Processing of Odor Mixtures in the Zebrafish Olfactory Bulb

Measurement of odor responses in sensory afferents to the OB

Odor-evoked activity in sensory afferents to glomeruli was measured by Ca²⁺ imaging after OSNs were loaded with calcium green-1 dextran (Friedrich and Korsching, 1997). Experiments were performed in a ventrolateral region of the OB that contains small, densely packed glomerular structures responding preferentially to amino acids and nucleotides (Friedrich and Korsching, 1997, 1998). Odor application to the nose evokes distinct patterns of fluorescence signals in the OB, reflecting changes in Ca²⁺ concentration in OSN axonal terminals (Friedrich and Korsching, 1997; Fuss and Korsching, 2001; Wachowiak and Cohen, 2001; Wachowiak et al., 2004) (see Fig. 2A1-A4). Application of the carrier solution alone did not evoke detectable signals (data not shown). As shown previously (Friedrich and Korsching, 1997), the time courses of odor-evoked Ca²⁺ signals were similar and stereotyped throughout the observed region. This is consistent with odor responses of single zebrafish OSNs recorded electrophysiologically: OSN responses follow a stereotyped phasic-tonic time course, and the pattern of activity across multiple OSNs does not change much during odor stimulation (Friedrich and Laurent, 2001, 2004). Maps of Ca²⁺ signals, therefore, were time averaged. Negative-going responses were very rare, consistent with the low abundance of inhibitory electrical responses in zebrafish OSNs (Friedrich and Laurent, 2004). Responses of glomerular units have odor- and glomerulus-specific thresholds and increase monotonically with concentration over several log units until saturation (Friedrich and Korsching, 1997).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Patterns of afferent glomerular activity evoked by binary odor mixtures and their components. A1-A4, Maps of afferent glomerular Ca²⁺ signals in the ventrolateral OB, evoked by two amino acid components at 1× and 2× concentration. Lateral is to the bottom; anterior is to the right. A5-A7, Lower bound, upper bound, and prediction (see Fig. 1) as calculated from A1-A4. A8, Measured response to the mixture. A9, Sum of component activity patterns at 1× concentration (A1 + A3). A10, Difference between mixture response and prediction (A8 -A7). A11, Difference between response maps evoked by two repeated applications of the mixture. The left color scale applies to difference images (A10, A11) and the right color scale to all other images. B, Distribution of pixel values in difference images, averaged over n = 18 experiments with binary amino acid mixtures in which the full set of stimuli was applied twice. Blue, Difference between response map evoked by the mixture and the prediction; cyan, same for a second set of applications; red, difference between response maps evoked by the mixture in repeated applications. C, Quantification of the width of distributions of pixel values in difference images. In each of n = 18 experiments the width of the distribution of pixel values in difference images was measured as the SD of a Gaussian fit. Error bars show the average SD for each type of difference map. The significance of differences in mean SD is shown by symbols in each bar. ^*p < 0.05; ^**p < 0.01; ns, not significant. The top row of significance symbols is for comparison between each of the bars to the first (left) bar and the bottom row for comparisons to the second bar. D, Same display as in A for a binary mixture of food extracts. E, F, Same plots as in B, C for binary mixtures of food extracts (n = 17). Scale bars, 100 μm.

