 started with the cloning of the insect odor receptors and then it's all kind of tropic stuff. And also she's done a lot of work with Andrea Scheller on human olfaction, which is what she's doing. Thank you so much. I love this place. I love this meeting. I've already learned a lot. I love the vigorous discussion in Tim Holi's talk. So my talk is modular. So vigorous discussion denounced me in public. Maybe we won't get through the whole thing. So my lab is most known for its work on insect olfaction. But we do have a component in the lab that studies human olfactory perception. As mentioned, spearheaded by my scientist, Andreas Keller, who actually retired from science. And this is going to be a retrospective view of all the work we've done in human olfaction. So just an introduction following on Stuart's beautiful introductory talk. So this is the problem. I love coffee. So what happens when you inhale coffee into your nose? You're actually, coffee isn't a single molecule. Most of the work done in olfaction, as we know because it's simple, is done on single molecules. But in the real world, Bill has so brilliantly done chemical ecology. The real world, it's multiple odors. And so how do we struggle with this? How do we predict what each of these individual molecules smells like? How do they come together in the brain to give you the percept of coffee? And we know that it's a much bigger problem because it's not just coffee. As this monolithic thing, there's hundreds and hundreds of different types of coffee, where the beans come from, how they're brewed. So the olfactory system is all about taking complex stimuli and trying to extract information from them. So we attempt to take these chemical cues and put semantic labels onto them. And I'll go over in my talk how fraught and maybe hopeless this is because words probably can't describe what we're feeling. A lot of these words are very culturally specific, individual specific. This is an attempt when you say, what does it smell like? I think is probably the most dangerous question in olfaction, because I don't think it really gets to the heart of what we do when we smell things. So again, Stuart laid out the problem. How do you go from chemistry to biology to psychology? And this goes all the way out from the natural source of the stimulus, coffee has hundreds of different molecules. You then are down to the molecule that exists that's going to be smelled. Most lab-based science uses monomolecular odorants in nature, it's all complex mixtures. Then we have the problem of detection. We had some beautiful, robust discussion about how sniffing, if you don't sniff, you generally don't smell anything. The strength and the direction of the sniffing directly governs what you'll smell. Then it goes into the olfactory mucus, and this is the horrible dirty secret of olfaction is that there are enzymes in the mucus that take the beautiful odorant and the mixtures and chew it up to create new odorants. So Tohara has done some work on this. So it's all well and good to purchase a molecule from Sigma, but once your animal has sniffed it, it's a little bit unclear if it's the original molecule from the vial or something that the enzymes and the mucus have done to that molecule. Then it hits the canonical buccanaxel receptors. Great privilege to have Linda in the room. And then there's signal transduction, which follows from that. Each of these points introduce some complexity. Then we get to the circuit, which in humans is extremely difficult or impossible for us to assess. Peter had this lovely work he just presented about trying to grapple with what does the circuit look like. Noam has done some early work on trying to image, on not trying, not imaging the olfactory system in behaving humans. But we completely lack the resolution that you'll ever get in an insect or a mouse because you can't. We're not yet allowed to go into humans and chop them up and do live imaging. But we know that the information courses from the epithelium to some pre-processing in the bulb to multiple olfactory cortices with strong influence of prior memories and associations. So then you get to the percept, from the molecule to the nose to the neural circuit to the perception. And olfactory perception isn't as perhaps simple as saying, like the simplest task of is this color red or green? Is this a line? Is it a tone? Each of these olfactory percepts is heavily influenced or contaminated by prior experience. Your memory of what the odorants, the first time you smell the odorant can overshadow an objective sense. Your state of emotion at the time that you smell the odor, whether or not you have a nasal infection can affect whether you can smell anything at all. And Peter just talked about the effect of neurodegeneration. So all of these things introduce complexity from chemistry to biology to psychology. So I'm going to talk today about molecules, odorant receptors, and perception in our work to try to tie these together. Realizing that all of these circuit elements at the moment we can't get to. So I'm going to talk about three stories. The first one is looking at variation in an odorant receptor and how that affects perception. That works out 10 years old, so I'll breeze over it. Then I'll talk about how you go from a molecule to a percept. How can we predict what a particular structure will smell like? So I have the privilege of having my major collaborator, Pablo Meyer, right there. Raise your hand. So Pablo Meyer was the director of this dream of faction prediction challenge where groups from around the world competed to try to figure out, given this molecular structure, this chemical identifier, can we predict what it smells like? One of the oldest and most difficult problems. And we made some progress. That involves machine learning. So you've been in talks where someone like biologists tries to get up there and explain a model, and you see they're struggling to try to figure out what was done. So I'm going to rely heavily on all the questions about the machine learning part go directly to Pablo Meyer. And then the last part, if there's time, we have used inspired by Noam Sobel's work on the perception of complex mixtures, developed what I