Skip to content

The example dataset

There is a public project you can open right now, without building anything:

uvx bigclust2@latest --from https://flyem.mrc-lmb.cam.ac.uk/flyconnectome/bigclust_data/examples/MaleCNS_FlyWire_hemibrain_central_brain_bigclust

That is the whole setup. No account, no credentials, no download step — BigClust streams the project over HTTP. First load pulls about 63 MB and takes a few seconds on a decent connection.

You can also paste the URL into File → Open Project (Cmd+O) if you already have the app running.

What it is

A co-clustering of central brain neurons from three different connectomes, laid out by connectivity:

Dataset Neurons
FlyWire 32,384
MaleCNS 32,164
Hemibrain 22,715
Total 87,263

The point of a co-clustering is that neurons from different brains are embedded in one space, so the same cell type from three specimens should land in the same place. Whether it does is exactly what the app is for.

The project ships:

  • An embedding — connectivity (UMAP, cosine).
  • A feature matrix — 87,263 × 8,346 connectivity features, split into upstream and downstream groups.
  • Meta data — label, dataset, side, hemilineage, mapping and cn_frac_used, with 11,677 distinct labels.
  • 3D sources — a per-neuron Neuroglancer source per dataset, with FlyWire and Hemibrain registered into MaleCNS space, plus a brain outline mesh.

That last point is what makes the 3D viewer worth looking at here: select a cluster and you get neurons from all three connectomes rendered together, in a common space, without any transform step of your own.

Features, not distances

This project has no distances file — a full pairwise matrix for 87,263 neurons would be a 7.6-billion-cell array. Everything that needs all pairwise distances is therefore unavailable: the Distance Heatmap stays disabled.

Everything that works from features works fine, and there is more of it — you can change the metric, use feature subsets, and rank features between groups. See the data model.

Things to try

Colour by dataset. In the control panel's General tab, set Color by to dataset. This is the first question to ask of any co-clustering: if the three datasets separate into their own regions, the co-clustering has failed and you are looking at batch effect. If they intermingle within clusters, it worked.

Select a cluster and look at it. Shift + Ctrl + drag a lasso around a tight group. The 3D viewer fills with those neurons — from all three connectomes at once. Matching morphology across specimens is the visual confirmation that a cluster is a real cell type rather than an artefact of one dataset.

Colour by hemilineage. Developmental origin is independent of the connectivity that produced the layout, so agreement between the two is meaningful rather than circular.

Colour by side. Left and right homologues of the same cell type should sit together — a cluster that splits cleanly down the middle by side is usually telling you about asymmetric reconstruction rather than about biology.

Find a type you know. Use the Search box in the General tab, or turn on labels with L and shift + double-click one to select every neuron sharing it. The Kenyon cells are the largest groups here — KCg-m alone has 4,121 neurons.

Ask what separates two clusters. Select one group, Copy Selection into Group A of the Feature Comparison widget, then do the same for a neighbouring group into Group B. Because the features carry upstream and downstream levels, you can ask whether two types differ by what they listen to or by what they talk to.

Re-embed on half the data. In the Embeddings tab, untick downstream under Feature Subset and press Re-calculate positions. That lays the same neurons out by their inputs alone. Clusters that survive both are on firmer ground than clusters that only exist in one.

Working comfortably at this size

87,263 points is a real dataset, not a toy, and two things are worth knowing.

Selecting a lot of neurons is expensive. Every selected neuron is a mesh the 3D viewer fetches and renders. If you want to explore the scatter plot quickly, untick View → Synchronize Viewer — selection then becomes instant, at the cost of not seeing morphology. Turn it back on when you have a group worth looking at.

Load a subset instead. The Filters box in the Open Project dialog is applied while reading, so a subset loads proportionally faster:

hemilineage == "MBp3"

gives you 2,863 neurons — 1,284 from FlyWire, 1,038 from MaleCNS and 541 from Hemibrain. Set Embeddings to calculate from features at the same time, so the layout is recomputed for the subset rather than inherited from a plot that was arranged around 87,000 points. See loading part of a dataset.

Check a column's values before you filter on it

A filter that matches nothing, or matches only part of what you meant, fails silently — an empty result looks exactly like a genuinely empty category. Open the Meta Data Explorer (Shift+Cmd+M) and look at the column first.

The columns here are already reconciled across the three connectomes: side is left / right / center throughout, so side == "left" means the same thing in all of them. That is not free — each source brain spells its annotations differently, and lining them up is exactly what column mapping is for. Expect to do it in projects of your own.

Then what

Once you have a feel for the app on real data: