Issue 30 · Pick 07 Neuroscience ✓ read
A neural circuit for olfactory motion detection
TL;DR: Fruit flies can tell which way an odor is drifting across their two antennae, and this paper finds where and how that computation happens: at the very first synapse of the olfactory system, in the antennal lobe. A single GABAergic local neuron, il3LN6, reads mostly from the opposite antenna and delivers a ~40 ms-delayed inhibition onto projection neurons that are already excited by the same-side antenna. That is, structurally, the Barlow–Levick motion detector from the retina—rebuilt in chemistry. The upshot: an "early" sensory area long thought to just relabel odor identity is actually running a canonical motion computation, and it does so across nearly every odor channel.
Why odor motion is a real computation, and why it's hard
Start with what any motion detector must have. Since Hassenstein–Reichardt (1956) and Barlow–Levick (1965), the recipe is fixed: two spatially separated sensors, a delay in one pathway relative to the other, and a nonlinear comparison of the two over time. Give one sensor a head start in processing, and a stimulus sweeping in one direction will line up the delayed and non-delayed signals (they reinforce), while the opposite direction misaligns them (they cancel). Direction selectivity falls out of that timing asymmetry.
Vision has this worked out in exquisite circuit detail. Olfaction seemingly shouldn't: odor plumes are intermittent blobs, not smooth moving edges, and for years the assumption was that flies orient to odor mostly by sensing wind. But Kadakia et al. (2022) showed walking flies infer odor-motion direction from the timing difference between their two antennae, independent of wind. That behavior demands a neural motion detector somewhere. Nobody knew where, or what it looked like.
The two antennae are the separated sensors. The open questions were: where do the bilateral signals first meet, is there a delay line, and what supplies the nonlinear comparison?
The clean localization: ORN in, ALPN out
The elegant move in this paper is a subtraction experiment across one synapse. In the Drosophila antennal lobe, olfactory receptor neurons (ORNs) synapse onto antennal-lobe projection neurons (ALPNs), one glomerulus per odor channel. Crucially, most ORNs project bilaterally, so left and right signals already converge at this first synapse.
So the authors imaged both sides of that synapse with the identical stimulus—optogenetically generated "moving bars" of odor, made by activating ORNs in one antenna then the other with a fixed inter-antennal delay (using CsChrimson, so no wind and no mechanosensation confounds). The result is a crisp dissociation: ORN axon terminals show almost no direction preference, while ALPN dendrites in the very same glomeruli are strongly direction-selective (Figure 3). The computation is created at this synapse, one step from the periphery.
And it isn't a specialist trick. Ranking glomeruli by direction-selectivity index (DSI, the normalized difference between "ipsi-first" and "contra-first" responses), the aversive channel DL5 hits 0.525 and the food channel DM4 0.482, with clear selectivity across pheromone, food, and aversive pathways alike, all sharing the same ipsi-first preference in each hemisphere.
The direction-selective signal doesn't stay local—it's carried by ALPN axons to both the lateral horn (innate odor behavior) and the mushroom body calyx (learning), so downstream circuits get a ready-made "odor motion" variable.
The mechanism: delayed contralateral inhibition through il3LN6
Here is the core picture. The ALPN gets fast excitation from its ipsilateral ORNs. Superimposed is a delayed inhibition whose source is the opposite antenna. When odor arrives ipsi-side first (preferred), the excitation fires before the delayed contralateral inhibition catches up—big response. When odor arrives contra-side first (null), the delayed inhibition arrives right on top of the excitation and cancels it—small response. That is exactly Barlow–Levick.
How they nailed down il3LN6 is worth appreciating because it chains four independent methods:
Pharmacology. Bath-applying GABA-A and GABA-B antagonists collapses DA1's direction selectivity toward zero (saline n=10 vs. drug n=12, p<0.0002). So inhibition is required, consistent with a delayed-inhibition (Barlow–Levick) rather than delayed-excitation (Reichardt) scheme.
Connectomics. Searching FlyWire for local neurons with (a) strongly contralateral-biased ORN input and (b) heavy output onto ALPNs, il3LN6 stands out sharply from all other local neuron types. (Nicely, this same neuron was previously implicated in odor gradient sensing—see the Discussion's argument that gradients and motion may share a lateralized code.)
Physiology. il3LN6 is excited by contralateral ORN activation and only weakly suppressed by ipsilateral—and it prefers contra-first moving bars, i.e., the null direction of the DA1 ALPNs it inhibits. That mirror-image tuning is exactly what starburst amacrine cells do relative to direction-selective ganglion cells in the mammalian retina.
