off policy

projects fly fast weights a brake in the compass

The direction of each step: the compass is wired with a possible brake

The direction of each step comes from the fly’s compass. In the standard model, the neurons that turn the compass read its state at one place and write it one step further round, so signals flow through them one way. We checked where these neurons actually receive their compass input. Unexpectedly, most of it arrives where they write, from the very neurons they drive. This forms a loop that the hand-built models leave out. Added to such a model at its measured strength, the loop stops the compass from turning.

What is known

The fly’s compass is a ring of about fifty EPG neurons in the ellipsoid body, a doughnut-shaped brain region. At any moment one patch of the ring is active, and its position around the ring is the fly’s heading. The ring holds a single patch in place by its own wiring: neighbouring EPGs excite each other, and inhibition spreads across the ring1,2. A network that holds its state this way is called a ring attractor.

When the fly turns, about forty PEN neurons move the patch. Each PEN reads the EPG activity at one position, in a second region called the protocerebral bridge. It then excites the EPGs one step further round the ring, back in the ellipsoid body. Half the PENs push clockwise and half anticlockwise. A turn drives one half harder than the other, so the patch moves the right way at the right speed3,4. The turning-speed signal reaches the PENs through GLNO neurons (finding 1).

The two cell types connect the two regions in opposite directions. EPGs receive their input mainly in the ellipsoid body and send their output to the bridge. PENs do the reverse: they receive in the bridge and send their output to the ellipsoid body. So in this model, signals flow one way through each PEN (Fig. 1a). It hears from the EPGs in the bridge and speaks to them in the ellipsoid body.

a What the standard model assumes in the bridge in the ellipsoid body EPG x PEN EPG x+1 reads writes signals flow one way b What the wiring shows in the bridge in the ellipsoid body EPG x PEN EPG x+1 reads writes 1,281 talks back · 2,545
Fig. 1 | Where a PEN receives its compass input. a, In the standard model, a PEN reads EPG x in the bridge and writes to EPG x+1 in the ellipsoid body. b, In the hemibrain, EPG x+1 sends PENa twice as many synapses back as EPG x sends it in the bridge. PENs that push the other way mirror this toward x−1.

Our goal was to check this picture against the wiring: where does each PEN actually receive its EPG input?

What we found in the wiring

We used the hemibrain, a map of a female fly’s brain in which each synapse is assigned to a brain region. The model predicts that EPG synapses onto PENs sit in the bridge. Instead, most of them sit in the ellipsoid body, where PENs are supposed to send, not receive: 2.4 to 3 times as many as in the bridge (3,844 against 1,281 for PENa, and 3,645 against 1,512 for PENb; Table 1). Most of these come from the very EPGs the PEN writes to (66% for PENa and 88% for PENb). So the EPGs a PEN drives talk back to it, more strongly than the EPGs it reads from (Fig. 1b). The male brain map, MaleCNS, shows the same pattern.

These synapses are not new. Turner-Evans et al.5 found them and called them unexpected, and Hulse et al.6 noted how densely EPGs and PENa connect in the ellipsoid body. What the counts add is their size: the input a PEN gets where it writes outnumbers the input it gets where it reads.

This makes a loop. Each PEN excites a patch of EPGs, and that patch excites it back. Both sets of PENs, clockwise and anticlockwise, have this loop. So it drives both sets from wherever the patch already is, and dilutes the difference between them that moves the patch. If the loop is as strong as the synapse counts suggest, it should act as a brake.

Table 1 | EPG synapses onto PENs in the hemibrain, by brain region.

EPG synapses ontoin the bridge (read)in the ellipsoid body (write)of these, from the EPGs it writes to
PENa (hemibrain)1,2813,8442,545
PENb (hemibrain)1,5123,6453,206

What we simulated

To see how strong the brake would be, we added the loop to a standard hand-built ring attractor, with 16 EPG positions and eight PENs on each side, and gave it a constant turn signal. Without the loop, the patch turned at 139° per second. With the loop at half the strength of the read input, it turned at 72° per second. At equal strength, and at the measured 2.7 times, it stopped. The full stop is partly an artefact of this small model, but the steep slowing is not: it also appears when the PENs’ total activity is held in check by normalisation (Methods). You can raise the loop’s strength yourself in Fig. 2.

