Nose Knows More than Math Pros Suppose
The aroma of coffee, of a steak, of cherries – these smells are all composed of dozens if not hundreds of separate molecules, yet our brains immediately recognize them each as a coherent whole. How does the nose and the brain process all this information? This is the subject of an article in the Caltech magazine Engineering and Science1 by Gilles Laurent, Caltech professor and neurologist, who studies olfaction and also “how single neurons perform nonlinear operations such as multiplication.”
Unlike vision and hearing, our olfactory sense does not allow us to decompose a composite input into its constituents. We perceive odors as single entities. Studies on insects by Laurent and his students show that this is because individual receptors fire in patterns that are mapped like a code to a large number of unique sensors called Kenyon cells. In insects, these cells reside in a part of the olfactory apparatus called the mushroom body (in vertebrates, it’s the olfactory cortex of the brain). Each Kenyon cell gets a very unique set of inputs from the receptors, and thus a distinct, composite signal from a highly diverse set of inputs. Laurent does the math to show the staggering number of possibilities for odor memory that this system permits:
The locust has 800 projection neurons connecting to 50,000 Kenyon cells. With such a large mismatch in numbers, how are these nerve-cell populations interconnected? When Ron Jortner, a graduate student in my lab, recorded simultaneously from both projection neurons and individual Kenyon cells to assess the probability of connection between them he found, surprisingly, that the probability was about 0.5. In other words, each Kenyon cell seems to connect on average to half of the input population, that is, to 400 projection neurons. The number of ways in which 400 neurons can be selected out of 800—the number of possible connection patterns—is about 10240. It’s an enormous number. To put it in context, there are about 1010 seconds in a century, and there have been about 1019 seconds since the beginning of the universe. With 10240 possible combinations of projection neurons to choose from—assuming random connectivity—almost every Kenyon cell is likely to sample a combination of inputs that is very different from that sampled by the other Kenyon cells. Each cell will therefore gain a picture of the state of the projection neuron population very different from that gathered by any other Kenyon cell. It follows that the responses of individual Kenyon cells will be very specific; a given cell should respond only to particular combinations of activated projection neurons, maximally different on average from those experienced by the other Kenyon cells. (Emphasis added in all quotes.)
Laurent noted at the beginning of the article that olfaction is a form of pattern recognition, and that “Brains solve pattern-recognition problems much better than any machine built today.” His lab tries to figure out “how brains solve these problems.” Most of the research by Laurent and his students is on insects, whose olfactory receptors are on their antennae. A fruit fly has 1300 receptor neurons, with 60 different receptor types, but some moths might have several hundred thousand receptor neurons. This gives them an amazing sensitivity to low concentrations of odors like pheromones.
A diagram and electron micrograph on p. 44 shows what receptor neurons look like. They have dozens of cilia projecting into the nasal cavity. The reason dogs have superior sensitivity to smells, he explains, is that their nasal cavity contains much more surface area where the receptors project from sponge-like tissue called turbinate bone. Dogs have ten times as much turbinate bone as humans. He provides a fragrant illustration: “In a medium-size dog,” he says, “the turbinates have a total surface area the size of a large pizza. In humans, they’re the size of a large cookie.” Each receptor neuron has a single sensitivity dictated by the order of the amino acids in its multi-folded receptor proteins. The amino acid sequences of receptor proteins show areas of both high conservation and high variability between species. They loop seven times through the cell membrane, providing pockets where the odor molecules bind.
Laurent describes something striking about how the receptor neurons map their inputs to ball-shaped structures called glomeruli (singular, glomerulus). “In an amazing feat of organization during development,” a picture caption states, “each type of receptor neuron… sends its axon to the same glomerulus….” He calls it a “surprise” that all the axons of the same receptor type (colored red in the diagram) converge so neatly to their exact counterparts. “By implication,” he continues, “this means there are about as many glomeruli as there are receptor types. And with the exception of the roundworm, this extraordinary organization is found in almost all the animal species that have so far been looked at.”
From the glomeruli, the information is passed on to a smaller group of nerve cells called projection neurons, which have no axons but connect with a dozen or more glomeruli. “With 100,000 receptor neurons converging on just 800 projection neurons, what is being computed?” he asks. Experiments show that the precise timing of firing creates a kind of code from the multiple inputs, a pulse pattern that can be mapped and analyzed. He likens the result to the unique arrangement that billiard balls take after the player breaks them with the cue ball; two very similar initial setups, but with slightly different angles of attack, can produce initially similar but ultimately divergent patterns of balls on the table. (The billiard game in the nose is super-fast. He notes on p. 48, “This happens so quickly that the representations are optimally separated within 100 to 300 milliseconds.”) As a result, differences between very similar smells can be amplified by the system. “That’s basically what we think is taking place in the olfactory circuit,” he says. “The remarkable thing is that this near-chaotic process is very sensitive to the input, but very reliable nevertheless.”
