cs.NESep 30, 2026

Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison

Authors: Gyujeong Park

Organizations: IonLabs

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Randomised copies of a connectome are the usual baseline for asking whether measured wiring matters, and the answer depends on what the randomisation preserves. We evolved embodied foraging agents whose brains are a compressed adult Drosophila connectome (FlyWire v783; 512 cell-type groups and 1,000 Kenyon cells) alongside agents built on randomised wiring, in pre-registered experiments with ten seeds, four ecologies and 600 generations. Two standard randomisations, a column shuffle and degree-preserving edge swaps, route 10.6 to 10.7 % of olfactory output directly onto descending motor groups, against 0.012 % in the connectome. On the registered primary endpoint, fitness averaged over the run, no difference was detected; at the last common-garden probe the connectome was behind both controls (-0.22 and -0.20 fitness units on seed means). Against controls that keep every sensory-output and motor-input edge and rewire only the interior, the seed-mean difference lay within a +/-0.10 equivalence bound (+0.002 and -0.074, unchanged under a calibration that also matches activity spread), although per ecology the interior column shuffle was ahead by 0.26 in one of four ecologies at ten seeds, a lead that ten further pre-registered seeds did not replicate. Rewiring the connectome so that it acquires the shortcut raised its fitness by 0.44 (10 of 10 seeds) and its dependence on olfaction from 0.15 to 0.99; graded doses raised both in step; at comparable swap counts the full dose was ahead of an interior-only sham by 0.53 (10 of 10 seeds); and a sham that rewired the same boundary edges without creating shortcuts matched the connectome (+0.007) while the full dose was ahead of it by 0.60. What a null preserves at the sensory-motor boundary can decide an evolutionary connectome comparison, and sensory-to-motor path statistics belong next to the degree statistics a null is said to preserve.

Figures & tables

Figure 1: What the null preserves at the sensory-motor boundary, and what it changes. (a) In the compressed connectome olfactory input reaches descending (motor) groups through several synapses; the standard column-shuffle and degree-preserving nulls route about a tenth of olfactory output directly onto them; the boundary-preserving nulls keep every sensory-output and motor-input edge and replace about nine in ten interior edges. (b) Differences in evolved fitness at the last common-garden probe with 90 % intervals: connectome minus each null on seed means over four ecologies (median), and the shortcut interventions over two ecologies (transplant: median; full dose minus sham and sham minus connectome: mean, as registered). The last two rows are the boundary-targeted sham (Section 4.5 ), from separate runs with their own connectome island. The grey band is the ±0.10 equivalence bound.
codeoperationadditionally preserved
N1column shuffleout-degree, per-source weight multiset
N2degree-preserving target swapsin- and out-degree
N4N2 on interior edges onlyas N2, plus every sensory-output and motor-input edge
N5N1 on interior edges onlyas N1, plus every sensory-output and motor-input edge
Table 1: Randomised controls. All keep edge counts and Dale signs.
conditionsmell: directsmell: hopsvision: directtaste: direct
connectome0.0001230.3810.121
N1 column shuffle0.10710.1080.113
N2 degree-preserving0.10610.1280.093
N4 interior swaps0.0001030.3400.125
N5 interior shuffle0.0001130.3690.123
Table 2: Sensory-to-motor structure of the calibrated networks (medians over 10 seeds). “Direct”: share of that sense’s output weight landing on descending groups; “hops”: median shortest path from that sense to a descending group.
controlseed meansP0T0P1T0P0T1P1T1
N1 (shortcut)−0.224−0.621−0.312−0.159+0.054
 [−0.34, −0.10][−0.66, −0.47][−0.44, −0.22][−0.21, +0.01][−0.05, +0.13]
N2 (shortcut)−0.199−0.403−0.338−0.017+0.015
 [−0.32, −0.13][−0.42, −0.31][−0.53, −0.07][−0.33, +0.01][−0.13, +0.14]
N4 (boundary)+0.002−0.064+0.021−0.031+0.120
 [−0.026, +0.045][−0.26, +0.00][−0.03, +0.04][−0.12, +0.09][−0.06, +0.28]
Table 3: Corrected grid, last probe (generation 550), connectome minus control: median over 10 seeds, with the 90 % interval below. Equivalence is assessed only on seed means across ecologies.
Figure 2: Intervention. Left and middle: fitness trajectories in the two ecologies (median and interquartile range over 10 seeds). Right: differences from the connectome at the last probe, seed means over both ecologies, with 90 % intervals. Adding shortcut-creating rewiring (AS) reproduces the control’s advantage; removing it from the degree-preserving control (N2R) removes the advantage where it was clear.
Figure 3: Graded shortcut transplant. Left and middle: fitness at the last probe against the share of olfactory output landing directly on descending groups, per ecology; grey lines are seeds, purple the mean with standard error, and the green square the interior-only sham (101 swaps) plotted at the connectome’s own share. Right: fitness lost when olfaction is blocked, against the same axis.
Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