Analysis of mixture interactions in sensory afferents to the OB

The monotonicity of concentration-response functions and the very rare occurrence of negative-going responses are reminiscent of concentration-response curves describing ligand-receptor interactions. We therefore used a simple algorithm based on models of ligand-receptor activation to predict afferent patterns of Ca²⁺ signals evoked by binary odor mixtures from the patterns evoked by their components (Fig. 1). This algorithm assumes that components (ligands) do not interact and activate a common effector (receptor) associated with each glomerulus. The strength of activation of a given glomerulus by a given ligand is described by the corresponding concentration-response function. Local information about concentration-response functions was obtained by measuring responses to components at the concentration used in the mixture (1×) and at twice that concentration (2×). The minimum signal expected in response to the mixture is the signal elicited by the more potent component at the 1× concentration (Max_1×; Fig. 1). Added to this signal is the signal increment expected from the second component (Superposition term; Fig. 1). This is approximated as the geometric mean of the signal increase between 1× and 2× concentrations for each component. The geometric mean was used because it yields correct values when the concentration of one component is below response threshold or saturating (i.e., when the concentration-response function has a slope near zero; see Fig. 1 and Materials and Methods). The calculation yields three values: a conservative lower bound (Max_1×), a conservative upper bound (Max_2×; the signal elicited by the more potent component at 2× concentration in each pixel), and the prediction (Fig. 1). The difference between lower and upper bounds is usually small relative to the absolute response magnitude, because a twofold concentration change is small relative to the dynamic range and affects activity patterns only slightly (see below). The prediction error arising from uncertainties in the concentration-response function therefore is also small. Nevertheless, the prediction by the geometric mean is an approximation; precise knowledge of the concentration-response function would be required to construct an exact algorithm. Moreover, the calculation of the prediction accumulates noise in the data, arising, for example, from shot noise and the low amount of bleaching. To minimize these effects, we included only those pixels in which the response (ΔF/F) was >0.4% in the analysis.

We first examined glomerular Ca²⁺ signals in response to single amino acids, which are natural odors for fish and other aquatic animals. Concentrations used were 10-40 μm, which are intermediate in the behaviorally relevant range (Carr, 1988). Figure 2A1-A4 shows Ca²⁺ signal maps evoked by two amino acid components at 1 and 2× concentrations in the ventrolateral OB, thresholded at ΔF/F = 0.4%. From these responses the bounds and the prediction were calculated pixel-wise and compared with the actual mixture response.

The lower and upper bounds (Fig. 2A5,A6) were only slightly different from each other. The prediction (Fig. 2A7) was very similar to the measured mixture response (Fig. 2A8). Responsive regions in the pattern evoked by the mixture corresponded to those in the prediction, and similar distributions of response magnitude were observed in the two maps. The sum of the component patterns (Fig. 2A9) did not predict response patterns as well. This was expected, because it did not take into account the nonlinearity of concentration-response functions.

Figure 2A10 depicts the difference between the mixture response and the prediction. Negative and positive deviations from zero were encoded by opposite colors. Deviations were small relative to the overall response magnitude (note different calibration of color scales) and appeared to be only slightly greater than deviations between maps evoked by repeated applications of the mixture (Fig. 2A11). Similar results were obtained in n = 33 experiments.

The deviation between the mixture response and the prediction was quantified in 18 experiments performed under conditions of low photon noise (see Materials and Methods). The differences in pixel values between the prediction and the actual mixture response are shown as histograms in Figure 2C. Similar distributions were measured in two consecutive series of applications (Fig. 2C, blue and cyan, respectively), indicating that responses were stable. The widths of these distributions, quantified by the SD of a Gaussian fit to the histogram, were not significantly different from each other (p > 0.65; Fig. 2D). These distributions were slightly but significantly broader than the distribution of difference values for repeated mixture responses (Fig. 2C, red, D). Activity maps evoked by repeated responses to the mixture were therefore more similar to each other than to the prediction. These results were confirmed by using the correlation coefficient as a measure of similarity. The mean correlation between the prediction and the mixture response was 0.89 ± 0.06 for the first series of stimuli and 0.88 ± 0.09 for the second set of stimuli (not significantly different; p > 0.8), whereas the mean correlation between maps evoked by repeated mixture applications was slightly but significantly higher (0.93 ± 0.06; p < 0.05).

