Learning about natural variation of odor mixtures enhances categorization in early olfactory processing

Abstract

Natural odors are typically mixtures of several chemical components. Mixtures vary in composition among odor objects that have the same meaning. Therefore a central 'categorization' problem for an animal as it makes decisions about odors in natural contexts is to correctly identify odor variants that have the same meaning and avoid variants that have a different meaning. We propose that identified mechanisms of associative and non-associative plasticity in early sensory processing in the insect antennal lobe and mammalian olfactory bulb are central to solving this problem. Accordingly, this plasticity should work to improve categorization of odors that have the opposite meanings in relation to important events. Using synthetic mixtures designed to mimic natural odor variation among flowers, we studied how honey bees learn about and generalize among floral odors associated with food. We behaviorally conditioned honey bees on a difficult odor discrimination problem using synthetic mixtures that mimic natural variation among snapdragon flowers. We then used calcium imaging to measure responses of projection neurons of the antennal lobe, which is the first synaptic relay of olfactory sensory information in the brain, to study how ensembles of projection neurons change as a result of behavioral conditioning. We show how these ensembles become 'tuned' through plasticity to improve categorization of odors that have the different meanings. We argue that this tuning allows more efficient use of the immense coding space of the antennal lobe and olfactory bulb to solve the categorization problem. Our data point to the need for a better understanding of the 'statistics' of the odor space.

Keywords: Categorization; Natural odors; Olfaction; Plasticity; Variability.

Fig. 1.

Each component and the complete PH1 blend elicits unique patterns of activity across glomeruli using calcium imaging of projection neurons in the antennal lobe. (A) Pictures represent false color-coded images of activation patterns (ΔR_340/380) induced by the six components and the blend in a representative bee. The activity shown for each component corresponds to the concentration of that component in the PH1 blend. Each picture is the average of the activity measured between 375 and 625 ms after odor onset. The white arrowhead points to glomerulus 17, which was activated by ocimene but suppressed in the blend in this example bee. (B) Raw fluorescence of the AL after staining the PNs with Fura-dextran and identification of 24 glomeruli on the dorsal surface. (C) Mean±s.e.m. activation across glomeruli elicited by each component and the PH1 blend (n=6 naïve animals). Glomeruli are ordered top to bottom from least to most activated during stimulation with the blend. The same ordering was used for each component to highlight the differences in the activation patterns elicited by different odors.

Fig. 2.

Natural variation in composition of blends from Antirrhinummajus varieties. (A) The variation projected onto the first two principal components, which extract over 70% of the natural variation. The small green and orange circles represent volatiles collected from individual flowers of Pale hybrid (PH) and Potomac pink (PP), respectively (Wright et al., 2005a). Factor loadings of each of the six original components into the first two PCs are shown in Table S1. (B) Relative proportion of the six components in each one of the artificial blends (large darker circles in A). Each blend was made by mixing components in proportions that mimicked the variability found in the natural samples. The concentrations of the components in the PH1 and PP1 blends corresponded to the average proportion of each component measured in natural samples. The PH2 and PP2 blends were made with concentrations that corresponded to the 25th percentile of the concentrations measured in the natural varieties. PP3 and PH3 were composed of concentrations that correspond to the 75th percentile. Blends 4 and 5 were random combinations of the average, 25th or 75th percentile of each component. PP6 and PH6 were selected from the natural samples as extreme examples. Act, acetophenone; Meb, methyl benzoate; Tmc, trans-methyl cinnamate; Lin, linalool; Myr, myrcene; Oci, ocimene.

Fig. 3.

Behavioral conditioning. (A) Two groups of bees were conditioned. Pre-exposure: 40 unrewarded trials with mineral oil (MO−) or pseudorandomized presentations of artificial blends of PP (PP2–PP6). Then, bees were differentially conditioned along 8 rewarded trials using the PH blends (PH2–PH6) intermingled with 8 unrewarded trials. Groups differed on unrewarded trials: the PH+/MO− group received mineral oil trials while the PH+/PP− group received trials with PP2–PP6 in randomized order. After training, all bees were subjected to a transfer test with PH1 and PP1, which had not been used during training, as well as with 2-octanone. (B) Percentage of proboscis extension across sequential acquisition trials in PH+/PP− (top; n=45 bees) and PH+/MO− (bottom; n=33 bees). (C) Transfer test results showing timelines representing latency and duration of proboscis extension (mean±s.e.) in the same bees as in B. Latency two-way repeated-measures ANOVA: training protocol F_1,77=6.233, P=0.01; test blend PH1 vs PP1 (repeated factor) F_1,77=6.121, P=0.01; interaction F_1,77=0.418, *P<0.05. Tukey's HSD test revealed significant differences in response latency to PP1 from group PH+/PP− in comparison with PH1 from the same group (P<0.01) and PH1 (P<0.01) and PP1 (P<0.01) from the PH+/MO− group. Duration two-way repeated measures ANOVA: training protocol F_1,77=0.270, NS; test blend PH1 vs PP1 (repeated factor) F_1,77=14.308, P<0.001; interaction F_1,77=0.267, NS. Tukey's HSD test, revealed significant differences in response duration to PH1 and PP1 in the differentially conditioned bees PH+/PP− (*P<0.01), whereas responses were not significantly different in PH+/MO− bees.