Behavior. Silencing il3LN6 (shibire^ts) strongly reduces flies' tendency to orient against moving odor bars—no significant difference between stimulus and blank periods, and significantly less oriented than either parental control (KL-divergence permutation test, p<0.0001 vs. empty control). Importantly, silenced flies still track a straight odor ribbon in wind, so this isn't a generic "can't smell" deficit—it's specific to the motion computation.
The tuning looks like a motion detector, down to reverse-phi
Two experiments make the "this really is a correlation-based motion detector" case convincing.
First, delay tuning. Sweeping the inter-antennal delay of the moving bar, DSI becomes significant around ~25 ms; using correlated-noise stimuli (binary noise presented to both antennae with an imposed lag), the directional response peaks at ~40 ms and falls off across the 20–120 ms range. That's a concrete effective delay for the circuit, and it matches behavioral timescales from Kadakia et al.
Second—and this is the satisfying part for anyone who knows the vision literature—they imported the reverse-phi trick. Positively-correlated noise oriented in the preferred direction gives a large response. Flip the correlation sign (delayed inverted copy at the other antenna) and the response reverses: now it's largest for null-direction orientation. A detector sensitive to the sign of pairwise spatiotemporal correlations is the algorithmic signature of a Hassenstein–Reichardt/Adelson–Bergen style estimator. The olfactory system shows the chemical analog of the reverse-phi motion illusion. They further confirm it responds to drifting sinusoids based on temporal frequency (velocity/wavelength ratio), like fly visual motion detectors—and, critically, that direction selectivity survives naturalistic, power-law-distributed "whiff" stimuli, so it's not an artifact of clean periodic input (whiff-response difference grows with imposed delay, r²=0.856).
The model: measured filters plus a squared rectifier
To close the loop, they measured DA1 ALPN and il3LN6 linear filters via white-noise analysis (with a timing-based trick to beat the calcium frame rate). The ALPN ipsilateral filter is fast excitation; its contralateral filter is transient activation followed by suppression—i.e., delayed inhibition, exactly as the circuit predicts. Feed moving-bar traces through these two measured filters, sum them, and pass through a squared rectifier y[n]=\max(0,s[n])^2. That minimal model reproduces the observed ipsi-first preference. No fitted scale factor or offset. It's a demonstration of sufficiency, not a high-precision fit.
What to believe, and what to hold loosely
The localization result (ORN ≈ null, ALPN selective) is strong and clean, and the four-method convergence on il3LN6 is genuinely convincing for the DA1 pathway. The framing—an early "identity" area quietly computing dynamics with the same motif vision uses—is the kind of cross-modal unification worth remembering.
Where to keep skepticism:
-
DA1 carries the deep mechanistic story; generality is asserted more than dissected. Direction selectivity is broad across glomeruli, but the delay tuning, filter extraction, and model are all in DA1 (a pheromone channel). il3LN6 innervates many glomeruli, so it's plausible the same neuron does the job broadly, but that's an inference, not a per-channel demonstration.
-
The behavioral effect is real but modest and it's a bulk readout. Silencing il3LN6 reduces orientation-against-motion and abolishes the stimulus-vs-blank difference; it does not eliminate odor navigation. That's the honest result—one local neuron shapes a cue, it isn't the whole steering system—but don't read it as "il3LN6 is the motion detector."
-
The model is qualitative. It shows the measured filters can generate the sign of the effect; the authors explicitly leave open contributions from the rest of the recurrent antennal-lobe network, and note that il3LN6 is itself direction-selective, so the ALPN selectivity may be inherited as much as computed locally.
-
All the "motion" here is timing between antennae, delivered largely by optogenetics that bypass real odor transport. The naturalistic-whiff experiment is the bridge to real plumes, and it holds, but the antennae-are-two-pixels abstraction is a strong simplification of a turbulent world.
If you read one section, read the connectome-to-physiology-to-behavior arc around Figure 5 (main text "Connectomics identifies a candidate inhibitory pathway" through the model): that's where the Barlow–Levick analogy stops being a metaphor and becomes an identified neuron with matching wiring, tuning, and a behavioral consequence. The broader implication the authors raise—that motion detection may occupy a narrow circuit design space, so delayed lateral inhibition keeps getting reinvented across modalities and phyla—is speculative but well-earned here.