hope will be a new olfactory test that will be really effective for diagnosing smell dysfunction in normal subjects who have smell dysfunction and people suffering from neurodegeneration. So we're excited about that. So this is what the Smell Lab at Rockefeller looks like. We don't have fancy olfactometers. We have amber vials that contain odorants. Modern molecular odorants, I would say, for the last 10 years, the previous 10 years, and then very recently, again, taking the Sobel complex mixtures. We sit people at tables. There's a computer screen. These are self-administered tests. They pick up a vial. They scan the bar code, and they do what we tell them. How intense is it? How pleasant is it? How much does it smell like sweat or coffee? And we can screen. We've screened thousands of people using this. So humans are great. Part of the reason that I have a human olfactory research program is because I became so frustrated at my inability to teach insects how to speak. It's so difficult to infer from animals what the hell they're thinking. What are they thinking? Are they ignoring? Are they generalizing to an odor or not generalizing? Because they just don't care. So humans, although we can't get into the circuit part of it, is the only, I would argue, because they're the only animal that we can maybe Diana Reese would differ with me. There are other animals you can communicate with. But with verbal sophisticated communications, humans are it. So I'm going to go through all of that. Yeah, so the descriptor is, like I said, like I think words. Putting words onto smells, we do it, but I think it's stupid. It's imperfect. So writes that we study human olfactory perception, the importance of the sniff, something that we can't really regulate. We're not intubating these people. We're not tracheotomizing them. It's really self-directed. So when you listen to my talk, keep in mind all of the imperfections. These people are not intubated. We don't control when they sniff. We can't control what the mucus is doing. We can't control if the odorant was linked to a traumatic experience. So the data are very noisy. Nevertheless, I believe that we've made some produced major insights into some of the questions. But just keep that in mind. Like we're not dealing with C. elegans where you can really control everything. So what do we do with our subjects when they come in? So intensity. This is pretty simple. Culture independent. People can say, I can't smell anything, too. It's extremely intense. So that's something. It's semantic free. And everybody can understand that task. And I don't think that's polluted by culture much. So one rose or many roses. Valence. So is it beautiful or is it disgusting? Again, on a 100 point scale, the most disgusting thing you've ever smelled are the most beautiful, orgasmately wonderful thing you've ever smelled. 0 to 100. This, I would argue, is very polluted by prior experience. So it really depends on whether you're Asian, where this fruit connotes a delicious, expensive experience. If you're from Europe, this cheese connotes a delicious, expensive experience. And vice versa is very different. So valence, pleasantness, unpleasantness, I think is already a complicated thing, keeping in mind the culture of the person. And I'm going to get onto my, the thing that makes me, gets me most agitated in olfaction is this difference between categorizing odors, smells like, and discriminating odors. So I think it's a key thing that confuses the literature. Categorization has an old history. Stuart introduced this idea that you try to divide whatever the total sum of all molecules that can exist, let's categorize them into the original ones were funny, like goatee, I can't remember them, floral. Yeah, ethereal. Yeah, ethereal. Words that we don't use anymore. Like goatee, goatee, smelling like a goat is not something that would, in this century, make the top seven list of categories, right? Yeah, yeah, exactly. But so again, the problem of culture and language, like in the, whatever that was, 1600s, goatee was like a major, let's divide, let's divide that. So here we're using pictures to try to categorize. You could make your own assumptions. You could have citrus, or you could have fruit. You can further categorize that into lemon, subdivisions of lemon. You could have plants, and you could have food, right? So you could have a food category. These things can fall into multiple categories, and all heavily overlaid by culture. If you don't know what you're looking at, then you can't categorize. If you haven't smelled it, you can't categorize. So we know that this is a violet. If you live in a world, part of the world where you have never smelled a violet, it's very difficult to do categorization. So these would all fall into a floral group, all bound by prior experience. If you haven't smelled these flowers, then that is very challenging to categorize them. So there's many of these categorization wheels out there. I've seen them for food, for flowers. This one is for body odor, which is really funny. So you start with segmenting an initial concept, body odor, into progressively finer and finer and finer gradations, but always limited by the words that you can use. So you can go from body odor to animatic to goatee. There we go, goatee. From body odor, vegetable to cabbage. So all of these are attempts to grapple with complexity in the olfactory system. I would argue imperfect, because they're heavily polluted by culture and prior experience. And so I would say that the field has benefited greatly from this list of descriptors. So these are the famous Dravniks descriptors from the mid-80s, where he sat down a series of highly trained people from the flavor and fragrance industry and said, look, let's come up with a list that helps us describe our products of all molecules that we could come up with so that we could have a common language. So if you make a molecule at Jiva Don, you could say it meets these criteria. You could communicate that to your customer. So this was not done with a view to really anything except commercializing food and flavor, ingredients, and some malodors. And so there's some things on this list that made sense to Dravniks in the