Fig. 4.

Projection neuron responses to artificial blends. (A) Calcium imaging responses (ΔR_340/380) elicited by artificial blends of PH and PP in a representative animal. (B) Graphs show the mean±s.e. activity measured in 24 identified glomeruli between 325 and 625 ms after stimulus onset in animals trained to PH+/MO− (n=10 bees). The glomeruli were ordered from lowest to highest response according to the blend PH1 and the same ordering was repeated for the remaining blends.

Fig. 5.

Differential conditioning increases separation between the representations of floral blends. (A) To quantify similarity among neural representations of odors within and between varieties, we calculated correlations between all possible pairs of patterns elicited by the 12 synthetic floral blends as well as between replicate measurements of the same blend for both groups of bees (Table S2). The correlation values obtained were further Fisher's Z-transformed to be used in graphs and in statistical analyses. The obtained values were grouped according to: replicate measurements of the same blend (dark green and dark orange); different blends from the same cultivar (light green and light orange); and any two blends from different cultivars (gray bars). The graph shows mean±s.e.m. correlation values for each of these categories for PH+/MO− trained bees (solid, n=10) and PH+/PP− trained bees (hatched, n=7). Tables on the right show results of two-factor ANOVA with training condition (PH+/MO− or PH+/PP−) as one factor and correlated pair (5 categories indicated in the abscissa and explained above) as the second factor. P-values in the lower table correspond to the post hoc contrasts that are indicated with brackets and corresponding numbers or letters on the figure. Only statistically significant differences are indicated with an asterisk in the graph. (B) Same data as in A shown as cumulative proportion of correlations with specific values (x-axis) for each of the test conditions. Data represent accumulated numbers of values that were used to calculate the means in each cell in a color code (green, orange or gray) of Table S2.

Fig. 6.

Differential conditioning separates the spatiotemporal activity patterns of rewarded and non-rewarded blends. (A) 10 PH+/MO− bees and 7 PH+/PP− bees were combined by group to create two ‘average’ bees. The 12 spatiotemporal trajectories in each graph correspond to each of the synthetic blends. Each trajectory represents the activity of 24 glomeruli during the evolution of odor elicited activation pattern. Thus, the whole dataset of both average bees (24 activation patterns) were aligned by glomeruli and subjected to principal component analysis. The trajectories correspond to each blend shown from 3 frames before odor onset to 1 s after odor onset (12 frames out of a total of 80) plotted using the first two principal components (89% variance explained). PH+/MO− and PH+/PP− bees were plotted in separate graphs, but both graphs were obtained from a common PCA to make them comparable. Red/orange colors correspond to PP blends and green/blue colors to PH blends. (B) Euclidean distance between PH and PP blends is higher for differentially trained bees for the duration of the stimulus. ED was calculated in the original 24-dimensional space, frame by frame, and for all possible pairs of PH and PP flowers (36 combinations). The 36 distances were averaged to obtain a unique value of PH-PP distance per honey bee. The traces represent the mean±s.e. of 10 PH+/MO− bees (black) and 7 PH+/PP− bees (blue). Gray bar along x-axis represents odor stimulation.

Fig. 7.

Contribution of individual glomeruli to the different activity patterns elicited by both odor varieties and differentially affected by the training condition. Two-factor ANOVA was performed for each individual glomerulus to identify the glomeruli that are responsible for the different patterns among varieties and to identify glomeruli that are differentially affected by training. Each group of four bars shows the mean±s.e.m. of PH+/MO− (n=10) and PH+/PP− (n=7) honey bees for PH and PP varieties for each specific glomerulus. Glomeruli are grouped according to significant differences in the ANOVA factors. No glomerulus showed a significant interaction between factors.