The Nose That Glows

David Pacchioli
September 01, 1999

It might be good to begin with what the Tufts artificial nose is not. It is not a proboscis. It is not a prosthesis. It is not a cleverly designed substitute appendage that a smell-impaired person might strap or screw or otherwise affix to his or her heretofore nose-less physiognomy. One cannot blow the Tufts artificial nose.

"The term 'artificial nose' is not in the grant," says Peter Jurs, preemptively. "That's kind of a flamboyant way to describe it." On the other hand, the thing does, after a fashion, actually smell. Or as Jurs, a professor of chemistry at Penn State who is working closely with the nose's originators, puts it: "Our scientific objective is to generate instruments which mimic the behavior of the mammalian nose in the sense that they can identify volatile organic compounds in an airstream."

In gross appearance, the nose is an assembly of fiberoptic components about the size of two shoeboxes—although it can, and no doubt will, be made smaller. At its heart is a bundle of 19 fibers, the same kind that carry telephone signals. This nose, it turns out, does its smelling with light.

"The tip of each fiber is coated with a different polymer," Jurs explains, and each polymer is infused with a fluorescent dye. When exposed to vapor molecules from various chemical compounds, the polymers react according to their individual natures: In response to benzene, say, one polymer may swell, another shrink, and a third undergo a change in polarity. These varied responses in turn create unique changes in each tip's fluorescence. In short, the tips are smell-activated optical sensors.

"You know how fluorescence works, right?" Jurs goes on. "You shine light on it, it shines light back at a different wavelength. So, we send continuous laser light down one end of this bundle. On the other end, we set up an airstream. Then we waft a sample across the tips of the fibers."

As the polymers react, and the returning fluorescence is altered, a video camera records the tips in a cross-section image: 19 circles, like a cluster of Christmas-tree lights. Over the course of a two-second pulse—the whiff of a vapor sample—these circles change dramatically in hue and brightness. Some go blue; others orange, yellow, red. "Some light up, others dim," Jurs says. "Some do it slow, some fast." This multicolored pattern, different for every chemical compound, is the unique signature of an odor.

The physiological rudiments of smell are fairly well known. The sensory organ —the nose—is lined in its upper reaches with some 50 million olfactory neurons, nerve cells designed to respond to the teeming variety of molecules that make their way into the nasal cavity, from the tantalizing scent of hot pizza to the less-inviting smells of brackish pond water or fresh manure. These cells are connected, by virtue of long, fiber-like axons, to the olfactory bulb—the pea-sized fore-brain region that processes incoming smell signals and moves them along to the cortex, where they are further distributed to other parts of the brain.

Beyond this bare outline, however, the picture gets somewhat hazier. While recent discoveries have clarified important aspects, they have also demonstrated that the process of smelling is more complicated than had previously been imagined.

John Kauer, a neurophysiologist at Tufts Medical School, has been studying the biology of olfaction—the brain mechanisms that underly the encoding of sensory information—for some 20 years. His principal model during that time has been, as Kauer puts it, "the lowly salamander."

With its simple nervous system and well-developed nasal organs (needed for finding both food and mates), the salamander makes a very attractive olfactory study. Its sac-like nasal cavity, Kauer explains, can be readily opened, providing easy access to the olfactory neurons. Using an optical recording system that he devised, he can make his observations non-invasively, "with special dyes and video cameras instead of probes. You can actually watch the brain in action. It's fantastic."

Over the years, Kauer says, what these observations have shown is that smell works by what he calls "a distributed combinatorial system. It doesn't work directly, one sensor to one odor," he explains. There isn't a particular sensor that lights up for the smell of hot driveway tar and another for the odor of new-mown grass. Rather, a given odorant triggers a number of sensors, and not necessarily all at once. It is this combination— and more, the unique pattern of activation over time—that codes information to the brain. "It's somewhat like a chord on a piano," Kauer says. "It includes several individual notes, and they can be played simultaneously or one after another, in various orders. The effect is different each time."

