Yes, far above chance. A classifier trained only on the male wiring ranks the 478 annotated sex-related cell types with a cross-validated AUC-PR of 0.759 against 0.041 by chance, and none of 500 randomized wirings matched it (p = 0.002). Most of that signal turns out to be anatomical location.
_{Dhruvin Sarkar. An independent analysis of public connectome data, not peer reviewed.}

Illustrated findings Atlas Circuits Guess the Neuron Technical report Poster

Abstract

HHMI Janelia and Google Research have released the first complete wiring diagram of an adult male fruit fly's central nervous system, about 166,700 neurons and 125 million synapses, and found that roughly 6% of the cell types matched between the male and female brains are dimorphic or sex-specific.¹ We asked whether that difference leaves a detectable signature in the male wiring alone, using no features from the female brain. A gradient-boosted classifier² ranks the 478 types annotated dimorphic or male-specific far above chance, with a cross-validated AUC-PR of 0.759 against 0.041. None of 500 degree-preserving randomized wirings reached that score (p = 0.002). Two types annotated isomorphic score in the top 1% and are flagged as candidates for further investigation. As a check on the graph construction, the well-established Giant Fiber escape circuit is recovered from the wiring alone.

Results

Important

These are statistical associations between a type's place in the male wiring diagram and an existing annotation. They are correlational and say nothing about causes or behavior. Neuropil location carries most of the signal: neuropil features alone reach AUC-PR 0.722, and randomized wirings that keep every non-topology feature still score 0.712.

Pre-specified checks

• Both null-model tests are significant at the Bonferroni-corrected level of 0.025 (p = 0.002 each)
• The strongest LPLC2 to TTMn route passes through DNp01, the Giant Fiber
• At least one of AOTU008, AOTU012, DNg13 or LoVP92 ranks in the top 20 (they rank 531, 591, 572 and 574; see where it falls short)

All headline values

	value
out-of-fold AUC-PR (chance)	0.759 (0.041)
ROC AUC	0.952
precision in top 100	0.99
randomized graphs, all features	0.712 ± 0.005, 0 of 500 ≥ real, p = 0.002
topology only, real vs randomized	0.489 vs 0.070 ± 0.004, p = 0.002
static features only	0.726
hemilineage-grouped cross-validation	0.701
male-specific / dimorphic types (AUC-PR)	0.832 / 0.376

_{Source: results/classifier_metrics.json, results/null_model_summary.json}

Figure 1. The annotated sex-related types rise to the top of the ranking. Top: all 11,751 cell types ordered by out-of-fold probability, with male-specific types in green and dimorphic types in cyan; below it, the top 10% enlarged. Bottom: precision against recall for the full model and three comparisons: static features only, hemilineage-grouped folds and topology features only. Out-of-fold means every type was scored by a model that never saw its label. The top 1% (118 types) holds 116 sex-related types, and the top 10% holds 85% of all 478.

Values behind Figure 1

flagged	types	sex-related among them	precision	share of all 478 sex-related types
top 1%	118	116	0.98	24%
top 5%	588	350	0.60	73%
top 10%	1,175	405	0.34	85%

folds	fold 1	fold 2	fold 3	fold 4	fold 5	pooled AUC-PR
stratified by label	0.743	0.753	0.772	0.787	0.750	0.759
grouped by hemilineage (1,124 groups)	0.683	0.709	0.751	0.627	0.746	0.701

class	types	AUC-PR	chance
male-specific	312	0.832	0.027
dimorphic	166	0.376	0.015

Each class is scored against isomorphic types only, using the same out-of-fold probabilities.

_{Source: results/model_diagnostics.md, results/classifier_metrics.md}

The real wiring against 500 randomized wirings

Figure 2. No randomized wiring reaches the real score. Each randomized graph keeps every cell type's exact in-degree, out-degree and outgoing synapse total, and shuffles its partners.³ All nine topology features were recomputed on each graph, and the identical model was retrained with identical folds. p-values are one-sided and empirical.⁴ Because no randomized graph reached the real score, p = 0.002, the smallest value 500 graphs can give. It is a floor, not a precise estimate.

	all 89 features	topology only (9)
real wiring	0.759	0.489
randomized, mean ± sd	0.712 ± 0.005	0.070 ± 0.004
randomized, range	0.696 to 0.725	0.060 to 0.082
randomized at or above real	0 of 500	0 of 500
z	9.6	111
empirical p	0.002	0.002

All 500 scores are in results/null_model_scores.csv, which GitHub shows as a searchable table.