To recap, the receptor proteins in the cilia of the receptor neurons react to molecules in odors. These neurons fire their axons to the glomeruli. The glomeruli then pass their encoded information patterns to the projection neurons. That noise-reduced information is passed in very unique ways to the tens of thousands of Kenyon cells, which have a near infinite way to respond to the myriad possible combinations of smells. “Kenyon cells are so specific that they only recognize one, or at most a few, odors,” a caption explains on p. 51. He summed it up earlier (p. 46): “In other words, each odor is defined by a certain combination of receptors; the code is combinatorial…. The perception of an odor must therefore result from the brain’s interpretation of combinatorial activity patterns.” Why, though, do a large number of receptors map to few encoders, and then those few to a large number of interpreters? There’s a reason for everything:
It seems wasteful that hundreds of thousands of olfactory receptor neurons converge on their respective glomeruli in an amazingly precise way, but that this precision is then thrown away when seemingly disordered patterns of activation are generated in the projection neurons. But there’s a good reason for it. A system that amplifies small differences in signals runs the risk of also amplifying noise, in this case the noise coming from the receptors. Noise fluctuations would make the output of the projection neurons unreliable: the averaging that results from this kind of convergent design is precisely one way to reduce such fluctuations.
(p. 49; for more on the problem of noise reduction, see 12/20/2004 entry). The sense of smell, obviously, is “quite complex.” It involves many more receptor types than other senses, like vision, which uses only four types of photoreceptor. How did the code in the nose, and all the apparatus in the circuitry, come about? Early in the article he speculated briefly about this question, but his answer assumes a remarkable convergence rather than demonstrating the evolutionary steps:
In parts of the looping receptor protein chain, the order in which the amino acids are strung together is so variable that some animals, such as the rat, have over 1,200 different receptor types. On average, mammals have about 1,000 types, fish and birds between 100 and 200, round- worms (Caenorhabditis elegans) 1,000, and fruit flies 60. Humans have only 600 different odorant receptor genes, but almost half of these are “pseudogenes” that no longer function, leaving us with only 350 receptor types in our nasal mucosa….
Interestingly, when the receptor genes of mammals, flies, and worms were compared, no sequence homology was found. In other words, the genes had probably not evolved from a common ancestor: different types of animals had come up with their own particular (but related) designs for olfactory receptors independently throughout evolutionary history. Such convergent evolution, as it’s called, happens a lot in biological systems. The single-lens eye design, for example, has evolved independently at least eight times in the animal kingdom.
How that happened is left as an exercise, but for Laurent, his job is in the here and now, studying the sensitive yet reliable olfactory computer: “Finding the rules of such nonlinear dynamical problems is one of our goals” (p. 49). Concluding, he says, “Our research into olfaction is…giving some valuable insights into how such kinds of high-level synthetic representations arise from the organization and dynamics of neural circuits” (p. 51). The nose shows that “Classifying and recognizing patterns is, after all, what our brains do best.”
1Gilles Laurent, “Olfaction: A Window into the Brain,” Engineering and Science (LXVIII:1/2), [summer] 2005, pp. 43-51 (PDF).
This article is a good companion to the next one (see 06/25/2005 entry). The language is similar: circuitry, computation, communication, codes, signals, and information. The lead-in photo shows a man with a very satisfied look savoring a cup of coffee, probably unaware that he is sensing a cocktail of two to three hundred compounds. Did you have any idea how much computation and circuitry make that pleasant feeling possible? We joke about our noses and don’t usually give them the same respect we pay the eye or ear, but each sense is more wonderful than we could possibly realize.
Werner Gitt, in his delightful book The Wonder of Man, elaborates on some wonders of our human sense of smell. We have between 10 and 25 million receptor cells where the odor molecules fit with the proteins like a lock and key. Each olfactory cell measures only 5 to 15 millionths of an inch. Past these cells waft about 12 cubic meters of air per day, as we inhale and exhale 12,000 times. Our olfactory sense is extremely sensitive, exceeding the capabilities of most technological measuring instruments. We can detect one ten million millionth of a gram of mercaptan, for instance, and even distinguish between left- and right-handed forms of the same molecule. Remarkable as that is, we all know how the animal kingdom relies even more heavily on the sense of smell, marking territory with scents, using scents for sexual attraction, and navigating by their noses. A dog has 220 million receptor cells, tenfold more than we do; think of how dogs can be trained to sniff out bombs in luggage and people trapped under rubble or avalanches, or how bloodhounds can follow the footsteps of a crook all the way from the crime scene to his shoes. Maybe it’s good we humans don’t have that TMI problem, but our olfactory sensitivity is nothing to sneeze at. Smells enhance the taste and flavor of our food, color our world, and influence the way we think and act in many subtle ways. They warn us of danger, or attract us to pleasurable sensations. “Our memory for odours is astounding,” Gitt says; “nothing can stir up old memories better than a certain scent.” The fresh air in a pine forest, the sunshine after the rain, the fragrance of a rose, the symphony of smells at a table of great food – how impoverished life would be without a sense of smell. Thank God for your nose.
Laurent’s brief side trip into Fantasyland with Tinker Bell (see 03/11/2005 commentary) provided some comic relief for this intense and thought-provoking look at a system of mind-boggling complexity. Did you enjoy the Fairy Godmother’s song, the Ballad of Convergent Evolution?
Impossible! for a random mutation to become a neural circuit;
Impossible! for an unguided process to produce a code so perfect.
And four DNA bases will never produce Code Morses,
Such fol-de-rol and fiddle-dee-dee of course is:
Impossible!
But the world is full of zanies and fools
who don’t believe in sensible rules
and won’t believe what sensible people say…
and because these daft and dewey eyed dopes
keep building up impossible hopes—
Impossible! things are happening every day!
Nothing like a magic wand named Natural Selection to do impossible things. Just wish… and believe. While the Darwinists are wishing upon a star in Fantasyland, design scientists are turning Frontierland into Tomorrowland.