The slightly higher similarity between repeated mixture responses could indicate a significant deviation of the prediction from the mixture response. However, it is likely that this deviation is at least partially a consequence of the calculation of the prediction, because the prediction is not exact and because the procedure accumulates noise. Moreover, the correlation between mixture responses and prediction was high, and the average difference between the SDs of the Gaussian fits was 0.05 ± 0.03% ΔF/F. This value is very small compared with the ΔF/F values in the response maps. We therefore conclude that afferent glomerular responses to binary mixtures of amino acids are predictable within narrow limits by a model assuming no interactions between components.

We next measured afferent glomerular activity in response to different food extracts, each containing many different compounds. These complex stimuli evoked widespread glomerular activity. As a consequence, patterns of Ca²⁺ signals evoked by different food odor stimuli overlapped greatly (Fig. 2D1-D4). Nevertheless, responses to binary mixtures of food extracts were in large part predictable from the component responses by the algorithm that was used. As for amino acids, lower and upper bounds were similar to each other and to the prediction (Fig. 2D5-D7). The prediction, but not the sum of the activity maps of the component, was a close approximation of the true response to a binary mixture of food extracts (Fig. 2D7-D11). Similar results with mixtures of food extracts were obtained in n = 24 experiments.

Quantitative analysis of the deviation of the prediction from the mixture response, as described above, was performed for 17 experiments performed under low-noise conditions (see Materials and Methods). The pixel value differences between prediction and mixture response were not significantly different for repeated series of stimuli (p > 0.8; Fig. 2E, blue and cyan, F). As for amino acids, these distributions were slightly but significantly broader than the distributions of differences between repeated mixture responses (Fig. 2E, red, F). The average correlation between the prediction and the mixture response was 0.92 ± 0.06 for the first series of stimuli and 0.91 ± 0.07 for the second set of stimuli (p > 0.8). These values were not significantly different from the mean correlation of responses to repeated mixture applications (0.95 ± 0.03; p > 0.06). Hence, as with amino acid stimuli, the correlation between mixture response and prediction was high, and the average difference between the SDs of Gaussian fits was very small (0.02 ± 0.05% ΔF/F). Thus, responses to binary mixtures of complex food odor stimuli are also predictable within narrow limits by a model assuming no interactions.

Finally, we tested whether the small deviation between the prediction and the response to the mixture depended on the overlap between the component activity patterns. However, the deviation of the mixture from the prediction, quantified by the difference in the width of difference distributions, was not correlated with the correlation between the component activity patterns (amino acids, r = -0.17; p > 0.5; food extracts, r = -0.19; p > 0.45). Together, these results indicate that mixture interactions between the tested simple and complex stimuli in the peripheral olfactory system of zebrafish are, at most, weak.

Mitral cell odor responses

We next examined odor responses of the output neurons of the OB, the MCs. MCs were recorded extracellularly in the loose-patch or cell-attached configuration or were recorded intracellularly by the whole-cell patch-clamp technique. Unlike OSNs, MCs frequently respond to odors with inhibition. Moreover, MC firing rates often are modulated over tens or hundreds of milliseconds and can comprise successive excitatory and inhibitory epochs. In addition, a subset of mitral cell responses showed subthreshold membrane potential oscillations of 20-30 Hz (Fig. 3) (Friedrich and Laurent, 2001, 2004). Mitral cell response patterns to the same odor do not change abruptly as a function of concentration and are usually similar over concentration ranges of more than one log unit (R. W. Friedrich, unpublished observations).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Examples of mitral cell odor responses to amino acids and food extracts. Shown are whole-cell recordings of mitral cell responses to different amino acids (left column) and food extracts (right column). Each response is from a different mitral cell. Both types of stimuli can evoke simple (phasic-tonic; top) and more complex (multi-phasic) response time courses. Fast subthreshold membrane potential oscillations are seen in a subset of odor responses. Odor stimulus is indicated by the horizontal bar. Vertical bars, 20 mV.

The complex response properties arise from synaptic interactions in the OB, indicating that MC responses do not simply relay their glomerular input but also reflect dynamic processing by neural circuits in the OB. The temporal patterning and the occurrence of inhibitory responses preclude the prediction of mixture responses by simple models of receptor-ligand interactions. We therefore compared MC odor responses by using methods that take into account their temporal response patterns.