The two tests answer different questions. Static features alone reach 0.726, and the randomized wirings keep them, so for the full model real wiring adds a small but consistent increment on top of location and transmitter. The topology-only test is where the gap is wide: the real graph scores seven times higher than randomized graphs with the same degrees.

Where the signal sits

Figure 3. Location carries most of the signal; wiring community carries most of the rest. Top: the share of mean absolute SHAP attribution⁵ by feature group (neuropil 68%, topology 29%, transmitter 3%). Bottom: AUC-PR of the same classifier trained on feature subsets. Filled marks were specified before any randomized graph was scored; open marks were run afterwards and only describe where the signal is. The largest single attribution is Leiden community membership:⁶ community 2 holds 256 of the 312 male-specific and 75 of the 166 dimorphic types, and its output centers on SMP, SIP and CRE.

Values behind Figure 3

features	count	AUC-PR	ROC AUC	precision in top 100	status
all features	89	0.759	0.952	0.99	pre-specified
all except community	88	0.740	0.945	1.00	exploratory
static only	80	0.726	0.946	0.98	pre-specified
neuropil only	79	0.722	0.945	0.98	exploratory
topology and transmitter	10	0.503	0.902	0.83	exploratory
topology only	9	0.489	0.894	0.78	pre-specified
community only	1	0.164	0.818	0.22	exploratory
topology without community	8	0.123	0.783	0.15	exploratory
transmitter only	1	0.049	0.559	0.05	exploratory

The 15 features with the largest mean absolute SHAP value (log-odds), each type explained by the fold model that scored it:

feature	group	mean absolute SHAP	mean SHAP, sex-related	mean SHAP, isomorphic
Leiden community	topology	0.828	+1.485	−0.101
output share in PLP	neuropil	0.462	+0.307	−0.003
output share in SIP	neuropil	0.455	+1.146	−0.026
output share in SLP	neuropil	0.367	+0.152	+0.004
in-strength	topology	0.352	+0.192	+0.007
output share in unassigned central brain	neuropil	0.267	+0.232	−0.020
in-degree	topology	0.264	+0.059	+0.002
PageRank	topology	0.250	−0.006	−0.005
output share in SCL	neuropil	0.227	+0.382	−0.053
output share in SPS	neuropil	0.218	+0.111	+0.002
output share in WED	neuropil	0.205	+0.063	−0.004
predicted transmitter	transmitter	0.203	+0.061	−0.001
output share in LegNp(T1)	neuropil	0.195	+0.062	−0.002
output share in AVLP	neuropil	0.192	+0.403	−0.022
output share in unassigned ventral nerve cord	neuropil	0.188	+0.043	−0.033

_{Source: results/model_diagnostics.md}

A circuit recovered from the graph

Figure 4. The Giant Fiber route, and the best route left without it. Each connection costs the negative log of its share of the upstream type's output synapses, so the cheapest route has the largest product of output shares. The validation pair was fixed before the search. The route from the looming detector LPLC2 to the jump motor neuron TTMn runs through DNp01, the Giant Fiber.⁷ Deleting DNp01 leaves a route twice as long, with a product about 190 times smaller (9.45 × 10⁻⁴ to 4.92 × 10⁻⁶). This describes redundancy in the wiring diagram, not what a fly without a Giant Fiber would do. Electrical synapses, which carry part of the Giant Fiber's output, are not in the data.

