Finding Subclonal Point Mutations in Spatial Transcriptomics by Pooling Reads Over Space

Tumors grow as clones. A cell that acquires a mutation divides, and its descendants tend to stay near each other, so a subclonal point mutation should occupy a contiguous patch of tissue. Spatial transcriptomics measures gene expression at known positions in a tissue slice. Its sequencing-based versions also produce RNA reads that span a few hundred bases of each transcript, so in principle they carry evidence of point mutations. At any single spot, though, that evidence is almost nothing: a handful of reads at a given base, most of them from the normal allele. Tools that call mutations from single-cell data, such as Monopogen, get around this sparsity by pooling reads across cells in the same expression cluster. They were built for dissociated cells, which have no coordinates. Spatial tools that do detect alterations, such as SpaCNA, only see large copy-number changes over one-megabase bins, never single-base mutations.

This note proposes that the right unit for pooling mutation evidence in tissue is a spatially contiguous patch, not an expression cluster. The reason is that true subclonal mutations are spatially clustered, while sequencing and alignment errors are scattered at random. A caller that scans the slide for patches with too many variant reads should find subclonal mutations that are invisible both to per-spot calling and to whole-slide pooling. Once such mutations are placed on the map, a second, deliberately modest step asks which of them sit at the center of the gene network in their own tissue neighborhood. The network is built from simple correlations inside each patch, and genes are ranked with a centrality score that accounts for how much each gene's expression changes near the mutant patch.

Core Claim

The claim is that spatial scan pooling detects subclonal single-nucleotide variants (SNVs, changes at a single DNA base) from sequencing-based spatial transcriptomics with better sensitivity than expression-cluster pooling, at matched false-discovery rate. The advantage should grow as the clone occupies a smaller fraction of the slide.

The statistic is a likelihood ratio computed over candidate regions. For a genomic site and a connected set of spots $R$, let $a_R$ be the number of reads showing the alternative base and $d_R$ the total number of reads covering the site. Let $\varepsilon$ be the site's background error rate, estimated from the whole slide, and let $\hat f_R = a_R/d_R$ be the local variant fraction. The score is

$$ S(R) = a_R \log\frac{\hat f_R}{\varepsilon} + (d_R - a_R)\log\frac{1-\hat f_R}{1-\varepsilon}, $$

maximized over regions $R$ grown on the spot-adjacency graph. Significance comes from permuting the variant reads across spots while preserving each spot's depth. This asks directly whether the variant reads are more spatially concentrated than depth alone explains. The form is the standard Kulldorff-style scan statistic for a binomial count. It is used because it has known behavior under multiple testing over regions.

The prior work splits into three roles. Monopogen supplies the pooling logic: sparse reads become callable once aggregated. It also supplies the use of population panels to recognize inherited variants. We replace its expression-defined pools with spatial ones. SpaCNA supplies the spot-adjacency graph and the evidence that subclones are spatially segregated in real tumors. That evidence motivates the approach, and we extend it from copy-number segments to single bases. MVP scores missense variants for pathogenicity from sequence and protein features alone. We keep it only as a context-free prior for ranking, not as something we try to improve.

Research Program

The caller handles platforms that yield reads: Visium, VisiumHD and Open-ST. Imaging platforms such as Xenium and MERFISH detect transcripts with probes, never read the sequence, and so fall outside the scope entirely. We state this plainly because it limits where the method applies.

Candidate sites come from pileups over expressed exons. The first filter removes inherited variants. Where a matched normal sample exists, sites present in it are removed. Otherwise sites are removed using linkage-disequilibrium panels from the 1000 Genomes Project, as in Monopogen. Two confounders remain, and both are spatially structured, which makes them the main threat to the idea:

  1. RNA editing is the cell's enzymatic conversion of A to G on RNA. It varies by cell type and therefore by region.
  2. Allele-specific expression means one parental copy of a gene is expressed more than the other. It can make an inherited variant look regionally enriched.

We handle them in three ways. We model A-to-G changes separately from other substitutions. We require the variant to be absent from the matched normal. And we report what fraction of calls would disappear without these filters.

The second stage is secondary and is validated on its own. For each called mutation, we build a gene co-expression network separately inside the mutant patch and inside a matched neighboring region. We use Pearson correlation and GRNBoost2 because geneRNIB found that these simple network-building methods beat foundation models and database-heavy pipelines at predicting the targets of perturbed genes. We then rank genes with Contextual Centrality. That score sums weighted random walks on a network, weighting each node by a contribution vector $y$. In the original work the network was social and $y$ encoded node payoffs. Here $y$ is each gene's expression shift between the mutant patch and its neighbor, so a mutated gene ranks highly only if its network neighborhood also changes.

This differs from the usual hub-gene analysis, such as the morusin aortic-dissection study. That study ranked genes by degree on a fixed STRING protein-interaction network and only afterwards looked up their expression in single-cell clusters. It also differs from Starfysh, whose "hubs" are tissue neighborhoods of mixed cell types rather than central genes in a network.

