 Yeah, I know, we will smell some. I hope. Good morning, everyone. Let's get started with the first of today's non-solver metric approachable factor, please. We had this PhD at Stanford, and he went to California where he went out with Bison, where he is. Thank you very much. First, I mean, this is sort of an automatic, but from the bottom of my heart, I want to thank the organizers and the Institute for both A, just organizing this, and B, making the mistake of inviting me. It's a real honor and a real pleasure, and it's been a lot of fun so far, and I hope it will continue to be fun. So thank you. I will say a word or two just to tell those in the lab who are not familiar with our work a few words about our other aspects of our life in lab, because our lab does a few sort of disparate things before I'll reach the topic that I'll talk about today in length. So just to tell you a bit about life in our lab. As you can see, this is a healthy user. There's a lot of different things, and a lot of it revolves around the fact that we love building stuff. So we're a group of quite a few engineers in lab. We like building things. The thing you're seeing here is actually a sniff-controlled wheelchair becoming commercially available this year, we hope, not the wheelchair, but the sniff-controlled device. We've developed a method to use sniffing to control any other device. You can control your computer with this. It works very effectively for significantly disabled individuals, and it's just a lot of fun building. We have a whole line coming out of sniff-controlled devices, and here you see our sniff-controlled wheelchair. And that's about all I'll be saying about our device building world that really occupies a big part of our life. Beyond that, about half of the lab is involved in the world of social chemo signaling. That is, the world of interaction between individuals using their sense of smell. Of course, to everybody in the room, it's completely intuitive that there's being information exchanged here between these two mammals. I have news for you. Humans are large mammals, and we all exchange information using similar methods, and so it's not to go un-updated on my slides. I updated my slides. And we invest a lot of effort. The Americans in the crowd are going crazy. I don't know why. The Europeans have Dominic Strauss-Kahn. I mean, he's not as dumb as this guy, but he's way more crazy. At least in the world of chemo signaling, I think he's more crazy, OK? And I'll just tell you, in one word, which is really doing injustice, but in one word of finding that came out from our work, that is a bit of our claim to fame in this subworld of social chemo signaling. In rodents, of course, the media that's most studied is a carrier of social chemo signals as urine. Humans do not typically walk around sniffing each other's urine, and therefore, other media that the body exhumed are typically studied as carriers of social chemo signaling, most typically sweat. So there's tons of work on potential social chemo signals carried in human sweat. We, however, hypothesize that there may be other carriers of signals. And the one that interests us initially was this liquid here, otherwise known as emotional tears. A liquid that the body exerts specifically in highly emotional settings where communication, and often nonverbal communication, may be paramount and important. So we hypothesize that this liquid may contain a chemo signal. In fact, we hypothesize that that might be the purpose, the functional purpose of crying. And through a long set of studies, which I will not detail here in any way, we found that indeed, there is some odorless airborne component in this liquid that we have yet to identify that drives distinct patterns of activity in the limbic system. For example, primarily in the hypothalamus, as you see here, and these changes in brain activity have a host of downstream or upstream, depends how you want to look at it, effects. Where the most pronounced was a marked reduction in testosterone in men. And now we've replicated the testosterone effect in women exposed to the airborne molecule, whatever it may be in emotional tears. So here you see, this is all the data on the unit slope line. But for simplicity, now you can look at the bar graph of the same data. And this is a free testosterone exposed before and after sniffing emotional tears. From conversation with doctors, as far as we understand, this happens to be the fastest and most effective way to lower testosterone that's out there. It lowers testosterone by about 15% in about 40 minutes. Amazingly, this has been replicated independently by others. So here you have a study by a group who I don't even know out of South Korea, who found a nearly identical effect in terms of its extent in reductions into testosterone after sniffing emotional tears. And here there's something that I like to sort of, I enjoy, I must admit, typically in studying humans, we often, especially in chemo signaling, we often go out to see if we can see effects in humans that we know that have existed for a long time in rodents and we see if they exist in humans as well. This is a rare case where this has taken the opposite path. And in fact, after we did this, a very, very similar effect was discovered in mice where you expose mice to what they're referring to as emotional tears from mice pups or otherwise known as lacrimal secretions. You have a marked reduction in sexual aggression by other male adult male mice exposed to this liquid. In this case, they also discovered the active compound. So a similar effect exists across mammalian species. However, lest we feel too good about ourselves that we discovered something and then it was found in mammals, of course that's because nobody reads papers anymore. And had we or they looked backwards, we would have all learned that way before us the exact effect was shown in the blind mole rack where in the blind mole rack also if you take, so the blind mole rack emits tears or lacrimal duct secretions. And if you wash this onto another mole rack then there's less aggression towards that mole rack. So the way we like to think now about these emotional tears is like a chemical blanket protecting you from aggression of others. And of course this raises endless thoughts on what does this do? Why do babies cry and children cry and what does that do to your interaction with them? And we're of course really trying hard to find the active components. We really hope that it's a component because if it's components we're hosed and we're doing our best to try and push on this front. I didn't look yet for that peptide. And that raises an issue right because the peptide is non-volatile. Although we're desperately trying to find a way where somehow the body might make peptides volatile. But the peptide would be non-volatile and the effects we observed were by a volatile. At least I know that. Yeah, okay I really don't wanna park on this for time but I'll be happy if we have time left to come back because this was just to tell people a bit about our