codeoperationpreserved
N1permute the row positions of each presynaptic column’s non-zerosout-degree, Dale sign, per-source weight multiset
N2Maslov–Sneppen target swaps (Maslov & Sneppen, 2002), 10·|E| attemptsin- and out-degree, weight–source pairing, sign
N3permute entries within each (super-class × hemisphere) block pairconnections between block pairs, block-level laterality
N4Maslov–Sneppen swaps restricted to interior edgesas N2, plus every edge leaving a sensory group and every edge entering a descending group
N5column shuffle restricted to interior edgesas N1, plus the same boundary edges
Appendix
Table 7
whenwhateffect on reported results
exploratory phase, before PROTOCOL.md27 changes to brain construction, calibration and world code (for example: mandatory inclusion of the gustatory-to-dopaminergic path groups; replacement of clustered Kenyon cells by individually wired ones; removal of an energy loophole that let agents survive without eating; a retracted learning-specificity result traced to a numerically unstable calibration)none of the frozen experiments were run before these changes
after PROTOCOL_ECO.md was frozen, before any runthe setting meant to remove the predator instead enabled an adaptive predator; fixed and re-frozen before the first gridnone
after the first gridstored genomes were population samples rather than elites, by an indexing erroraffects only the post hoc assays of Appendix H , as stated there
after the first grid (AI-assisted review of draft 1)one-sided heading noise, Appendix Dthe first grid’s generation-0 result is not used; all later experiments corrected
after the first gridthe standard nulls carry olfactory-to-motor shortcuts, Section 4.1motivated the corrected grid and the intervention
Appendix
Table 8
Figure 4: First grid: common-garden fitness over 600 generations in the four ecologies, median and interquartile range over 10 seeds. This experiment was run before the noise correction (Appendix D ) and uses controls that carry sensory-to-motor shortcuts.
Figure 5: Corrected grid: connectome, shortcut-carrying controls (N1, N2) and boundary-preserving controls (N4, N5).
cellcontrolmedian90 % CIsmallest boundδ = 0.050.0750.100.150.20
seed meanN4+0.002[−0.026, +0.045]0.045yesyesyesyesyes
seed meanN5−0.074[−0.096, +0.002]0.096nonoyesyesyes
P0T0N4−0.064[−0.255, +0.002]0.255nonononono
P0T0N5−0.264[−0.368, −0.179]0.368nonononono
P1T0N4+0.021[−0.026, +0.038]0.038yesyesyesyesyes
P1T0N5+0.067[−0.067, +0.090]0.090nonoyesyesyes
Appendix
Table 11
conditionbetween-group SDsilent groupsKenyon-cell mean
A0.11870.050
N10.11830.050
N20.1186.50.050
N40.1184.50.050
N50.1184.50.050
Appendix
Table 12
contrastglobal (registered grid)distribution-matchedverdict
A − N4+0.002 [−0.026, +0.045]+0.055 [+0.035, +0.097]EQUIVALENT
A − N5−0.074 [−0.096, +0.002]−0.009 [−0.076, +0.046]EQUIVALENT
A − N1−0.224 [−0.337, −0.103]−0.274 [−0.306, −0.123]BEHIND
A − N2−0.199 [−0.321, −0.133]−0.146 [−0.231, −0.037]BEHIND
Appendix
Table 13
controlchangeverdict
N1+0.010 [−0.112, +0.072]UNDETERMINED
N2+0.101 [+0.016, +0.133]CHANGED
