Neural signal propagation atlas of Caenorhabditis elegans

Nature 623(7986):406–414. Received 15 July 2022; accepted 27 September 2023; published online 1 November 2023 (issue 9 November 2023). DOI: 10.1038/s41586-023-06683-4. Open access, CC BY 4.0.

Updated 15 January 2024. Source data for Figs 2–6 and Extended Data (ED) Figs 2–7 and 9 had been "omitted owing to a production error" and were then added. Crossref records this as an "Update" with no separate correction DOI.

Peer-reviewed primary experimental paper. It maps responses to single-neuron optogenetic activation across the worm head in 113 wild-type-background animals and 18 unc-31 mutants, and compares them with predictions from the connectome.

Read 30 September 2026 by a subagent (claude-opus-5-5, reasoning effort max). The main session re-ran the source-data check script (all checks passed). It verified by eye that Fig. 6's rightmost bar (~0.61) disagrees with its source data (0.23), while the other four bars match (0.02, 0.12, 0.12, 0.21).

Authors (verified against PDF page 1, the article metadata and Crossref, which lists ORCIDs): Francesco Randi, Anuj K. Sharma, Sophie Dvali and Andrew M. Leifer (corresponding).

• All four are at the Department of Physics, Princeton; Leifer also at the Princeton Neuroscience Institute.
• Randi is now at Regeneron Pharmaceuticals.
• The article thanks Mei Zhen and anonymous referee(s). No peer-review file is published.

Labels used throughout:

• [V] verified. Checked against the paper's own text, figures, supplements or source data: for a statement, that the paper makes it; for a number, that it matches or reproduces.
• [I] inferred. Our reasoning from verified facts.
• [G] guess. Plausible but unchecked.

The Method and Results sections report the paper's statements and are [V] unless tagged otherwise. Every number marked "our check" was recomputed from the published source data.

What was read

The article page lists sixteen supplementary items. All sixteen, the version-of-record PDF and the article HTML were downloaded on the first attempt (HTTP 200) and read as follows.

• Main PDF (version of record, 28 pages).
  • It was generated 17 January 2024, after the update (PDF metadata). The November 2023 original was not obtained; the change notice says only the source-data files changed.
  • Layout: article pp. 1–9 (journal pp. 406–414); Methods, references and declarations pp. 10–15; ED Figs 1–9 pp. 16–24; ED Table 1 p. 25; Reporting Summary pp. 26–28.
  • The reading-order text (3,380 lines) was read line by line.
  • All 28 pages were rendered and inspected. Figure panels were re-rendered at 200–600 dpi. Methods equations 1–5 were read on a 220-dpi render and re-derived by hand.
  • Gap: the large matrices (Fig. 2a; ED Figs 3a, 4a–b, 5a–b, 6a, 6d, 6e, 7a) were inspected for structure only. Their neuron labels overlap even at the embedded 600 ppi, so cell values were read through the source data, not by eye.
• Supplementary Information (MOESM1, 9 pages: a duplicated cover, then pages i–vii). Read in full: background, validation of the actuator and of targeting, interpretation of the extrasynaptic screen, supplementary discussion, refs 75–97 and the legends of the other items. It is text only, with no figures or tables.
• Reporting Summary (MOESM2, 3 pages). Read, with checkbox states inspected. Its text is identical to PDF pp. 26–28 (programmatic diff), and the renders match to within about one grey level on average.
• Supplementary Table 1 (MOESM3). Read in full. It lists 53 neuron pairs: 174 peptide–receptor rows cover 36 pairs, and 17 pairs carry a "no compatible combination" note.
• Supplementary Video 1 (MOESM4; 30.1 s, 904 frames). Inspected as 30 frames at one-second spacing plus four frames at full size. It was not watched in real time.
• Source Data (MOESM5–16; 12 files).
  • The structure and headers of every sheet were read, and every sheet was parsed.
  • The matrices (up to 185 × 168) were not read cell by cell. They were used to check the paper's numbers by script.
• Article HTML. Used for metadata, the supplementary inventory, the change history and the data, code and peer-review statements. Its body was not read separately. By a programmatic check, 430 of 446 sampled 60-letter windows of its paragraphs occur in the PDF text. The misses are "Full size image" labels, equation markup and page breaks.

Checks by script.

• The script is papers/consciousness_connectomics/randi2023_propagation_checks.py. It is local and git-ignored like the sources; its hash is in the provenance file.
• It asserts every source-data number quoted in this summary, including the paper's numbers marked "reproduced" and ours marked "our check".
• Not in the script: values read from figure pixels (Fig. 6; ED Fig. 2j), the ED Fig. 9b list transcribed at 600 dpi, and the HTML-to-PDF text comparison.
• Command, run on CPU: uv run --no-project --with pandas --with openpyxl --with xlrd --with scipy python randi2023_propagation_checks.py. Output: all checks passed.

Not audited:

• the OSF dataset (doi:10.17605/OSF.IO/E2SYT);
• the browsable sites funconn.princeton.edu and funsim.princeton.edu;
• the analysis and acquisition code (leiferlab/pumpprobe, wormdatamodel, wormneuronsegmentation-c, wormbrain, pump-probe-acquisition, and francescorandi/wormneuroatlas);
• CeNGEN and the connectome datasets;
• the bioRxiv/arXiv preprint, which was not fetched;
• the cited papers, including the anatomical model (Kunert et al. 2014) and the polarity data (Fenyves et al. 2020).

