#155 Alex Wiltschko: From Molecules to Memories & The AI Approach to Scent Detection

0:00There's molecules that are very important for some of our memories that need replacing, that need upgrading. And so what we're doing is we're applying our artificial intelligence platform to discover these molecules that not only smell great and are affordable to use so that they can be used in many products. If you just rank the senses, you say how much brain matter is dedicated to each of the five senses. Vision is number one. Vision is so important for us. Smell is number two. Hi, this episode is sponsored by Salonis, the global leader in process mining. AI has landed and enterprises are adapting, giving customers slick experiences and the technology to deliver.

0:41The road feels long, but you're closer than you think. You see, your business processes run through many systems, creating data at every step. Salonis reconstructs this data to generate process intelligence, a common business language. With process intelligence, AI knows how your business flows across every department, every system, and every process. With AI solutions powered by Solonis, enterprises get faster, more accurate insights, a new level of automation, and a step change in productivity, performance, and customer satisfaction. Process intelligence is the missing piece in the AI-enabled tech stack.

1:26Search Celonis, C-E-L-O-N-I-S, to find out more. Hi, I'm Craig Smith, and this is Eye on AI. In this episode, we explore the fascinating world of fragrance with Alex Witsko, CEO of Osmo. Alex's lifelong passion for neuroscience, beginning at the University of Michigan, through to his PhD at Harvard, and his journey into the entrepreneurial world with ventures, Alex is pioneering the intersection of neuroscience and technology, particularly focused on our olfactory experiences. From a childhood fascination with perfumes to advancing AI in scent detection and creation, Alex's story is a deep dive into the rich complexity of how we perceive the world around us.

2:30Join us as we unravel the mysteries of smell and its digital future in this. I hope you find the conversation as fascinating as I did. My name is Alex Wolchko. I'm the CEO of Osmo. in school. I studied neuroscience at the University of Michigan. And then I studied neuroscience again at Harvard, or I did my PhD. And then I ended up doing some startups. I went to Twitter and then Google Brain, where I worked on neuroscience again. And then I've spun out Osmo as a new company. We're a little bit more than a year old. And we do neuroscience as well. So it's been olfactory neuroscience all the way through.

3:16This is my one very, very deep obsession. And it's been a straight line through basically the entire way. That's great. And when did you get focused on olfactory senses? I mean, the olfactory sense. Yeah, the olfactory sense for me. The obsession started when I was very young. I think different people have different senses that they're attuned to. So some people just have that window of their house, the sight window that lets in information to their insides really wide open. I don't, not for vision, not really for hearing, but smell has always been this very vivid slice of reality for me. So just from a very young age, I was always drawn to it.

4:05I was attuned to it. not that I knew how to talk about it or describe it or create them at all. It just was very interesting to me. And then I started collecting perfumes when I was in high school, roughly. And I grew up in a very small town in Texas, and I was already a computer nerd. And then to add perfume collecting on top of that, not necessarily a recipe for popularity, but these have been my obsessions. And I've tried to combine those over my entire life. And I feel very fortunate to work at that intersection of computers and of all faction. And I think there's just there's so much depth and richness here at this intersection.

4:43And that's that's what I've dedicated my life to. And that's what Osmo is all about. And when you were saying whole factory, what what are you referring to? Just I'm just punning with you. I mean, the space of smells is filled with puns. And so throughout our conversation this hour, and I'm sure we'll be able to drop some more. Okay. you know I know a little I've talked to one other guy about the smelliverse as he called it in AI a guy named Alexei Kulikov do you know him? I do I'm a fan of his work oh okay yeah we talked about he was classifying volatile molecules according to their smell or predicting their smell using supervised learning.

5:38So tell us one other thing I have to ask if you're particularly attuned. I spent many years in France, and because I was a correspondent for a big newspaper, paper all the wine people sought me out and I had one guy in particular who would take me out to these uh these meals and do the you know guess where the wine came from Trek he and his colleagues and it was remarkable that they uh could get within a a few kilometers of where the wine came from and he explained in his view that you know the the mind of a sommelier is like a card catalog and and you just begin narrowing it down as you focus on different elements of the essentially the odor right as much as the taste and how does that maybe you can talk about the neuroscience.

