis, but its direct targets are uncharacterized.
In soybean, GmCHI4A and GmCHI4B are root-predominant. Their overexpression in hairy roots upregulates GmSOS1, GmNHX1, GmSOD1 and GmAPX1 and raises isoflavones about 5.1-fold. Isoflavones are metabolites that can move between cells, so a non-cell-autonomous route is plausible but untested.
The SiNHX family likewise shows root-predominant expression. All of these studies used bulk tissue, which averages away the differences between root zones where sodium exclusion, vacuolar storage and osmolyte build-up actually happen.
Machine learning has reached plants mainly as image analysis. StateSpace-SSL trains a Vision Mamba encoder on leaf photographs for disease classification. CountMamba counts crop organs using four raster scan directions fused with a convolutional branch. Both stop at a visible phenotype label and never touch gene expression.
Our method draws on a separate set of computational work:
- Celcomen is the closest existing method. It is a generative graph network over spatial transcriptomics that separates within-cell from cell-to-cell gene interactions and predicts knockout effects in mouse lung tumors. Its reach, however, is limited to a few graph hops, and it has never been applied to plants.
- VORTEX predicts 3D expression from imaging with Transformers. It models correlation, not regulation.
- KINDLE and UGRN infer gene regulatory networks from dissociated single cells, so they have no spatial information at all.
Our proposal differs from Celcomen in three ways: a whole-section receptive field at linear cost, scan orders tied to root anatomy, and validation against a physically mosaic perturbation in a plant.
Three results constrain the design:
- MambaOut shows that gated convolutions without any state-space layer beat Visual Mamba on short sequences, reaching 84.1% ImageNet top-1 for MambaOut-Small. Mamba overtook them only on dense tasks with roughly 4,000 or more tokens. We therefore commit to whole-section inputs with tens of thousands of positions, and we carry a gated convolutional network as a mandatory baseline.
- geneRNIB found that correlation and GRNBoost2 outperformed foundation models and multiomic pipelines for regulatory network inference.
- A PerturbCausal study found that no deep virtual-cell model significantly beat a sparse ElasticNet at recovering causal structure, and that deep models inflated perturbation effect sizes about 23-fold.
These last two results mean our model must beat sparse linear models and report calibration, not just ranking.
Research Program
Data. The empirical core is spatial transcriptomics of longitudinal cotton root sections that include the tip, elongation and maturation zones. We collect sections under control and NaCl conditions from three groups of plants: empty-vector VIGS plants, GhANN4-VIGS plants, and unsilenced plants. We choose VIGS because its silencing is spatially patchy. Within one section, some cells lose GhANN4 and their neighbors do not. Each spot can be scored for GhANN4 transcript and for viral vector transcript, so the section itself records where the perturbation landed. That makes it possible to ask how expression changes as a function of distance from silenced cells.
A second system provides a held-out check with independent ground truth: soybean hairy roots overexpressing GmCHI4A under 100–150 mmol/L NaCl. For this system the published marker set (GmSOS1, GmNHX1, GmSOD1, GmAPX1) gives partial ground truth.
No plant spatial perturbation benchmark appears in the literature we have, so these data must be generated. That is the largest cost of the project and its main contribution beyond the model.
Model. Each spot is assigned root-intrinsic coordinates:
- a longitudinal coordinate, measured as distance from the root tip;
- a radial coordinate, given as a tissue-layer index (epidermis, cortex, endodermis, pericycle, stele) inferred from marker genes.
Spots are then ordered into four scans: tip-to-base, base-to-tip, outer-to-inner and inner-to-outer. For radial scans, ties are broken by longitudinal position, and vice versa.
The contrast with raster scanning is deliberate. The known weakness of state-space models on images is that a one-directional recurrence imposes an order the image does not have. A root does have an order. Development runs from the tip upward, and water and ions cross the root from the outer layers inward. The scan order therefore encodes plausible directions of influence rather than an artifact to be averaged away.
Architecturally, we reuse the multi-directional scan groups of CountMamba and the idea of fusing a global state-space branch with a local convolutional branch. The model is trained by masked-expression reconstruction. We hide the expression of a random subset of genes at a random subset of spots, and the model must predict them from everything else in the section. Condition (salt or control) is supplied as an input token.
Virtual silencing. For in-silico perturbation, we set GhANN4 input to its silenced level inside a chosen patch and record the predicted change in every other gene at every other spot. Each downstream gene then receives two readouts: its response inside the patch and its response as a function of distance and tissue layer outside it.
To guard against the effect-size inflation documented for virtual-cell models, the model output is converted into effect sizes that are checked against observed effects in held-out real mosaic sections. We report the ratio of predicted to observed effect size directly.
Baselines
Naive reference. The do-nothing reference is pseudo-bulk differential expression. We sum all spots per section and compare GhANN4-VIGS with empty-vector sections under salt. This is what the existing cotton study effectively did with qPCR. It yields a target list with no spatial attribution. Any spatial model must at least recover its top targets, and must add correct distance-resolved predictions to count as progress.
Sparse linear with spatial lags. This is the strongest simple competitor. It is an ElasticNet predicting each gene at each spot from all genes at that spot plus distance-binned neighborhood averages along the same root coordinates. It receives the same inputs, splits and hyperparameter search budget as our model. The PerturbCausal study reports only that "no deep VCM significantly outperformed a sparse linear comparator"; no numeric F1 is given in our file, and no reported result found on Lacuna for ElasticNet on plant spatial perturbation data. Beating it means lower held-out error on real mosaic sections at distances beyond the nearest neighbors, together with an effect-size ratio closer to one.
