[music]
0:00:05.7 Sarah Crespi: This is the Science Podcast for September 1st, 2023. I’m Sarah Crespi. First up on this week’s show, physicist and education researcher, Laird Kramer discusses his work revealing how improving calculus instruction at universities might encourage more students to stick with STEM fields. We also hear in this segment from some science staffers about their calculus trauma, from the fear of spinning shapes to thinking twice about majoring in physics. Next on the show, what it takes to teach an AI to match molecular structures to odors. Can a computer predict what something will smell like to a person from just its chemical structure? Researcher Emily Mayhew talks about how this was accomplished using a panel of trained smellers and what the next steps are for digitizing the sense of smell.
[music]
0:01:00.4 SC: When I first heard about the science paper on improving the teaching methods for calculus, I immediately thought of all these calculus horror stories that I had heard, and I even have my own. I did have to take it twice. I was asking people, as I went about my day, what was calculus like for you? Did you like it? Was it hard? Did your teacher do a good job? And I brought some of those voices here today. We have Meagan Cantwell. She’s another producer on the podcast. We have Lizzie Wade. She’s a contributing correspondent for News. We have Paul Voosen. He’s a staff writer who’s frequently appeared on the show. And we also have Kevin McLean. He’s a producer who works very closely with me on the podcast. And so we’re going to actually open with Kevin’s experience with calculus.
[music]
0:01:51.3 SC: All right. So, Kevin, did you have to take calculus in high school and college?
0:01:56.4 Kevin McLean: Yeah, I took it in high school and I took it in college. I had a really good calc teacher in high school, which sort of tricked me into thinking I really liked calculus as opposed to in college, they’re like, you guys already know calculus, right? Okay. Alright. Here’s the test. Bye.
0:02:14.4 SC: Yeah. So what was good about the good class?
0:02:16.8 KM: I think at the beginning of the year, I decided that like, this is going to be a lot of work. This is going to be hard. So my friend Shana and I would go to class and then we would go to the library after school every day and do our homework in the presence of each other to make sure that we understood it. [laughter]
0:02:33.0 SC: Wow.
0:02:34.8 KM: And it was just that habit. 0:02:36.9 SC: Yeah.
0:02:37.7 KM: I came up with a system to make sure that I was setting myself up for success. And then it ended up being actually legitimately enjoyable by the end of it.
0:02:46.6 SC: That actually has some similarities to this paper. Like, they do this in class learning. It’s active learning. So they’re solving problems in small groups. So they are doing this kind of communication about the problem as it’s being solved instead of lecture and then watching your instructor perform a problem. And then you go off and do it by yourself later. All the thinking is done by yourself. So, yeah, it sounds like you kind of came up with this on your own. And actually, that’s what the researcher that I’m going to talk to, Laird Kramer, found with a pretty strong study that really got to the heart of this. How do we teach calculus in a way that doesn’t drive students away, that gives them a good experience and, takes advantage of the way that we actually learn? So, of course, even though most of us know what calculus is, maybe you’ve forgot, maybe you’ve blocked it out because it was traumatic. So the first thing I asked of Laird was, what is calculus?
[music]
0:03:48.2 SC: What is calculus? Why do we need it? Why is it being taught for all STEM students?
0:03:54.2 Laird Kramer: So calculus is really the mathematical study of change in processes and systems. So it’s about using derivatives and integrals to understand how functions change in terms of does a function increase or decrease?
0:04:08.1 SC: Calculus is used all over the place, I assume.
0:04:09.3 LK: Yeah, mathematics, sciences, engineering. Science and STEM is really about
understanding change and predicting change. 0:04:17.0 SC: Yeah.
0:04:17.7 LK: And so, it’s kind of a core building block of that understanding of being able to predict the future.
0:04:23.8 SC: What’s wrong with the way that I learned calculus and so many of us learned calculus in undergraduate or even in high school levels? What are some of the problems with the way it’s historically been taught?
0:04:35.1 LK: Big challenge is that a lot of faculty at universities and colleges aren’t really developed in how to teach well. Sometimes us faculty are used to surviving despite the system. A lot of faculty replicate what they’ve done as they were students. And sometimes that’s about trying to identify the best of the best. But really, it’s how do we democratize and how do we let every student be successful and learn and enjoy and find the excitement in calculus or science or engineering?