other sort of existence. And sort of pulling us towards the topic of today's talk we of course study olfactory perception and I'll use this opportunity for a quick plug Coby Snitz from our lab is starting in a cross-cultural comparison where we're gonna address the question of whether olfactory perception is indeed different across cultures where our work hypothesis is that it's not. We can at the end talk about why we think that in contrast to what is typically thought about this question. This is the current geographical coverage of labs involved in the project and we'd really be happy to have more. So if you wanna join especially if you somehow exist in some of these barren areas then we'd be very happy to have you join the project. Okay, now what is the question that drives the other half of our lab the half that brings in here and what I'll talk about from here on. So you've in a way seen this quote already throughout this meeting and by none other than Alexander Graham Bell and I will actually take the trouble of actually reading through it. So 103 years ago Alexander Graham Bell said, can you measure the difference between one kind of a smell and another? It's very obvious that you have very many different kinds of smells all the way from the order of violets and roses up to Asa Fetida but until you can measure their likenesses and differences you can have no signs of odor. So what is Bell telling us here? So he went on to study inferior sensory systems but what is Bell saying here or to recapture in other words and this has been shown in various graphics throughout this meeting, this notion whereby the sense of smell is a set of transforms from some sort of physical world of molecules on through some sort of world of neural processing and onto some sort of world of perception obviously heightened link to this world of neural processing but we can think of it as this separate rule and to date there is no scientist or perfumer for that matter who can consistently look at a novel molecular structure and predict its odor or smell something and predict its molecular perception. Of course we say there is no of course they're starting to be more and more cracks to this rule of no and the most notable current crack is the one set by Leslie and her group where they now can in fact predict various perceptual aspects of a molecule by its structure alone but they're still and I assume you would agree with me with the statement that there's still full sort of a total ability to look at any molecule and say oh that's gonna smell like X or like Y and we know the study of the sense of smell has taken an odd path in this way because what's governed our efforts to try and when I say our I mean the community what's governed our efforts to try and solve this question is the very profound understanding we first obtained about this level right about the neural mechanisms or underpinnings of the sensory system indeed a lot thanks to the work of Linda and so on and so forth we have a very deep understanding of what happens here and oddly this has guided our effort to try and understand what's happening all together I say oddly because this is very different from the history of other sensory systems right you look at vision there was first a very deep understanding of profound understanding of the psychophysics of vision take the example of color vision right we had home hoes who gave us the trichromatic theory and this profound understanding of how you can combine and this is somewhat giving generalize but you can mix any three perceivable you can mix three perceivable lights to obtain any other perceivable color of light and lo and behold neurobiologists later found three options right where we now know those are fourth right but they weren't looking for four they were looking for three because they knew that there should be three because of the way the behavior works now here we did not have first the understanding of behavior and we know that we have roughly a thousand receptor subtypes so everybody's saying wow this has to be really high dimensional complex because we have a thousand receptor subtypes but that's an odd way to approach in my view a sensory system to try and understand the sense by its neurobiology rather than first understanding the sense and then see how the neurobiology supports this right because it could be those thousand receptors could be 990 of them are doing something totally different and the way I may end up in trying to argue that so we said okay let's first look at perception let's first try and make some put some order into perception and under the notion that what we're aiming to do is to measure perception somehow measure structure somehow and measure neural activity in some way and then link these measures to generate a predictive framework which is in a way what we're all partially after so how do you measure perception so we like many others relied extensively initially on the work of others done before us and most specifically Andrew Dravenickson colleagues who published a very helpful data set first in a paper in science in 1983 and then in the Atlas Outer Character Profiles in 85 and to those in the room who are unfamiliar with this even after the talks we've had in brief what they did is they sent 138 monomolecular species it says you're 144 because some of them were repeated at different intensities so they sent 138 monomolecular species by mail to about 150 participants across the United States these were all smell professionals of one kind or another many scientists and had each one of them rate each one of the 138 monomolecules along 146 verbal descriptors so these are some examples so you would get if you were participant in this you would get a molecule on the mail and you would rate it at how lemony it is and how burnt milk it is and how wet dog it is and so on and so forth on the scale from zero to five so on the face of it you have here a 146 dimensional space of odor perception because there's these 146 dimensions and you potentially have this huge space now I would of course like to well one would assume first of all that the space the actual dimensionality of the space is much lower than that apparent dimensionality and if only because of auto correlations because how lemony and how citrusy something smell will probably be one in the same those are not really different entities and so they'll definitely auto correlations here and the underlying dimensionality is probably much lower than 146 and there are various statistical methods to exhume actual from apparent dimensionality from large data sets where the most pens free simplistic approach would be a method called PCA or principal components analysis now in the sake of about three minutes or four minutes of our time how many people in the room have a profound very good understanding of PCA please lift your hand okay a certain amount and how many people would be very happy if I spend three four minutes telling them what PCA is a good amount great so here comes our one slide course of PCA remember I wanted