Nothing was simulated or re-modelled; only the published source-data files were re-analysed. CPU only. Source URLs, SHA-256 hashes and derived files are in papers/consciousness_connectomics/randi2023_propagation.provenance.json.

Question and contribution

How well does anatomy predict how signals actually propagate through a nervous system? The paper measures propagation directly: it activates one neuron and images the calcium response of every other head neuron. It contributes four things:

1. A signal propagation atlas of the C. elegans head. Responses to two-photon activation of single neurons are pooled over 113 animals. Each pair gets a q value for "functionally connected" and a q_eq value for "functionally non-connected".
2. A comparison with anatomy, in two forms:
   • the minimum path length in a permissive binary connectome;
   • a connectome-constrained biophysical model, run with anatomical or with fitted weights.
3. Evidence that extrasynaptic (dense-core-vesicle, peptidergic) signalling evokes acute calcium responses. It comes from unc-31 mutants, the neuron RID, and a screen for pairs connected in WT but not in unc-31.
4. A test of whether the measured kernels predict correlations in spontaneous activity better than anatomy does.

The paper measures effective, not direct, connections: responses "represent the propagation of signals over all paths in the network between the stimulated and the responding neuron". The authors propose that extrasynaptically released neuropeptides "serve a similar function to that of classical neurotransmitters" on sub-second to second timescales.

Method

Animals, stimulation and imaging

• Strain AML462 ("TWISP"), hermaphrodites. Every neuron expresses nuclear GCaMP6s, the NeuroPAL colour code for identification, and the purple-light actuator GUR-3/PRDX-2.
  • Actuator expression is switched on with dexamethasone from larval stage L2/L3, to avoid developmental toxicity.
  • The strain is not behaviourally wild type: it moves at 0.03 mm/s off drug and 0.02 mm/s on drug, against 0.15 mm/s for WT.
• unc-31 knockout (AML508). A CRISPR deletion of 12,476 of 13,169 bp removes 18 of 20 exons. Our arithmetic confirms the 699-bp remnant shown in ED Fig. 1b [V].
• Mounting. Worms are immobilised on 10% agarose with polystyrene beads and levamisole. They are "largely immobilized … while still allowing it to slightly move", and pharyngeal pumping stays visible.
• Imaging. Spinning-disk confocal at 2 volumes/s. Excitation is 505 nm at 1.4 mW/mm², below the level that activates GUR-3/PRDX-2. Under these conditions "animals exhibited very little spontaneous calcium activity".
• Stimulation.
  • Every 30 s, a neuron in the current volume is picked at random by its tagRFP-T signal.
  • It is targeted with a two-photon spot (850 nm, 500 kHz, 1.2 mW at the sample; temporal focusing; about 3.1 µm FWHM in x, y and z).
  • Duration is 0.5 s in WT and 0.3 s in unc-31, chosen so that maximal autoresponse amplitudes match.
  • RID and AFD were sometimes targeted by hand. A small AFD set used 0.1–0.5 s pulses.
  • Identities were assigned after the experiment with NeuroPAL. Each recording lasted up to 40 min.
• Coverage (our check). 1,889 included stimulation events make up the whole WT atlas, over 168 stimulated neurons, with a median of 8 events per neuron. ED Fig. 3b and the diagonal of ED Fig. 4a agree exactly [V].

Response definition

• Response measure. ΔF/F0 uses F0 = F over the 30 s before stimulation. The atlas value ⟨ΔF/F0⟩ is the mean over a 30-s post-stimulus window, averaged over trials and animals.
• Inclusion requires an autoresponse. The stimulated neuron's own transient must exceed thresholds on ΔF/F0 and on its second derivative for a contiguous 4 s.
  • The thresholds were "originally set to match the human perception of a response above noise" and are applied uniformly.
  • A trace is excluded if its identity is unknown, or if more than 5% of the response window is missing.
• RID exceptions (Fig. 4, ED Fig. 7c).
  • Trials are kept even without an RID trace.
  • Any neuron missing more than 50% of time points gets a "matchless" trace, placed from its neighbours' deformation.
  • ΔF/F above 2 is excluded.
• Preprocessing.
  • Missing points are interpolated and photobleaching is corrected with a double exponential.
  • Outliers beyond 5 SD are removed.
  • A causal Savitzky–Golay filter (6.5 s) is applied for display; fits use the unsmoothed traces.

Statistics

• Functionally connected (q < 0.05).
  • Test: two-sided Kolmogorov–Smirnov tests of all observations of a pair, on ΔF/F0 and on its second derivative. The null is empirical, from recordings without stimulation.
  • The two P values are combined by Fisher's method, then converted to Storey–Tibshirani q values across pairs.
• Functionally non-connected (q_eq < 0.05).
  • Test: two one-sided tests (TOST) of equivalence to the null within ε = 1.2σ of the control distribution, on both features.
  • Again Fisher-combined, with q_eq from Storey–Tibshirani.
  • The authors call ε "arbitrary" and "conservative".
• Undetermined: pairs that pass neither test.
• Pooling. Trials from different animals are pooled per pair. The Reporting Summary marks "hierarchical and complex designs" as not applicable.