mid-80s, so carbonic and kippery. I think you have to be at a specific, kippery refers to a Jewish, like a kosher kipper fish that you can get, yeah, maybe kippers, but it's like a specific, I think it's difficult to find anybody under the age of 50 who knows what kippery is. And then carbonic, I think, is a cleaning solution that was used in the 30s and 40s. So these are things that are, you have to refresh these lists every frequently to try to have them stay current. Just like language evolves quickly, these terms evolve quickly. And then a lot of them, because they were invented in the USA, have a very American vernacular. So root beer, gingerbread, cheddar cheese, dill pickle, and wintergreen are all words that describe olfactory percepts that are highly specific to Americans. So if you try to give someone a smell test to say, how much does this smell like root beer? They have no idea because they don't know what root beer is. And so then discrimination, which again is a semantic free, it's not polluted by culture, can you tell the difference between different things? And so we published the most hated paper in biology in 2014 dealing with this problem of discrimination. And I'm gonna try to make an attempt to figure out what I can salvage from the most hated paper in biology. But it was based on Noam Sobel's group's discovery that if you mix together multiple components, starting with, when you mix, this is from Lang's work, that when you mix a couple things together, you can still discriminate them. But once you exceed a certain number, whatever it is, three, four, they all start to blend into something where you cannot, you can't put a word onto it, which I think is great because then you can access that part of olfactory perception that's independent of experience. And what Noam did is to say like let's, we're gonna train these people, we're gonna make a new category called Lorax, we're gonna train them. These things that we've mixed together into a big pot that don't really smell like anything, we're gonna say this is Lorax. And he could train them quickly to put those things into the Lorax category, showing that you can make a new category. And I think that the thing that interested us here is that the more things you mix together, the more difficult they are to discriminate because you're just heaping all this stuff together. Nevertheless, I would say, I don't know if you agree that they are readily discriminable. So they all are this olfactory white, like white noise from auditory modeling to this olfactory white. So we became inspired. This is the Bush did paper, the most hated paper in biology, where we thought like let's make a lot of olfactory whites. Let's start with a palette of 128 monomolecular odors that the SOBA lab intensity matched so that people would not be learning. This one odor is super intense so that they could discriminate it. And we made mixtures of 10, 20 and 30 different molecules. The key thing is that you start with 30 molecules that are totally different and you progressively reduce the overlap. So the point where you have 29 molecules are identical and one is different. And we ask how good are humans at picking out this one molecule from this mixture, this kind of toxic mixture thing that smells like Lorax. And then doing some math, we figured out, given 128 individual components, how many mixtures could you make in theory? And this is where the paper starts to become the most hated paper because we assumed that each of those 10 to the 22nd mixtures would create a new olfactory white that would be discriminable from all the other ones. So that's the first assumption is that when you take mixture 10, billion and two that it doesn't smell like mixture 30. So then we do a simple triangle test. So triangle tests again are non-semantic. A person gets three bar-coded vials. In the case where I said, let's say two vials have 29, two vials, let's go with this one. So you would have two of the blue that has 29 bluish molecules and a single one that has 30 reddish molecules. And the person has to pick the reddish one. That's easy. Everybody in this room should be able to say, this one is an olfactory white that's different from this one. Then we have this progressive overlap. Here is the case where you have 29 that are identical. Pick the one that has the single blue or red. It's all blinded. And what we show here is that for mixtures of 10, 20 or 30 that humans are incredibly good at doing this. This is the second point where the most hated paper in the world was discussed. Because how do you decide if someone can discriminate it? So this is chance. So you have three vials. You have by chance the ability to pick one. And so you can see, I think everybody would agree that for these data sets like everybody gets it where it's pick the blue versus the red. As you start to have progressively more overlap, you do have people, this red dot that is able to discriminate it, right? So I wish I'd been able to bring these because they're fascinating things to smell. And to me, the fact that we find people who can reliably distinguish the single blue and the red odor and a mix of black, I think is astonishing, right? So these are our empirical data. And so then we went to a physicist to try to grapple with, given the assumption that starting with a palette of 128 monomolecular odors, we mixed together 10 to the 22, we didn't test 10 to the 22 mixtures, or we tested what we could test realistically. How many possible mixtures of that kind could exist? And given the empirical psychophysical data with these cutoffs that we selected, how many different odors could humans discriminate? Because this was a question, every Gilbert wrote this really engaging book about olfaction, repeating this, trying to get to the bottom of this fact that humans can detect 10,000 odorants, right? And it ends up being, it's just a mistake in the literature, someone sketched a model and then extrapolated to 8,000 and it was rounded to 10,000. So it's out there that humans can either detect or discriminate 10,000 odorants. So we were trying to figure out, what is the discriminative capacity of humans? So I have the donuts here, because we went to Marcelo's office and