to get the actual underlying dimensionality of this data set now first I'll say in words what PCA does this may be slightly obscure then I'll give you an analogy which should make it reasonably clear so PCA takes your entire data cloud right in this case again 150 experts rating 138 orders on 146 descriptors so lots of lots of data points right and basically rotates this cloud until it finds the best single line that you can pull through the cloud that is the one line that explains the majority the most of the variance that you can explain by one line through this data and that by definition is PC1 or principal component number one okay then you go on rotating the cloud again until you find the line that best explains the remaining variance in the data this is called PC2 or principal component number two which because of the underlying assumptions of PCA is by definition orthogonal to PC1 and you can go on doing this till you find a set of axes that explain the variance in your data theoretically up to 146 in this data set if they were all totally independent which is highly unlikely to give you a better sense of this forget measuring odors and imagine our physically measuring people in the room and forget 146 verbal descriptors let's imagine eight physical descriptors of people in the room okay so you have height, weight, shoe number, thinness which may appear like an odd term but soon you'll see why I use it thumb diameter, head circumference, nose length and nostril diameter because we study smell now once again I would like to show you all the people in the room on an eight dimensional graph, I cannot so I will just choose two of the dimensions let's say height and weight and a typical room would look something like this where this room may have some people who are fairly happy people he's not here so can I afford to do this and fairly unhappy people but as a rule a room would look like this and it's fairly easy to sort of see what the line that explains the majority of the variance in the data here this would be that line now pulling the line through the data then allows you to do various things that can be meaningful when you try to later understand your data one thing you can do is you can ask to what extent each one of these descriptors weighted on the line or influence the line so you can end up with something like this where these are not typos these bunches up here because indeed if you're high up on this line if you have a high PC1 value in this case you'll have a lot of weight and a lot of height so they'll be positively weighted on your PC1 but no, thinness is not unimportant if you have a lot of thinness you'll be negatively weighted on the line so thinness is negatively weighted on PC1 or you can think of PC1 here as some sort of line that goes from having a lot of thinness to having a lot of height and weight whereas all these others are bunched up in the middle they're not really meaningless meaningful I'm sorry for this line now so that's one nice thing that this can do for you is see how these things influence this and often also you can give this especially if this happens to explain a lot of variance you can give it a meaningful name in the world that we can think of so for example what would be something something that has a lot of weight and a lot of height has a lot of what? In one word, mass would be size perfect yeah so size right so you can give a meaningful name in this case to PC1 if you see something that goes from here to here that's size, it doesn't always work but often it works at the intuitive level so now forget all these measures again forget people in the room and forget these measures think of 150 odors, 146 verbal descriptors this is the method we applied and this is the initial result we get and what you see here so these are the first 10 principle components in blue you see the percent of the variance in that data explained by each one of the components and in red you see the cumulative variance by a number of PCs and what you see here right off the bat is that indeed this is likely not 146 dimensional space because PC1 of odor perception alone explains slightly less than 30% of variance remember that number slightly less because I'm going to show you a strikingly seemingly similar graph later on but you'll know it's different because it'll be slightly more but this is slightly less than 30% of the variance for PC1 and you'll notice that four PCs alone explain well over half of the variance okay so this is not 146 dimensional space it's a space of some lower dimensionality and now I can use this for something that's helpful what I can do now is represent those 138 odorants not in 146 dimensional space I'll choose a four dimensional space four because it's more than half of the variance this is an arbitrary decision where each odor is now represented by four values it's PC1, 2, 3 and 4 value so I know this captures more than half of the variance now again I'd like to show you a four dimensional graph I cannot so I will show you a two dimensional graph here's PC1 here's PC2 and here are 138 odorants represented in this space and I'm allowing myself to call this now a meaningful perceptual odor space now that's already making a claim if this is an understanding question and I'll address it now if it's a general question I'll let's do it a bit later just because I wanna capture some ground so these are 138 odorants represented in olfactory perceptual space I'm already making a claim here and if you wanna make a claim you have to validate it and one underlying assumption of this claim that it can be validated is if this is a valid perceptual space that would imply that odors that are close together here should smell similar to each other and odors that are far apart here should be dissimilar from one another so we pseudo-randomly selected nine odorants that span the space then you can just measure the pair-wise Euclidean distance between these odor pairs and then we brought in subjects to the lab presented them with these pairs of odors from an olfactometer where every time after they smell the two they have to cross a line anywhere between extremely different to identical so it's a fair prediction we predict that odors that are close and our space will smell similar and odors that are far in our space will smell dissimilar and here's the result this is 30 individuals where each point is a pair-wise odor comparison and here you have the PCA-based distance and the actual rated similarity and as you can see this is quite straightforward odors that are close in the space smell similar odors that are far in the space smell dissimilar so in this sense the space is perceptually valid now remember in our sort of brief PCA course we said one thing that's cool is that you can sometimes understand the nature of PC1 at least by observing how the various descriptors weighted on it okay what what were its two ends and then sometimes you can give that a nice meaning so of course it would be nice to name PC1 so let's see let's see what its two ends were and as you can