Mixture interactions in mitral cell responses: simple odors

We first examined responses to amino acids and their binary mixtures. Figure 4A shows responses of one MC, recorded extracellularly. In this example the response to component 1 was clearly excitatory, whereas the response to component 2 was inhibitory. The response to the mixture was excitatory, with a magnitude and time course almost identical to the response to component 1. Hence, the response to the mixture was dominated by the response to the excitatory component. Figure 4B shows data from another MC that also responded with clear-cut excitation to one and with clear-cut inhibition to another component. In this case the response to the mixture was dominated by the inhibitory component. In each of 18 experiments testing mixtures of one clearly excitatory and one clearly inhibitory component, one component dominated over the other in the response to the mixture. In 83% of the cases (15 of 18) the dominant component was the excitatory one, whereas in the remaining 17% of the cases (3 of 18) the inhibitory component dominated.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Responses of mitral cells to binary mixtures of amino acids. A-C, Responses of three different MCs to different sets of amino acids and their binary mixtures. Action potentials measured in repeated trials are indicated by ticks in subsequent rows. PSTHs show average firing rate in 100 msec time bins. Horizontal bar indicates stimulus presentation. Gray lines in right panels superimpose PSTHs in response to components onto the PSTHs of responses to the mixture.

In another 18 experiments that used amino acid stimuli, components evoked more complex responses that often differed mainly in their temporal patterns. In the example shown in Figure 4C, the responses to both components were excitatory but had different, odor-specific time courses. The response to the mixture was dominated by one of the component responses and was clearly distinct from the response to the other component. A dominance of one component also was observed in most of the remaining experiments with components eliciting temporally complex responses.

The same amino acid could elicit a dominant response in one MC and a nondominant response in another MC, indicating that dominance is not a general property of the stimulus. It was never observed that a MC responded to one or both components, but not to their mixture. In summary, MC responses to binary mixtures of amino acids generally were dominated by one of the components throughout the response. When the component responses were of clearly different sign (excitatory or inhibitory), the mixture response was more often excitatory.

Mixture interactions in mitral cell responses: complex odors

MC responses to food odors were qualitatively similar to responses evoked by amino acids. They could be simple (mainly excitatory or inhibitory), or they could consist of more complex firing rate modulations, including successive excitatory and inhibitory epochs within the same response (Fig. 3). One difference between responses to simple and complex stimuli was that food odors did not evoke firing rates >70 Hz, whereas responses to amino acids transiently could reach >100 Hz (for example, Fig. 4B,C). This is unlikely to be attributable to lower activation of OSNs by food extracts, because peak Ca²⁺ signals evoked by food extracts and amino acids in OSN axon terminals were similar.

Responses to binary mixtures of food extracts, measured in 21 experiments, often differed from the responses to both individual components. Figure 5A shows an example of a MC that responded with phasic-tonic excitation to component 1 and with inhibition to component 2. The response to the mixture was initially excitatory, thus resembling the response to component 1. However, the initial excitation evoked by the mixture was smaller in magnitude and delayed compared with the excitation evoked by component 1 alone. Subsequently, the response to the mixture became inhibitory and, thus, similar to the response to component 2. Another similar example is shown in Figure 5B. The response to a mixture of complex odors therefore is not described accurately by the response to one of the components.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

Responses of mitral cells to binary mixtures of food extracts. A, B, Responses of two different MCs to different sets of food extracts and their binary mixtures. Action potentials in individual trials and PSTHs (100 msec time bins) are shown. Horizontal bar indicates stimulus presentation. Gray lines in right panels superimpose PSTHs in response to components onto the PSTHs of responses to the mixture.