In 1991, Linda Buck and Richard Axel of Columbia University gave the combinatorial idea a big push forward when they identified the precise mechanism that the olfactory neurons employ. On the surfaces of these cells, it turns out, reside a class of proteins, now known as odorant receptors. In humans, these receptors come in roughly a thousand different types, each type occurring in only one of four distribution zones within the nasal cavity. Again, these receptors work not singly but in combination to identify odors: Buck, now at Harvard, has likened them to the letters in an olfactory alphabet. Thus it is that only a thousand types of receptor are needed for the human nose to be able to discriminate between the roughly 10 times that many odors that most people can discern. What goes on in the brain, at the most basic level, is a form of pattern recognition.

Looking at all the data that he and others had accumulated over the years, Kauer says, "we thought that what had been learned about olfactory signal processing should be put into some sort of formal framework. We wanted to write equations that would explain, for example, how animals can make such exquisite sensory differentiations between odor molecules—which govern their lives, after all. And it occurred to us, if we could put this set of equations together with a sufficiently sensitive array of sensors, we would have an artificial olfactory system."

He joined with Tufts colleague David Walt, a chemist whose lab specializes in the building of very sophisticated molecular sensors, and together the two created the Tufts artificial nose. "The system is able to identify individual vapours at different concentrations with great accuracy," they reported in the British journal Nature in 1996. "What we get is a spatio-temporal pattern that represents an odor," Kauer says now. "In order to 'read' this odor, we need a brain of some type. So we have tried to replicate the signal processing mechanism in the biological brain. That's how we hooked up with Peter Jurs."

My expertise is in computational chemistry," Jurs says. "That's my whole game." He is a youthful 50-something; sturdily built; authoritative but also wryly boyish. As he talks, one hand scissors absently through his wavy gray-brown hair.

As computational chemists, he explains, what Jurs and his students spend most of their time doing is developing and testing algorithms: software programs that will enable them to sort through huge and complex piles of chemical data and pull out significant patterns. "We do experiments in our lab," Jurs says, "just not with glassware, and not in lab coats."

One type of algorithm they are using with increasing success is the computational neural network, a program that works by mimicking the brain's parallel-processing capability. Like the brain, a neural network consists of a large number of interconnected nodes (in the brain's case, neurons), each of which does only relatively simple processing. The key to its power is that the nodes work all at once. Even a huge task, spread across such a network, is much more rapidly processed than if it were tackled in sequential steps, as is the case with conventional computer processing.

Jurs, who wrote his doctoral dissertation on a related subject in 1969, is one of the pioneers in the use of neural networks in chemistry. A neural network, he says, is perfectly suited to serve as the "brain" for an artificial nose.

What those 19 optical sensors present, he explains, is a huge amount of data, in a very complex form. Each image contains a wealth of information: both the color and intensity of individual sensors changing across time and the combinatory pattern made by the entire array. A conventional rule-based program simply couldn't handle all that data. At best, such a system would be severely limited in the number of odorants it could recognize. It would easily be stumped by mixtures of compounds, competing odors, varying concentrations —smells that were close to but not exactly what it already knew.

The beauty of the neural net, in contrast, is that in effect it makes up the rules as it goes along. Its superior processing capacity allows it to be "generically" equipped: set up to recognize patterns, no matter what form they might take, and to make associations. It is thus adaptable to many different kinds of problems. Instead of being programmed, it is "trained."

"It's a bootstrap method," Jurs explains. For each set of odorants to be analyzed, the neural net is first fed a large helping of sample odors to train on. "At the very beginning it's ignorant, but as it goes along, it builds up a set of prototypes or patterns, an internal model of the 'world' it has experienced." The training process is reinforced by positive feedback, through numerous iterations: after each run, the net's 'answers' are checked against reality. Where necessary, they are corrected, and the patterns are refined. "It's supervised learning," Jurs says. "Similar to teaching a child how to spell."

Only when training is complete is the nose given actual "live" samples to identify: odorants that it hasn't seen before, but which fall somewhere within the "world" it has learned to negotiate. The new samples are compared against existing patterns. If they match, they are slotted accordingly. If not, the program either adapts an existing pattern or creates a new one.