6:46Well, first I wanted to ask you whether you're a wine connoisseur, but then how does that work in the cerebral cortex? It's a great question. So I enjoy wine. I would absolutely not pass the test that you've just described. So if I were to guess, I would be within maybe thousands of kilometers, basically throwing darts. I don't have any special skill when it comes to wine because I think it's exactly what getting really

7:17being talented at associating words or properties with smells is a matter of practice. I really don't think it's a matter of innate talent. Certainly your nose has to be open. You can't have it closed via allergies or something anatomical that prevents you from smelling things, but I've worked with perfumers and master perfumers over the years. And what isn't the case is that they can smell vanishingly small amounts of stuff, right? They don't have a lower detection threshold on average than the average person, but they have phenomenal abilities to associate smells with words. And they have a phenomenal catalog of experiences with what the world smells like.

8:00And so their ability to use language and to reason about the contents of something that they're smelling is truly extraordinary to watch. And I've never spent time with a master sommelier, but I imagine that's the experience that you're getting as well. So if we were to break it down, like what's happening there, when you're smelling wine, even when you're tasting wine, because when you're tasting it, the liquid rolls off your tongue, you're experiencing the five or so different tastes that exist but then when you get to flavor flavor is a collaboration of the tongue and the nose and 90 of flavor is actually smell what happens is what's called the chimney effect whereas liquid rolls down your tongue down your throat vapors are released that actually go backwards so instead of smelling in through your nostrils their smells enter backwards from your throat into your your nasal passageway and you're smelling so So the molecules that are in wine are there because of an entire history of a plant being grown, of the fruits being picked, of them being processed and fermented and placed in a bottle and aged.

9:06And there are many, many thousands of different types of molecules in there. And the exact identities of those molecules and their ratios tells the story of how that sip of wine arrived into your mouth. And none of that's there by accident. None of it. It's all an indication of the history of that sip, right? From its very birth, from the sunlight and the nutrients in the soil, all the way through how it was processed and stored to your enjoyment of it. And so to me, it is not a surprise that if you were to look at the identities of those molecules, so forget the nose, if you were just to actually know like what the molecular structures were, that you could probably put together a case of, oh, this is from Southern Spain or from Eastern France or whatever it might be.

9:54And it's this grape and it was processed in this way because that has to show up in the data, right? And the things that you do to living things and how you extract them and work with them, it's going to have an effect on the chemistry by definition. Now, the amazing part, the miracle is that your nose can do that to some extent and that your nose can actually do that pretty well. And once you start to gain experience of what things smell like, which is this incredibly powerful compression of one of the richest sources of information on the planet, which is what are the molecules in the air around us or in the things that we eat and drink.

10:30It's incredible that that comes through and that we can even talk about it. So the neuroscience there in terms of how we actually process that and how that rises to consciousness, still a work in progress. We don't know. But the fact that we can do it, I just find to be miraculous and wonderful. Yeah. Yeah. And there's a particular part of the brain that's associated with, that's activated by smells or does that move around? No, just like there's a visual cortex in the back of your head and there's an auditory cortex kind of on the sides above your ears, several other places that process the senses.

11:07There's an area of the brain called the piriform cortex. And the piriform cortex is kind of underneath the jaw here. And that's where smell gets processed. And actually, if you go to a wonderful project called the Allen Brain Atlas, and the Allen Brain Atlas is meant to map the different parts of the brain responsible for different things. If you look at, if you just rank the senses, you say, how much brain matter is dedicated to each of the five senses? Vision is number one. Vision is so important for us, smell is number two. When I looked that up, I was surprised, actually. I would think that hearing or something would come to be number two, but according to the Allen Brain Analyst, smell is right behind vision, which I think is phenomenal.

11:49Wow. Yeah. Yeah, that's fascinating. I have to say, particularly for somebody like me, who unfortunately has lost his sense of smell. Oh, no. Yeah. Yeah. Was it from COVID or a trauma? it happened around covid but wow it happened before i got coven so maybe i had some some other some strain of covid that i wasn't aware but anyway yeah i'm sorry to hear that i'm hoping it comes back um you know there's there's there's some tricks for bringing it back um there's been no foolproof way that I've seen that like definitely, uh, regenerates olfactory receptors. But what I've seen relatively consistently, at least help a little bit, but no guarantees is to start smell training.