Anatomy

• Minimum path length. Two neurons count as directly connected if any of four connectomes (three adults and one L4; refs 1 and 6) contains "at least one synaptic contact" between them. The authors call this definition "permissive".
  • Gap junctions evidently count [I]. AVDR→AVDL has l = 1 in the Fig. 2f data, although the text says the pair's only wiring is "a very weak gap junction" in one of the four connectomes.
  • Any two head neurons are at most four hops apart (SI).
• Anatomy-derived model. The Kunert et al. 2014 network model: differential equations for electrical and chemical synapses, with a nonlinear synaptic-activation variable.
  • Weights are a scaled average of the four connectomes.
  • Polarities come from a gene-expression analysis (Fenyves et al. 2020 S1 data; CeNGEN), merged per cell class; ambiguous polarities are set excitatory. Other parameters are "commonly used" values.
  • The predicted response is the peak ΔV of neuron i after current is injected into neuron j.
• Fitted variant (Fig. 3d). Weights and polarities are "allowed to float and were fitted from the functional measurements", with no new connections allowed. The fitting objective, procedure and any held-out data are not described [V].

Kernels and simulations

• Kernel definition. k_ij satisfies ΔF_i = k_ij ∗ ΔF_j. It is fitted per stimulation event as a sum of at most two terms. Each term is a chain of convolved decaying exponentials (equation 1), convolved symbolically by the rules in equations 2–5.
  • Our re-derivation confirms equations 2–5. Rules 3 and 4 state their conditions as γ_i = γ_μ and γ_i ≠ γ_μ where γ_n is meant [V].
  • The authors say this form "is exact for linear systems". It represents only impulse responses with real poles, so it is not exact for oscillatory linear systems [I].
• Rise time. The time from 1/e of the peak to the peak, after removing the saturation terms.
• Stereotypy. Kernels of one pair are convolved with a common set of inputs, and the mean Pearson correlation of the outputs is reported.
• Spontaneous-activity prediction (Fig. 6).
  • Procedure: a driver neuron gets a representative transient; activity is propagated through trial-averaged kernels, with pairs at q > 0.05 set to zero. The resulting activities are correlated, and the correlations are averaged over the driven set ("all" or "top-n").
  • Recording conditions: spontaneous activity was recorded in AML320, which lacks the actuator, at about 10 mW/mm² — seven times brighter. That setting produced "population-wide slow stereotyped spontaneous oscillatory calcium dynamics".
  • The caption refers to "an immobilized animal"; no other n is given.
• Extrasynaptic screen.
  • A pair qualifies if q^WT < 0.05, q_eq^unc-31 < 0.05 and q^unc-31 > 0.05.
  • Peptide–receptor combinations come from CeNGEN (threshold 4) and published ligand–receptor screens, via the Worm Neuro Atlas package.

Results

Coverage and the map

• Scale. 113 animals; 23,433 "pairs of neurons (66% of all possible pairs in the head)"; activity from 186 of 188 head neurons (never observed: AIAR and I5, our check). Pairs were "imaged at least once, and some as many as 59 times".
  • Our check: these totals include each stimulated neuron's response to itself (see "Our checks", item 1).
• Connections. 1,310 of 23,433 pairs (6%) are functionally connected at q < 0.05.
• Bilateral homologues are "eight times" more likely to be connected than random pairs: 0.469 against 0.054 in the Fig. 2c source data, a ratio of 8.6 [V].
• Inhibition. 11% of connections are inhibitory (Fig. 2d). The authors call this a likely lower bound, because inhibition can only be seen in neurons that are already tonically active.
• Variability. Downstream neurons respond to some stimulations of the same upstream neuron but not others, and kernels vary across trials and animals. The authors attribute this "presumably" to state- and history-dependent effects, neuromodulation, plasticity and inter-animal variability. They do not separate these sources.
  • Our check:
    • The median fraction of stimulations that evoke a response is 0.50 among connected pairs (IQR 0.24–0.86).
    • Median kernel stereotypy is 0.63 among connected pairs.
    • The ED Fig. 6f histograms give a mean kernel correlation of about 0.46 within a pair, against 0.28 across all kernels (approximate, from binned densities).

How well anatomy predicts propagation

All rows use pairs of distinct neurons, from the source data, unless noted.

The authors summarise this as "fairly poor agreement" between anatomy-based predictions and measurement. They add that discrepancies also appear without any model: wired pairs that are functionally non-connected (Fig. 2g), and RID (Fig. 4). AFD→AIY is their example of a sign that anatomy gets wrong: AFD carries glutamatergic inhibitory machinery but excites AIY through peptides (citing Narayan et al. 2011).