0:05:07.6 SC: Yeah, be teachers, be educators, lift people up and get them to understand calculus somehow without lecturing for an hour and then sending them home with problem sets.
0:05:18.3 LK: How humans learn is extremely well established. It’s a very messy, it’s a very engaged process. You’ve got to mess around, you’ve got to make mistakes, you’ve got to wrestle with the material. But that really can’t happen in a traditional classroom that features lecturing at people.
[music]
0:05:37.0 SC: So when I found out about this paper, it kind of became my goto question when I was talking to somebody on Zoom for the podcast or for meetings. And so I did talk to our staff writer, Paul Voosen, who covers physical sciences, earth science, about his experience with calculus. A lot of people suffered through it, did not enjoy it. And that’s what the study is about, making calculus education better and less painful. And that it’s kind of like a gatekeeper class. A lot of people are like, well, if I can’t do calculus, then I shouldn’t stay in STEM fields.
0:06:14.3 Paul Voosen: Yeah, that kind of happened to me. 0:06:16.2 SC: Really?
0:06:16.6 PV: And I think I skipped the kind of first calculus collegewise ’cause I went in as a physics and computer science major. And so I had this really horrible professor for calculus. I got the first C+, I think, in my life [laughter] from that. It was my start of freshman year and lots of other stuff was going on, but it made me rethink my interest in physics wisely, I think.
0:06:43.4 SC: Yeah, well, I mean, you’re really good at what you do now, so I would not question it.
[music]
0:06:47.5 SC: That gatekeeper function that Paul and I talked about was something that was a big concern for Laird and the other authors on the paper.
0:06:55.1 LK: Calculus is an essential course that all STEM majors need to take in order to graduate and take on STEM professions. And failing calculus is really one of the mechanisms that can lead to students departing their degree programs. Only about 40% of students that enter universities with an intended STEM major actually graduate with a STEM degree. And even more concerning for us is that the number of women that decide to switch outta stem after calculus class is about one and a half times higher than that of comparable male students. And the same thing is happening for Hispanic, black AfricanAmerican students that are leaving at about a 50% higher failure rate than white students in calculus.
0:07:34.8 SC: So let’s get into what your research found and how you connected it. What did you test as an alternative to kind of this more lecturebased style of calculus teaching?
0:07:44.4 LK: The idea is that we’re really trying to have students replicate the practices of mathematicians in the classroom so they can learn in the classroom time, taking advantage of the resources around them, their colleagues and the faculty in the courses. They work in groups, working on kind of predesigned activities to identify patterns, have some argumentation with themselves and each other, discuss ideas, solve problems, and then really communicate what they’ve learned to each other to kind of help make that more concrete inside their heads.
0:08:15.9 SC: When I did this kinda class in college, we had problem sets that we would take away from the class and we would do them either in small groups or by ourselves. And this is kind of saying that’s the part where you’re doing the learning, you should put it in the classroom.
0:08:27.6 LK: It’s very hard to learn a challenging mathematics problem, calculus problem at two
o’clock in the morning, you’re probably not gonna call up your professor.
0:08:37.0 SC: How did you know? [laughter]
0:08:38.2 LK: You probably shouldn’t call up your professor. Really what we wanna do is have that learning happening in a classroom and allow people to make mistakes because mistakes are kind of essential to learning.
0:08:51.6 SC: I also reached out to one of the learning assistants that Laird mentioned. She was teaching during the research project for this paper. These are students that are at a higher level in the school and they come and help teach earlier students that are taking Calculus 1.
0:09:07.7 Carolyn Marquez: So my name is Carolyn Marquez. I am a mathematics student. I’ll be graduating this upcoming December. I’m a student at Florida International University and I am a learning assistant for Calculus 1.
0:09:18.9 SC: Okay. So what does it mean to be a learning assistant in Calculus 1 these days?
0:09:23.8 CM: From my experience when I was taking calculus 1, our learning assistant was more in charge of grading and occasionally coming to the groups and asking like, oh, do you understand? And if nobody answers or everyone was like, oh, we got it, they just move on. Now it’s changed where we don’t really grade the assignments anymore, but we’re more involved in the classroom. So now we ask students questions and even if students are like, oh yeah, I’m totally fine, we try to make sure that they actually understand the material.