spent an hour saying like, how will we do this extrapolation? And so based on that one hour conversation, we published the world's most hated paper to extrapolate. Given this small sample of molecules that we tested, in principle, if this continued and the odorants didn't loop back and repeat, how many possible olfactory whites could be discriminated and we came up with a very large number that exceeded the normal discriminative capacity of people to detect differences between tones and colors. Now, here's an object lesson. Spend more than an hour over donuts with a physicist, kind of going over what it was. And so we did issue a correction to this paper, kind of showing our math, like the things that actually were underlying our model. So you can check it out. So I already talked about some assumptions. This idea that each of the 10 to 20 second mixtures was unique is an assumption. We assume that olfactory dimensionality is high. We believe that with the 400 buck and axle receptors and the ability to have combinatorial coating, that olfactory dimensionality is high. And then some nuances of the sphere packing model. What I take away from this as a psychophysicist is people are, in the things we tested, incredibly good at detecting kind of differences. So I think that can humans discriminate, can they discriminate 10,000, a million, a hundred billion or a trillion, obviously, to do the psychophysics is impossible. But I think it's a fruitful ground for additional discussion. So, and I see a couple posters here use these data to play with it. So anyway, how do we go from these molecules to a perception? We take three visual receptors and contact all visible colors. And of course, as already mentioned many times, the problem is difficult in olfaction because we have an unknown number of molecules. And they're all chemically different in some confusing way. Linda really opened up the field by identifying these odor receptors. There are many invertebrates and they're relieved and there's excellent evidence to show that they combinatorially encode different molecules. The thing we're still grappling with is how are the molecules interacting with these receptors? We don't have a crystal structure, although there's dozens and dozens of high resolution crystal structures of deporting coupled receptors that bind neurotransmitters. We don't have a structure. There's been some modeling experiments and here an empirical experiment by Tohar are to try to guess where the molecule is bound. But I'm really excited about somebody cracking this open eventually and showing really how are these molecules interacting with receptors? And I wanna reinforce that really the first person to do this, I think this is, Stuart's paper doesn't get enough important recognition. So Stuart was the first person in this field to link, I think, an odor receptor with the molecule. So that's I7. So we're working with incomplete information about what are the number of molecules out there and how are the receptors binding them. And we know that both early work from Linda Buck and then later work from Mainland and Matsunami showed that there's a lot of promiscuity and crosstalk. It isn't like the serotonin receptor which binds serotonin and really nothing else. The odor receptors are really built for diversity. Which is why I continue to be a big believer that humans have great discriminative capacity because there are so many receptors and there are so many commentarial possibilities and there's such an elaborate neural circuit for decoding them. And so what can we learn about the link, the perceptual link between a single receptor and the molecule that it binds and the percept. And so here inspired by Duran Lancet who sort of early after the identification of the odor receptor started genotyping humans to figure out given that you have 400 receptors is everybody identical or is there something like color blindness? So color blindness is due to genetic variation in visual receptors. Is there a genomic correlate of that? And he found quite immediately in a small sample that every human to a first approximation is unique. So nobody in the room here if we sequence your olfactory subgenome will have the same set of receptors which is an astonishing fact. And so we made use of that fact and about a century or probably many centuries of the observation that people perceive odor is differently. So these are these sex steroid derived odors where you take sex steroids and metabolize them and they come out of your sweat. And they have, depending on who you are either smell nasty. So sweaty urinus, sweet floral or odorless. And so this has been known for decades to try to estimate how many people out there are truly an osmic is a challenge. I think no one you have a lower probably more precise number but depending on who you ask between one and 10% of humans on earth can't smell these molecules. And so we figured out this is called specific anosmia where you have an otherwise normal sense of smell but you can't detect these odors. And so we used our very large database of human psychophysical studies to try to get to the bottom of it. What's the underlying genetic cause for people unable to smell and drostinone? And you can see immediately how much people disagree. So vanilla, the top rated descriptor for vanilla is vanilla. There is no top rated descriptor for and drostinone. It's all over the place from cleaning fluid to celery to musky to fragrance. And so teaming up with Hiromatsunami he had found a single odorant receptor in humans that was very selectively activated in cell based assays by and drostinone. And so we simply did a genetic association study to try to figure out our mutations in this receptor associated with this different perception of and drostinone, and drostinone, not 10 years old. But in red here is the response of a receptor in a person that has an altered perception of and drostinone, and drostinone. And you can see that this receptor doesn't respond to these molecules. And so this maps to these two polymorphisms in the receptor. If you have WM, you have an altered percept of the odor. In vitro, the receptor is functionally dead, so the people are walking around with a receptor that doesn't