When responses to both components were of the same sign (excitatory or inhibitory) in a given time bin, the sign of the response to the mixture was never of the opposite sign. Hence, the sign of the mixture response corresponded to that of at least one of the component responses in each time bin. However, the mixture response could differ considerably in magnitude from each component response, and the sign of the mixture response could resemble different component responses in different time bins. Responses to binary mixtures of food odors therefore are still related to their component responses, but they are not dominated as strongly by one of the components as responses to binary amino acid mixtures.

Analysis of component dominance

The dominance of a component in the response to the mixture was analyzed quantitatively. Each odor response was described by the PSTH of firing rates during the 2.4 sec of odor presentation, averaged over repeated applications, and binned in 100 msec windows. The similarity between the response to the mixture and the responses to the two components was measured by the Pearson correlation coefficient. In each experiment the higher correlation reflects the similarity between the mixture response and the more similar component response. This value therefore was used to quantify component dominance. For binary mixtures of amino acids most component dominance values were high, whereas component dominance values for binary food odor mixtures were distributed more broadly (Fig. 6A). The average component dominance was 0.73 ± 0.24 for binary mixtures of amino acids and significantly lower for mixtures of food odors (0.46 ± 0.40; p < 0.01; Fig. 6B). The average correlation between the nondominant component response and the mixture response was not significantly different for amino acid and food odor stimuli (-0.10 ± 0.40 and -0.03 ± 0.40, respectively; p > 0.5). A higher component dominance may be expected if the responses to the individual components are already more similar to each other. However, in our data set the average correlation between the responses to the two components was lower for amino acids (-0.10 ± 0.46) than for food odors (0.17 ± 0.41). We therefore conclude that component dominance is significantly more pronounced in binary mixtures of amino acids.

• Download figure
• Open in new tab
• Download powerpoint

Figure 6.

Quantification of component dominance in mitral cell responses to binary mixtures. A, Distributions of component dominance values in experiments testing binary mixtures of amino acids (top) and food extracts (bottom). B, Mean component dominance values for binary mixtures of amino acids and food extracts. Error bars show SD; ^**p < 0.01.

The correlation coefficient underlying the component dominance measure compares the shapes of PSTHs (i. e., response time courses) independently of the overall firing rate. We therefore also measured the difference in firing rates, averaged over the PSTH, between the responses to each binary mixture and the respective dominant component. The firing rate differences for mixtures of amino acids and food odors were not significantly different from each other (amino acids, 3.8 ± 4.0 Hz; food odors, 2.4 ± 2.6 Hz; p > 0.1). Hence, for mixtures of food odors the responses to the mixture and those to the dominant component are more different with respect to the time course, but not with respect to the total firing rate.

Finally, we explored whether the component dominance in MC responses to binary mixtures of amino acids depends on the similarity of afferent activity patterns evoked by the components. The similarity of input patterns was assessed by the correlation between amino acid-evoked activity patterns across OSNs, which were determined previously in zebrafish (Friedrich and Laurent, 2001). As shown in Figure 7, component dominance and the correlation between afferent activity patterns evoked by the components were weakly, but significantly, correlated (r = 0.42; p < 0.05). Hence, component dominance tends to become less pronounced as the overlap of input activity patterns evoked by the components increases.

• Download figure
• Open in new tab
• Download powerpoint

Figure 7.

Component dominance as a function of similarity between OSN activity patterns evoked by amino acid components. The correlation between OSN activity patterns evoked by each pair of amino acid components used in the mixture was taken from Friedrich and Laurent (2001) and is plotted on the x-axis. The component dominance in mixtures of the same pairs of amino acids is plotted on the y-axis. Line shows a linear regression.

Mixture interactions in the peripheral olfactory system

Patterns of afferent glomerular Ca²⁺ signals evoked by binary odor mixtures were predictable within narrow limits from the patterns of the components. The elements of the pattern evoked by the mixture were also active in at least one component response and vice versa. The small differences in response intensities between prediction and mixture response may reflect mixture interactions, the inaccuracy of the prediction, or both. We therefore conclude that mixture interactions in the peripheral olfactory system of zebrafish between the tested odors are weak or negligible. These results are generally consistent with cross-adaptation of glomerular responses in zebrafish (Fuss and Korsching, 2001) and results from intrinsic signal imaging in mice (Belluscio and Katz, 2001).