0:09:50.2 SC: Do you see a difference in how the students react to material or understand the material?
0:09:55.9 CM: I think now that because allays are more involved in the discussion, students are becoming more comfortable with asking questions. So I think because they’re becoming more comfortable and asking questions, they’re understanding the material more.
0:10:08.5 SC: Do you like doing it this way better than if you were just grading?
0:10:12.4 CM: Oh, for sure. I’m actually a mathematics student, but I’m in the education track, so I actually really like being involved in the learning process. I like seeing students struggle, but I like being in the process of seeing them struggle and then be like, oh my God, I understand it now. I’m actually helping a student. Even if it’s just one student. I’m making a small change. And I feel like, especially calculus 1, there’s a lot of freshmen taking that class. So I feel like helping them not feel so overwhelmed in such a tough subject, it’s like pretty fulfilling.
0:10:44.3 SC: Well, this sounds great to me. Honest face, it sounds good. But you also sought evidence that this is a better technique for learning. So how did you set up your study to kind of support this claim?
0:10:54.8 LK: We carried a large scale trial that featured randomized student allocation to our state treatment and control conditions. So we could really rigorously test this evidencebased active student engagement calculus course and compare it to a traditional lecturebased calculus instruction. The aim was really to elevate the quality of the research, so that would become
challenging to try to refute the results that we expected. We’ve been working in institutional change and faculty change for many years. It’s a bit frustrating because there’s a lot of evidence that comes from many different fields that all point to active engaged learning is how humans learn. But the adoption and adaption of that effective instructional practice has not really happened the way we would hope for a plethora of reasons. And so we wanted to take the best strategies that we could and put it into a study to kind of put a very strong level of evidence in front of them.
[music]
0:12:00.3 SC: So Lizzie, did you have to take calculus for one of your… I don’t know, high school college?
0:12:04.9 Lizzie Wade: I took it in my last year of high school as an AP class. 0:12:08.6 SC: Did you feel like it was, I guess, unnecessarily difficult? [laughter]
0:12:14.3 LW: Overall, I enjoyed the experience and the part I remember the worst, like my calculus trauma, was definitely about where you sort of have to imagine a 2D shape, like spinning around an access to create a 3D shape. And I remember studying so hard and feeling like I got it and studying with friends and just really, really practicing and sitting down for the test and being like, oh my God. It’s like I’ve never seen this before. And [laughter] that didn’t happen to me that often in high school. And it was pretty shocking. I do sort of feel like… I went on to college, I made it in literature. I never used calculus again. I liked that glimpse into math that was sort of about itself and not about describing concrete things in the world, if that makes sense.
0:13:02.9 SC: Totally.
0:13:03.0 LW: I didn’t really go further with that, but I did think it was interesting.
0:13:05.4 SC: Agreed. That was Lizzie Wade. She’s a contributing correspondent for our news team. So let’s get to your results. How did the students do? How did these two different arms compare in your work?
0:13:16.9 LK: In the end, what we found was medium or large effects size in the treatment condition compared to controls, over three semesters of taking data with about 800 students in controlled and treatment conditions.
0:13:31.1 SC: What did you look at specifically to say the student got more out of a class than this other student?
0:13:37.3 LK: The idea was we kind of took very standardized problems that one often sees in a calculus course. Calculate the limit, calculate the derivative, calculate the interval, kind of the standard problems that come up in a traditional course and then basically insert those questions into the final exams. And then we arranged this so you could not identify which section the students were in. And so we could blindly assess and evaluate all of these exam questions. And so we kind of had a team of undergraduate research students and some others that participated in essentially blind grading of everything. We had multiple people grading them. We established a rubric that was pretty consistent. And then we essentially graded across three semesters and then put it all together and found the medium to large effect sizes.
0:14:26.5 SC: For the study, there was more than just a difference and in how the students did in their calculus exams within the class. The impact went further than that for the students that were in the experimental arm of the trial. What about the grades that, how did the students do on paper, if you will?
