A Worm Taught an Algorithm About Flies. It Was Right.
We trained a generative model on a 302-neuron worm brain. Then we pointed it at a 3036-neuron fly brain it had never seen. It correctly predicted which neurons would be hubs -- with cosine similarity above 0.94. Pre-registered, confirmed.
The Experiment
In topology research, a hub node is a node with disproportionately high connectivity -- the neuron every other neuron talks through. In neural connectomes, hubs are implicated in signal integration, resilience, and disease vulnerability.
The question we pre-registered: if you compress the structure of a small connectome into a generative model, does that compression capture something universal -- something that transfers to a completely different organism?
We trained on C. elegans: 302 neurons, fully mapped, the most studied connectome in biology. Then we applied the learned model to Drosophila larva: 3,036 neurons, a brain 10x larger, a different phylum.
The model had never seen the fly. It only knew the worm.
The Method
IRDME's core algorithm fits a multilayer graph model to a real network and outputs a compressed generative representation -- a set of parameters that can reproduce the network's structural statistics.
For this experiment, the compression was fit to the C. elegans connectome (chemical synapse + gap junction layers). The resulting parameter set was then used to generate a synthetic network at Drosophila scale (n=2,952 after matching the connected component).
We measured hub transfer using rank-vector cosine similarity: sort both networks' nodes by degree, take the top-k rank vectors, compute cosine. This is a distribution-level metric -- it asks whether the shape of the hub distribution is preserved, not whether specific named neurons match (which would be meaningless across species).
The Result
The pre-registered threshold was cosine >= 0.85 on all three layer comparisons.
| Layer comparison | Cosine | Verdict | |---|---|---| | Chemical synapse (h1) | 0.985 | CONFIRMED | | Gap junction (h2) | 0.908 | CONFIRMED | | Combined (h3) | 0.947 | CONFIRMED |
3/3 CONFIRMED. Average cosine: 0.947.
The worm's compression told the algorithm something true about the fly.
Why This Matters
This result is not about C. elegans. It's not about Drosophila. It's about the geometry of biological connectivity.
If hub structure is transferable across species -- if the same generative rules that describe a 302-neuron worm also approximate the hub distribution of a 3,036-neuron fly -- then biological connectomes may share a deeper structural grammar than their surface diversity suggests.
For network science, this implies that generative compression is not just a compression technique. It may be capturing conserved biological design principles.
For medicine, it raises a question: if hub geometry is conserved, are hub vulnerabilities also conserved? Could a worm model predict which human neurons are most fragile?
The Pre-Registration
This result was pre-registered before analysis at the public preregistrations repository (commit timestamp: May 26, 2026, 24+ hours before the run). The prediction, threshold, and metric were locked. We did not adjust anything after seeing the data.
The pre-registration hash is cd2ed080 (full SHA256 in the repository). The analysis was run May 27, 2026.
This matters because a cosine of 0.985 is an extraordinary number. Without pre-registration, a skeptic would be right to ask: did you tune the method after seeing results? The answer here is verifiably no.
What's Next
F14 will test the same transfer on adult Drosophila (FlyWire connectome, ~130,000 neurons). If the transfer holds at that scale -- from a 302-neuron worm to a brain 400x larger -- the case for conserved hub geometry becomes much harder to dismiss.
We are also exploring whether the compression parameters themselves encode interpretable biological quantities, or whether the structure is opaque. Early signs suggest the layer-interaction terms correlate with known anatomical properties -- but that is a separate pre-registration.
All code, data, and pre-registration files are publicly available. The generative model is available via the IRDME calculator.