The Ca²⁺ indicator selectively reports the activity of sensory inputs to glomeruli. OSNs converging onto the same glomerulus in mice have very similar response properties (Wachowiak et al., 2004), and OSNs respond to odors with simple, stereotyped temporal firing rate patterns in zebrafish (Friedrich and Laurent, 2001, 2004). It therefore is assumed that individual OSN responses are similar to the glomerular population responses detected by Ca²⁺ imaging. Indeed, spiking responses of single zebrafish OSNs to mixtures of 16 amino acids were similar to the response to the most potent stimulus alone (Friedrich and Laurent, 2001, 2004), showing that single OSNs responding to at least one component also respond to the mixture, as found for afferent glomerular Ca²⁺ signals.

It is possible that mixture interactions occur between odorants not tested in this study. Indeed, interactions have been observed between carboxylic acids and amino acids in zebrafish (Fuss and Korsching, 2001) and in OSNs of other species (Ache et al., 1988; Laing et al., 1989; Ache and Zhainazarov, 1995; Kang and Caprio, 1997; Cromarty and Derby, 1998; Spehr et al., 2002). The occurrence of mixture interactions therefore may depend on the odor and species.

In zebrafish it is unknown currently whether each glomerulus is innervated by idiotypic OSNs as in rodents. Activity patterns evoked by mixtures were predicted with little error by an algorithm modeling each afferent glomerular response as a process mediated by a single receptor. This would not be expected if the relation between receptors and glomeruli was different from that in rodents: if populations of OSNs responding to different sets of odors converged onto the same glomerulus, the response to the mixture should be close to the sum of the component responses. If each OSN expressed multiple odorant receptors with different response profiles, the mixture response would become unpredictable. Hence, our data suggest that the functional connectivity between OSNs and glomeruli in the zebrafish OB is similar to that in rodents.

The glomeruli responding to a mixture were the combination of the glomeruli activated by each component. The density of responding glomeruli therefore should increase with the number of components in a mixture. Indeed, food extracts containing many different compounds evoked widespread and highly overlapping activity patterns. As a consequence, information resides more in the relative response intensity, rather than the identity, of active glomeruli. This may be an important property of representations of natural odor stimuli.

Mixture interactions in the olfactory bulb

Mixture interactions have been observed in the OB of vertebrates (Bell et al., 1987; Kang and Caprio, 1995; Giraudet et al., 2002; Wilson, 2003), but it is unclear whether they arise peripherally or centrally. We found that not all MCs excited by at least one component are also active in response to the mixture. Hence, MC activity patterns evoked by mixtures are not a simple combination of the component patterns, indicating that components interact. Because such interactions were not observed in afferent glomerular activity patterns, they most likely arise in the OB. Possible mechanisms include dendritic integration in MCs and circuit interactions. Because mixture interactions often affect inhibitory and temporally patterned MC responses, circuit interactions are likely candidates (Friedrich and Laurent, 2004).

The component dominance in binary mixtures of single compounds also has been observed in temporally patterned responses of rat MCs (Giraudet et al., 2002). Because it could not be predicted which of the components in a mixture will dominate, the activity pattern across MCs evoked by a binary mixture cannot be predicted accurately from the component patterns. Nevertheless, the pattern evoked by the mixture is related to the component patterns, because the response of each MC resembles the response to one or the other component.

MC responses to binary mixtures of food extracts were dominated less by one component than amino acid mixtures but acquired some novel properties, indicating more complex mixture interactions. Patterns evoked by binary food odor mixtures therefore are related less closely to their component patterns. Each food odor component itself contains many individual compounds. However, at the level of input to the OB, the main difference between food extract and amino acid stimuli is the more widespread and dense activation of glomeruli by food extracts. Differences between mixture interactions among amino acids and food extracts therefore have to be explained by these different properties of glomerular input patterns.