Baselines

The naive reference points are per-spot calling and whole-slide pseudobulk calling. Per-spot calling runs a binomial test at each spot with no pooling. Whole-slide pseudobulk calling sums all reads and applies the same likelihood with $R$ fixed to the entire slide. Per-spot calling bounds how much pooling helps at all. Pseudobulk is the practical default a lab would use.

The strongest competitor is Monopogen-style expression-cluster pooling. We cluster spots by expression, pool reads within each cluster, and apply the same variant model with identical error-rate estimation and inherited-variant filtering. Every method gets the same reads and filters. Every method is also tuned to the same empirical false-discovery rate on the semi-synthetic data, so differences reflect only the pooling unit.

There is no reported result on Lacuna for somatic SNV calling from spatial transcriptomics. The nearest reported number comes from a different task. Monopogen reports genotyping accuracy above 95% against matched whole-genome sequencing, and that figure is for inherited variants in dissociated cells. It is not a target we can compare to directly.

Beating this baseline means higher recall at 10% false-discovery rate on held-out real tumors with matched whole-exome truth. The gain must concentrate in mutations whose truth-set variant fraction implies a clone covering a small fraction of the tumor area.

Two ablations isolate the key ingredient. The first uses spatial pooling that ignores expression, taking contiguous regions from the adjacency graph alone. The second pools within the copy-number subclone regions that SpaCNA outputs. If the SpaCNA-region ablation matches the full scan, then scanning over regions adds nothing beyond reusing copy-number segmentation. That would be a genuinely useful negative result, and we would report the simpler method.

For the hub stage, the baselines are degree and betweenness centrality on the same patch-specific network and on STRING. A second ablation runs contextual centrality with $y$ set to all ones, which removes the expression-shift weighting.

Why not just…?

Why not just call mutations on the whole-slide pseudobulk and then find where the variant reads fall?

This is the real threat. For clonal mutations present in most tumor cells it will work about as well as our method, and we expect it to. It fails when a clone is small. Pooling everything dilutes a variant fraction of, say, 20% within a patch down to a few percent across the slide. That is near the error floor.

The semi-synthetic experiment settles the question by varying clone area. If pseudobulk plus post-hoc mapping matches the scan at every clone size, the contribution shrinks to a visualization step, and we would say so.

Why not just use Monopogen on expression clusters of spots?

Subclones within one cell type usually look alike in expression. SpaCNA found subclones separated by tissue boundaries rather than cleanly by cell-type identity. Expression clusters will therefore mix mutant and wild-type spots of the same type, diluting the signal.

This is precisely the head-to-head comparison in the Baselines section. There is one case where expression pooling should win: when a mutation drives a distinct expression state that is spread across the tissue. We will report this case separately rather than average it away.

Why not just microdissect regions and run whole-exome sequencing?

That is the gold standard where tissue allows. It requires choosing regions in advance, and it cannot find a clone nobody suspected was there. It also consumes tissue that archival spatial studies no longer have. We use such data as ground truth, not as a competitor.

Why not rank hub genes by degree on STRING, or use a foundation model?

Degree on a fixed database network gives the same ranking in every tissue neighborhood. It therefore cannot express the claim that a gene is central in this niche. geneRNIB also reports that database priors and foundation models such as scPRINT underperform simple correlation networks. We test contextual against structural centrality on perturbation ground truth before trusting either one in tissue.

Proposed Experiments

Semi-synthetic calibration. We take real Visium and Open-ST read files and implant variant reads at chosen sites. The implanted mutations form spatially contiguous clones of varying area, at variant fractions within the clone between 5% and 50%, in genes spanning a range of expression levels. As a negative control we use inherited heterozygous sites. After normal-sample filtering these should yield no calls, and before filtering they should show no spurious spatial concentration. This control checks that the permutation test is calibrated. The outputs are recall and false-discovery rate by clone area and depth.

Real tumors with exome truth. We use the liver, breast and colorectal cancer and head-and-neck squamous cell carcinoma samples from the SpaCNA evaluation, which have matched whole-exome sequencing. Their availability at the raw-read level has to be confirmed. Truth is somatic SNVs in expressed genes with adequate RNA coverage. We report precision and recall for every method in the Baselines section. We also report the fraction of false positives attributable to A-to-G changes, before and after editing-aware filtering.

Validating the hub score where truth exists. On the Norman, Replogle and Adamson perturbation sets used by geneRNIB, we compute contextual centrality on correlation networks from control cells, with $y$ given by each gene's expression response. We then ask whether centrality predicts the size of the downstream effect of perturbing that gene, better than degree, betweenness, or the $y$-equal-to-ones ablation. If it does not, the hub stage is dropped from the tissue analysis.

Decisive check. The proposal stands if the spatial scan achieves higher recall than expression-cluster pooling at matched false-discovery rate for small clones on real tumors. It must also not be matched by pooling over SpaCNA regions. If either comparison fails, the finding is that existing pooling units suffice, and the mutation-mapping idea reduces to applying Monopogen to spot clusters.