0:14:44.4 LK: Grades are not maybe the best measure. These are not as independent, but by and in large, most of the faculty had a very similar strategy in the syllabus and in the grading policies. And so we saw about a 0.4 increase in their overall GPA in the students with treatment compared to traditional and control, which turns into about 11% additional students passing the course. And we also saw that the odds of females passing the course are about 58% higher in treatment than in control. Hispanic students doing better was about double. And really what we saw across all kind of demographic breakdowns, we saw that there was more success in all of them in treatment versus control.
0:15:26.2 SC: Do you expect that to translate into retaining them in STEM majors as well?
0:15:30.3 LK: Yes. I mean part of our hope is that they only did better in the course, but also understand calculus more and appreciate calculus more and kind of see the opportunities of calculus and maybe aren’t quite so mathphobic.
0:15:42.5 SC: Yes, that would be good. Do you see this approach as something that can be used in many different institutions at different universities, different colleges?
0:15:50.6 LK: This was a new curriculum that hadn’t been around before and so in some ways is reflective of what one should expect to have happened at the institution that adopts this in the first few semesters. And we would hope that it continues to get better as expertise develops. And that’s kinda what we also saw across the three semesters that actually the effect got bigger. At FIU, we teach about 2000 students taking calculus for the first time.
0:16:16.4 SC: Wow.
0:16:16.9 LK: Every year. And so our 11% increase adds about 220 additional students passing and succeeding in calculus also technically reduces the instructional needs by about five sections at our institution. There’s about 300,000 students taking calculus every year in the nation. If this were to become the standard, that should translate to about 33,000 additional students passing the courses each year. And that will save them both time and opportunities for their careers. But it can also save, you know, we guessed about $24 million in tuition could be saved kind of using the average price for a three credit course across the nation. We tried to provide some very strongly and very compelling arguments that this should be adopted elsewhere.
0:17:00.1 SC: Calculus obviously is not just taught in college, but also in high school. I talked to one of our producers, Meagan Cantwell, about it.
0:17:07.1 Meagan Cantwell: Well, in your calc class in college, had most people already taken it in high school?
0:17:12.0 SC: I don’t know. I don’t think so. It was like 1995.
0:17:15.9 MC: Well, I can tell you for sure my calc in college, the professor on the first day asked, how many of you had taken calculus before? And almost everybody raised their hand except for maybe like two people.
0:17:28.4 SC: Wow.
0:17:28.9 MC: And he said, if you didn’t raise your hand, please come see me because I don’t know if this is gonna be a good fit for you. So the fact that the expectation is for you to have taken calculus already is just already like, oh God.
0:17:42.6 SC: What about high schools? Did you feel like this is a good fit for that classroom as well for those classrooms?
0:17:47.7 LK: Absolutely, because the content is similar. In some ways, high schools are even more appropriate because there actually is more instructional time with students in a high school setting. And it also make them more competitive for college.
0:18:02.4 SC: You are studying change in many different ways. [laughter]
0:18:05.1 LK: Yes.
0:18:06.0 SC: Thank you so much, Laird.
0:18:07.1 LK: Thanks. It’s really been a pleasure being able to talk about this paper. It’s really exciting. Thanks.
0:18:11.5 SC: Laird Kramer is a professor of physics and faculty in the STEM Transformation Institute at Florida International University. You can find a link to the paper we discussed at science.org/podcast. Special thanks to Kevin McLean, Meagan Cantwell, Paul Voosen, and Lizzie Wade for answering my calculus questions. And thanks again for sharing your embarrassing but also insightful calculus stories. Up next, training in AI to smell odors like people do, or even better than the average person.
[music]
0:18:55.6 SC: We’ve digitized sound and sight, we can input commands that will generate output of color, image sounds, music, single notes. We can also go the other way and use a device to capture images or sounds and it will save them and play them back to us. Now imagine doing that with one of our other senses, with odor, with smell. We are so not there yet. Unfortunately, there are some steps left before we can digitize the sense of smell or digitize odor. One is mapping the space that these odors take up. Why do certain molecules or certain chemicals evoke certain perceptions in people? The pathways are just not well mapped. This week in Science, Emily Mayhew and colleagues use a big set of odor data and machine learning to create a map like this. Hi Emily. Welcome to the Science podcast.