respond to and drostinone. And in Andreas' psychophysical data we find that people who have a broken copy of 74 find steroids less unpleasant, less sickening, more vanilla, and less intense. So this was, I love this paper. So this was the first demonstration that you can tie a specific anosmia to a particular receptor. So there is a heterozygous effect. So if you have one functional copy and one non-functional copy, you have this perturbation. This is for heterozygotes. So homozygotes are exceedingly rare in the population. We had a few homozygotes. I think we had three or four homozygotes, and they are much more affected. We have a few people who had a gain of function mutation, and they are much more sensitive. The receptor is much more sensitive, but the gain of function people, there were three subjects who were gain of function. So we can't, it's just like a story I tell you over coffee. So does this mean that there were any data? Oh, yes. Oh, yes, because of the monolithic important point. So because of monolithic expression, there's many cases where people are heterozygous at low side. There's many receptors that look like they're intact that might actually be functionally dead. As this one, I think that some people are walking around with a WM. So the diversity gets even more insane and interesting. All right, so that was, and many people have continued to work in this area. So Joel Mainland continues. Geron Lancet has done work on isoballaric acid. So trying to use perception to work backward. How do you go from the percept back to the receptor? Because we can't do optogenetics or make mutations in humans. So the next part is trying to grapple with his problem of, how can we have a quantitative science? How can we make some rules about how the sense of smell works to try to infer something about the stimulus to percept problem? And as already said, by many people, people study color vision how it's so easy. It's just like, so it took them a while to figure this out. But to go from wavelength to color perception is like a lookup table. You could give it to a two-year-old and they could do it. Tone perception in the auditory system, you could give to a dog and they could do it. We should be jealous of these sensory systems. In olfaction, unless you've memorized what these molecules are, anybody know what these are probably billmills? What's the molecule here? Nice. That's this 3-hectanol, so it's grass. But I have to, so this is beta-ionone and this is some kind of rose oxide. But even unless you're a trained perfumer or chemist and you really think about this, you have to have a lookup table of what does someone say it smells like is very challenging. And the bigger challenge here is this is like a computational chemistry paper to say how many molecules are there? How many molecules could you make on planet Earth? Given what we know about chemistry, not every bond can form without exploding. How many molecules could you make? And it's a really, really big number, $166 billion. Now I'm going to make a couple people mad here. I'm going to say, I think we're just exploring here. I think this is where, I don't know where the olfactory space is that we've been exploring, but let's say it could be scattered, but it's going to be a small part of this. So the way we've been trying to address this problem of how do you figure out what a molecule smells like is to get a different data set. So Draveniks produced this incredibly useful data set for the industry of molecules that are either maloders, like sweaty things that you want to get rid of if you're making a perfume or beautiful things that you could put into food. And so it was very constrained in the odor universe. So we wanted to have some weird outliers. Just like when you do a genetic tree, a phylogenetic tree, you need to have an outlier to try to root the tree. So this is what we did here. I love this Dr. Seuss book on beyond zebra. So we have the A to Z alphabet. I would argue that the Draveniks data set is A to Z. And what we really need to solve the problem is have letters like yuz, zaps, vru, and yek, where no one smell them and the words are nonsense, right? And so we published this a couple of years ago. So the idea would be to have a really big data set, get people to come in and smell a thousand molecules. So this is a lot of smelling. They did, they come in for 10 visits, 100 molecules per visit, two or three hours. To try to boost up the number of molecules that exist and put perverse things in there like water, L-cysteine, right? Things that you say like, I don't have an odor or if it does it's garbage, glycerol, just put some things out there. And so this is what these poor people did. We did pay them, but of course. No, no, go for it. Denounce me, I've got 25 minutes. Really important point, right? People can, people can, their performance can drop. So we have exactly to try to get some insights into if they're having problems is we introduce that regular intervals, odors that repeat. So we have a set of the beginning and a set at the end to try to figure out if, so you have odors one through six or repeated at the end, odors one through six. So yeah, yeah. I mean we are, we try to impose, in some cases we impose a timeout of two minutes, but the subjects won't do it. It just drives them crazy. So everything, all the psychophysics, I'm telling this group here is it has these built in flaws. There's not an olfactometer, the SNFs aren't controlled. In this case, we couldn't control the interpulse interval because they would just quit. So they just, the subjects won't do it, but so that's, it's a factor, right? And so the philosophy has always been like, you get these people in and you just have to average out over the group because it's the only way to do it. It's the only way to do it. We have not done that. Even though I know a lot of perfumers, the main thing, maybe Stuart can speak to this is like try to try to get them to, try to get them to spend 20 hours sitting in a lab smelling glycerol. It's just, it's just, so we haven't, it's just complicated. You know, in some sense for this study, I don't think perfumers will be any better because we're not giving them for the most part