Strongest existing method. This is Celcomen, run on the same sections with its k-hop spatial graph, with k tuned over the same compute budget. Its reported evidence is a knockout correlation of p=0.0079 on mouse Perturb-map, running 500 genes across 2,500 cells in 11 seconds. There is no reported result found on Lacuna for Celcomen on plant tissue. Beating it means better distance-stratified prediction of real silencing effects, especially across tissue layers from the silenced patch.
Local-only control. The MambaOut-style gated convolutional network uses the same depth and width as our model, with the state-space layers removed and 7×7 depthwise kernels. It tests whether a whole-section view is needed at all. For context, its reported ImageNet number is 84.1% top-1 for MambaOut-Small, which says nothing about tissue.
Key-ingredient ablation. The ablation isolates our specific contribution by comparing two versions of the same state-space model: one scanned along root coordinates, and one scanned along CountMamba's four raster directions. If the raster version performs equally well, anatomy-aware scanning adds nothing, and the contribution reduces to applying existing vision state-space models to a new data type.
Why not just…?
Why not just use correlation or a sparse linear model, given that these win on existing regulatory-network benchmarks? This is the most serious threat. The evidence from geneRNIB and the PerturbCausal study is that simple methods match or beat deep ones on dissociated-cell data. Those benchmarks, however, contain no spatial structure, and our question is specifically about effects at a distance. We give the linear model spatial-lag features, so it can use distance too. If it matches our model at every distance bin, we will report that and use it for the biology.
Why not just use a gated convolutional network, since MambaOut showed Mamba is often unnecessary? MambaOut itself found that the state-space model helps only on long, dense inputs. Whole root sections are in that regime, but the real question is whether the relevant biology is long-range. The distance-stratified comparison settles this. The global model's advantage should be near zero close to silenced patches and grow with distance. A flat difference, or none, would argue for the convolutional model.
Why not just deepen Celcomen's graph to more hops? Deeper message passing is the natural fix, but it tends to over-smooth, blurring signals across many hops. More importantly, it offers no ordered notion of "upstream in development" or "outer to inner." We tune k generously, and the comparison across tissue layers will show whether depth closes the gap.
Why not just run a Transformer or a single-cell foundation model? Full attention over tens of thousands of spots is quadratic in cost. StateSpace-SSL reports 8.7 GB memory and 9 hours of training against 17.8–22.6 GB for its Transformer competitors. Existing foundation models also use mammalian gene vocabularies, and UGRN reports a baseline AUPRC of only 0.584 without extra machinery. We therefore exclude these for cost and vocabulary reasons, not on performance grounds.
Decisive Checks
The first decisive test is distance-stratified prediction on held-out GhANN4-VIGS sections. We measure per-gene Pearson correlation between predicted and observed silencing effects, binned by distance and by tissue-layer separation from silenced spots. We also report the predicted-to-observed effect-size ratio. Claim one holds only if the root-axis state-space model beats the gated convolutional network, spatial-lag ElasticNet and Celcomen in the far bins, and does so beyond seed-to-seed variation.
The second test is transfer to soybean GmCHI4A hairy roots. The model is fine-tuned only on control sections and then asked, without seeing any overexpression data, whether overexpression raises GmSOS1, GmNHX1, GmSOD1 and GmAPX1, and in which zones.
The third test is small but necessary wet-lab validation. We take the five genes the model most confidently calls "responds at a distance" and the five it calls "responds in place." For each, we measure expression in dissected outer versus inner tissue of VIGS roots. Claim two stands if the model's in-place versus at-a-distance calls agree with what the dissected tissue shows. If the calls disagree, the model's mechanistic attributions are not trustworthy, whatever its prediction scores.
References
- Genome-Wide Identification of the GhANN Gene Family and Functional Validation of GhANN11 and GhANN4 under Abiotic Stress
- Comprehensive Genome-Wide Characterization of the GRAS Gene Family and Their Role in Salt Stress Tolerance in Punica granatum L.
- Overexpression of Chalcone Isomerase-like Genes, GmCHI4A and GmCHI4B, Enhances Salt Tolerance of Cotyledon Hairy Roots and Composite Plant in Soybean (Glycine max (L.) Merr.)
- Genome-Wide Identification of the SiNHX Gene Family in Foxtail Millet (Setaria Italica) and Functional Characterization of SiNHX7 in Arabidopsis
- StateSpace-SSL: Linear-Time Self-supervised Learning for Plant Disease Detection
- CountMamba: Exploring Multi-directional Selective State-Space Models for Plant Counting
- Estimation of single-cell and tissue perturbation effect in spatial transcriptomics via Spatial Causal Disentanglement
- AI-driven 3D Spatial Transcriptomics
- KINDLE: Knowledge-Guided Distillation for Prior-Free Gene Regulatory Network Inference
- Towards Universal Gene Regulatory Network Inference: Unlocking Generalizable Regulatory Knowledge in Single-cell Foundation Models
- MambaOut: Do We Really Need Mamba for Vision?
- geneRNIB: a living benchmark for gene regulatory network inference
- Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery