← Blog
✎blogbericht

A complete worm wiring diagram did not flinch, and the electrical links were why

A translucent glass lattice of a nerve ring resting on a dark steel plate, with a warm amber glow in one central node, in teal and steel, the AMARBARO mark and wordmark centred over it.

The full C. elegans wiring reproduced its published sheet edge for edge but gave a touch reflex with the wrong sign, and the reflex passed only after the electrical gap pairs from PLM and PVC were removed.

We built the full wiring diagram of the C. elegans worm from the Cook 2019 data, and it matched the published sheet edge for edge. Then we ran a touch reflex on it. The posterior response came out at +0.050 millivolts, the wrong sign: a touch should push that response down. The reflex passed only after we removed a set of electrical links from the model.

That removal makes the passing model a different network from the connectome. It tells us which path breaks the reflex in our model. It does not tell us what the worm does.

The wiring matched, the behaviour did not

The diagram has 473 nodes, which are neurons plus muscles and end organs. It has 4,879 chemical edges and 2,883 gap edges. An edge is one connection. A chemical edge passes a signal through a messenger molecule. A gap edge is a direct electrical link between two cells.

The test compares the built wiring with the sheet, connection by connection. It skips when the spreadsheet is not on disk, so a skipped run proves nothing. This run passed.

The reflex is a different test. The anterior response came out at +0.023 millivolts, which was acceptable. The posterior response was the wrong sign. Two of five degree-shuffled versions also passed. A degree shuffle rewires the cells at random and keeps each cell's count of connections. When shuffled wiring passes the same test, the test cannot tell the real wiring from the shuffled one.

At rest, the chemical links sat at about 98 percent of their maximum. A link near its ceiling cannot respond to a touch much more. Our reading was that the chemical synapses went inert, and the electrical links carried the touch.

The PVC to AVA gap pairs then carried posterior touch straight to AVA. Sign fixes did not solve it. With the touch inhibition and command cross-inhibition sign rules applied, the best posterior response was +0.001 millivolts, and it never went negative.

Removing the gap pairs passed the reflex

We removed the symmetric gap pairs from PLM and PVC to AVA and AVD. Each pair runs both ways, so one removal cut two links at once, a bigger change than it sounds. The reflex then passed at every gap strength we tried, with a posterior response of -0.03 to -0.09 millivolts.

This is post hoc. We removed the pairs after we had diagnosed them as the cause. A passing test on the same data we used to explain the failure is weaker evidence than a passing test on new data.

Fitting the strengths did not work

We also tried a fit: per-cell reversal potentials plus one global gap strength. The loss went from 0.99 to 0.95, the gap strength hit its floor, and nothing passed on the real wiring or on the shuffles. Changing reversal potentials alone cannot repair an electrical path, which is where the problem sat.

What the reduced model does not show

After removal, our record reports the posterior response only. It does not report the anterior response again, and it does not rerun the degree-shuffle test. So we cannot say the reduced model still separates real wiring from shuffled wiring. A reviewer should ask for both before trusting the pass.

What we do not know

  • The source files give no transmitter sign for each cell. Our sign rules are a fit, not data. The neuron table labels some cells in ways we did not trust, such as naming two cells GABA and one FMRFamide.
  • Whether real worms need the gap pairs we removed. The next test is per-synapse strengths from published fits, with the pairs kept.
  • Whether the reflex result says anything about the worm's other behaviours. We tested one reflex.

How to prove this wrong

Fit per-synapse strengths from a published source, keep the gap pairs, and run the reflex on the full graph. If it passes, the claim that the gap pairs cause the failure is wrong, and the removal was a fit to our own diagnosis. Also rerun the degree-shuffle test on the reduced model. If the shuffles pass again, the reduced model is no better at telling real wiring from random wiring.

Provenance

  • ~/Brain/fly-connectome/2026-10-06-worm-connectome.md: the node and edge counts, the reflex results, the chemical saturation cause, the failed fit, and the removal result.
  • ~/Projects/science/fly-connectome/tests/test_worm_csr.py: the edge-for-edge test against the sheet, which skips when the spreadsheet is absent.
  • ~/Projects/science/fly-connectome/tools/worm_reflex_diag.py and tools/worm_fit.py, named in the note as the diagnosis and fit scripts.

Reacties

Nog geen reacties.

Inloggen om te reageren.