12:43So if you pick household items or buy essential oils, just get a set of like 10 to 20 things that are consistently the same, consistently available. So like eucalyptus oil or your detergent or your dishwasher cleaner, whatever has a consistent smell that's been added to it. And you just try to train yourself and maybe blind yourself. So put it in canisters that where you can turn it around and not see the answer, something like that. And then try to quiz yourself. That has helped people that I don't know if they've completely lost their sense of smell, but folks that have kind of a jumbled sense of smell after COVID, that training process has accelerated their recovery.

13:22Again, no guarantee that that's going to work, but it's the only thing that I've heard that's helped. So you might try that. I certainly will. I certainly will. Let me know how it goes. Yeah, yeah. So tell us about your work with AI and smell. And I'm particularly interested in the architecture, whatever AI model you're using. Is this straightforward supervised learning or something more complicated? Great. Let's dig in. So I think the first point to touch on is why smell an AI? Why use these techniques at all, given that we understand vision perfectly well, well before anything that resembled artificial intelligence came on the scene, actually well before we had computers?

14:15And the answer is that the other senses have had what smell has not for a long time, and that is a map. So color has had RGB, three numbers that can help you map and do arithmetic on all the colors that we can see, even more than we can see, actually. And sound has had a map, low to high frequency. We call it the Fourier basis, but that's the map of sound, right? And with it, you can describe any sound and then also describe any music. So these maps have enabled the entire ecosystem of the digitization of these senses. So without a map, you can't have a camera and you can't have JPEG and Photoshop and printers and displays.

15:02You can't have microphones and sound editors and speakers. You need a structured understanding of the sense in order to do anything else. Smell hasn't had this map. And it's not crazy that it hasn't, because smell is different in some important ways. The first is if you just look at how complex it is right at the beginning of the sensing process, there are three different channels of color, roughly corresponding to red, green, blue, in your eye, and then one kind of black and white channel. And in your ear, there's really just one sensor, which is kind of a loudness sensor, and it's arrayed in your ear such that you get this map of low to high frequency.

15:46But smell has 350 distinct channels of information, right? And that's 100 times more complex in terms of the dimensionality of the input than vision. And so how are we going to make a 350 dimensional map with our map making tools from 100 years ago? The answer is no, like we were never going to succeed at doing that. You can put the maps of color and the maps of sound on a flat piece of paper or a globe, but you can't do anything close to that with smell. It's just too complex. Let me just jump in. When you say 350 channels, are those different kinds of uh, neurons, neural receptors, neurons, or, or are those regions or how, what does excuse me, what distinguishes, uh, those 350 from each other?

16:46It's a great question. So the surface in your nose that senses smell is called the olfactory epithelium, which is like the retina, but for smell. And in us, it's about the size of a postage stamp. And it's covered with olfactory sensory neurons that have sensing surfaces. And each olfactory sensory neuron only expresses one type of olfactory receptor. And so it's the same neuron, but with different receptors in it and there's 350 different types of receptors. And each of those receptors is sensitive to different parts of chemical space. Now, the thing that's been mysterious about smell is we don't really know how to talk about what part of chemical space each receptor is sensitive to.

17:37That's been really at the root of the mystery is like, what's the relationship between what a molecule structure is and what it's going to smell like when viewed through this sensory system? We So that's where the 350 number comes from. It's these 350 different types of receptors that are expressed in neurons in the olfactory epithelium. So the issue is finding this map, right? And the map making tools of however many years ago weren't going to cut it. So that's where AI comes in. This is a pattern recognition problem. This is a data understanding problem. And this is exactly what the tools of artificial intelligence are.

18:19built to do is to extract structure from very complicated, high dimensional data. And so that's what we have been doing. And that's what we've done is focus on this one problem to start, which is if you can draw the structure of a molecule, can you predict what it's going to smell like? And this is something that lots of people have worked on before, including Alexei Kulikov. I love the work that he's done. The threshold that we crossed in the study we recently published in science is we got really good at that prediction problem. In fact, we got so good that you'd prefer the predictions of our neural network that are just working with the digital structure of a molecule than a person that has access to the physical molecule smelling it and talking about it.