see PC1 ranged on the positively weighted end from sweet perfumy aromatic floral and light in the negatively weighted end it had sweaty sharp pungent acid rancid putrid foul decayed and sickening okay quite straightforward again I will ask you for one word that describes things that are floral and perfumy at one end or putrid foul decayed and sickening at the other pleasantness thank you so so so so odor pleasantness would seem to be a good word and pleasantness here in the notion of ranging from very unpleasant to very pleasant to this primary dimension of human olfactory perception once again intuition is nice but not enough so we repeated an experiment where now we have rating of pleasantness from extremely pleasant to extremely unpleasant and here you have the weighting of that on PC1 okay PC1 of the perceptual data is pleasantness this has been since been replicated by by many and this is a robust finding so to quickly reiterate what we've done in perception we've taken this old perceptual data to try and reduce its apparent dimensionality and find the axes of perception that shape this space and we've tested the validity of those axes as an axis for a spatial organization of voters what's slightly cool about this is note that the tests we did are in subjects mostly in a different country 30 years after the original data was collected right so this retained validity across those distances that's kind of nice in turn I'll point out that we've not really discovered anything new here I'll just share this with you so the notion that pleasantness is the primary dimension of human olfactory perceptual space or perception has been long known and to attribute it to anyone then I'll attribute it to Susan Schiffman and her colleagues working in the 70s so this is not novel in that sense it's a rediscovery of something known previously where the nice thing about it is that unlike other methods that were used before PCA although one can discuss its disadvantages and advantages that it's hands off right so I didn't introduce priors to the data I didn't select weightings this is the story the data tells us and that's a nice aspect but to sum up where we're standing on order perception we have a perceptual space we've identified its primary access so we can at least do something meaningful in this respect yes yeah so of course there's there are two selection questions here the descriptors and the odorants I'm large well I should be careful I'm reasonably confident that it's not extremely dependent on the selection of the odorants because within this data set and that's in this paper we did subsets of the data that were randomly selected in various ways it comes out the same thing but more importantly we've since done huge amounts of data and others have as well and it always turns out this way so we have this now on 1500 molecules instead of 138 others have done it on different sets of molecules now you might of course say okay it's the set of molecules that exist in old-fashioned research granted so the extent where I can tell you it's not dependent on it is that it's worked on much larger datasets and it works on cutting of those datasets randomly and arbitrarily which is all I can do right and so to the extent that we can address the question we've addressed it but I agree with you that this can you know maybe if you'll bring a selection of odorants from Mars this might fall apart or descriptors but we'll get to that again only if these are fundamental understanding questions because otherwise you know for debate questions if I can delay you a bit but yeah so with that in mind but again that would be addressed by the things we have done and shown of you know randomly splitting your descriptor set and all this you always end up with this PC1 I mean it's the same PC1 again and again and again no matter what you do with the descriptors or with the odors well okay I see where so in a way I think where this presentation is going we'll sort of address that in a fundamental way because I think this is an I don't think this is related to humans I think it's related to the structure of the world MDS no but MDS you're going to select weights you could in any case I should say you know again this is going to the discussion that I'd really prefer to have at the end but I will say one word because it's also meaningful and this is reiterating a point made by Arun Ravia from our lab we'll see his work soon is that we've over time tested and retested this using really fancy non-linear methods and etc etc etc and you know name your flavor of the month right you know Bayesian statistics whatever you know whatever works that month right nothing works better than the simple linear approach which is interesting about the system right I mean you know you but this is again let's leave this for for later on uh... so we've said something about perception now let's see if we can we can relate it uh... in some extent uh... to neural structure and then to to neural responses and to structure and the neural part here I'm just going to do really I'm not going to go into this in depth in this presentation I'm going to concentrate on this link structure to perception parking here for a second on the neural just to to use you ensure with you a bit of of the things we do uh... and and so humans as you know already from this uh... meeting i don't to go into detail have a full-blown all factory system seem beautiful work on their olfactory uh... bulbs right here they have an olfactory feeling with about six million o r n's in it probably uh... and and and full-blown olfactory system in their brain for us this is a unique opportunity for those of us who study humans this is a unique opportunity because these neurons these olfactory in fact have their their uh... endings uh... outside of the skull in the news where we can reach so we do that and here you have one of our happy participants uh... we invest a lot in developing non-invasive stereotactic devices this is our generation two or somewhere generation five right now so it doesn't look like this anymore uh... it looks worse uh... but but key things key things to see here uh... are the order line uh... entering the nose right here uh... and this is an lfp electrode going and and so how does it look when you go there uh... this i think is my nose all while not sure uh... so we have a fiber optic on this as well uh... that's one of one of the nice things we were talking about this before about my city uh... humans have a big nose uh... and and and you can put in a fiber optic with your with your electrode uh... this is the middle nasal turbinate which already has receptors on it and and this is our electrode arriving uh... this is about just to give you an idea this is about two hundred seventy uh... microns across on the electrode in the world most of the road people live here this is an utterly crude recording device in the world of measuring human neurons this is you know for us this is the best that it gets right because we're recording uh... an lfp uh... in a live behaving human