Extrasynaptic signalling

• unc-31 (18 animals, ED Fig. 7). A smaller proportion of pairs is functionally connected than in WT, among pairs measured in both: 7.8% against 4.5% (ED Fig. 7b).
  • Our check:
    • After repairing garbled neuron labels in the unc-31 sheet: 7.7% against 4.5% over 8,253 common pairs, self-pairs included. Excluding self-pairs: 6.6% against 3.7%, a 44% relative reduction.
    • Of the 534 WT connections between distinct neurons also measured in unc-31, 59 (11%) reach q < 0.05 there.
  • ED Fig. 8 shows four pairs whose responses persist, with 2–4 unc-31 trials each.
  • Fig. 3d: unc-31 agrees "better" with anatomy (R² −0.047 against −0.116).
• RID (Fig. 4, ED Fig. 7c). RID is thought to signal through neuropeptides and has "only few and weak" outgoing wired connections.
  • Anatomy predicts almost no response in URX, ADL and AWB. These neurons respond strongly in WT, and the responses are reduced or absent in unc-31.
  • All three express receptors for RID peptides: NPR-4 and NPR-11 for FLP-14, and PDFR-1 for PDF-1.
  • Their shortest wired paths rely on "fragile single-contact synapses that appear in only one out of the four individual connectomes".
  • Our check:
    • Trial counts, WT against unc-31: ADLR 3 against 9; URXL 6 against 16; AWBR 18 against 25.
    • One-sided rank tests, which the paper did not run, give P = 0.0045 (the smallest possible with 3 against 9 trials), P = 0.0009 and P = 0.24 respectively.
    • Under the alphabetical ordering verified for the Fig. 2f sheet, the Fig. 4 source data give anatomy-derived responses of 0.0006–0.0013 V for these three pairs [I].
• Screen (Fig. 5, ED Fig. 9, Supplementary Table 1). 53 pairs are "putative candidates for purely extrasynaptic signalling". The authors call this "likely to be a lower bound".
  • Examples. M3L→URYVL, with FLP-4→NPR-4 and FLP-5→NPR-11. AVDR→AVDL, with NLP-10→NPR-35 and FLP-6→FRPR-8, a predicted autocrine loop in a predicted peptidergic hub neuron.
  • Our checks:
    • All 53 have WT q < 0.05, and 52 have unc-31 q > 0.05.
    • 44 of 53 have no direct contact in any of the four connectomes: 26 at two hops, 17 at three, 1 at four.
    • 36 of 53 (68%) have at least one compatible peptide–receptor combination.
    • The upstream neurons concentrate on a few cells: M3L in 12 pairs, AVDR in 6, IL1DL in 5.
    • The 53 pairs are 4.6% of the 1,151 WT connections [I, arithmetic].
• Timescales. Candidates have rise times "similar to that of all functional connections" (Fig. 5a). The abstract says "including on timescales of less than a second".
  • Our check: rise times exist for 21 of the 53 candidates; 15 are under 1 s, with a median of 0.49 s. The "All" histogram holds 281 values, with a median of 0.46 s.
  • Imaging runs at 2 volumes/s, and the SI reports up to one volume (0.5 s) of timing jitter. The SI says the jitter does not affect kernels, because kernels relate upstream to downstream activity. Sub-second rise times are therefore within about two imaging volumes [I].
• The authors' case against a pure-modulation reading (SI).
  • The upstream autoresponse is required, so reduced upstream excitability cannot explain a missing response.
  • For 85% of the candidate pairs, the downstream neuron still responds to another input in unc-31. Our check: 45/53 = 84.9%, after repairing the labels.
  • They cannot rule out unc-31-dependent up-regulation of specific synapses, but find it "less compelling".

Validation of the method

• Targeting.
  • The two-photon spot is about 3.1 µm FWHM in all three axes (ED Fig. 2a).
  • Photobleaching stays restricted to the targeted neuron, both near and far from the objective (ED Fig. 2c–d).
  • One in vivo example: no response with the spot 4 µm off-soma along z, and an autoresponse on target (ED Fig. 2e).
• Off-target bound. The SI states that about 75% of stimulations "did not show obvious activation" in neighbouring neurons, "which places a worst-case-scenario extreme upper bound on off-target stimulation at ~25%". Our check contradicts the numbers ("Our checks", item 6).
• Actuator in AVA (ED Fig. 2h). GUR-3/PRDX-2 in AVA (strain AML546) evokes more reversals under blue light than in WT: n = 11 against 8, p = 0.025.
  • Our check: Student's t test of the two blue-light rates reproduces p = 0.0249. The increase over each strain's own control does not differ between strains (p = 0.14).
• Pharynx (Supplementary Video 1). M1 stimulation evokes pharyngeal contractions in four examples.
• Literature agreements (ED Table 1). Four selected agreements with earlier functional work: ASH→AVA, ASH→AVD, AFD→AIY (linear) and reciprocal AVA–AVE.

Our checks against the paper and its source data

All [V] unless marked. The script reproduces each source-data number.

Reproduced exactly or closely:

• Path-length panels. The Fig. 2e and 2g probabilities by path length reproduce exactly once the one pair lacking a q_eq is excluded, as the paper evidently did (5,967/10,887 at two hops).
• Headline statistics.
  • the mean of 2.07 hops;
  • the bilateral ratio;
  • Fig. 3b's P;
  • ED Fig. 2h's p;
  • the ED Fig. 2i counts (1,894, 11,780 and 21,211);
  • ED Fig. 7b (closely);
  • the SI's 85%;
  • the same 53 pairs in ED Fig. 9b and Supplementary Table 1.
• Example pairs. Trial counts in Figs 1, 4 and 5 and ED Fig. 4c match the observation matrix:
  • AVJR→AVDR: 24 trials;
  • AVER→AVAR: 11;
  • SAADL→OLLR: 5;
  • M3L→URYVL: 12;
  • AVDR→AVDL: 20;
  • AVJR→ASGR: 9, with q = 0.075 against "q = 0.07" on the panel.