N4+0.062 [−0.002, +0.117]UNDETERMINED
N5+0.076 [−0.017, +0.128]UNDETERMINED
Appendix
Table 14
ecologycontrolglobaldistribution-matched
P0T0N1−0.621 [−0.661, −0.469]−0.437 [−0.563, −0.235]
P0T0N2−0.403 [−0.421, −0.308]−0.478 [−0.677, −0.052]
P0T0N4−0.064 [−0.255, +0.002]+0.139 [−0.128, +0.230]
P0T0N5−0.264 [−0.368, −0.179]−0.066 [−0.320, +0.124]
P1T0N1−0.312 [−0.441, −0.197]−0.489 [−0.664, −0.275]
P1T0N2−0.338 [−0.530, −0.074]−0.164 [−0.483, −0.013]
Appendix
Table 15
ecologycontrolseeds 0–9seeds 10–19n = 20p_holm (n = 20)inside ±0.10 (R1)
P0T0N1−0.621 [−0.661, −0.469]−0.304 [−0.513, −0.161]−0.491 [−0.624, −0.298]0.001no
P0T0N2−0.403 [−0.421, −0.308]+0.100 [−0.068, +0.332]−0.214 [−0.375, +0.042]0.248no
P0T0N4−0.064 [−0.255, +0.002]+0.090 [−0.109, +0.217]−0.011 [−0.114, +0.080]0.841no
P0T0N5−0.264 [−0.368, −0.179]+0.031 [−0.068, +0.152]−0.104 [−0.208, +0.016]0.265no
P1T0N1−0.312 [−0.441, −0.197]−0.417 [−0.522, +0.067]−0.333 [−0.482, −0.197]0.004no
P1T0N2−0.338 [−0.530, −0.074]−0.101 [−0.335, +0.103]−0.203 [−0.430, −0.032]0.046no
Appendix
Table 16
controlmedian [90 % CI]seeds with connectome aheadinside ±0.10
N1−0.401 [−0.497, −0.241]3/20no
N2−0.155 [−0.273, −0.060]4/20no
N40.000 [−0.087, +0.083]10/20yes
N5−0.059 [−0.093, +0.018]7/20yes
Appendix
Table 17
quantity (seed mean over P0T0 and P1T0, n = 10)mean90 % CIregistered rule
P: AS10 − BS10, last probe+0.600[+0.414, +0.797]lower bound > +0.10 → shortcut-specific
R: AS10 − BS10, fresh worlds (4 × 8 lives, generation-599 elites)+0.639[+0.490, +0.792]same category as P → robust
S2: AS10 − BS10, P0T0 / P1T0+0.568 / +0.631[+0.315, +0.832] / [+0.428, +0.831]descriptive
S3: BS10 − A+0.007[−0.070, +0.086]±0.10 reference: inside
S4: olfaction-ablation cost AS10 / BS101.175 / 0.354difference +0.821 [+0.563, +1.079]descriptive
S5: AS10 − A+0.606[+0.435, +0.798]lower bound > 0 → replicated
Appendix
Table 18
endpointglobalpergroupdistmatch (post hoc)
generation 0A > N1, N2, N3 (10/10, 10/10, 10/10), p_holm 0.006A > all (9/10, 10/10, 10/10), p_holm 0.006–0.008A > all (10/10, 9/10, 10/10), p_holm 0.006
olfaction ablationA −0.03 vs +0.31 / +0.34 / +0.30; p_holm 0.006–0.014A −0.08 vs +0.16 / +0.18 / +0.22; p_holm 0.006–0.008A −0.10 vs +0.25 / +0.32 / +0.28; p_holm 0.006–0.008
vision ablationA +0.30 vs +0.13 / +0.14 / +0.07; p_holm 0.06–0.11A +0.33 vs +0.09 / +0.13 / +0.17; p_holm 0.041A +0.47 vs +0.17 / +0.13 / +0.05; p_holm 0.012
generation 275no difference vs N1, N2 (p_holm 1.0); vs N3 p_holm 0.059no difference vs N1, N2; vs N3 p_holm 0.111no difference vs N1, N2; vs N3 p_holm 0.252
AUC (primary)A > N1, N3 (p_holm 0.006); N2 n.s.A > N3 only (p_holm 0.012)A > N3 only (p_holm 0.012)
Appendix
Table 19
Figure 6: Common-garden fitness over 300 generations in paper-v1, per calibration scheme (median and interquartile range over 10 seeds).
Figure 7: Paper-v1 endpoints per seed: generation 0, AUC (primary), generation 275, and ablation costs.
Figure 8: Circuit metrics before evolution under three calibration schemes.
Figure 9: Generation-250 samples of paper-v1 tested in harder worlds without further evolution.

