Vector-Based Olfactory Navigation Strategy Revealed in Drosophila Flies

Flies hunting airborne odor plumes do more than follow scent—they “edge track,” repeatedly steering along the moving boundary between odor and no-odor. In a new Nature study, researchers probe how small insects build and refine a memory of plume direction while crossing that boundary, focusing on the angular information flies likely extract from each exit and re-entry.

Although odor corridors can span a 180° range, the team finds that flies adopt a narrow set of angles that depends on plume geometry. This consistency suggests a simple yet powerful mechanism: whenever a fly crosses the plume’s boundary, it stores its heading relative to wind as an angular memory. Subsequent boundary encounters then use these memories to bias trajectories toward efficient re-tracking.

To formalize the idea, the authors build a switching state-space model in which simulated flies alternate between two functional regimes: “leaving” and “returning.” These states are not merely inside-versus-outside plume epochs; instead, transition rates depend on odor presence. Crucially, the model’s velocity is generated by an autoregressive process whose correlation timescale and noise depend on the current regime.

The model also incorporates memory: in the leaving state, motion is biased by an exit-angle memory updated each time the simulated fly exits the plume, whereas in the returning state, motion is biased by an entry-angle memory updated each time it re-enters. Each update blends the prior memory with the current heading, making the remembered direction gradually adapt with experience.

Using variational inference, the researchers infer latent leaving and returning states and latent goal directions from individual trajectories across plume geometries. The resulting parameter sets vary across simulated individuals, matching behavioral heterogeneity observed in real flies—some adhere tightly to the boundary, while others take longer, more circuitous excursions before returning.

When the authors simulate trajectories with average learned parameters, the model reproduces key experimental outcomes: path efficiency relative to perpendicular distance from the plume boundary, and similar behavioral statistics across plume orientations. The first outside bout is excluded, reflecting the model’s need for an entry-angle memory before a meaningful return goal can form.

Testing memory necessity reveals a striking asymmetry. Removing entry-angle memory severely disrupts edge tracking across all plume orientations, indicating that dynamic entry angles actively instruct an angular goal for returns to the plume boundary. By contrast, continuous updating of exit-angle memory is less critical—while a purely upwind or resampled exit memory degrades performance in some geometries, it does not abolish edge tracking outright.

Finally, operant-style training experiments connect directly to the model: a single session enhances the ability to track a later plume segment oriented oppositely, consistent with rapid strengthening of crosswind components in the inferred entry-angle goal. Together, the work supports a vector-based navigational strategy in which stable exit bias helps steering, but fast, experience-driven entry memories provide the decisive instruction for progression along odor boundaries.