Even before the system can be trained, however, Jurs has to deploy other algorithms to render the nose's complex optical data into digestible form. Each sensor's response is first plotted as a curve: 19 curves per odorant. Next, each curve is converted into a set of numerical descriptors, each number corresponding to one of its attributes—its slope, maybe, or its highest or lowest intensity. It may take 15 numbers to describe a curve adequately, Jurs says. Unfortunately, 15 numbers per curve—285 per odorant—is too many numbers to process efficiently. To sift through all the descriptors and find which ones carry the most descriptive weight, Jurs uses something called a genetic algorithm.

The name is no caprice, or hardly one. There is indeed a Darwinian parallel, as Jurs explains: It's survival of the fittest numbers. Ask the genetic algorithm to choose the best set of 10 descriptors from a pool of 100, and it will first generate a number of random solutions, "maybe 50 different sets of 10." A sub-routine then kicks in, to evaluate the fitness (in this case, the descriptive power) of each of these sets. The less fit are allowed to "die;" the better ones survive to be "mated," paired off and recombined to form "progeny." The fitness of the offspring is in turn determined, and the process is repeated. "With each iteration," Jurs says, "you end up with a population of better-quality descriptor sets. We let the algorithm run and run and eventually we end up with a very good set of features." This well-honed set, finally, is used to re-construct, mathematically, the odorant that has just been "smelled." The model is compared against the patterns established by training, and the program predicts its identity.

"It works amazingly well," Jurs reports. In a recent experiment, what he calls "the needle in the haystack problem," the artificial nose was asked to determine the presence or absence of the cleaning solvent trichloroethylene, or TCE, in air or in the presence of other compounds. After training the nose on almost 800 samples, the researchers threw it a set of over 100 samples of 19 odorants, from pure compounds like benzene and carbon tetrachloride to unholy mixtures like kerosene and Coleman fuel. Some samples contained small amounts of TCE, others none. The nose classified correctly 94.6 percent of the time.

Theoretically, John Kauer says, there are any number of applications for an artificial nose. "One area would be for physical diagnosis. There are subtle odors associated with certain disease states—cancer and diabetes, for example. This could be a non-invasive screening method." Another area with potential, he adds, is environmental monitoring—"not just for obvious odors, but for those our noses don't detect. The smell of bearings going bad in an elevator, say, or of the breakdown of pipes. There may be a subliminal olfactory world in our environment that an artificial device could uncover."

A more immediate objective, however, is to adapt the artificial nose to the urgent task of detecting buried land mines. Toward this end, with funding from the Defense Advanced Research Projects Agency, Kauer and colleagues are working to make the artificial nose behave more like one of evolution's finest dedicated smelling machines: the nose of a dog.

"What we're talking about is these little anti-personnel mines, about the size of a tuna-fish can," Kauer says. "They are a terrible problem in many parts of the world. Because they contain no metal, they can't be located with traditional detectors. "Currently, there are only two ways to find them," he continues. "One is to have a guy crawling along on his stomach with a stick out in front of him, probing ahead. This is terribly dangerous, of course, but this is how it's done in most of the world. The other way is to use dogs." (Penn State's Jeff Schiano is working on a third way. See Research/Penn State September 1998.)

A trained dog, with its vastly super-human smelling ability, can pick up what Kauer calls the "signature" of a buried mine: "most likely a complex mixture of various odors, from the explosive material, the plastic casing, and the local disturbance of the earth." This scent is extremely rarefied; it taxes even canine ability to its limits. Still, where they are available, dogs do very well at land-mine work. But dogs are expensive to train and to keep, and they have woefully short attention spans. An artificial device, attached to some sort of simple robot, could do the job much more cheaply and efficiently, and the only risk would be to machinery.

Trying to make an artificial nose sufficiently dog-like, Kauer acknowledges, first requires understanding how the dog manages to smell so well. "Yes, the nose itself is much more sensitive than a human's," he says. "But there are other things. Dogs are built for smelling. If you and I were to get down on our hands and knees, close to the ground, we would probably enrich our olfactory experience considerably. Also, dogs attend more to odors than we do. Smells are more important to them." Finally, the canine system apparently does a much better job of delivering odors to the sensors. "It turns out the sniffing process is not trivial," Kauer says. "It's not just how much odor is received, but how much is received in a given time frame. They're two completely different questions."

