This is a preprint of a paper published in Proceedings of the National Academy of Sciences, USA (2007) 104:1953-1958.

Biological Sciences: Neuroscience

Relational representation in the olfactory system

Thomas A. Cleland¹, Brett A. Johnson², Michael Leon^2*, and Christiane Linster¹

¹Dept. Neurobiology and Behavior, Cornell University

²Dept. Neurobiology and Behavior, University of California, Irvine

* Corresponding author: Michael Leon, Dept. Neurobiology & Behavior, University of California, Irvine, CA 92675. Phone 949-824-5343; fax 949-824-2447; mleon@uci.edu.

15 text pages

5 figures (+2 in supplementary text)

0 tables (+2 in supplementary text)

Abstract: 175 words

Manuscript: 34,915 text characters + figure/margin allotments (11,940) = 46,855

Abstract

The perceptual quality of odors is usually robust to variability in concentration. However, maps of neural activation across the olfactory bulb glomerular layer are not stable in this respect; rather, glomerular odor representations both broaden and intensify as odorant concentrations are increased. The relative levels of activation among glomeruli, in contrast, remain relatively stable across concentrations, suggesting that the representation of odor quality may rely on these relational activity patterns. However, the neural normalization mechanisms enabling extraction of such relational representations are unclear. Using glomerular imaging activity profiles from the rat olfactory bulb together with computational modeling, we here show that (1) global normalization preserves concentration-independent odor quality information, (2) perceptual similarities, as assessed behaviorally, are better predicted by normalized than by raw bulbar activity profiles, and (3) a recurrent excitatory circuit recently described in the olfactory bulb is capable of performing such normalization. We show for the first time that global feedforward normalization in a sensory system is behaviorally relevant, and that a center-surround neural architecture does not necessarily imply center-surround function.

The accurate, replicable perception of stimulus quality over a wide range of intensities is a central issue in sensory neuroscience. Fundamental limitations in the physics of sampling and unpredictable sources of interference generate variance in primary sensory maps that subsequent processing layers must deconstruct in order to synthesize a useful stimulus representation. Indeed, percepts of stimulus quality are generally synthetic, arising from directed compilations of sensory information. For example, visual stimuli are represented in unpredictable coordinates on the retina, with the visual percept of an object ultimately being synthesized from numerous sequential primary images. The timbre of a clarinet – the stationary component of which is largely derived from the ratios of amplitudes of the overtones of a fundamental frequency – can be identified irrespective of the pitch or volume of a given note. Features of texture can be identified across a range of absolute pressures. In each of these cases, information necessary to identify critical features of stimulus quality must be contained in the relative, or relational, response profiles of activated sensory channels, rather than their absolute activity levels.

Relational representations depend upon the broad patterns of primary sensory activation that are typically observed in response to elemental stimuli. For example, the movement of a single vibrissa evokes responses across a broad region of somatosensory cortex (1), a single point of light evokes widespread activity in visual cortex (2), and a single pure tone elicits a response across large areas of auditory cortex (3). Odorants evoke comparably broad responses in the mammalian olfactory system. Primary olfactory sensory neurons (OSNs) express a single olfactory receptor (OR) species in rodents, yet are broadly selective for odorant stimuli (4-6), such that even monomolecular odorants can evoke responses across much of the epithelium (7) and the glomerular layer of the olfactory bulb (OB) (8). Due to the convergence of OSNs expressing the same OR species onto common glomeruli, the pattern of activated glomeruli on the surface of the OB reflects the pattern of activated ORs, thus clearly identifying the constellation of chemical qualities that constitute the presented odor (8-11).

We investigated the possibility that relational representations enable the maintenance of odor quality over large differences in odorant concentration and the mechanism by which the olfactory system can maintain these relational representations. It has been well established that the degree of overlap among different glomerular activity maps reflects odorant structural commonalities and corresponds to the perceptual similarity of odors (12-15). However, this relationship breaks down when odorant concentration is included as a variable. OSN activation profiles and the corresponding glomerular activity maps universally become broader and more intense when odors are presented at higher concentrations (Fig. 1A) (16-20), simply because higher odorant concentrations recruit additional OSN populations with progressively lower affinities for the presented agonist(s).

Normalization of glomerular activity maps mitigates these disruptive effects of concentration. In contrast to absolute maps of glomerular activity, normalized glomerular images reveal odor-specific maps that are substantially insensitive to concentration (Fig. 1B) (17). By implication, this normalized output would be mediated by mitral cells, second-order principal neurons that integrate inputs from primary sensory neurons and interneurons within the OB.

The computation of intensity-independent stimulus representations in sensory systems depends on the existence of associated inhibitory systems that are more broadly activated by sensory stimuli than are the principal neurons receiving direct sensory inputs (23-25). Specifically, inhibition that is uniformly distributed across the input field – perhaps representing an average or sum of all input activity – will tend to preserve the pattern of relative activation levels across the field irrespective of total intensity, while inhibition that is delivered in proportion to local activation levels, or in an otherwise biased manner, will retain concentration-dependent distortions in the resulting output patterns. Recent work in the OB (26, 27) has identified an interconnected network of excitatory juxtaglomerular interneurons – external tufted (ET) and short-axon (SA) cells – that delivers inhibition onto OB principal neurons (mitral cells) via local inhibitory neurons (periglomerular, or PG, cells), and it has been suggested that this recurrent excitatory network mediates the uniform inhibition required for concentration-invariant secondary representations and other odor processing mechanisms (23). We show here that this interglomerular excitatory network, via activation of local inhibitory neurons and with estimates of synaptic densities and connectivity derived from experimental data, is not only capable of computing the required normalization, but seems optimized to perform this operation. Using a full-scale, 2200-glomerulus model to process and analyze 2-deoxyglucose imaging data from the rat OB, along with behavioral data, we demonstrate the feasibility and potential mechanism of perceptual concentration invariance in olfaction.

Results

Because normalization is effected at the level of mitral cells, which are often inhibited below baseline by odorant presentation, we normalized with respect to the mean using a z-score transformation (z_i = (x_i-m_x)/s_x). To measure the similarity between glomerular activation patterns, we calculated pairwise indices of dissimilarity (normalized Euclidean distance) between the raw and normalized patterns of activation evoked by four odorants – 2-hexanone, methyl valerate, n-pentanal, and n-pentanol – at vapor-phase concentrations of 25, 75 and 250 ppm. The indices of dissimilarity between the normalized patterns evoked by any given odorant across concentrations were significantly reduced compared to those of the corresponding raw patterns (p < 0.05 in all cases; Fig. 1C, D), indicating that normalized patterns represent concentration-invariant quality information significantly better than do raw response patterns. In contrast, dissimilarity indices between pairs of different odorants at the same concentration were not consistently reduced by normalization; effects varied from a marginal reduction to a substantial increase in dissimilarity (Fig. 1E).

If normalized glomerular activity patterns represent the output of bulbar computations performed on raw input patterns, then they should be better predictors of perceptual similarity than the raw patterns. Indeed, quantitative comparisons between glomerular activation patterns and olfactory perception have generally been based on normalized data