Internal inconsistencies:

1. The headline counts include each stimulated neuron's response to itself. The source matrices hold two kinds of entry:

   • pairs of distinct neurons: 23,265 measured, of which 1,151 are connected;
   • self-pairs: 168 measured, of which 159 are connected.

   Hence 23,433 and 1,310. The "as many as 59 times" is also a self-pair: RMDVR was stimulated 59 times, while the maximum for distinct neurons is 54. Autoresponses are required for inclusion and drawn black on the diagonal, yet they are counted as "pairs of neurons".

   • Corrected values:
     • 4.9% of measured distinct pairs are connected, not 6%.
     • 13.1% of connections (151/1,151) are inhibitory, not 11%. The Fig. 2d curve also divides by k − 1, giving 151/1,309.
   • Knock-on effect on Fig. 3. The same inclusion puts 166 self-pairs into Fig. 3b's "predicted large" group and 157 into Fig. 3c's "connected" group. A neuron's simulated response to its own stimulation is large, so this inflates both results.
     • Without self-pairs, Fig. 3b's P is about 2 × 10⁻²⁶ instead of 10⁻⁸⁸.
     • In Fig. 3c, 9.7% of connected against 1.7% of non-connected pairs are predicted ΔV > 0.1 V, instead of 20.5% against 1.8%.
     • Both effects keep their direction.

2. "Connected" and "non-connected" are not exclusive.

   • 327 pairs of distinct neurons pass both tests, 28% of the connected pairs.
   • 11 of 168 autoresponses (mean ΔF/F0 0.19–0.68) pass the non-connection test, so those neurons are "functionally non-connected" to themselves.
   • The text presents three exclusive categories. The screen's extra condition q^unc-31 > 0.05 exists to handle exactly this overlap.

3. The "non-connected" verdict tracks sampling. The share of pairs with q_eq < 0.05, by number of observations:

   Observations	1	2	3–4	5–8	9–16	≥ 17
   Share with q_eq < 0.05	1%	46%	75%	89%	94%	94%

   • For pairs observed at least twice, the share still rises with path length within each stratum. Among pairs observed once it stays at 0–1.4% whatever the path length.
   • The mean number of observations is similar across path lengths (4.4–4.9), so the path-length trend is not a sampling artefact [I].
   • But the level of "52% of directly wired pairs are functionally non-connected" is set mainly by how often pairs were observed. Two-thirds of measured pairs were observed only 1–4 times [I].

4. The Fig. 3b caption swaps the group sizes.

   • The caption gives "ΔV > 0.1 V, n = 23,454 pairs" and "ΔV < 0.1 V, n = 614 pairs". The source data hold 614 values for ΔV > 0.1 and 23,454 for ΔV < 0.1.
   • The groups total 24,068, more than the 23,433 measured pairs. 866 of the values have no exact match in the published ΔF/F matrix.

5. Fig. 6: the plotted best bar is not in the source data.

   • Measured in pixels against the axis ticks, the printed top-n kernel bar reads 0.61; the source data give 0.23. The other four bars agree (0.02, 0.12, 0.12, 0.21).
   • The text's "markedly improved performance" matches the plot.
   • The package cannot decide which value is right. Fig. 6 is the paper's only test of predicting dynamics.

6. ED Fig. 2j is offset from its source data.

   • Digitised curves sit above the empirical CDFs at the same ΔF/F0. For 4–10 µm neighbours at 0.09, the plot shows 0.755 and the data 0.594.
   • The plotted levels match the data about one histogram bin (≈ 0.14) further right [I, mechanism].
   • In the data, 61.6% of 4–10 µm neighbour responses are ≤ 0.1, not "~75%". The SI's own rule would therefore give a worst-case off-target bound of about 38%, not about 25%.
   • The authors argue most neighbour responses are genuine downstream responses. Neither side of that argument is tested here [I].

7. The ED Fig. 2g source data do not match the plot.

   • The source sheet has 10 rows, one an exact duplicate, including an AIY value of −5.8.
   • The panel shows 9 points, all positive, on a smaller scale: AFD up to about 4.4 against 8.7 in the sheet.
   • In the atlas, AFDL→AIYL has q = 0.55 from one observation, and AFDR→AIYR was never measured.
   • So the AFD→AIY example rests on the separate variable-duration experiment and on earlier literature, not on the atlas.

8. ED Table 1 misstates one entry. It says "ASHR→AVAR is functionally connected (q<0.05)"; the source data give q = 0.062 from 6 observations. ASHL→AVAL has q = 0.019.

