Insects rely on olfactory cues to navigate toward or away from odor sources during flight. In a natural environment, odor plumes can fluctuate very rapidly, implying that fast classification of olfactory stimuli is necessary to support adequate behavioral decisions. However, how the brain achieves this is still unclear.

In drosophila, odors are represented by the activity of ~50 glomeruli in the antennal lobe (AL), the first olfactory center of the brain. Prior research suggests that hedonic valence (attractive or aversive) may be coded at the glomerular level, and that behavioral responses can be accounted for by the contributions of single or a group of glomeruli. However, these observations cover a limited range of odorants and glomeruli, and it is unclear whether this mechanism can support the fast behavioral choices made during flight.

Here, we monitored behavioral responses during flight to a variety of odors and mixtures of odors, using a flight-simulator setup. In parallel, we used two-photon Ca²⁺ imaging to probe neuronal responses in most AL output neurons. We found that flies displayed responses ranging continuously from attraction to aversion, and were able to make behavioral decisions within a few 100s of milliseconds. Morphing attractive odors into aversive ones evoked gradual changes in AL representation that were mirrored by changes in behavior. These results argue for a model in which rapid odor classification is achieved by a linear readout of the onset of the AL response. Consistent with this hypothesis, we found that a linear model can not only recapitulate the observed behavior, but can also predict responses to novel odors not used in the fitting procedure. This analysis revealed glomeruli that were strongly associated with attractive or aversive behavior, indicating that certain glomeruli indeed convey information about the valence of a stimulus. In a next step, we will test the causality between AL activity and behavior by manipulating the activity of these glomeruli and asking whether the behavioral outcome agrees with the predictions of the linear model. These manipulations will shed light on how AL activity is processed to shape behavioral responses.