molecules that they care about or know. So a lot of perfumers are, have strong specific anosomias from musts. So it isn't like all perfumers are super smellers, they're just super experts. So, but we don't know. But I mean that they're trained to, through the training that they're training, I would love, I would love to know I've never been able to convince a real professional to come in and there's, you know, there's some. Which, yeah, but you need more than one of them, right, to have the power, IFF. I know, yeah, just for fun. But it is, it's like, it's punishing, it's 30 hours. Not at the same day, but this is not discrimination. Yeah. So I guess, and for humans, I'd do that kind of way. Yeah, yeah, yeah. What we'd be sure is that the thread, because they know it right away. If they, if they, I don't want to, I mean, because I remember seeing a paper that I thought that the R-score value for the repeats was actually quite low. And they're within the same subject. I remember that I'm looking at the data and it was 11.3 for the repeats. It varies depending on the spectrum. Yeah. Yeah. Some are really high, so we should go with that. And it's because, I mean, we've, the people, they couldn't spend the whole, they couldn't spend 30 hours continuously. So we bring them in and there's enormous inter-individual variability, both due to genetics and when they've had a cold and experience. There's also a lot of inter-individual variability that you see the same subjects, testing molecules. Yep, yep. And that's, and that's not unique to us. So every study that we've done, there's like, there's this variability. So yeah, it's not low for everything. Yeah, yep. This is the overall, we take all the data and smoosh it together. It's pulled down by descriptors that people don't know what to do with. But intensity and balance, the reproducibility is very intensity and pleasantness, the reproducibility is extremely high. Yep. This is why I hate when computational people take all the stuff and smoosh it together. I just think it's stupid. I don't know, Pablo's gossiping, but I don't know why, I don't know why we smooshed all the shit together and got a correlation, because I wouldn't do that, but. So anyway, we have them tell us how strong it is, how pleasant, and then we use 19 of the 146 Dravmix descriptors because most of the Dravmix descriptors are never used by normal people. Kipri Fresh is never used, Carbonic is never used, just doesn't ever come up. So these are ones that people actually use in our history of tenure history of doing work. So these were the people, diverse, male, female, different ethnicities were blessed to be in New York City. And I think the first discovery is that, again, familiarity has a huge effect on how people perceive odors. We ask them, do you know what the odor is or not? Like does it smell like something? The more familiar something is, the less likely they were to say, I don't know what the odor is. So highly familiar odors said mostly, I know what the odor is. So this identification task relies a lot on familiarity. Most of the things we smell are pleasant. So if you look at the correlation with familiarity, edible sweet bakery fruit, people are really fast to use those words because they're familiar with those things. Some of these other ones are used less frequently. So the cool thing is that when people say, I have no idea what the odor is, they tend to just put it into a bucket of chemical. I view this as Lorax, right? They just use chemical as this descriptor that has no tethering in the world. So they just put it into the chemical bucket. Again, showing that we don't have enough words to describe. If we train people to take the hundreds of molecules that were unfamiliar, train them to say this is Lorax, this is zip, this is lap, people could probably do it. So unfamiliar odors are neither pleasant nor unpleasant, again, because experience, if you're like, okay, fecal is gross, watermelon is good, so you already have peg to them valence. If they're unfamiliar, the unfamiliar ones are sort of in the mushy middle here. Then we let the subjects free to pick whatever the words they liked, untether them from any suggestions. What's cool here is that women were much more verbal, so women described many more odors than men. And then going from the, how experts, back to Linda's question, how experts describe these things versus how the people describe it. When people use words, they're very susceptible to marketing. So this camphor smells like this American product called Vix vapor rub, you put it on your chest. This carbon, this stereoisomeric carbon, Colgate toothpaste appeared a lot. Colgate, yeah. When people are just given a pen and they can write whatever, but they're pretty good, right? So this is this stinky thing and people come up with all sorts of disgusting things. Vomit, wet sand, food that went bad, bad shellfish, rotten food, feces. So people are pretty good at doing this. So what are some low level correlations we can make? This has been known for ages that the higher the vapor pressure in general, the greater the intensity because you get more molecules into the nose. This is also inversely correlated with molecular weight. So obviously, so the heavier, bigger molecules have lower intensity. And then an easy low-hanging fruit was the more sulfur atoms you have, the more likely you'll smell like garlic, fish and decay. And so at this point, Yeah, but most of these things are correlated. Highly, highly cross-correlated, which is the interesting, yeah. The interesting thing in the Dravonics is that again, you could have fruit, fruit, lemon, grapefruit, those are all highly correlated because lemon and grapefruit are a subset. So there's a lot of inter-correlation. So, and so this has been known. So to try to make an attempt to have something that took a more global view with the chemistry of these molecules and their percept, we use this program as already mentioned. So this, a dragon is an Italian program vented by Italian scientists to try to come up with thousands and thousands of different chemical, chemie-informatic descriptors, simple things like molecular weight, number of sulfur atoms, but also like van der Waals' interactions, dipole