A browser viewer for the Cook 2019 connectome — static, offline, and every view is a link

[Dev-cmyser]

I put together a small web app for reading the hermaphrodite connectome in a browser. Showing it here mostly because I would like a second pair of eyes on the data handling.

🔗 https://b-on-g.github.io/worm/ · source and converter: https://github.com/b-on-g/worm

What it does

• 302 neurons plus 148 muscles and other end organs, 4681 chemical synapses and 1359 gap junctions, all on one canvas
• Two layouts: layered (sensory → polymodal → interneuron → motor → muscle, ordered by ganglion inside a column) and a deterministic force-directed one
• Click a cell or type a name and it lights up with its first-order neighbours. ASH selects ASHL and ASHR together, and the side panel lists every connection sorted by weight
• Filters by cell type, connection kind and minimum contact count
• Signal paths: pick two cells and get up to three shortest chains of chemical synapses, five hops at most. ASH → AVA gives the escape circuit, ASH → dBWML8 follows it all the way to a body wall muscle
• The whole view state lives in the URL, so a link restores exactly what you were looking at. A service worker caches the app, so it keeps working with the network off

Where the data comes from

All published, most of it through OpenWorm repos. Nothing is fetched at runtime — the dataset is compiled into the page.

• synapses: herm_full_edgelist.csv from openworm/c302 (Cook et al. 2019, doi:10.1038/s41586-019-1352-7)
• function classes and neurotransmitters: owmeta_cache.json from the same repo
• ganglia and soma coordinates: wormneuroatlas
• cell name expansions: all_cell_info.csv from openworm/ConnectomeToolbox

Judgement calls I would like checked

This is where I am most likely to be wrong:

1. Neuron classes are derived from the name set, not from a lookup table. Drop a trailing L/R when the mirrored partner exists, then a trailing D/V when both a dorsal and a ventral stem exist and the base is at least three characters. That keeps AVL, RID and PVR as their own class and folds IL1DL into IL1. It lands on exactly 118 classes, which is why I trust it, but the rule is mine.
2. DB7, RMHR and VA11 are absent from the ganglion table I used. Each takes the majority ganglion of its own class, which puts DB7 and VA11 in the ventral nerve cord and RMHR in the ventral ganglion. Is that right?
3. Gap junctions are listed from both sides and I collapse them into one undirected edge. Twenty entries appear from one side only, fourteen of which join two processes of the same cell (M4, M5, ASKR, PVM, AIAR and others). Thirteen pairs disagree on the contact count between directions, by one or two except PVPL–hyp, which is 47 one way and 41 the other. I take the larger. Is there a convention for this?
4. I kept the muscles and other end organs from the same edge list instead of trimming to the 302 neurons, so a path can end where it biologically ends. They can be filtered out.

The converter is a single Node script with the sources vendored next to it, and it refuses to emit a dataset that does not come out at 302 neurons in 118 classes with synapse counts inside the published range.

Stack, in case it matters: TypeScript on $mol, no graph or charting library. About 440 KB of JavaScript in total, data included.

If you have five minutes, open the demo and tell me where the data reads wrong to you. Beyond that the obvious next steps are the neuropeptide connectome as a second layer, a layout using the real soma coordinates, and a White 1986 versus Cook 2019 diff, but I would rather hear what would actually be useful.