fernandofernandes/fly-connectome-malecns-166k

Fly Connectome — MaleCNS 166k

The whole fruit fly central nervous system, as plain typed arrays. 166,700 neurons, 25,582,938 directed connections, 124,177,617 synaptic contacts — raw signed weights, nothing normalized, in a destination-major CSR you can np.fromfile and use in about four lines.

This is the complete graph. If you want only the 49,393-neuron central-brain subset that the Fly LLM language models run on, with measured soma positions, that is fly-connectome-49k. The two use the same conventions and compose.

Measured biological connectivity. Nothing here was learned or fitted.

Contents — 209.5 MB

15,768,940 excitatory and 9,813,998 inhibitory edges, none zero. Mean in-degree 153.5, median 98, maximum 11,526. Weight magnitudes run from 0 to 712.5 with a median of 0.55.

The largest superclasses: ol_intrinsic 89,403 · cb_intrinsic 32,164 · vnc_intrinsic 13,161 · visual_projection 9,201 · vnc_sensory 6,370 · ol_sensory 6,098 · cb_sensory 4,868 · ascending_neuron 1,846 · descending_neuron 1,314 · vnc_motor 708.

Load it

import numpy as np
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download('fernandofernandes/fly-connectome-malecns-166k', repo_type='dataset'))
offsets = np.fromfile(root / 'graph/edges_offsets.i32', dtype=np.int32)   # 166,701
source  = np.fromfile(root / 'graph/edges_source.i32',  dtype=np.int32)   # 25,582,938
weight  = np.fromfile(root / 'graph/edges_weight.f32',  dtype=np.float32)

# Everything presynaptic to neuron 12345, as (index, signed weight) pairs:
lo, hi = offsets[12345], offsets[12346]
incoming = list(zip(source[lo:hi], weight[lo:hi]))

As a sparse matrix, W[postsynaptic, presynaptic]:

from scipy.sparse import csr_matrix
W = csr_matrix((weight, source, offsets), shape=(166700, 166700))

Drive it like a rate network — visual input into populations/retina.i32, read motor intent out of populations/descending.i32:

retina = np.fromfile(root / 'populations/retina.i32', dtype=np.int32)
descending = np.fromfile(root / 'populations/descending.i32', dtype=np.int32)

x = np.zeros(166700, dtype=np.float32)
for _ in range(steps):
    drive = np.zeros_like(x)
    drive[retina] = stimulus                      # your input, however you encode it
    x = 0.9 * np.tanh(W @ x + drive) + 0.1 * x    # your dynamics, this is only an example
motor = x[descending]

The recurrence above is an example, not part of the dataset. The connectome fixes the wiring; every dynamical choice — gains, leak, nonlinearity, timestep — is yours.

Verify it

Every file's SHA-256 is in manifest.json. `verify_connectome_package.py` runs 17 structural checks with nothing but NumPy, and works on both this dataset and the 49k subset:

All 17 checks passed: 166,700 neurons, 25,582,938 edges, 11,752 cell types, 4 named populations.

What the weights mean, precisely

A weight is a signed synaptic count: magnitude from the flat connectome at minimum confidence 0.5, sign from the neurotransmitter prediction. It is not a conductance, not a measured physiological strength, and not calibrated across cell types. No normalization of any kind has been applied — normalization_applied is false in the manifest, and any scaling you need is yours to choose and to record.

Edges are not coalesced. Cell type is an empty label where the source carries no official type. The two derived/ hop arrays were computed by this project, not by the upstream release, and are -1 where a neuron is unreachable.

Soma coordinates are not included here. The 49k subset carries measured soma positions for its neurons.

Provenance

Assembled from the three publicly hosted MaleCNS v1.0 flat-connectome files, via Doomfly at commit 71ecf53d:

Full URLs and digests are in manifest.json under provenance.upstream_files, so you can rebuild this package from the original release and check it against ours.

Licence and attribution

CC BY 4.0. Connectome data: FlyEM / HHMI Janelia Research Campus, University of Cambridge, MRC Laboratory of Molecular Biology, and Google Research. Packaging code MIT.

Used by fly-wordbrain; hear the 49k subset write and read a story aloud at Fly Recital.