Disorder and the neural representation of complex odors: smelling in the real world

Kamesh Krishnamurthy, Ann M. Hermundstad, Thierry Mora, Aleksandra M. Walczak, Vijay BalasubramanianbioRxiv 2017 · DOI

TL;DR

This paper proposes that the olfactory system leverages "disorder" at multiple stages to overcome the challenge of identifying a vast number of odors with a limited number of receptors. By using diffuse receptor binding, densifying/decorrelating neural signals, and employing disordered projections, the system creates a representation that is both compact and flexible for learning odor associations.

What Problem Does This Paper Address?

The olfactory system faces a fundamental challenge: how can animals distinguish between potentially billions of different odor molecules using a relatively small repertoire of olfactory receptors (e.g., ~300 in humans, ~1000 in mice)? This is a classic problem of mapping a high-dimensional input space (odor molecules) to a low-dimensional sensor space (receptor responses). Furthermore, this representation needs to be flexible enough to learn arbitrary associations between odors and their associated valences (e.g., pleasant or unpleasant). Traditional approaches often focus on specific receptor tuning, but this paper argues that the olfactory system's solution lies in exploiting "disorder" at various processing stages.

Methods

The authors propose a three-stage model of olfactory processing and test it using data from Drosophila.

  1. Olfactory Receptor Neurons (ORNs): They hypothesize that ORNs use "diffuse sensing," meaning each receptor binds to a broad range of odorants. This compresses the high-dimensional odor space into a lower-dimensional receptor space while preserving similarity between odors. They analyze existing Drosophila ORN response data to monomolecular odorants and model responses to complex mixtures.
  2. Glomerular Transformation (Antennal Lobe): The responses from ORNs converge in the Antennal Lobe (analogous to the olfactory bulb in mammals). Here, lateral inhibitory interneurons perform a "densification" and "decorrelation" of the ORN responses. This process is modeled as a divisive normalization.
  3. Disordered Projections (Mushroom Body): The transformed signals are then projected to the Mushroom Body (analogous to the piriform cortex). This stage involves "disordered" and "expansive" projections, which are hypothesized to reformat the information for flexible learning. They simulate Kenyon cell responses based on these projections and test their ability to support classification.

Datasets

The primary dataset used is from Hallem and Carlson (2006), which provides firing rate responses of 24 Drosophila ORN types to a panel of 110 monomolecular odorants. This dataset is used to model ORN responses to mixtures and to simulate downstream processing.

Key Insights and Ideas

  • Disorder as a Feature, Not a Bug: The central idea is that disorder (diffuse binding, decorrelation, random projections) is not a limitation but a crucial computational strategy for the olfactory system.
  • Compression via Diffuse Sensing: The initial diffuse binding of ORNs to many odorants effectively compresses the vast odor space into a manageable receptor space, preserving essential similarity information.
  • Densification and Decorrelation for Robustness: The Antennal Lobe's processing enhances robustness to noise by densifying and decorrelating neural responses.
  • Disordered Projections for Flexible Learning: The expansive and disordered projections to higher brain areas are key to reformatting the compressed information, making it suitable for learning arbitrary associations between odors and valences.
  • Any Subset is Good Enough: The disordered nature of projections means that any sufficiently large random subset of neurons in the Mushroom Body can support flexible learning, implying robustness to neuronal loss or variability.

Key Results

  • Drosophila ORN responses can accurately decode complex odor mixtures, even with a small number of receptors, supporting the diffuse sensing hypothesis.
  • The divisive normalization in the Antennal Lobe increases the disorder (densifies and decorrelates responses) and significantly improves the decodability of odor mixtures compared to raw ORN responses.
  • A linear classifier trained on simulated Mushroom Body responses (representing disordered projections) can classify odor mixtures into arbitrary categories with much higher accuracy than classifiers trained on ORN or glomerular responses.
  • Removing disorder (either by removing densification or introducing structure in projections) significantly degrades the system's robustness to noise.

Limitations and Future Work

  • The authors acknowledge that their decoding algorithm is a benchmark and the brain may not use this exact scheme.
  • The study focuses on Drosophila, but the authors suggest that similar principles likely apply to mammalian olfactory systems due to analogous anatomical and functional motifs.
  • Future work should involve more complex behavioral tasks that reflect real-world olfactory challenges to further test these predictions.
  • Experimental manipulation of specific components of the olfactory pathway (e.g., blocking inhibitory neurons, altering projection patterns) could directly test the proposed roles of disorder.

Summary updated

Abstract

Animals smelling in the real world use a small number of receptors to sense a vast number of natural molecular mixtures, and proceed to learn arbitrary associations between odors and valences. Here, we propose a new interpretation of how the architecture of olfactory circuits is adapted to meet these immense complementary challenges. First, the diffuse binding of receptors to many molecules compresses a vast odor space into a tiny receptor space, while preserving similarity. Next, lateral interactions “densify” and decorrelate the response, enhancing robustness to noise. Finally, disordered projections from the periphery to the central brain reconfigure the densely packed information into a format suitable for flexible learning of associations and valences. We test our theory empirically using data from Drosophila. Our theory suggests that the neural processing of olfactory information differs from the other senses in its fundamental use of disorder.
BibTeX
@misc{krishnamurthy2017disorder,
  author = {Kamesh Krishnamurthy and Ann M. Hermundstad and Thierry Mora and Aleksandra M. Walczak and Vijay Balasubramanian},
  title = {{Disorder and the neural representation of complex odors: smelling in the real world}},
  year = {2017},
  doi = {10.1101/160382},
  howpublished = {bioRxiv},
}
Figure