here as you will see here we've gone all the way up to the crew reform plate so this right here basically your brain is two millimeters from here to three millimeters from here this is the pinnest part of your skull uh... and and where parked here right at the crew reform plate uh... with the recording electrode uh... to wake everybody up because in the early phase i'll share with you in amusing uh... historical that happened to us so we we we we we presented this project uh... in in front of the um... we were asked to present it in front of the political uh... gathering of the european union this is because it this was funded by the by the by the nrc and they selected it as as a flagship not because of the science right but because this led to this project uh... that that involved uh... assisting disabled individuals so they they were happy to sort of show it off and they asked us to come to the political meeting which that year happened to be held in in austria uh... and and so i'm standing right in in a room full of like three thousand suits basically right including you know heads of states foreign ministers uh... you name it uh... in in uh... in austria and and one of the graduations in a lab offer pro uh... equipping uh... with the video dad to my talk what where where when i reached showing this i said that uh... our approach to entering the nose was inspired uh... by an austrian renaissance man and i showed this what happened is that nobody in the room even smiled because because apparently uh... Arnold Schwarzenegger is a sensitive issue in austria i was not fully aware of this and and and the talk sort of recovered after that but this was a touchy moment uh... but now back to it's a real videos so here you have these frames are video is well and they're here just to to portray the person is not really moving the shefa and you're hearing those words in Hebrew uh... shefa and and the critical point is that this is the order line and this is the electrode line and what you see here here's the order and here's the the drop uh... in the negative deflection uh... in in the intranasal electrode where on one hand you can look at it as it died you know there's nothing very exciting here but on the other hand this is this is recording a neural response in a behaving human who gives you behavior in real time right there is no equivalent of that that i know of right and so from that aspect it's cool and as i said i won't go into the depth of this part of our work it would be a total separate talk but i'll say in one word remember that our our uh... primary perceptual uh... access is pleasantness uh... we tried any way we could to explain the neural recordings with something meaningful in terms of order and the best explanation we could have found uh... was a fitting uh... to order pleasantness so this is each point here is a comparison of two orders this is the difference in pleasantness versus the difference in response measure in the nose and it's the best explanation we found of the various explanations uh... we tried again not explaining an overwhelming portion of the variance but on the other hand very real so there's something in there uh... that's being picked up so if you give you give so you're giving lots of orders here in succession right and and each order gives you a different e o g response and you're trying to understand explain those differences in response so you can look at every pair of orders you can compare there and ask okay what's the difference between those orders that best correlates with the difference between the responses so with me so it is just like the subjects so it depends on which study your pleasantness was just i think uh... later or before not during uh... but but it is judged by the same subjects of the of the things we have like intensity like pleasantness like familiarity like structure that we tried all sorts of things the one best line that we could find to explain those differences was over pleasantness now this does imply that in a way you could think of the feeling as as patchy along a line of pleasantness but again this is going into a study that i really don't want to spend time on here uh... eighteen different orders sometimes yes but but not what you were looking yes more or less i i wouldn't say a different study because it's the same subjects right and and the differences in pleasantness are not that big across subjects but it's the same subjects so the same individuals rated the pleasantness of the order not while you were seeing this being recorded shortly thereafter or shortly before and the difference in response was correlated with the difference in pleasantness to a to a degree i mean you saw it's not a correlation of one no you rate how pleasant it is on a scale from zero to a hundred it's a difference so so you so this isn't that this is an area under the curve measure so you have uh... this is a response which is a few thousand receptors together probably and you have an area under the curve right so you have area under curve for order a area under curve for order b on to sixteen or eighteen okay so you have eighteen areas under the curve for each two orders there's a difference in the area under the curve and there's that difference versus the difference in pleasantness is is correlated which again does mean again we're parking in at the wrong place but that but so sorry let's go back to this later yeah i'm sorry of course not yeah sadly so right so that so this again desperately not wanting to talk about the study at length but because because i really want to reach new stuff but but uh... uh... in a word right so there's a question which in ways what the study tested there's an underlying assumption of whether the receptor subtypes are randomly and uniformly distributed if they're randomly and uniformly distributed you're right if they're patchy you're wrong our data suggests therefore that they're patchy this is really where we're not but you know let let me because we're i'm way behind i want to talk about new stuff and this is all published stuff you know extreme so so let me try to move ahead so to reiterate where we're standing we've measured perception and and and we came up with this notion of some primary dimension i'm not going to claim much about what we've done here i'm just saying that it seems like at some level this primary dimension is reflected at the earliest stage of neural responses reflected is all i'm saying i'm not saying it's you know i'm not saying more than that but now let's see if we can relate all this to order structure uh... where once again how do you measure order structure we have chemists for this chemists have come up with with basic measures let's say that even i understand and and they've come up with way more complex measures that probably nobody in the room understands and and luckily for that we identified a software called dragon i think that if we'll end up having any contribution to the field of of all faction at the end of the day uh... from our lab it'll be that we brought dragon into