Explore similar work

CardsList
Sep 29, 2026cs.NE

How much of fly walking is written in the wiring?

Connectome models of the fly nerve cord generate walking-like motor rhythms, but oscillation alone does not show that the specific wiring matters. Here we provide, to our knowledge, the first test of which features of motor output depend on the specific wiring. We simulated the leg motor systems of two independent Drosophila connectomes, with synapse counts as fixed weights and glutamatergic synapses treated as inhibitory, and compared each with six families of rewired networks that preserve progressively more of its structure, using pre-registered criteria. We find that rhythm is generic but antagonist coordination is not: many rewired networks were more rhythmic than the real ones, yet the real wiring coordinated antagonistic motor pools more strongly than every rewired network, most of all at the thorax--coxa joint. We trace this specificity to how premotor input is allocated between antagonistic pools. Both connectomes carry Sherrington's reciprocal innervation---neurons that excite one pool inhibit its antagonist---and no rewired network does. Reassigning premotor inputs between the pools abolished coordination even when motor neurons' typical input strength changed little (all pre-registered criteria met in one connectome; same direction in the other). Coordination, not rhythm, therefore reveals whether a connectome model's wiring matters.

Isabel Guan, Yuntian Zhao, Dingyuan Zhang +2

The Hong Kong University of Science and Technology · ZENBOT · Nanyang Technological University +1

Neural CircuitsNeuronsMusculoskeletalWiring
AskPDF
Jun 5, 2026cs.NE

The Whale That Outswam Evolution: Swarm Intelligence Maximises Memory in Connectome Reservoirs

Reservoir computing exploits the fixed dynamics of a recurrent network for temporal processing, requiring only a trained linear readout. Biological neural connectomes, shaped by millions of years of evolution, may encode computational structure beyond what random reservoirs provide, yet whether that structure can be further enhanced by principled optimisation remains an open question. We address it by applying four gradient-free, bio-inspired optimisers (Particle Swarm Optimisation, Differential Evolution, Grey Wolf Optimiser, and Whale Optimisation Algorithm) to the edge weights of connectome-based echo-state networks across six species spanning six orders of magnitude in neural complexity: C. elegans (279 neurons), Drosophila (49 nodes), mouse (112), rat (73), macaque (29 regions, continuous FLNe synaptic strengths), and human structural MRI connectivity (83 parcels). Each connectome is evaluated on four canonical reservoir computing benchmarks: Memory Capacity (MC), Lorenz attractor prediction, NARMA-10 system identification, and Mackey-Glass chaotic time-series prediction. All four optimisers consistently outperform unoptimised biological baselines across every task and species when initialised from biological weights. WOA achieves the largest gains on every task: up to a 17x MC improvement (C. elegans: 1.39 to 23.91) and up to 89% NRMSE reduction (Mackey-Glass, human), corresponding to an average 214% improvement across all species and tasks. Crucially, random initialisation on the same topology reliably underperforms biology, establishing biological weight values as an essential inductive bias that topology alone cannot recover. These results position bio-inspired, biologically-initialised optimisation as a principled and broadly effective strategy for connectome reservoir computing across the animal kingdom.

Anmol Guragain, Savvas Kakalis, Juan Ignacio Godino-Llorente

ETSI de Telecomunicación, Universidad Politécnica de Madrid, Madrid, Spain

Reservoir ComputingConnectomeSwarmsRobust OptimizationEvolutionary Optimization Methods
AskPDF
Jun 21, 2026cs.RO

FLYNN: Robust Neural Network for Robot Navigation using Fly Brain Topology

While deep learning models achieve state-of-the-art performance in complex tasks, they remain brittle when faced with new environments or sensory deprivation. In contrast, biological systems exhibit remarkable tolerance to these challenges. We address this vulnerability by developing a recurrent neural network (RNN) whose architecture is directly derived from the synaptic-resolution brain connectome of the fruit fly Drosophila melanogaster. We demonstrate the feasibility of training the fly connectome neural network (FLYNN) to perform vision-based navigation in MuJoCo, achieving performance comparable to modern hand-crafted networks of similar parameter counts. Crucially, FLYNN exhibits superior resistance to out-of-distribution (OOD) data and tolerance to sensory loss without further training. It remained functional even under total vision loss while hand-crafted networks largely failed, even when specifically trained with camera dropout. Principal Component Analysis (PCA) of the internal state of FLYNN suggests that it exhibits a particularly high degree of representational modularity, which might be related to its robustness. Our work provides a new direction for designing resilient artificial agents following the topology of biological brains.

Benquan Wang, Jingdao Chen

Department of Computer Science Engineering, Mississippi State University, Mississippi State, MS 39762, USA

Brain-Inspired Synergistic FrameworkRecurrent Neural NetworksHuman Connectome ProjectSensorimotor ExperienceRobot NavigationTopology
AskPDF