18:57So our predictions have passed a kind of an odor Turing test, which is exciting because that means we can find totally new fragrance molecules from the enormity of chemical space, which is, we can get into that topic of why that's interesting and important for society, but that's the threshold that we crossed with the science that we've done. wow that's fascinating so and so you're you're you're you're taking the the physical shape of the molecule or or the the uh molecular composition the the different atoms and yeah you've got it exactly what we're taking as input is the molecule phrased as a graph So if you imagine each atom being a node and each bond being an edge, just like a social network graph, right?

19:45Where instead of friends being nodes and friendships being edges, it's atoms as nodes and bonds as edges. And these are small graphs, right? Because the molecules that we can smell typically have five, 10, 20 atoms in them. Whereas social network graphs are big. These are small graphs. So we take the graph structure and we provide that to a neural network as input. And you asked about the structure of what we use. we use supervised machine learning. And so there's inputs, which are the molecular structures, and there's outputs, which is a perfumer description of what that molecule smells like.

20:18Those are the pairs that form training data for this particular neural network. And we then can run that neural network on new structures for which we've never smelled those structures. And the type of architecture is called a graph neural network. And this was actually something that didn't really exist in a way that was easily usable, even for practitioners in the field. And when we started all this work, maybe five, six years ago, graph neural networks were just starting to show promise as an architecture that worked. And of course, things have evolved since then quite substantially. But this notion of using graphs as input to machine learning was actually kind of an interesting and new thing, because most of the inputs to machine learning usually are rectangles.

21:02It's images, grids, or it's text, it's strings of characters or tokens. And so having something that's a graph structure was actually a bit of a challenge to provide the machine learning, but that's been largely solved. And we're grateful for the work that we're building on top of and the um so your system the the pairs that you're trained on uh are you using is there a dictionary of smell for perfume perfume a's i guess they're called perfumers perfumers yeah or they're sometimes referred to as noses noses yeah uh is there a dictionary that they use so that there's a consistent vocabulary otherwise you're into the subjective realm of semantics yeah uh so the answer is yes but there's no single taxonomy so you mentioned wine before.

22:05There are odor wheels for wine. I don't know if there's one odor wheel for wine. There's a lot of overlap between different ones. So it's the same in perfume where there are a lot of different taxonomies that cover most of what you're going to find in a fine fragrance or even in like hand soap or detergent, but it won't cover everything. And generally different odor wheels disagree. What we did is we had to standardize this for machine learning. So we created, we actually have a few taxonomies. The one that we use the most has about 130, 140 different words in it. And this is something that we created after a very careful study of all the words that we found used to describe the ingredients in the fragrance industry.

22:46And it turns out that 140-ish words really capture the bulk of what you find in the industry. Now, we've since used a smaller taxonomy that we use for higher throughput labeling. So in the in the science paper, we used a 55 word lexicon because that was what was economical to train people on very quickly. Now we have since worked on a much larger, more comprehensive taxonomy that's meant to actually be the end all be all. It'll be evolving, but we have very large comprehensive taxonomies as well, but you're completely right. You need to standardize. Otherwise, you know, you don't have a target to shoot for.

23:23Now, what turns out to be the case is you and I, if we sit there with our noses and after you do your smell training and your smell recovers, hopefully you would be able to do this in a few hours, we can learn to associate smells and words in a few hours. So when a perfumer says fresh cut grass, they mean something quite precise. And there's actually a reference for that. There's a reference molecule called CYS3 hexanol. And if you smell something that smells like CYS3 hexanol, it means you're smelling fresh cut grass. and that's the term that they use. So it's not subjective at all in the sense that there are references.

23:58People can be trained to label in the same way that you and I can take a Crayola crayon box and know, okay, this is salmon, this is orange, this is mauve, this is black, this is white, whatever. And we can learn to get pretty good at that. And we might disagree about shades, but we can agree. Modulo color blindness and things like that, we can agree. and um so so take me through then to to once you've trained the the uh graph network then uh what do you do just are you presenting because the world of molecules of possible molecules is endless. So yeah, how then do you start applying this to build your...