Component dominance decreased with increasing similarity of afferent activity patterns for amino acid stimuli. This trend is consistent with the low component dominance in binary mixtures of food extracts and suggests that mixture interactions become more complex with increasing overlap of afferent component representations. Similar effects have been observed in a simple circuit model of the OB (Linster and Cleland, 2004) and were explained by lateral inhibitory interactions in the OB, which are expected to become denser with increasing pattern overlap. As a consequence, the output of each MC deviates more from its sensory input because the relative strength of inhibitory inputs increases. Hence, the dependence of component dominance on the overlap of input activity patterns may reflect nonlinear interactions mediated by inhibitory interneurons in the OB. However, additional insights are required to understand the mechanisms underlying mixture interactions in the OB.

Mixture interactions also have been described at the next level of the olfactory pathway. These mixture interactions are experience-dependent and appear to contribute to the establishment of complex receptive fields of neurons in the anterior piriform cortex (Wilson, 2003). Mixture interactions in higher olfactory areas therefore also may contribute to the processing and the synthetic perception of odor mixtures (Wilson and Stevenson, 2003).

Mixture representations in the olfactory bulb and odor perception

Our results suggest tentative rules describing some properties of binary mixture interactions in the OB that are similar to conclusions derived from the model by Linster and Cleland (2004). First, patterns of glomerular input activity evoked by a mixture are given by the combination of the patterns of the components. Second, MC responses to mixtures of dissimilar compounds are dominated quantitatively by the response to one of the components. Third, for mixtures of similar components, the sign, but not necessarily the magnitude, of the response corresponds to that of at least one of the components. Additional properties of mixture interactions may exist. For example, it is possible that the relative strength of components influences mixture interactions.

The properties of mixture interactions in the OB lead to predictions concerning odor representations by activity patterns across multiple glomeruli or MCs. First, the predictability of a mixture representation across glomeruli from the representations of the components is high. Much of the information about the components is therefore, in theory, available at the level of glomerular inputs. Second, the predictability of mixture representations by MCs is generally lower than the predictability of afferent mixture representations, indicating that processing in the OB discards component information. This may contribute to the synthetic nature of odor perception and constrain the analytical power of olfactory processing. Third, MC representations of mixtures of simple dissimilar components combine parts of the representation of each component, because the response of each MC represents one or the other component. Hence, the mixture representation neither is close to the average of the representations of the components nor is it a novel representation, but the response of each MC is predictable with a certain probability. This organization retains some information about the components. Fourth, with increasing similarity of the components the representation of the mixture is still related to those of the components but becomes less predictable. Information about the components therefore decreases, and the representation acquires novel properties. This effect is pronounced for mixtures of food extracts, suggesting that mixture interactions play an important role in the processing of natural stimuli. One function of mixture interactions also may be to limit the total MC activity when the number of active glomeruli increases.

Our results give rise to experimentally approachable predictions about the effect of mixture interactions on odor perception that can be compared with existing psychophysical or behavioral data from other vertebrates. The component information in the representation of a mixture should decrease with the number of components, because fewer responses of MCs should be dominated by each component and because component dominance should decrease in general. Indeed, human psychophysics demonstrated that the segmentation of a mixture into its components becomes impossible when the number of components exceeds 3-4 (Laing and Francis, 1989; Livermore and Laing, 1996). Furthermore, component identification should become more difficult and the perception of a binary mixture should acquire a more novel quality with increasing similarity of the components. Indeed, rats cannot identify components reliably in binary mixtures of similar, but not dissimilar, compounds. Rather, mixtures of similar compounds are perceived as different from both components (Wiltrout et al., 2003). Hence, mixture interactions in the OB may play an important role in odor perception.