moments, all this detail pharmaceutical chemistry. And Andreas started to try to figure out some ways. So you could say like, okay, sweat correlates with this dragon descriptor, musky with carboxylic acids, which makes some sense, but we weren't making any headway just as psychophysicists. And so we teamed up with Pablo Meyer, who's sitting there in the third row to try to figure out, can we harness the power of machine learning of scientists who care nothing about alfaction in general just to see if they can solve this problem? And so this was the data set that the challenge groups were given. So just over a million psychophysical data points. I already discussed this. So about 49 people sniffing these thousand stimuli, and then just over 2 million dragon descriptors. So the idea is, you have this molecule and you have these dragon descriptors, and then you have the percept. Can you come up with some model that can predict, given this chemical, how sweaty, how intense, how pleasant is it going to be? And so just like in these challenges, you don't give people the whole data set. So Pablo gave them just data from 338 molecules out of the 476. He hid away 69 that he would use to test their work halfway through. And then importantly, 69 that no one had ever seen. So at the very end, you have a model that goes from chemical to percept. If the model works, you should be able to have it work on molecules that no one's ever seen. And so the two winners of the challenge were Rick Kerkin, who is an alfaction expert and Juan Fanguan, who wins every challenge. He's a really talented machine learning guru. And this is the task that people did. So you have this slider from can't smell it to very intense, pleasant, unpleasant, pleasant, and then these different sliders. So here's a molecule that's not very intense, pretty pleasant, it's sweet, fruit, a little bit bakery and warm, and all the other sliders are at zero. And here's where the noise comes in because these other ones, how cold is it, how acid is it, these descriptors aren't used very much because how do you decide something is acid? When they are used, they inject an enormous amount of noise into the data sets. That's a big interpretation. Okay, so here's my one quick slide. I'm gonna go over really quickly. What is machine learning? You have linear models, which is what Rick Kerkin used, and then this random forest where you go through all these iterations to try to describe what something is. In this example, does it have an ester, no, yes. Okay, has an ester, is it small or large, small? Does it have a nitrogen atom, no. And so you generate these profiles that at the end you can then match up to perception. That's all I'm gonna say. Don't ask me any questions. And so how do these people do? So the top-winning team here, this is Wang Fengguang, here is really good at predicting intensity. Intensity is easy because the more volatile, the smaller, so intensity is pretty easy to predict. Pleasantness is also well-predicted. And then the descriptor is batching, all the descriptors together is harder. It's just much harder, a harder task to ask people how acid, how bakery. This is looking at individuals. Individual data, as already discussed, are very noisy. And so if you batch all the data together for groups, we do much better. So intensity is very well-predicted, pleasantness. These are existing pleasantness models, and so some of them do quite well, ours do better. And then these descriptor models, to our knowledge, no one has attempted to make a model for how you get to bakery, sweaty, or acid. And so how well do we do? So we're able to predict garlic and fish, intensity and pleasantness to very high correlation. So these models that were developed by this group of 20 people that Pablo shepherded along in the process to test and refine the models, do a really good job of giving a molecule so that you can predict the percept. So which molecules are easy and hard to predict? So as you would imagine, glycerol, cysteine, and lysine are outliers. People would say like they don't have an odor, and so they're very difficult to predict because people don't know what to do with them. Some of these other ones have a high prediction correlation. And so what's the goal of this? So we did this because it was an open problem in affection, and we also did it because we thought it would be an interesting way for people in the industry to try to figure out what a molecule's going to smell like. And so what the industry people will do is they'll have this spider plot and we want something that's very intense, woody, kind of musky, a little bit garlic fruit. Like this is their target molecule to make a new perfume. They would typically, Stuart showed some of this musk structure activity work where they make a lot of molecules and hope for the best. Many of those don't smell like musk. So we were trying to come up with a model that you throw these 500 molecules at it. Can you come up with a model that will predict what a given molecule will smell like? And so here's a real world example where the model is cranking through. You want to figure out, this is the profile of butyric acid, which is really disgusting. The model will go through iterations and say, we think it's this molecule, no, this, no, no, no, no. In the seventh iteration, it hits it to butyric acid. It's not perfect, it has to go through seven iterations, but it was the seventh molecule out of 496 molecules. So I think that this vastly accelerates the search to constrain the search for molecule to percept. So this is what we did. So what's the applicability of it? So I think the question is, can you use the model for other molecules? This is something that a reviewer asked for, so we went through additional data sets and we're really good at predicting, I believe in open science, so I'm gonna out you. So we can predict intensity and pleasantness on molecules that didn't see, that weren't part of the model. And so we're optimistic that the model, either as an idea or as a model, will be useful to people. It's so hot in this room. So I'm gonna just try to do the last little bit, so the last, talked about how you go from genes to behavior. I talked about how