All three routes before and after removal

DNp01 is the only intermediate type on the LPLC2 → TTMn route (probability 0.024), so it is the type removed. Deleting it disconnects none of the three routes.

route	intermediates, intact	intermediates, DNp01 removed	output-share product
LPLC2 → TTMn	DNp01 (2 hops)	DNp103, IN07B054, IN21A027 (4 hops)	9.45 × 10⁻⁴ → 4.92 × 10⁻⁶
LC4 → TTMn	DNp01 (2 hops)	DNp11, IN21A026 (3 hops)	1.64 × 10⁻³ → 1.55 × 10⁻⁴
LPLC2 → DLMn c-f	DNp01, IN18B034 (3 hops)	LPLC4, DNp31 (3 hops)	1.39 × 10⁻⁴ → 2.83 × 10⁻⁵

Synapses on each connection, in route order:

route	types in order	synapses
LPLC2 → TTMn, intact	LPLC2, DNp01, TTMn	4,862; 90
LPLC2 → TTMn, DNp01 removed	LPLC2, DNp103, IN07B054, IN21A027, TTMn	5,065; 402; 75; 313
LC4 → TTMn, intact	LC4, DNp01, TTMn	6,362; 90
LC4 → TTMn, DNp01 removed	LC4, DNp11, IN21A026, TTMn	3,666; 112; 507
LPLC2 → DLMn c-f, intact	LPLC2, DNp01, IN18B034, DLMn c-f	4,862; 75; 1,233
LPLC2 → DLMn c-f, DNp01 removed	LPLC2, LPLC4, DNp31, DLMn c-f	3,476; 2,022; 1,128

_{Source: results/ablation_report.md}

Candidates for further investigation

Figure 5. The two annotated isomorphic types the classifier scores highest. The rule was fixed in advance: a type annotated isomorphic whose out-of-fold probability is in the top 1% (0.993 or higher). Two qualify, each of two neurons. For both, the largest contributions are membership of wiring community 2 and output shares in SIP and EPA. fru and dsx were not model inputs, yet both carry a fru annotation, as do 587 of the other 11,271 isomorphic types (one-sided Fisher exact test, p = 0.0027). fru-expressing neurons are common in the circuits the model keys on, so this is not independent confirmation.

cell type	probability	rank of 11,751	fru/dsx	sex-related partners
CL062_b3	0.995	84	fru_low (2 of 2)	63%
CL062_b2	0.993	111	fru_low (2 of 2)	69%

_{Sex-related partners is the share of a type's strong input and output partners in the graph that carry a sex-related annotation. Source: results/candidates.md}

A high score means a type is wired like known sex-related types. It is not evidence of dimorphism; that needs the female connectome or light-microscopy anatomy.

Where it falls short

Caution

One pre-specified check fails. It required at least one of four sex-related types named in public announcements to rank in the top 20: AOTU008 (the example in Google Research's announcement), AOTU012, DNg13 or LoVP92. They rank 531, 591, 572 and 574. The check is reported unchanged, and make checks exits non-zero while it fails.

Figure 6. Dimorphic types are hard, the named types rank near 550, and the probabilities overstate the rate. Dimorphic neurons exist in both sexes and differ in arbor or partners. That is harder to see from one sex than membership of a male-specific cluster. The probabilities rank types well, but the mean predicted value is 0.074 against a prevalence of 0.041. Brier score 0.0244, against 0.0390 for predicting the prevalence.

• Label provenance: no feature comes from the female brain, but the labels themselves come from Janelia's comparison of the male and female connectomes.
• Anatomy carries most of the signal: neuropil output shares alone reach AUC-PR 0.722; real wiring adds a small but consistent increment on top.
• Lineage: keeping each hemilineage inside one fold lowers AUC-PR to 0.701, so part of the score reflects sibling types that share a lineage.
• Scope: one animal, at cell-type resolution, with chemical synapses only.
• Fixed choices: the 1% edge threshold and the hyperparameters were fixed before evaluation; other choices would change the topology features.
• Precision of p: with 500 randomized graphs, p = 0.002 is a floor rather than a precise estimate.
• Label confidence: the labels include lower-confidence "potentially" annotations.
• Review: none of this has been peer reviewed.