9. Denominators differ between panels.

   • Measured pairs are counted as 23,265 or 23,433 (matrices), 24,068 (Fig. 3b), 24,075 (Fig. 2c's "All" = 1,310/24,075) and 24,598 (the ED Fig. 4b axis).
   • All possible pairs are counted as 35,156 (= 188 × 187, Fig. 4 source), 35,344 (= 188², Fig. 2f sheet) and 37,111 (the ED Fig. 4a axis).
   • Fig. 2f's "All possible pairs" histogram in fact contains only the 23,265 measured pairs.

10. Screen details.

    • SAAVL→AVAR has q^unc-31 = 0.0037, which fails the stated condition q^unc-31 > 0.05.
    • The ED Fig. 9 source data hold 55 entries, while the text, ED Fig. 9b and Supplementary Table 1 list 53.
    • Fig. 5a's "All" histogram holds 281 rise times, although ED Fig. 6d gives rise times for 921 connected pairs of distinct neurons. How the 281 were selected is not described.

11. Source-data labelling errors.

    • Fig. 2a's "alpha=(1 - q_val)" sheet actually holds clip(1 − 2q, 0, 1).
    • ED Fig. 6f's columns are labelled "wt" and "unc31" but hold the within-pair and shuffled distributions shown in the panel, with the "bin_edges" and "y" headers swapped.
    • ED Fig. 6a's diagonal holds fractions above 1 and infinity; the figure masks them.
    • Some neuron labels are not neuron names: "IL2V4" (Fig. 2a), and "ABAGR", "M24", "LI", "NSM4" and "12R" (ED Fig. 7a). The latter are most likely BAGR, M2R, MI, NSMR and I2R [I]. With that repair the SI's 85% reproduces, and the two candidates the garbled labels had hidden satisfy the screen's unc-31 condition.

12. Wrong panel or item references.

    • Rise times are "shown in Fig. 5c" (Methods) and in "Fig. 5b" (SI); they are in Fig. 5a.
    • The SI's "AVDR->AVDL appears in our screen (Fig. 5d)" refers to Fig. 5c.
    • The Reporting Summary cites "Supplementary Figure S5a", "S6" and "S11", which the published SI does not contain. They are probably now ED Figs 4a, 6 and 2g [G].

13. Minor slips.

    • The AVA strain is AML546 in the Transgenics text, ED Fig. 1a and the ED Fig. 2h source data, but "AML564" in the cross-validation Methods. ED Fig. 1a adds that it also expresses the actuator "in some pharyngeal neurons (likely I1, I4, M4 and NSM)".
    • The text puts URXL two hops from RID; the Fig. 2f data give three (URXR is two).
    • "0.2-nm fluorescent beads" should presumably be µm [I].
    • Refs 62 and 67 are the same paper with different years.
    • The pumpprobe Zenodo DOI differs between the paper (8312985) and the Reporting Summary (8247256).
    • "URVYL" is a typo for URYVL.
    • SI ref 96 (Taylor 1988, basic fluorescence microscopy) is cited for gene-expression predictions of peptide signalling. Taylor et al. 2021 (ref 38) was probably meant [G].

The effectome referee's figure is not in this paper. The phrase "as much as 30 % of the total 'neural variance'" (see the Pospisil summary) appears nowhere in the main text, SI or Reporting Summary (text search). No figure reports a variance fraction [V]. It is the referee's own characterisation.

What rests on measurement, and what on modelling

Limitations

Stated by the authors:

• The map is a lower bound on functional connections.
  • Dim or hard-to-identify neurons have fewer observations: AVB, ADF, RID and the ventral ganglion, and AWCon and AWCoff cannot be told apart.
  • Only nuclear calcium is imaged, so compartmentalised dynamics are missed.
  • Inhibition is visible only in tonically active neurons, and the statistics are stringent.
• The strain is not behaviourally wild type, so its "signal propagation might therefore also differ from the wild type".
• Connections are effective, not direct. Inferring direct connections "might require a higher signal-to-noise ratio".
• Only single neurons are stimulated. Nonlinearities and multi-neuron interactions need simultaneous stimulation.
• State and history may matter. Plasticity, neuromodulation, network state and experience may explain some of the variability and of the mismatch with anatomy. Some "latent connections" may become functional only in certain internal states.
• The biophysical model's assumptions would add to the disagreement if they are wrong. But the authors show discrepancies without any model (Fig. 2g; RID).
• The screen has blind spots. It misses pairs with parallel synaptic paths or too few observations. unc-31 might have other defects or change wiring, and neuromodulatory up-regulation of specific synapses cannot be excluded.
• Timing and kinetics. Stimulus timing jitters by up to 0.5 s. Dynamics are slow: graded potentials, nuclear calcium, GCaMP6s and actuator kinetics.
• Species scope. Peptide diffusion may suit the worm's small size; larger brains might limit its speed, strength or range.

