Reviving a model fly brain: anoxia and resuscitation in a connectome simulation

Bottom line: In a model of the whole adult fly brain (139,255 neurons, FlyWire FAFB v783), switching oxygen back on abruptly after a 180-second outage traps the network in a self-sustaining high-activity state, and its stimulus-response identity does not return. A 20-second total activity block at reperfusion resets it, and the identity returns.

This is a simplified simulation: leaky integrate-and-fire neurons on a real wiring diagram, with a phenomenological K⁺/ATP layer whose constants were not fitted to Drosophila. It is a source of testable predictions, not evidence about living tissue. No real brain was revived.


Author	Srihari Muralikrishnan (ORCID 0009-0004-0220-0361), Kalaris Labs
Published	2026, preprint with code
Evidence	Results come from simulation.
Data	FlyWire Codex FAFB v783, MCNS v1.0, BANC v888, MANC v1.2.1
License	Code MIT; connectome data under the FlyWire principles
Theme	Cognition and neuroscience

Paper and code

Paper, supplement, simulator, experiment definitions and figure scripts.

Agent tooling used

The study used Kalaris Labs Myriad and Research Agent Skills.

How the model works

Key results

#	Result	Evidence	Confidence
1	After 180 s anoxia, abrupt reperfusion traps the model in a self-sustained runaway	48/48 FAFB parameter sets; 5/5 MCNS runs	High
2	After 30 s anoxia, the trap depends on whether antennal-lobe local neurons are inhibitory	FAFB 0/3 runaway; MCNS FII_norm 1.000	Moderate
3	A 20 s total block at reperfusion restores the stimulus-response map; 2–5 s blocks fail	FAFB FII_norm 1.001 / 0.953 at 30 / 180 s	High
4	Releasing the block as soon as K⁺ and ATP normalise (11 s) is not enough	MCNS 5/5 runaway	Moderate
5	The trapped state is ionic, not synaptic	109/134 Hz with synapses off; 2.4/2.3 Hz with K⁺ clamp	High (model)
6	Model brain death needed near-total neuron death	All 92 such runs had ≥ 97.9% dead	High (model)
7	Protecting the top 1% most-connected neurons is far more efficient than whole-brain protection	~30× FAFB; 84× MCNS	High
8	Graded reperfusion did not beat abrupt reperfusion plus a block	Pre-registered H3 falsified	High (model)
9	Hub-targeted neuron loss is far more destructive than random loss	FAFB f_c 0.800 random vs 0.033 hub-targeted	High (direction)

Limitations

The model cannot represent oxidative stress, inflammation, glia, the blood-brain barrier, synaptic plasticity or memory storage. “Identity” means the stimulus-response map of a fixed wiring diagram, not memories or experience. The repository includes a proposed real-fly validation protocol (optogenetic silencing at reoxygenation, proboscis-extension and escape readouts).

Reproduce

git clone https://github.com/KalarisLabs/fly-brain-resuscitation.git
cd fly-brain-resuscitation
conda env create -f environment.yml && conda activate flybrain
pytest -q

Connectome tables are not redistributed. Download them from FlyWire Codex.

The Felt Edge: As machines commoditise processing, recall and problem solving, this paper argues that felt metacognition is the residual human edge, and that it is trainable.
Inventing the Variable: The HiddenLaw benchmark tests whether AI can introduce a variable its vocabulary lacks. Passive data blocks every learner; with interventions, Claude Opus 5.5 succeeds in 9 of 9 worlds.

Cite this study

@misc{muralikrishnan2026flybrain,
  title  = {Anoxia, brain death and resuscitation in a connectome-constrained simulation of the adult Drosophila brain},
  author = {Muralikrishnan, Srihari},
  year   = {2026},
  note   = {Preprint, Kalaris Labs},
  url    = {https://github.com/KalarisLabs/fly-brain-resuscitation}
}

Research

Self-improving agents

Evaluation and scientific discovery

Agent economies

Efficient inference

Cognition and neuroscience

Reviving a model fly brain: anoxia and resuscitation in a connectome simulation

Paper and code

Agent tooling used

How the model works

Key results

Limitations

Reproduce

Cite this study

Paper and code

Agent tooling used

How the model works

Key results

Limitations

Reproduce

Related research

Cite this study