0:19:48.5 Emily Mayhew: Hi Sarah.
0:19:49.3 SC: So why has it been such a big challenge to digitize the sense of smell the way we have with sound or with vision?
0:19:57.5 EM: Yeah. The relationship between odor stimuli, which are small molecules and odor percept is a lot hairier than the relationship between stimulus and perception for other senses, right? If you think about vision, it’s just one property, the wavelength of the light that corresponds to the color that we see. But for smell, what we’re perceiving is way more complex. So if we smell coffee, there are hundreds of small molecules that we are smelling that are activating our olfactory receptors and triggering that recognition of the smell of coffee and the relationship between the structures of those molecules and how we smell them, is often very counterintuitive. So we’ll see cases where there are two molecules that have very different structures, but they smell almost identical to us. And then we’ll see cases where molecules have very similar structures, but they have totally different odor percepts. So it’s a real puzzle. How can we look at this molecule and predict what that molecule will smell like to a person?
0:21:03.7 SC: This has been a really big challenge lining up what a molecule structure looks like with how we perceive it. What you wanted to do in this study was get a really big data set and train an AI on it. Where would you get a data set like this that would match up an odorant and how it’s perceived by people?
0:21:21.6 EM: Yeah, so it’s a very slow process to collect data on what molecules smell like. You have to go buy all those molecules. And so the best way to get a large set of this data is to go to the flavor and fragrance industry, these are the people who’ve been doing this professionally for a very long time. And so we pulled two professional data sets that are used by the flavor and fragrance industry that together had about 5000 unique molecules in them and descriptions of what they smelled like. So something like floral, fruity, green, or herbaceous. So we use those verbal descriptors for these 5000 or so molecules to train a graph neural network model to look at the structure of any molecule and predict what odor descriptors would apply for that smell. Or in other words, what that molecule will smell like.
0:22:13.9 SC: A little background on that dataset. Is it expert smellers that did this? Are these people who have kind of a background in odor detection?
0:22:22.8 EM: Yeah. So the provenance of those descriptions is not known to us because these are just old websites that are used by the industry. And so we don’t know if it was one person who smelled that molecule and wrote those words down, or if it was consolidation of many descriptions. We know that they are people in the industry. And so they have some degree of expertise, but the amount of data going into those descriptions is really opaque to us.
0:22:50.0 SC: But you do have a good starting point here. You have that massive data set that’s really gonna set you up for testing and honing how smells relate to molecules.
0:23:00.8 EM: Exactly. And one of the beauties of machine learning is that it can handle a lot of noise in the training data as long as you have enough training data. And so we have this imperfect data set, but it’s really, really large.
0:23:12.4 SC: The next step for you was to test how well the machine learning went, like how this neural net responded to the dataset with a fresh new dataset that no one had seen before. Can you talk a little bit about how you built that?
0:23:25.8 EM: Right. So as I mentioned, machine learning models can do really well on noisy
training data, but if we wanna know how reliable our model is, we need to test on clean data, right? If we have noisy data that we’re testing on, we can’t trust our measure of how reliable that model is. And so we set out to collect a brand new dataset on the odor character of molecules, and we purchased 400 molecules that to our knowledge, had never been smelled before. So these are brand new odorants, which is pretty exciting. That’s a large [laughter], a large expansion. When you think about the fact that we could only find 5000 known odorants and now we’re adding 400 new ones.
0:24:07.7 SC: Wow.
0:24:08.2 EM: That’s like a 10% increase. 0:24:10.2 SC: So, fascinating.
0:24:11.3 EM: Yeah. So, we purchased these 400 novel odorants molecules that based on their structure, we thought would have a smell. And then we recruited a panel of 15 human subjects, just normal individuals with a normal sense of smell. And we trained them to use a 55 word odor lexicon. So words like fruity, floral, grassy, musky, ozone. And we taught them using a odor reference kit, how these words mapped to odor percepts.
0:24:42.7 SC: How would you train someone how to smell?
