Research · talking flies
Flybook
Both flies are the full MaleCNS connectome, 166,700 neurons, frozen. Nothing in between is scripted. Fly A's song is its own wing motor neurons firing, and fly B hears it through its own ear neurons. We fixed the tests before running, ran the experiment three times, and report every result, including the one that failed.
Where it landed
In the most detailed brain (run 3), a threatened fly's song makes the listener react three times out of four, and its escape neurons fire three times harder than in silence.
Three runs
The brain as FLYBRAIN uses it. Smell neurons excite each other into a runaway loop and fire flat out, so an odour can add nothing.
Synapses onto sensory neurons are removed, which is how whole-brain spiking models of the fly treat them. Food now reaches its own relay neurons.
Ten times finer steps, with a 4 ms refractory period. We chose that period before running, as the one that matched the calibrated brain's resting rate without runaway neurons.
| Measure | Run 1 | Run 2 | Run 3 |
|---|---|---|---|
| The song | |||
| Song tells what happened to the singer | 41.1% · 0.30 bits · p 0.02 | 46.7% · 0.63 bits · p 0.02 | 55.0% · 0.83 bits · p 0.02 |
| Threat songs recognised | 163 / 180 | 180 / 180 | 120 / 120 |
| Same test, scrambled wiring | 23.5% · 0.01 bits · p 0.18 | 22.6% · 0.01 bits · p 0.45 | 47.1% · 0.87 bits · p 0.02 |
| Loudness alone * | 0.29 bits | 0.43 bits | 0.64 bits |
| Spectrum shape alone, loudness removed * | 0.05 bits | 0.14 bits | 0.54 bits |
| Food odour reaches its relay neurons | no · pinned at 50 Hz | yes · 28 vs 16 Hz | yes · 125 vs 4 Hz |
| The listener | |||
| Listener's brain, real song | 40.8% · 0.27 bits · p 0.02 | 45.8% · 0.41 bits · p 0.02 | 55.6% · 0.73 bits · p 0.02 |
| Listener's brain, time-shuffled song | 43.3% · 0.31 bits · p 0.02 | 40.4% · 0.49 bits · p 0.02 | 52.5% · 0.75 bits · p 0.02 |
| Listener's brain, silence | 26.7% · 0.02 bits · p 0.41 | 27.5% · 0.03 bits · p 0.25 | 21.9% · 0.03 bits · p 0.55 |
| Listener reacts after a threat song, vs silence * | 23.1% vs 16.7% · p 0.18 | 21.2% vs 13.8% · p 0.13 | 77.5% vs 17.5% · p < 0.001 |
| Listener escape firing after a threat song, vs silence * | 0.12 vs 0.09 Hz · p 0.34 | 1.12 vs 0.70 Hz · p < 0.001 | 3.95 vs 1.35 Hz · p < 0.001 |
Chance is 25% (four situations). p-values come from 50 label shuffles per decoding test, so 0.02 is the lowest possible; the reaction tests use 5,000 shuffles. Rows marked * are follow-up controls we added after seeing run 2, so they were not fixed before running. Runs 1 and 2 have 720 songs each, run 3 has 480.
The checks we set before running
This held in every run, and it got clearer each time: 0.30, then 0.63, then 0.83 bits. Threat was always the loudest and clearest. In run 3 a mate in view became partly readable too.
At 20 ms, a brain with the same neurons and scrambled connections sang nothing useful (0.01 bits), even when its wings were louder than the real brain's. At 2 ms it carries as many bits as the real brain (0.87 vs 0.83), so the test as we set it fails. The brains still differ in which situations they keep apart, a difference we noticed afterwards and report only as a description: the real brain keeps mate and threat almost completely separate, while the scrambled brain lumps them together, and it lumps food in with nothing.
We can read what fly A went through from fly B's descending neurons, even though B only heard it. In silence the same readout stays at chance, which rules out the labels leaking into the analysis.
The listener doesn't care about timing: in every run, the song with its time steps shuffled works as well as the real one. The song itself does have some shape beyond loudness, but in run 3 the threat song's strongest rhythm is 125 Hz, exactly one spike every 4 steps. That points to the model's timing grid, not fly song.
This was weak at 20 ms. In run 2 the listener's escape neurons responded (p < 0.001), but rarely enough to count as a reaction. At 2 ms, 77.5% of listeners react to a threat song, against 17.5% in silence. A mate song (26%) and a food song (6%) don't reliably trigger anything.
Real vs scrambled wiring
Rows are what really happened to fly A. Columns are what a decoder read from its wings. A clean diagonal would mean four distinct words.
The feed
Twelve flies are paired at random each round. A fly posts only when the translator hears a word in its wing song, and its partner's reaction is the comment.
Loading the feed…
Not there yet
The smell fix worked: in run 3 a food odour drives its relay neurons at 125 Hz, against 4 Hz without it. But the signal never reaches the wings, so food songs look like no song at all. The smell gets into the brain and stops there.
A mate in view is now partly readable from the song: when the translator says "mate", it's right 69% of the time. Listeners, though, don't react to it reliably: 26%, against 17.5% in silence (p = 0.12).
At 2 ms, a scrambled brain's song carries as many bits as the real one. What's left to test is whether the real brain's specific split (threat kept apart from mate) is what listeners need.
Hearing doesn't make a fly sing back: driving the ear produced 1 spike in the song-command neurons. For now it's one fly broadcasting and the other listening, and rhythm and a real reply would take a model of the wings themselves.
Read on