Sensory processing reformats odor coding around valence and dynamics

Extracting relevant features of a complex sensory signal typically involves sequential processing through multiple brain regions. However, identifying the logic and mechanisms of these transformations has been difficult, due to the challenges of measuring both activity within and long-range connectivity between multiple neural populations. Here, we investigate the reformatting of odor information across two stages of the Drosophila olfactory system. We measure the odor tuning of 20 types of anatomically-defined third order lateral horn neuron (LHN) and compare to predictions based on the odor tuning of second-order projection neurons (PNs) and PN-LHN connectivity. We find that LHNs reformat PN activity in two distinct ways. First, LHNs selectively discard information about odor identities with similar valence (i.e., attractiveness or aversiveness). This emerges from a precise alignment of PN odor tuning and PN-LHN connectivity, as well as odor-specific inhibition and boosting of LHN activity. This creates a population code for valence that is more explicit than in PNs. Second, a subset of LHNs selectively discard information about continuing odor presence, by responding only transiently to odor onset. This creates a population code for odor dynamics that is more explicit than in PNs. Across LHNs, valence and dynamics are independent of each other. Thus, feedforward connectivity and local inhibition combine to extract two orthogonal dimensions of olfactory information.