In order to get a clearer picture of—in grant-writing parlance— "the aerodynamics of canine olfaction," Kauer turned to Gary Settles, a professor of mechanical engineering at Penn State.

Settles, director of the University's gas dynamics laboratory, uses lasers, high-speed cameras, and what he bills the world's largest optical flow-visualization system, to create and analyze images of airflow and heat transfer from ventilation hoods, automobile exhaust systems, and air conditioners (See R/PS, May 1997). For this project, he trained his sophisticated equipment on the equally sophisticated schnozz of Bailey, the one-year-old golden retriever of his lab manager, Lori Dodson.

"We trained Bailey to place her head in a v-block," Settles says, "to keep her steady while we filmed the dilation of her nostrils with a high-speed video camera." ("It took a couple of months, every day," Dodson adds. "Now, at home, she'll lay her head on the coffee table and expect a treat.") They also did light-scattering studies with talcum powder placed around a scent source—a chunk of carrot, a piece of moss ("We were racking our brains to find stuff that would keep her interested")—that Bailey was trained to approach and sniff. The results of these efforts are apparent in a series of somewhat surreal video clips.

"You can see that she puts her nose right up against the scent source," Settles narrates. "She uses her whiskers as contact sensors—she can't see around that long snout." As the dog sniffs repeatedly, colored plumes show that her side-slitted nostrils—an evolutionary advantage shared with cats, rabbits, deer, and bears, among other mammals—direct each exhalation so that it doesn't interfere with the subsequent intake. "Here you can see the powder being blown back and to the sides," Settles says at one point."By tracking the particles we were able to establish the velocity of inhalation."

He and Dodson also isolated two distinct modes of sniffing: the long sniff, of a second or two in duration, used "when access to the scent source is limited," (i.e., your head is in a v-block); and the snuffling short sniff, "where the dog reads across a scent like she's reading a line of text. We found that dogs use short sniffs exclusively in land-mine detection," Settles concludes. Now, he adds, they're working on incorporating what they've learned into a design for a "front-end device" for the artificial nose—a mechanical sniffer.

No matter how proficient the Tufts artificial nose becomes at whatever specific scenting operations it is assigned, when it comes down to real-world smelling it will never come close to the authentic canine, or even the human, variety of nose. No one who is involved in the artificial-nose field seems to have any illusions about that.

"Making a chemical sensor to detect a chemical molecule is not making a sense of olfaction," admits Nathan Lewis, a Cal Tech chemist who spoke at Penn State in March about his own exploits in artificial-nose technology.

From Jurs's perspective, training a neural network to recognize just a handful of odors is already a giant task. "To train a nose for the general purpose of smelling . . . " His face splits with a 'that's impossible' grin. There are, after all, 700 odorants in a single glass of wine.

For Kauer, who has spent a career searching out the intricacies of olfaction, "any artificial device is going to be extremely simplistic in comparison to the biology, which is wonderfully elegant and sophisticated."

More than for any of its potential applications, in fact, Kauer prizes the artificial nose for what it may yet tell him about the sublime sensory apparatus that nature has perfected. "We continue to move back and forth between the two systems," he says. "We learn something from one and we carry it back to the other."

Peter C. Jurs, Ph.D., is professor of chemistry in the Eberly College of Science, 152 Davey Laboratory, University Park, PA 16802; 814-865-3739, or pcj@psu.edu. Gary S. Settles, Ph.D., is professor of mechanical engineering and director of the Gas Dynamics Laboratory, College of Engineering, 301D Reber Bldg., University Park PA 16802; 814-863-1504, or gss2@psu.edu. John Kauer, Ph.D., is professor of neuroscience at Tufts University Medical School, Boston, MA. Funding for the artificial nose project is provided by the Defense Advanced Research Projects Agency.

Last Updated September 01, 1999