RT @razoralign: Bonsai: Tree representations for distortion-free visualization and exploratory analysis of single-cell omics data https://t….

RT @DucheneJohan: Popular methods like UMAP & t-SNE are stochastic and distort data structure. Bonsai - a novel method - builds trees to….

RT @RoganAGrant: Radical approach to the UMAP debate: offering an actual alternative.

Or play with the visualizations of example datasets. A full Bonsai-scout tutorial is available at: . Single-cell omics data are amazing and deserve more reliable tools for visualization and exploratory analysis. We hope to have made a difference here!.

Remarkably, famous players show up as outliers on the Bonsai tree, e.g. Messi is on the longest branch!. Try it on your own data! Our webserver allows you to upload UMI count tables after which all analysis is performed automatically.

Bonsai should be applicable to any high-dimensional data with continuous features. To show this, we for fun applied it to statistics of professional football players. Bonsai finds clusters of players that indeed reflect groups of players with different roles and marker features.

It was known that NK cells can be derived in vitro from myeloid precursors, but now Bonsai discovers that in vivo there are indeed both lymphoid and myeloid derived NK cells. Bonsai also pinpoints the marker genes that distinguish myeloid-NK from lymphoid-NK cells.

But just inspecting the tree we immediately find novel predictions as well. For example, Bonsai finds NK cells divide into two groups: one deriving from the lymphoid lineage and one from the myeloid lineage. Indeed both groups express known NK-cell markers.

So what do we see on real data? To test Bonsai we applied it to a dataset of blood cells for which quite some lineage relationships are known. We find that the Bonsai tree automatically recovers not only the known cell types, but also the known lineage relationships between them.

Bonsai-scout allows different layouts, zooming in.and out on areas, visualizing gene expression or annotations on the tree, finding well-separated clades, and finding marker genes that distinguish any pair of clades. Check out our tutorial videos!.

@daandegroot91 made great efforts to make the inference computationally efficient so that we can analyze datasets with >100'000 cells. Moreover, we developed an app, Bonsai-scout, that not only visualizes the trees but also allows interactive exploratory analysis of results.

Now we find that Bonsai's tree representation vastly outperforms Sanity on kNN identication. Forcing the cells into a tree structure automatically regularizes the inference of gene expression states.

We also find that Bonsai automatically regularizes the measurent noise in scRNA-seq data. In our previous Sanity paper (10.1038/s41587-021-00875-x) we showed that finding true nearest-neighbor cells is actually very hard, and developed a specific method that outperforms others.

Here is a Bonsai visualization for a realistic dataset. In contrast to UMAP or PCA, Bonsai accurately reflects the structure in the data. Not only visually, but the correlations between the true distances and those in the Bonsai representation are close to 1 for almost all cells.

This holds both for random data and also for scRNA-seq data. Moreover, for cells from a single organism, we KNOW that gen expression has actually diverged along the branches of a tree, so trees are in fact the most natural way of representing lineage relationships between cells.

Why does this work? You may have heard of the curse of dimensionality, but we discovered that there is also a blessing of dimensionality: As the dimensionality of the space becomes large, distances between sets of points can GENERICALLY be well approximated by a tree structure.

So what does Bonsai do? Given estimated positions of objects in a high-dimensional space, with separate error-bars for each feature of each object, Bonsai reconstructs an optimal tree representation so that the distances along the branches reflect the true distances.

Some of you have heard me rant about t-SNE/UMAP. ( And I always have to hear how there is nothing better. Well, now there is! Bonsai allows visualization and exploratory analysis of scRNA-seq data WITHOUT distortion of the structure in the data.

Here it is! Bonsai. No more excuse to use t-SNE/UMAP. Bonsai not only makes cool pictures of your data. It actually rigorously preserves its structure. No tunable parameters. Absolutely incredible work of @dhdegroot.bsky.social. I'm so excited about this.

RT @biozentrum: The slower cells grow, the more sensitively they respond to their environment!.A new study by @NimwegenLab reveals that the….