Calibration and named types

Types in ten equal-count bins of out-of-fold probability:

probability range	types	mean probability	observed sex-related rate
0.00003 to 0.00055	1,176	0.000	0.000
0.00055 to 0.00125	1,175	0.001	0.001
0.00126 to 0.00237	1,175	0.002	0.003
0.00237 to 0.00413	1,175	0.003	0.002
0.00413 to 0.00723	1,175	0.006	0.006
0.00724 to 0.01292	1,175	0.010	0.003
0.01293 to 0.02375	1,175	0.018	0.009
0.02376 to 0.05157	1,175	0.035	0.009
0.05158 to 0.17483	1,175	0.095	0.029
0.17506 to 0.99814	1,175	0.569	0.345

named cell type	rank	probability	label
AOTU008	531	0.569	dimorphic
DNg13	572	0.520	dimorphic
LoVP92	574	0.518	male-specific
AOTU012	591	0.502	dimorphic

_{Source: results/model_diagnostics.json, results/model_diagnostics.md}

Explore the site

The live site is an illustrated version of everything above. Each image opens the view it shows.

Atlas: every cell type and neuropil, with the classifier score, SHAP reasons, synapse locations, graph-position percentiles, strongest partners, and a reconstructed neuron of the type drawn inside the stained brain.

_{Looking up CL062_b3: probability 0.995, rank 84 of 11,751, drawn inside the stained brain next to the features that raised its score. A candidate for further investigation, not evidence of dimorphism.}

Circuits: 17 curated routes between sensory, courtship, descending and motor cell types, with node removal.

_{Removing DNp01 in the findings article's Figure 12. The best remaining route takes four hops instead of two and the product of output shares falls from 9.45 × 10⁻⁴ to 4.92 × 10⁻⁶. The Circuits page offers the same removal for all 17 routes. A structural result, not a behavioral one.}

• Findings: a long-form illustrated article with 12 interactive figures: the type census, a wiring similarity map, a precision-recall explorer, a degree-preserving rewiring demo with the null distributions, a SHAP attribution beeswarm, the feature-set comparison, wiring communities, neuropil agreement, topology distributions, a limitations panel, the candidate types with 3D skeletons, and the Giant Fiber circuit with node removal.
• Guess the Neuron: a 10-round game that pairs a sex-related type with an isomorphic type of the same superclass and dominant neuropil, answered with a tap or the A and B keys.
• Methods: the full technical report on the web, with tables of all features, classifier settings, folds, the null model, calibration, all node removals, the pre-specified checks, limitations, software versions and references.

_{Findings, Guess the Neuron and Methods at phone width.}

Methods

1. Data. The male CNS connectome, male-cns:v1.0, queried through neuPrint.⁸ The database holds 176,422 neuron records, of which 164,506 carry one of 11,751 cell types. No FlyWire data, no cross-dataset matching field and no female-derived quantity is used as a model input.

2. Labels. Janelia's per-neuron dimorphism annotations are reduced to one label per type, in priority order: male-specific (including potentially male-specific), then dimorphic (including potentially dimorphic), otherwise isomorphic. That gives 312 male-specific, 166 dimorphic and 11,273 isomorphic types, so the task is binary: 478 sex-related types (4.1%) against the rest. fru and dsx expression annotations are excluded from the features, because they would leak the label.

3. Type graph. Neuron-to-neuron synapse counts between typed neurons sum to 3.83 million type pairs carrying 122.3 million synapses. An edge from type $u$ to type $v$ is kept when it supplies at least 1% of $v$'s input synapses, and self-loops are removed, which leaves 243,439 edges carrying 81.3 million synapses:

   s_{uv} \ge 0.01 \sum_{w} s_{wv}
   

   Here $s_{uv}$ is the number of synapses from $u$ onto $v$, and the sum runs over every type $w$ that synapses onto $v$. The threshold was chosen from graph statistics before any model was trained.

4. Features. Nine topology features computed from the graph: in-degree, out-degree, in-strength, out-strength, PageRank, betweenness, Leiden community membership,⁶ hop distance from the nearest sensory type and hop distance to the nearest descending or motor type. Eighty static features describe a type without reference to topology: its predicted neurotransmitter and the share of its output synapses in each of 79 bilateral neuropils.