Ours:

• The headline numbers need correcting (items 1–3 above).
  • The corrected values are 4.9%, not 6%, connected, and 13%, not 11%, inhibitory.
  • "Non-connected" is neither exclusive of "connected" nor independent of sampling.
• No anatomy from the imaged animals.
  • Function is pooled over 113 animals. Anatomy is a union (for path length) or an average (for the model) of four other animals, one of them a larva.
  • Individual differences in either cannot be separated from trial-to-trial variability [I].
  • Counting single contacts in any of four animals makes "directly wired" a low bar [I].
• Only one anatomical model was tested.
  • With fitted weights and signs its R² stays below zero, and the fit is not described. So whether any connectome-constrained model could do better is untested [I].
  • R² below zero for a line through the origin means the fit is worse than predicting the mean. A calcium response and a voltage prediction may need a nonlinear link [I].
• The atlas measures one quiet state.
  • It is recorded under levamisole, at a light level that leaves "very little spontaneous calcium activity", in a slowed transgenic strain. Levamisole is a cholinergic agonist; that is background knowledge, not stated in the paper.
  • Fig. 6's spontaneous data come from another strain in a brighter, oscillatory state. Fig. 6 therefore tests transfer across states without saying so [I].
  • Whether levamisole changes propagation is not addressed [I]. That it matters is a guess [G].
• Trials are pooled across animals in the KS tests and TOST. The q values treat trials as independent, and animal is not a level of inference [I].
• The unc-31 contrast has no sample-matched control. It uses 18 against 113 animals, and detection rises steeply with the number of observations. Fewer detected connections in unc-31 is expected even without a biological difference.
  • A random 18-animal WT subset would be the fair control [I].
  • unc-31 is a constitutive whole-animal mutation, so developmental effects are possible [I].
• The screen is modest evidence [I].
  • It declares absence in unc-31, and that verdict depends on sampling.
  • It concentrates on a few upstream neurons (M3L in 12 of 53 pairs).
  • Peptide–receptor compatibility has no base rate: how often does a random pair have a compatible combination? The same broadly expressed receptors (DMSR-1, DMSR-7, FRPR-8) recur across many pairs.
• The RID examples rest on few trials. ADLR has 3 WT trials and URXL 6, under relaxed inclusion. The AWBR difference is not significant by a rank test (P = 0.24; our check) [V].
• Fig. 6 is weak evidence [I].
  • It rests on one recording ("an immobilized animal"), with no uncertainty.
  • The top-n subset is chosen as "optimal" by a procedure that is not described, apparently on the evaluation data [I].
  • Its best value conflicts with the source data [V].
• ED Table 1 lists agreements only. There is no systematic comparison with earlier functional studies [V].
• Code and raw data were not audited. Kernels, q values and the anatomical model were not recomputed from raw data.

Relevance to our research (our inference)

Answer to the fourth synthesis's OQ3

The fourth synthesis asked: "In a real nervous system, how well does synaptic anatomy predict measured perturbation effects, and how much of the measured influence is non-synaptic?" The answer below holds for the worm head, in this quiet, immobilised preparation.

• Anatomy predicts better than chance and poorly overall [V numbers; I reading]:
  • direct wiring raises the chance of a detectable effect 2–10-fold over multi-hop pairs;
  • but at the typical 1–4 observations only 11% of wired pairs show one, and about half do among the best-sampled pairs;
  • 77% of detected effects are between neurons with no direct contact;
  • the anatomical model called 90% of detected effects weak, and its strong predictions are right about a quarter of the time;
  • no amplitude agreement survives (R² < 0, even with fitted signs and weights).
• No variance fraction exists for the non-synaptic share. Three lower-bound or confounded estimates exist [V; I on weight]:
  • 53 unc-31-dependent candidate pairs, 4.6% of connections;
  • strong RID responses in neurons it barely wires to;
  • 44% fewer connections in unc-31 among pairs measured in both, confounded by sampling.

For the replica goal (carry-on goal)

The goal is a replica that carries on as the person. Change is acceptable only if the person slot drives it (Q1–Q3). Everything below is [I]; the worm-to-person step is an analogy.

• A structural record under-determines responses; a measured response map predicts better. In the worm, measured perturbation–response maps carry predictive content that wiring does not (Figs 2–3). Fig. 6 makes that claim for dynamics, but its evidence is weak.

  • The slot's analogue of wiring is the static record: facts, traits, stated views. The analogue of the atlas is the person's measured reactions to probes.
  • This supports Q1's option that the slot must hold the person's way of changing, estimated from perturbation-like interview probes, not only contents.

• Responses are probabilistic. A typical real connection fires on about half of trials.

  • One elicitation of a person's reaction is a sample, not a trait. The multiple interviews the user allows are needed.
  • Replica and person should be compared as distributions over repeated probes, matched for state.

• Exclude "self-pairs" from agreement scores. The paper's counts are inflated by trivially guaranteed entries, a neuron's response to its own stimulation.

  • The replica analogue is scoring answers that the slot directly encodes.
  • Those items should be reported separately from genuine prediction. This is a concrete acceptance check to add.

• Declaring "no drift" or "no slot effect" is an equivalence claim, and it inherits the paper's two failure modes:

  • the verdict tracks sample size;
  • "changed" and "stable" can both be declared for one item.

  Q2 designs should fix the equivalence margin in advance, report verdicts by sample-size stratum, and make the categories exclusive by rule.

• Knockout contrasts need matched samples. Own-, other- and empty-slot arms are knockout contrasts like WT against unc-31. An arm evaluated on fewer items or sessions will show "fewer effects" from power alone, so arms must be matched or subsampled.