24:50Yeah, you're right about that, which is the space of possible molecules is effectively endless. It's huge, the number of possible molecules that you can get to. What we did to first understand whether or not what we were doing was working was we needed to do a double blind trial. So we needed to take our GNN, our graph neural network, and predict the smells of molecules that nobody's ever smelled before. Some had never even been made before. And we needed to do this in a double blind setting so that we weren't fooling ourselves, right? We wanted to make sure that this was actually working and would generalize to new structures, to combinations of smells, excuse me, to combinations of smells that hadn't been seen before.

25:35And so that's what we did. And this was a very close collaboration between myself and Then my group at Google Brain and Joel Mainland and his laboratory at the at the Monell Center. And what we did is we digitally sniffed hundreds of thousands of molecules. So we downloaded a catalog of molecules that we could purchase so we could have synthesized. And we got rid of all the molecules, first of all, that were unsafe. That was an easy filter to run. And then we got rid of all the molecules that were too expensive or that would never have a smell. they're too big or otherwise some problem. And then we predicted the smell on what was left.

26:14And that was hundreds of thousands of molecules. And then what we did was we picked a set of 400 molecules that structurally looked nothing like what we'd trained the model on. So very, very different structurally. We selected these molecules to be structurally different from each other so that there was a diverse cloud of molecules. And then we also predicted their smells to be very different from each other. And we kept the identities of these molecules secret. And we actually, we bought them or we had them synthesized. So we now have these molecules physically, people could smell them. And then we presented these molecules to a panel that we trained, double blind.

26:53And we assessed, okay, the best you can do is the average of a panel. For all things in supervised machine learning, adjudication of a group of people or an average of a group of people is always the best, right? And then any one individual has some distance to that panel mean. And the question was, where are our model's predictions in the pack of the panelists? Is it worse than the worst? Is it indistinguishable from folks or is it somehow superhuman? And it turns out for most of the molecules, most of the time, our model's predictions are closer to the panel mean than the median panelist. What that means is you'd prefer to ask our neural network what something smells like than any one individual panelist most of the time, which nothing like that had ever happened before.

27:36That's superhuman-sent prediction ability, which we were just astounded by that this works. And just going back to the 350, so that's built into the neural network, those 350 categories or within your odor map and you're placing a particular molecule somewhere in those, some combination of those 350 or, yeah, related back to the 350? Great, great question. So this is a really interesting topic and actually is the location of a leap of faith that we took. So the way I was trained to approach this problem was first, you have to understand the receptor biology. And the first thing you should do is figure out, okay, here's a molecule, which receptors does it activate?

28:36It turns out that information is not available to us for a bunch of reasons that we don't have to get into. But we just, as a field of olfactory science and of molecular biology, we can't actually access that information right now. Maybe we can in five years. I hope we can. So we didn't wait for that information. What we did is we trained a neural network that just took as input the structure and took as output our spoken or written rating of what these molecules smell like. And in between those is the entire human brain, the entire thing, right? So we're not breaking down the problem at all. That was the leap of faith.

29:11Now, it turns out that all neural networks in solving a prediction problem build a map implicitly. And this is called an embedding. And I'm sure your listeners are familiar with the concept. So there is an embedding in our neural networks. And the dimensionality of that embedding is about 250 dimensions. Now, that's not a special number in any deep biological sense. That's what the engineering tells us it has to be. If it's much less than that, performance degrades. if it's much more than that performance kind of plateaus so 250 ish now that's suspiciously close to that 300 number that i mentioned to you and these this kind of like few hundred number keeps showing up in many different kinds of analyses that we do so i think there's something deep there i can't prove anything i can't prove that there's a relationship between what our neural network kind of what the dimensionality of our embedding in our neural network is and the true biology here But our odor map is about 250 dimensional.

30:06And that's what we need to solve the problem. Yeah. And then just to continue on the map, do you do then actually a 2D map or a 3D map in some rendering? Yeah. Or are you using the term map sort of metaphorically? Well, it's a map, but in its true form, it only lives in software and doesn't live in two dimensions. But we make visualizations of it in two and three dimensions. And it's a two-dimensional shadow of a much higher dimensional object. So you're going to miss information. But even when you do that, you see incredible structure. So one thing that popped out to us very early on, kind of shocked me actually, was how structured this map is, even when you view it in two dimensions.