Two-dimensional embeddings for single-cell visualisation

Almost every single-cell study shows its cells as a two-dimensional map made with t-SNE, UMAP or a relative of them. This benchmark asks how faithful such maps are, scoring 2D embeddings on how well they keep the neighbourhoods, distances and densities of the original high-dimensional data.

Task

A method receives a single-cell RNA-seq dataset with thousands of genes per cell. It must return a two-dimensional embedding: one pair of coordinates per cell. The embedding is compared with the original data by ten metrics of structure preservation.

Background

Single-cell datasets have tens of thousands of features and up to millions of cells, and 2D visualisations are central to how they are explored and presented. Popular methods such as t-SNE and UMAP emphasise local neighbourhoods and do not preserve the relative density of cell populations, which motivated densMAP, a UMAP variant with an extra density-preserving term (Narayan et al., 2021). The Open Problems task, by Luke Zappia, Michael Vinyard, Michal Klein, Scott Gigante, Ben DeMeo and Robrecht Cannoodt, compares methods under one set of quality measures.

Data and evaluation

• Data: four datasets in release v1.0.0: 5k peripheral blood mononuclear cells (10x Genomics), mouse haematopoietic stem and progenitor cells (Nestorowa et al., 2016), mouse myeloid differentiation (Olsson et al., 2016) and zebrafish embryos from the first day of development. They are downloaded by the task's dataset loaders from public sources.
• Split: none; the whole dataset is embedded and evaluated.
• Metric: ten metrics, all higher-is-better: trustworthiness and continuity; co-ranking measures (co-kNN size and its area under the curve, local continuity meta-criterion, and local and global properties); Spearman correlation of pairwise Euclidean distances and of diffusion distances; and density preservation. For each dataset and metric the raw value $x$ is rescaled between the worst and best control method (random 2D coordinates, a 1,000-dimensional diffusion map, and the original features), $s=\min\{1,\max\{0,(x-c_{\text{worst}})/(c_{\text{best}}-c_{\text{worst}})\}\}$, and the overall score is the mean of $s$ over the 4 × 10 pairs, with failed runs counted as 0. The leaderboard shows it to two decimals; methods with several parameter sets are listed by their best and worst set.
• Code: release v1.0.0 lives in the legacy openproblems repository; the task now continues in openproblems-bio/task_dimensionality_reduction.

What is known

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

• densMAP leads at 0.36 (0.3554 before rounding, recomputed from the published results) with its best parameter set, followed by densMAP on PCA input (0.31) and t-SNE (0.29).
• PyMDE, NeuralEE, UMAP, PHATE and PCA all fall between 0.25 and 0.28 with their best settings.
• Using the original features as the "embedding" (a control that is not two-dimensional) scores 0.76, which marks how much structure every 2D method loses.

What would count as progress

• An overall score above 0.3554 on the v1.0.0 datasets and metrics, computed with the release's evaluation code, with code that reproduces it.
• A method that improves the global measures (distance correlations, global co-ranking property) without losing the local ones (trustworthiness, co-kNN size), checked metric by metric on the leaderboard.
• The method contributed as a component of the maintained task_dimensionality_reduction repository, so that it can be rerun on the task's datasets.

References

1. Narayan A, Berger B, Cho H. Assessing single-cell transcriptomic variability through density-preserving data visualization. Nature Biotechnology 39, 765–774 (2021). doi:10.1038/s41587-020-00801-7
2. van der Maaten L, Hinton G. Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579–2605 (2008). jmlr.org/papers/v9/vandermaaten08a
3. 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

Source

Imported from Open Problems in Single-Cell Analysis: Dimensionality Reduction for Visualisation (dimensionality_reduction), 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.