Compass ring (16 EPG wedges), constant turn signal write-position feedback: 0.0× read input bump rotation: – °/s 0× is the textbook model. 2.7× is what the hemibrain measures. Same turn signal throughout. The red line is the bump's position. Watch how fast it goes round as you raise the feedback.

Fig. 2 | The loop slows the compass. A hand-built ring attractor under a constant turn signal. Raise the loop’s strength, relative to the read input, and the patch slows and stops.

Real flies turn normally, so something in the real circuit must hold the loop in check. Models fitted to the full hemibrain wiring include these synapses and still turn the compass correctly7. So the full wiring contains an answer, but which connections provide it is not known. The usual candidate for inhibition in the compass, the Δ7 neurons, is unlikely: their transmitter, glutamate, excites PENs in the bridge rather than inhibiting them8. The other possibility is that these synapses are weaker than their number suggests.

Three experiments would settle it. Stimulating the EPGs at a PEN’s read position and at its write position, and comparing the PEN’s response, would show whether the loop drives it. Locating the loop synapses on the PEN, on its input branches or its output terminals, would show what they can do. Refitting a compass model with Δ7 exciting the PENs would show which connections let the patch turn at the right speed yet hold still in the dark.

Methods

Data. Hemibrain v1.2, which gives the brain region of each synapse. Read-position synapses are EPG → PEN synapses in the protocerebral bridge. Write-position synapses are those in the ellipsoid body, with the subset from the EPGs each PEN projects to counted separately. MaleCNS v1.0 does not give a region for each synapse in the graph used here, so there we separated the two by position on the ring: the same position as the PEN (read) or the position it writes to.

Simulation. A rate-based ring attractor with 16 EPG positions and eight PENs on each side. A constant drive to one side stood in for a turn. The loop was added as excitation from each PEN’s write-position EPGs, at a strength relative to its read input. The measured strength, 2.7, is the hemibrain’s ratio of ellipsoid-body to bridge synapses for both PEN types together. With divisive normalisation of the PENs, rotation also fell, from 102° to 34° per second at equal strength and to zero at 2.7 times.

Code (in the fly-circuit-exploration repository): scripts/compass_recurrence.py, scripts/hemibrain_compass_comparison.py, simulations/eb_epg_pen_loop.py, simulations/eb_epg_pen_loop_variants.py. Notes are in docs/compass-recurrence.md.

References

  1. Seelig, J. D. & Jayaraman, V. Neural dynamics for landmark orientation and angular path integration. Nature 521, 186–191 (2015).
  2. Kim, S. S., Rouault, H., Druckmann, S. & Jayaraman, V. Ring attractor dynamics in the Drosophila central brain. Science 356, 849–853 (2017).
  3. Turner-Evans, D. et al. Angular velocity integration in a fly heading circuit. eLife 6:e23496 (2017).
  4. Green, J. et al. A neural circuit architecture for angular integration in Drosophila. Nature 546, 101–106 (2017).
  5. Turner-Evans, D. B. et al. The neuroanatomical ultrastructure and function of a biological ring attractor. Neuron 108, 145–163.e10 (2020).
  6. Hulse, B. K. et al. A connectome of the Drosophila central complex reveals network motifs suitable for flexible navigation and context-dependent action selection. eLife 10:e66039 (2021).
  7. Duan, S., Dong, L. L. & Fiete, I. From Synapses to Dynamics: Obtaining Function from Structure in a Connectome Constrained Model of the Head Direction Circuit. bioRxiv (preprint) (2025).
  8. Eddy, J., Shenasa, A. H., Alfaro, P. M., Viveros, M. F. & Turner-Evans, D. B. Divergent excitatory and inhibitory signaling in a head direction circuit. bioRxiv (preprint) (2026).