the world of all faction i think that's the only maybe valuable thing at the end but and and they're happy that they're italian by the way and they're really nice if you communicate with them and and uh... so dragon is a software that you can plug in any molecule and it spits back at you a bunch of physical chemical descriptors uh... the work i'm showing you here is with v2 of dragon which was one thousand six hundred sixty four descriptors per molecule current dragon is about seven thousand descriptors from out molecule uh... for those that wonder it makes no difference which you use in terms of what i am going to show you once again we have a very potentially high-dimensional space of auto structure right we have molecules with potentially one thousand six hundred sixty four dimensions to them uh... and since our lab is a one-trick pony at this stage uh... we said okay let's see if we can reduce this apparent dimensionality to some more meaningful actual dimensionality so we applied pca to this data some pca of shinados in the room which i know of at least a few might argue that you could not apply a potentially one thousand six hundred sixty four dimensional world two one hundred thirty eight potential data points and with that in mind we mind one thousand five hundred odorants at this stage which we applied the analysis to and here you see the initial result and remember i told you there would be a strikingly similar graph but it's not a copy so here you go uh... here are the first ten pcs of order structure and you will notice that pc one of structure explains in fact slightly more than thirty percent of the variance and once again four pcs of structure explain well over half of the variance in the structural data uh... with that in mind you can play the same set of games namely we can re-present our data in reduced dimensionality space so here are one thousand five hundred odors in pc one and pc two of structure and of course we can ask ourselves the enticing question what is pc one of structure and and again the intuitive way of doing that is looking at the descriptors that flanked both ends so let's look at the descriptors i'm sure everybody in the room will jump with intuitive understanding of a uh... pc that ranges from something like average eigenvector coefficient sum from electronegativity weighted distance matrix at one end that's one descriptor right to the sum of the atomic monorail volumes which may be slightly more meaningful some of us at the other end right you know obviously uh... you know i don't know maybe some people in the room that this speaks to in any way uh... we did not claim that we were one of those people now with that in mind uh... in collaboration with chemists uh... real chemists we've recently actually sort of selected a name uh... for this so this is actually not there's actually text underlying this this black outline but but i outlined it i i i i i i i raised it here for for a reason and and the reason being is that the name the one word name we gave to this chemical entity is arguable and and uh... and poorly argued by me right so i'm not not a chemist so to not put myself in that perless position uh... what i will continue doing here is referring to this access is pc one of structure which is a statistical statement and not a chemical statement okay which i can sort of stand by yeah yes and i don't think i know we've done it you know sometimes one also even if you do it again on the same set is you know you won't always get the same thing right but that more or less the same set you get the same pc one so if you use a different set of odorants in the correlation between pc one and pc one values is like point nine six like things like it's the same okay i'll say it once and and but then don't ask me questions about that now uh... the term that's underlying there is is compactness and it's something that captures both the size in the outdanced that i think where things that are more compact are less pleasant but you know i'm not in the position to defend that uh... but it's compactness perfect thank you thank you for uh... in he would there's uh... there's a term i i don't it doesn't it doesn't have a good translation uh... when you lift for somebody for a sir right so thank you for lifting for for the sir we've done these things independently right we've looked at we looked at at uh... the structure of perceptual space with this structure of structural space there's nothing since we've done this independently there's nothing to prevent us from asking whether these axes aren't anyway related right and this brings me to to to the result that for me till this day means my humble point of view the most you know interesting result in this entire story and as i've pointed out it's kind of sad that the most exciting moment in one's life is captured by this point black and white graph so so to try and give it some some life uh... a lot of color uh... and and here's what you're seeing you're seeing the independent correlograms between perceptual pc one pc two pc three pc four and onto potentially a hundred forty six right and each one of the structural pcs pc one two three four five on to potentially one thousand six hundred sixty four and one correlation statistically stood out and that was the correlation between the primary percept perceptual dimension and the primary structural right again not explaining in the insanely huge amount of the variance but statistically significantly different from any other correlation uh... obtained in this data with any uh... severity of statistical test you want to ask if this is really different from the others it is now what it what why am i saying you know that i think this is what does this this means that this notion of of of odor pleasantness that at least i told we did this work i thought it was entirely subjective i thought that it's merely reflection of you know of our culture of what our mother made us for breakfast of our first love you name it right this means no this means that that thing we call pleasantness is written into the the molecular structure right that it's that it's a it's a descriptor of the natural world it's how the world is organized pc one of structure is the access that explains the structural variance in the set of molecules pc one perception is the access that explains our primary perceptual space and they're linked which in a way is unsurprising right it means we've evolved to extract the most meaningful access in the world around us and we happen to call it pleasantness it could be your mice call it something else that doesn't matter we predict it would be the same access only with a different name right more or less that that of course is a strong statement and and and it's a statement that can be tested because what this means that we think we can predict pleasantness from structure and moreover that the prediction would would survive uh... across uh... cultures and and i will share with you that this is already a summation of lots of data published in separate studies where basically what