5. Classifier. LightGBM² with hyperparameters fixed before evaluation, scored by five-fold stratified cross-validation (seed 20260914). Every probability reported is out-of-fold. The primary metric is AUC-PR, whose chance level equals the positive rate of 0.041.

6. Null model. 500 randomized graphs, each the real graph after 10 × |E| degree-preserving swap attempts.³ Every type keeps its in-degree and out-degree, and its outgoing synapse counts are shuffled across its new edges, so out-strength is preserved too. The nine topology features are recomputed on each graph and the identical model is retrained with identical folds. The empirical one-sided p-value⁴ is one plus the number of randomized graphs that score at or above the real graph, divided by 501:

   p = \frac{1 + \#\{\, b : \mathrm{AUCPR}_b \ge \mathrm{AUCPR}_{\mathrm{real}} \,\}}{1 + 500}
   

   The two tests, all features and topology features only, are each assessed at a Bonferroni-corrected 0.025.

7. Routes and node removal. Dijkstra's algorithm on an edge cost equal to the negative log of the share of the upstream type's output synapses, where $S_u$ is the total output synapses of $u$:

   c_{uv} = -\log \frac{s_{uv}}{S_u}
   

   A route's cost is the negative log of the product of its output shares, so the cheapest route has the largest product. The intermediate type with the highest classifier probability on the LPLC2 to TTMn route is then deleted, and the search runs again for three routes.

8. Candidates. A candidate is a type annotated isomorphic whose out-of-fold probability is in the top 1% of all types, a rule fixed before candidate counts were inspected. Each is explained by its three largest SHAP contributions⁵ from the fold model that scored it, and the candidates are cross-checked against fru/dsx annotations with a one-sided Fisher exact test.

Classifier settings

setting	value
trees	400
learning rate	0.03
leaves per tree	15
minimum samples per leaf	20
row subsampling	0.8, every iteration
column subsampling	0.8
L2 penalty	1
positive-class weight	the class ratio
cross-validation	five folds, stratified, seed 20260914

The hyperparameters were fixed before evaluation and never tuned. Weighting the positive class makes the probabilities rank types well but overstate how often a type is sex-related.

Software

package	version
Python	3.12.12
neuprint-python	0.6.3
python-igraph	1.0.0
LightGBM	4.7.0
scikit-learn	1.9.1
SHAP	0.52.0
navis	1.12.0
NumPy / pandas / SciPy	2.5.3 / 3.0.5 / 1.18.1

The poster

One-page summary, 3508 × 4960 pixels. Open the full-size poster

Data and outputs

Every number on this page comes from a file in this repository. These are the ones worth opening directly, with nothing installed and nothing run.

file	what it holds
web/public/data/types.json	all 11,751 cell types with their out-of-fold probability, dimorphism label, superclass, predicted transmitter, wiring community and dominant neuropils
web/public/data/explanations.json	the largest SHAP contributions behind each type's score, written out in words
results/null_model_scores.csv	the AUC-PR of all 500 randomized wirings in both tests, which GitHub renders as a searchable table
results/model_diagnostics.json	calibration bins, feature-subset scores, SHAP attributions and breakdowns by superclass and community
results/candidates.md	the two candidate types with their features, fru/dsx cross-reference and Fisher test
web/public/data/routes.json	the 17 curated routes with synapse counts, output shares and the result of every node removal

Reproduce

Requirements: Python 3.12, Node 22, GNU Make, and network access to neuPrint and Janelia's public data bucket. The paper target also needs pandoc with typst.

python -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
make reproduce   # data, model, validate, export, test and checks, in that order
make paper       # paper/report.pdf
make web         # web/dist

The stages can also be run on their own: make data (schema, labels, type graph, features), make model (classifier and the Giant Fiber route), make validate (null model, candidates, node removal, diagnostics, hero image), make export (site data), make test (unit tests) and make checks (every check script).

Set NEUPRINT_APPLICATION_CREDENTIALS to a neuPrint token to query with your account. Without it, the public dataset is queried anonymously.