• Hidden channels. Worm peptides are influences invisible in the structural record that only perturbation reveals. The analogue is influences on a person's reactions that are absent from the interview record, such as mood, context or body state. That is a guess about people [G].

• Population is not individual. The atlas pools 113 animals, and nothing in it bears on individual-specific propagation. This agrees with the fourth synthesis: the slot cannot come from connectomics now.

For the validation design

This is the library's first real-organism test of anatomy against measured causal effects at scale. Its declaration rows, in the design's terms [I]:

• Biological unit. Population function against the population anatomy of different animals; no same-animal pairing.
• Structural resolution. A permissive union of connectomes, electrical contacts included, for path length; a weighted average for the model.
• Prediction target. Existence of a connection (q), its absence (q_eq), the mean 30-s ΔF/F0, kernel shape, and spontaneous correlations.
• Evaluation distribution.
  • Anatomy-to-atlas comparisons are out of sample, since the model was not fitted to these data; they are weak.
  • The fitted variant is in sample and still weak.
  • Fig. 6's kernels come from other animals, but its subset selection is in sample.
  • By the fourth synthesis's proposed decision rule, Fig. 6's top-n result scores "not tested" and its all-driven result "weak".
• Operating state (proposed new row). The atlas is recorded in a quiet state and Fig. 6 in an oscillatory one. This is the clearest case yet of an undeclared state change between model source and evaluation.
• Baselines and nulls (proposed new row). Present: the anatomy baselines and a pair base rate (Fig. 2c). Missing: a rewired-connectome null for Figs 2e and 6, and a sample-matched control for unc-31.

Additions the reading suggests:

1. Exclude a stimulated unit's own response from pair statistics.
2. For any claim of absence, fix ε in advance, show verdicts by sample-size stratum, and make the categories exclusive.
3. Match or subsample the arms of knockout contrasts.
4. Check plotted values against the source data where they exist. This paper has three plotted-versus-source conflicts: Fig. 6, and ED Figs 2g and 2j.
5. Reconcile denominators by script. This extends caveat C11: the paper and its source data use at least seven different pair counts.

Dataset option. The atlas, whose OSF data and code were not audited, could serve as a second test dataset alongside Shiu's fly setting. Examples: score anatomical models and sign rules against measured perturbation effects, or try the effectome's instrumental-variable (IV) idea on real stimulation data. Any such run needs its own frozen design.

For the synthesis's claims on static wiring, state and identifiability

• Static wiring does not fix which effects are possible once extrasynaptic signalling exists [I from V]. The fourth synthesis wrote that a wiring map "fixes which direct effects are possible" and that anatomy and effect "coincide only in the zero pattern" under a chemical-synapse model (Pospisil).

  • In the worm, unc-31-dependent effects occur between neurons with no direct contact (44 of 53 candidates) and from RID through only fragile multi-hop paths.
  • So the connectome's zero pattern is not the effectome's zero pattern.
  • Caveat: the paper measures effective connections. A direct non-synaptic link is inferred from unc-31 dependence plus the absence of short wired paths, not measured.
  • The converse also holds: most wired pairs show no detectable effect at this sampling.

• A connectome prior would be wrong in both directions here [I]:

  • it would pin non-synaptic effects near zero;
  • it would expect effects on many weak or absent ones.

  That is the case in which Pospisil's ten-neuron simulations found a confident wrong prior to be worse than none. It is now an empirical reason, not only a simulated one, to fix prior strength in advance and test it.

• Signs are inferred labels. Polarities from gene expression give R² < 0 even after refitting signs. AFD→AIY carries inhibitory machinery but acts as excitatory. This supports the synthesis's T1 and T7 [I].

• State determines what can be measured.

  • Inhibition is visible only in tonically active neurons. This is the worm counterpart of Shiu's disinhibition miss at a 0 Hz baseline: from rest, inhibition is invisible in both the model and the animal [I].
  • Responses fire on about half of trials, and kernels are only moderately stereotyped. The authors attribute this to state, history, neuromodulation and inter-animal variation without separating them [V].
  • This supports caveat C10: declare the operating point and treat pooled estimates as mixtures [I].

• Identifiability.

  • Single-neuron stimulation with calcium read-out identifies effective, not direct, connections. The authors say direct connections would need more signal-to-noise. This is consistent with Pospisil's argument that direct effects need designed perturbation and a model [I].
  • Absence is identifiable only with enough sampling (item 3), and the q versus q_eq overlap shows that "no effect" must be tied to a declared effect-size margin [I].

• Shiu's Usnea miss. Shiu's model missed Usnea, which appears to act through neuropeptides. That is consistent with this paper's evidence that peptides can evoke acute responses [I]. Shiu cites this paper for exactly this point (its ref. 50) [V].

Library status

• Cite it for:
  • the first measured whole-head perturbation atlas in the library;
  • weak prediction of effective propagation from anatomy, with the corrected numbers above;
  • evidence that non-synaptic signalling evokes acute responses where wiring is absent or fragile (RID; the unc-31 screen);
  • variable, state-dependent transfer.
• Do not cite it for:
  • a variance share of peptide signalling (there is none);
  • Fig. 6's 0.61 as a verified value;
  • individual-level prediction;
  • "functionally non-connected" as meaning no effect.