you see here uh... on on the x-axis is our physical chemical access which is made i should say not only of pc one but it's a pc-based access uh... and this is the rated pleasantness and this is the typical correlations we obtain so we can now uh... and we've done this across three cultures so far uh... no difference uh... and and one of them at least incredibly uh... you know way out so we've done this on native Ethiopians way off right culturally and and and uh... same thing actually you can say almost in a word for cultures i'm counting americans in israelis is the same which is uh... yeah so so uh... so so so you know what we can we can very significantly explain a reason reasonable proportion of the variance in order pleasantness uh... using this approach now remember i'll bring us back to where we set out to go right where this entire thing is to answer that challenge by by uh... uh... what's his name uh... audition uh... bell right that can we compare uh... uh... roses violin and and i think he got and as it's been pointed out here uh... several times throughout this meeting up till now i've been uh... working here with monomolecule the real world is not monomolecule rose violent are not monomolecule i would like to say okay since we know from other work that that when humans rate that the similarity of odorants they're really relying to a large part on the pleasantness similarity i could say okay let's use this is a similarity metric and predict how similar rose by the nasa fatigue are but but they're mixtures and these are monomolecules so so we need to go to the world of mixtures uh... and and this is work now uh... led by by coby snitz who's here in the room where we asked ourselves okay can we how can we use this metric that we have for monomolecules to understand the behavior of mixtures where in in general you can think of two approaches right one would be that let's say we have one mixture made of these two molecules and one mixture made of these three molecules and wanted to compare them you could think of taking this pleasantness metric doing all the pair-wise comparisons between each two possible molecules in the mixtures however many may have and somehow something that and reaching some sort of measure or alternatively you can think of not doing this pair-wise approach but rather first treating each mixture uh... generating one vector of its pleasantnesses right and and and then representing it is as one vector and then the other mixtures another vector and then somehow measuring the relation between the two vectors for example the angle between them what i'll refer to as the angle distance metric so you can think of these two alternatives and for the sake of time i will share with you that this alternative and not go through all the results that this alternative fails completely fails completely uh... and this alternative uh... works powerfully i don't really know how to answer a question of is it just obvious uh... but it wasn't obvious to us hey guys i have nine minutes to go through a ton of data so i'm just gonna say that that uh... this works uh... really powerfully and and here you have what we now call the optimized angle distance metric uh... here you have our predicted similarity here you have the rated similarity where each point here is a comparison between two order mixtures that range in their size from two to forty molecules and this works really really well you would say okay we can now go and answer bells question but wait there's more because here there was one condition in the development of this algorithm and that is that when we made these mixtures we first equated the perceived intensity of all components so we made all components equally intense because intensity adds a bunch of problems in the real world that's also not the case you have mixtures and the components are different intensities in the in the mixture so we set out to solve that problem where first we ask okay what happens if you apply this this existing algorithm to mixtures where the components are not equated intensity their natural mixtures and here's the one example of a result there's lots of data you're looking at here that will soon be publicly available uh... again each point is a mixture are rated by around twenty five people and you can see that sometimes this algorithm works pretty well for real mixtures that were not equated for perceived intensity right i mean this is pretty good where we're happy with this but sometimes uh... it works less good this is another set of mixtures and and sometimes it works less good right it's still statistically significant there's lots of data here but this is not already as impressive in predicting the actual similarity of these mixtures so to address that uh... what allen has now done is added an intensity uh... sort of measure for each component in the mixtures where you can do this in two ways you can either try to predict the intensity of the component by its structural own or actually measure it with people so what we call the expensive in the cheap way uh... and i'm gonna show you the only the result of the expensive way or the less effective way in the long term but basically what we do here is we first have individual subjects rate the intensity of each component of each mixture and we add that as a component in what's now the modified algorithm and here you have the result on that same data so we recover this now so we cannot really do this for natural mixtures okay that have different components with different intensities we can predict their similarity this brings us full circle to the question we set out to ask using the now modified snits raviya algorithm here are the predicted uh... differences uh... between rose violet violet asafetida and rose asafetida okay so this is the prediction um... and here's the result so i think you know i think it's fair for us to say that we've answered uh... the challenge right so we we can now uh... predict the perceptual similarity of rose violet and asafetida what is the implication of all or what could be the implications for all this uh... given uh... the restricted time left i'm i'm gonna actually so so i'm gonna thank leslie for for presenting olfactory white because that's gonna save me i'm not going to go into it i'm just gonna say one i'll just give the introduction to this that one potential outcome right is that now that we have a metric uh... a continuous metric to to to measure a continuous a continuous metric uh... uh... to measure uh... smell uh... since in the world of color vision at least if you make a mixture where you span the continuous metric or an audition where you span the continuous metric it doesn't matter what you put into your span as long as you've spanned you end up with the same percept called white light or white noise we predicted that if we span the perceptual space in all faction uh... even though we use mixtures with no overlapping components they should smell we predicted identical they don't smell identical they smell highly highly similar uh... we call this