Effect of Circuit Structure on Odor Representation in the Insect Olfactory System

The 50% connectivity does not maximally separate KC representations when PN inputs are static. a, The threshold model of KCs. The left-most vector represents the PN activity. This is combined through a connectivity matrix to give the input seen by each KC (a 50,000-element-long vector). Thresholding is then applied to define spiking KCs. b, The Hamming distance between inputs seen by two KCs is calculated for all possible pairs and averaged and plotted as a function of the PN–KC connectivity. c, The mean (±SD) normalized Hamming distance between the activity of KC networks driven by two different inputs is plotted on the y-axis as a function of the PN–KC connectivity. Different shades plot the distance between odor representations that differed in 5–80% of the active PNs.

Simulation of temporally patterned PN inputs to a KC network. a, The matrix on the left represents the activity of a set of 900 PNs. Each row shows the activity of a single PN during a 3000 ms time period. Blue dots show the time of a spike. The red region represents the time during which the odor was presented. Top, A summation of the activity of the entire PN network is shown clearly indicating the oscillations in the net PN activity. This input was used to calculate T and I _syn (the synaptic input to KCs). The differences between the population representations of two inputs were calculated using the Hamming distance. b, The mean population response of 900 PNs projected onto the first three principal components for three odors is shown by the black traces. Individual trials are shown by the colored traces. c, The mean membrane potential of all KCs shows a 20 Hz oscillation. Bottom, The response of two KCs (in red and black traces) to two different odors. Only the first odor evokes a consistent response from this particular KC across five odor trials (middle). The second odor does not lead to reliable spiking in this example KC.

PN temporal patterning reveals the functional differences between connectivities. a, Distance between odor representations. The mean ± SD normalized Hamming distance between the KC representations of odor pairs is shown as a function of the PN–KC connectivity value. Here KCs are modeled as described in Figure 3. b, Classification accuracy decreases with increasing PN–KC connectivity. A k-medoids clustering algorithm that used the distance between 25 KC activity vectors (five trials × five odors) was used to categorize each vector as one of five odors. The percentage of correctly classified odor representations is plotted on the y-axis as a function of the connectivity of the PN–KC network. c, Odor representations become indistinguishable with increasing PN–KC connectivity. Five odors that differed from each other by 5% PN input were mapped to a plane using multidimensional scaling. Different trials of a given odor are plotted using a single color. Different odors are plotted using different colors. The PN–KC connectivity is shown in the title of each subplot. d, Hamming distance between static odor representations. The mean ± SD normalized Hamming distance between the KC representations of odor pairs is plotted as a function of PN–KC connectivity. Here, the PN odor representation did not change in time.