# Detecting spatially variable genes

Genes whose expression follows the layout of a tissue, rather than varying at random from place to place, point to spatial domains, gradients and signalling. This benchmark asks a method to rank genes by how spatially variable they are, on realistic simulated data where the true degree of spatial variability of every gene is known.

## Task

A method receives a spatial transcriptomics dataset: a count matrix with the spatial coordinates of each spot or cell. It must return a spatial-variability score for every gene. Each dataset contains 21 copies of every source gene, generated with known spatial variability from 0 (none) to 1 (fully spatial) in steps of 0.05, and the score is judged on how well it orders these copies.

## Background

Many methods for finding spatially variable genes (SVGs) exist, from Moran's I to Gaussian-process models such as SpatialDE and nnSVG and non-parametric tests such as SPARK-X, but they had rarely been compared on data with a known answer. Li et al. built such data with scDesign3 and benchmarked 14 methods; their Genome Biology paper (2025) found that SPARK-X performed best overall, that Moran's I was a strong baseline, and that most methods were poorly calibrated. The Open Problems task, set up by Zhijian Li, Zain M. Patel, Dongyuan Song, Guanao Yan, Jingyi Jessica Li and Luca Pinello, keeps this comparison open to new methods.

## Data and evaluation

- **Data:** 50 simulated datasets derived from public spatial data (Visium, Slide-seqV2, MERFISH, seqFISH, STARmap, Stereo-seq, DBiT-seq and Slide-tags). For each real dataset, scDesign3 fits a Gaussian-process spatial model to the top 50 variable genes; a non-spatial version is made by shuffling the fitted means across locations, and each simulated gene mixes the two with weight $\alpha \in \{0, 0.05, \dots, 1\}$, which becomes its true score. Files are in `s3://openproblems-data/resources/task_spatially_variable_genes/datasets/`. The task code is MIT-licensed.
- **Split:** there is no train/test split. The simulation uses seed 2024.
- **Metric:** for each source gene, Kendall's $\tau$ between the predicted scores of its 21 copies and their true $\alpha$, averaged over genes. Per dataset this is rescaled between the random-ranking control (0) and the true-ranking control (1), clipped to $[0, 1]$, and the overall score is the mean over the 50 datasets, with failed runs counted as 0. The leaderboard shows it to two decimals.
- **Code:** [openproblems-bio/task_spatially_variable_genes](https://github.com/openproblems-bio/task_spatially_variable_genes), release v1.0.0; the original study's code is at [pinellolab/SVG_Benchmarking](https://github.com/pinellolab/SVG_Benchmarking).

## What is known

On the v1.0.0 leaderboard (as of 2026-10-20):

- SPARK-X leads at 0.89 (0.8869 before rounding, recomputed from the published results).
- Spanve (0.79), nnSVG (0.78) and Moran's I (0.78) follow; SpatialDE2 scores 0.67 and SpatialDE 0.62.
- Some methods are far behind, for example SpaGCN at 0.13 and BOOST-GP at 0.01.

## What would count as progress

- An overall score above 0.8869 on the v1.0.0 datasets, computed with the task's `run_benchmark` workflow, with code that reproduces it.
- A method that beats SPARK-X on each technology family (sequencing-based spots and imaging-based single cells), not only on average.
- A gain that does not come at a large cost in run time: SPARK-X is also the fastest method on the leaderboard, at about 9 seconds per dataset on average, and the leaderboard reports time and memory alongside scores.

## References

1. Li Z, Patel ZM, Song D, et al. Systematic benchmarking of computational methods to identify spatially variable genes. *Genome Biology* 26, 285 (2025). [doi:10.1186/s13059-025-03731-2](https://doi.org/10.1186/s13059-025-03731-2)
2. Zhu J, Sun S, Zhou X. SPARK-X: non-parametric modeling enables scalable and robust detection of spatial expression patterns for large spatial transcriptomic studies. *Genome Biology* 22, 184 (2021). [doi:10.1186/s13059-021-02404-0](https://doi.org/10.1186/s13059-021-02404-0)
3. Song D, Wang Q, Yan G, et al. scDesign3 generates realistic in silico data for multimodal single-cell and spatial omics. *Nature Biotechnology* 42, 247–252 (2024). [doi:10.1038/s41587-023-01772-1](https://doi.org/10.1038/s41587-023-01772-1)
4. Luecken MD, Gigante S, Burkhardt DB, Cannoodt R, et al. Defining and benchmarking open problems in single-cell analysis. *Nature Biotechnology* 43, 1035–1040 (2025). [doi:10.1038/s41587-025-02694-w](https://doi.org/10.1038/s41587-025-02694-w)

## Source

Imported from [Open Problems in Single-Cell Analysis: Spatially Variable Genes (task_spatially_variable_genes)](https://openproblems.bio/benchmarks/spatially_variable_genes), curated by the Open Problems in Single-Cell Analysis consortium. Open according to the source as of 2026-10-20 (leaderboard release v1.0.0). The statement above is written in our own words; see the source for the original wording, history and discussion.