you go from molecule to percept and the last little bit, how do we figure out if people can smell? What's a good smell test? So we devised just long story short, devised a new smell test. We think this is important because nobody in the real world cares about the sense of smell, but it's really important because people who are anosmic or have smell dysfunction, they have really problems with food, like they eat spoiled food, they either under eat or overeat. These are surveys that Andreas Keller put together worldwide. So he had 1,000 people fill in surveys who volunteered information about losing their sense of smell. There was one person who put paprika instead of cinnamon on their cereal because they're spices that look the same, but taste very different or smell very different. Here's some people, here's like an older woman who found a dead rodent in her house that her daughter discovered. Here's a mom who didn't know when to change her kid's diapers because you couldn't smell it. And then social isolation is a huge thing. People get very self-conscious about whether they have body odor because they can't smell it, so they'll either not go out or they'll shower obsessively. And I love this one. It's very difficult for me to make plans, feel desire, feel good and happy. I live in a permanent present. I have lost the sensations, linked to memories. I have no particular desire for the future. This, my friends, is why we exist, right? It's null faction, this is why we're important. So this is the current, this is the world's, there are two very important smell tests. This is one of them, Upset, devised by Dick Doty. You can see deeply rooted in semantics, right? So it's scratch and sniff. You take a pencil, you scratch, you sniff. Alternative force choice. Lemon, chocolate, root beer or black pepper. Give this to someone anywhere outside the U.S. And the answer is root beer, you're screwed, right? So, but it's a good test. It's quite good, but it's very difficult to translate it outside of the U.S. So when Doty went through and ran it on people in Europe, Asia, and South America, these are all of the descriptors that they just can't do. Like, what is fruit punch to someone in Germany? What is pumpkin pie to someone in Russia? They're just, they're very tied to American culture. So they've had to go through and adapt this. So they go to Taiwan and they say, all right, well let's put durian in. Let's put other words that make sense. So we've been looking for a way to test people where there's no words, no words. Where we're just purely testing sensitivity and discrimination. And so again, we use Noam's olfactory whites. We mix together a bunch of stuff and we dilute it a little bit and we ask people, same or different. So most people can solve this and we dilute it, dilute it, dilute it. Here where it's diluted like a million fold and we ask, can you pick out the odd vial? So here's an example of we tested someone, they couldn't do any of these and then at about this level, at about level eight between six and seven, they keep getting it right and so this is their threshold. So this is just conventional threshold testing. Threshold testing typically uses monomolecular odorants and I already told you that androstinone is an example of an odor selective insensitivity and so our test fixes that. Here's a discrimination test just like the trillion odors test where we make molecules, we make these mixtures more and more similar and people are good at detecting these. So the world standard of testing threshold is monomolecular odorants like phenylethyl alcohol, it's a rose-like odor. The frequency of specific anosmia to phenylethyl alcohol is unknown but non-zero as we show here. So if you test these normal subjects, you get this distribution of sensitivity of phenylethyl alcohol including these people down here. You then use two different versions of our smell test. I should do a disclaimer that we filed a patent. So this part's like a sales pitch for a smell test and so you find only this one subject actually has a problem with olfactory sensitivity, only the other people we would argue are normal. So that's what this test does and here's just another example of it and so then the problem of discrimination. So the upset is used as a way to look at discrimination. So we tested a bunch of people in Taiwan which are the people up here and a bunch of people in the US using the US version of the upset. And so you can see these are the wrong answers here so everybody has a problem with pine. The Taiwanese often mistake it for lilac. Here's an example of turpentine, the Americans and the Taiwanese mistake it for soap. So these are again normal subjects but you can see that overall there's much more shading here. So the Taiwanese do a terrible job on this test because we've taken the American test that has root beer, pumpkin pie, cheddar cheese, they have no idea what that stuff is. So they would be diagnosed as the entire country of Taiwan is terrible smellers which is kind of summarized here. So if you look at the upset score somewhere here is normal and so you can see that the distribution of Taiwanese is shifted to the left. So overall by this gold standard they are poorer smellers than Americans. We then use our discrimination test and it aligns quite well. If anything the Taiwanese are better than the Americans. So a test developed with Americans works really well on Americans, does not work on Taiwanese but it's a test that's semantic free works on Taiwanese and Americans and we hope will work with anybody. So Peter got us thinking about neurodegeneration so we're excited about using this test with Alzheimer's patients because we can deploy it around the world having them to remember what is root beer. If you really wanna figure out what the primary olfactory function is you don't wanna call in this declarative memory. It's another aspect of it and so that's that. So we tested almost 3,000 people over the last 12 years. Regrettably Andreas retires. This will probably be like one of my last talks on humans. We had a great collaboration with Pablo and his colleagues. Many productive years with Matsunami and then these are the folks who did the smell test and thank you.