olfactory white even though i'm terribly running out of time i still have to tell you this one amusing story because it's the only thing you'll probably remember uh... from this talk when we ran the experiments on olfactory white so we gave subjects this novel order and we we said let's give it a name we'll give it some name that sounds like a french perfume or something we called it l'orant after she learned right and and and so what subjects actually did was study l'orant but then when you come to the paper was accepted for publication and i said i i shouldn't get in trouble i emailed gil and said look gil we have this paper coming out with the novel order we've invented we called it l'orant is that cool with you and she answered back said no and what do you even know come on i'd love it if somebody would call an order after me don't be there you know and i emailed impressing again i said come on what do you care and said no call it axel you did so l'orax l'orax is l'orant axel so either we insult nobody or we insult everybody equally i mean you can choose it but that's the only thing you'll remember but that's that's the name of that's the source of the name l'orax and not dr seuss is what everybody else thought is where we brought it from uh... and indeed out again for the sake of time uh... i'll just point out that these these can these mixtures end up uh... it's really smelling similar now because this took longer than intended all the really interesting stuff is basically going to be in the four minutes here is our algorithm applied uh... to leslie's data okay so this is what she presented previously as her triangle test uh... and here's our similarity algorithm uh... the distance angle uh... metric applied to her data explains it kind of nicely and it also uncovers uh... phenomenon and that is that you'll notice that there's this there's this lower boundary uh... to this there's a lower boundary or in other words there's some sort of angle distance metric right some sort of measure in our world of angle distance metric we're given some variants around it this is the border of discrimination in human olfaction okay or what's referred to in in in psychophysics in in the world of sensory systems this is the just noticeable difference okay so this is the j in d of olfaction in the angle distance metric now uh... on this data it's somewhere at this range of of of angle distance metric this data was not collected in order to find this right in order to really find this you need a lot more mixtures right around this number so we now conducted an insane amount of work uh... each dot here is a full study uh... comparison of two mixtures you have a hundred different mixtures here rated each one by twenty five subjects where this is the the uh... angle this is the the result of the triangle test and this is the angle distance uh... in in radians and uh... this is the data where you can see it it summed together really seems to fit a nice psychophysical uh... uh... function uh... without any fitting uh... and and we're currently claiming that the jnd is somewhere around zero point zero five uh... uh... radians in angle distance metric so so this is uh... the jnd what can we do with the jnd i will tell you in two minutes uh... and and uh... to smell the result uh... maybe we'll we'll be able to in a second so what can we do with jnd uh... here you have a bunch of a hundred twenty eight orders uh... and and i'll just point out we've now already done this with two thousand seven hundred but i'm gonna show you the results of hundred twenty eight is that's what i have in the slides that basically span odor space you can take these hundred twenty eight molecules and make any possible mixture that you could make ranging size from two to forty five molecules now it's not really any because anybody who can quickly do math in their head will say that the whitesman super computer will spend the next ten years on doing that uh... so so we we sort of jump the spaces but this is using whitesman you know really powerful computers uh... so it's not all uh... of the mixtures and so you can end up with this huge amount of potential odor mixtures from all these but then we apply the jnd to it and we ask okay leave only those that are discriminable from all others okay only those that are discriminable from all others using our jnd metric and this should be the orders that humans can smell this has a ton of implications which are in the next thirty slides that have twenty two seconds to show you so i will not i will just say in words what they mean one thing this means is that you could potentially take the perimeter of this and the size of the perimeter depends on how you want to define your jnd and mix the perimeter to get any result in the middle right this means that you can choose orders from around here and mix them to generate any other order okay or in other words you'll have a set of other primaries that you can build any order in the world and they'll be indiscriminable by jnd to do that we we we worked together with christophe le Damien, the perfumer known to many here he made for us now a set of orders of he chose whatever he wanted out of the two thousand seven hundred source molecules and we're now recreating the same orders using none of the molecules he's using okay and what you have here is a christophe rose and a rose of ours that are designed to smell the same although they're not and although there's one glitch here, one component that we can talk about if we get to it so A, you have a set of primaries that can make any perceivable order B, this sets up prediction on how many discriminable orders there are, a question that's been a major source of interest in our field and C, this predicts how many perceptual dimensions the system should really have so with some notion of risk I'll give you the numbers that we apply to this right now and I'm doing this, this is a bit risky behavior on my side because by the time this is the paper these numbers can change but they won't change, you know, by orders of magnitude, I think so we're predicting that this model implies that humans can discriminate about twenty million orders that we need two hundred or less primaries to generate any perceivable order that you have and that the dimensions of human olfactory perceptual space are about six I'm going to end with that because even though I wanted to show you more stuff but I still must thank the people who were the heavy lifters in this project and I'll start off with historically Rehan Khan Rafi Haddad who's here in the room Haddas Lapid and Adya Blanka generated important aspects to this sort of tale and currently the people who are carrying the load are Kobi Sneets and Arun Raviya both of them in the room both of them way better equipped than I had answering hard questions and finally our funding which is primarily European both Horizon 2020 and ERC Advanced carrying major portions of this load including some additional funders and finally thank you very much for your attention I know this has been dense