The 500-graph null model is the slow step: about 1.5 hours on 12 cores. make checks exits non-zero while the named-type check above fails.

Pipeline modules and outputs

step	module	output
neuPrint schema	pipeline/inspect_schema.py	local schema notes
type labels	pipeline/ground_truth.py	312 male-specific, 166 dimorphic, 11,273 isomorphic types
type graph	pipeline/build_type_graph.py	11,751 types, 243,439 edges (each ≥1% of the target's input)
features	pipeline/compute_features.py	9 topology + 80 neuropil and transmitter features
classifier	pipeline/train_classifier.py	results/classifier_metrics.md
null model	pipeline/null_model.py	results/null_model_validation.md, plot
pathfinder	pipeline/pathfinder.py	LPLC2 → DNp01 (Giant Fiber) → TTMn
candidates	pipeline/candidates.py	results/candidates.md
node removal	pipeline/ablation.py	results/ablation_report.md
diagnostics	pipeline/model_diagnostics.py	calibration, SHAP attributions, and breakdowns by feature subset, superclass and community: results/model_diagnostics.json, summary
hero image	pipeline/render_hero.py	assets/hero.png, results/neuropil_scores.csv
README figures	pipeline/render_readme_figures.py	assets/readme/fig-*.svg
site data	export/build_static_json.py	web/public/data/
demo site	web/ (React, three.js, Vite)	GitHub Pages via .github/workflows/pages.yml

verify/ holds one check script per step, and tests/ holds the unit tests. The README figures are drawn with python -m pipeline.render_readme_figures --font-dir <folder>, where the folder holds static TTF instances of Newsreader and Atkinson Hyperlegible Next; without it, matplotlib's default fonts are used.

Credits and citation

The data are the male adult Drosophila CNS connectome from HHMI Janelia FlyEM and Google Research, released under CC-BY 4.0. Please cite the original work:

Berg S, Beckett IR, Costa M, et al. (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell 189(18):5504–5526.e15. doi:10.1016/j.cell.2026.08.015

Also see Google Research's announcement, A connectomics milestone: Mapping the complete male fruit fly brain. The data were accessed through neuPrint, with skeletons processed by navis. Citation metadata for this repository is in CITATION.cff.

Code is released under the MIT License.

Citing this work

@software{sarkar2026wired,
  author = {Sarkar, Dhruvin},
  title  = {Wired Different: predicting sexual dimorphism from male {Drosophila} connectome wiring},
  year   = {2026},
  url    = {https://dhruvin-sarkar.github.io/ConnectomeLens/},
  note   = {Independent analysis of public connectome data, not peer reviewed}
}

Footnotes

1. Berg S, Beckett IR, Costa M, et al. (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell 189(18):5504–5526.e15. doi:10.1016/j.cell.2026.08.015 ↩

2. Ke G, Meng Q, Finley T, et al. (2017). LightGBM: a highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems 30. ↩ ↩²

3. Maslov S, Sneppen K (2002). Specificity and stability in topology of protein networks. Science 296(5569):910–913. doi:10.1126/science.1065103 ↩ ↩²

4. Phipson B, Smyth GK (2010). Permutation p-values should never be zero: calculating exact p-values when permutations are randomly drawn. Statistical Applications in Genetics and Molecular Biology 9(1):39. doi:10.2202/1544-6115.1585 ↩ ↩²

5. Lundberg SM, Lee S-I (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems 30. arXiv:1705.07874 ↩ ↩²

6. Traag VA, Waltman L, van Eck NJ (2019). From Louvain to Leiden: guaranteeing well-connected communities. Scientific Reports 9:5233. doi:10.1038/s41598-019-41695-z ↩ ↩²

7. von Reyn CR, Breads P, Peek MY, et al. (2014). A spike-timing mechanism for action selection. Nature Neuroscience 17(7):962–970. doi:10.1038/nn.3741 ↩

8. Plaza SM, Clements J, Dolafi T, et al. (2022). neuPrint: an open access tool for EM connectomics. Frontiers in Neuroinformatics 16:896292. doi:10.3389/fninf.2022.896292 ↩