0:24:45.7 EM: Yeah. So a lot of us, I mean, we smell every day, but we don’t have practice using words to describe what we smell. And so in past olfactory research, we’ve had this issue where people sometimes don’t know the precise word to describe what they’re perceiving or they conflate two words. So there’s research where people made ratings for how musky an odor smelled, and they would conflate that with musty. So thinking that a perfume [laughter] smells like your dirty socks.
0:25:20.3 SC: Okay, so attic smell versus like the smell of a base note of a perfume. Yeah. That’s a big difference.
0:25:27.3 EM: Exactly. So we need to help people by teaching them what are the right words to use to describe the different kinds of things that you can smell.
0:25:35.9 SC: Okay. So you have these 15 people that got trained on kind of a smell kit. 0:25:40.0 EM: They got trained on a smell kit.
0:25:41.4 SC: They have the words that they need to describe smells. And then you gave them all 400 novel smells?
0:25:46.4 EM: Yes. In phases, in batches of 25, they received all 400 of these novel smells. And then they chose which of those 55 words in our lexicon applied and described those smells and to what extent. And so that formed our new much cleaner dataset.
0:26:03.8 SC: And so then you took those molecules over to your AI and you said, these are molecule structure. What do you think that they’re going to smell like? How are they going to be described by humans?
0:26:13.9 EM: Exactly.
0:26:14.7 SC: And how did it do? Does it do a good job describing odors?
0:26:18.3 EM: Surprisingly well, so I think often when we’re evaluating a model, we think about its performance with respect to the performance of a person. How well can the model do this task compared to a person? And what we found is that our model came closer to the average description of that odorant smell than the median panelist. So in other words, if you want to know what a new molecule would smell like, should you ask a random person to smell that and tell you what it smells like or should you ask this computer or this model to tell you what that molecule will smell like? We found basically in the majority of cases, the computer is going to give you a better answer. So you should ask this model how humans will perceive this scent, which is kind of amazing because a person can actually smell it and output their perception. And this model is just looking at the structure and yet its prediction of what that molecule will smell like is closer to the human average than the median human subject.
0:27:17.5 SC: So that’s it. We digitized smell. [laughter] Just kidding. What can we do with this now that we have a machine train like this? Does this open doors for designing smells? What do you see as the next steps with this area?
0:27:31.7 EM: Yeah, so now we get a lot faster at predicting what new odorants will smell like. So I told you that we had 15 people smell 400 molecules, and that took about a year [laughter] to generate all of that data. So now we have a model that can read through the hundreds of thousands of structures and output a pretty good guess about what that molecule is going to smell like. So we’ve gotten just infinitely faster at surveying chemical space for potentially interesting odorants.
0:28:06.5 SC: So if you took a sample of the environment and sent the structures over to me and said, this is all the structures that we saw, this would be able to help tell me what it smelled like in that space?
0:28:16.9 EM: So that’s a really interesting question because now you’ve brought in the idea of odour mixtures.
0:28:21.3 SC: Oh yeah, that’s right. These are pure chemicals versus…
0:28:25.3 EM: These are pure chemicals. So we are asking the model to tell us what one molecule at a time would smell like. But of course, most of the things that we smell in our environment are mixtures of lots of different molecules. And so that’s kind of the next step. Once we know what a single molecule will smell like, can we predict what a certain mixture of molecules will smell like.
0:28:45.8 SC: That’s going to take even more time. Even though you’re faster on the AI side, you’re still going to be mixing things and exposing people to mixtures.
0:28:54.6 EM: And that data is really not available in the same way. We’re gonna have to generate a lot of that data ourselves.
0:29:00.7 SC: Wow, your smellers are going to have a lot of work to do. 0:29:02.8 EM: They’re going to have a lot of work to do. [laughter]
0:29:05.7 SC: So let’s take it past the mixtures and more training. What is the amazing future direction? What do odour scientists dream of when it comes to digitizing smell? What would the future look like?
0:29:18.6 EM: Well, I think digitizing smell is the perfect phrase to use. In order to match what we’ve achieved for sound or for vision, we need to be able to capture, so an image or a sound or a smell, and then we need to be able to store it and then reproduce it. We’re getting to the point where we can capture and map based on these molecules that are present. What do we think that percept would be? We can kind of place it in that perceptual space, but then there’s still the translation back out. So how do you reproduce that smell? And so I think there’s another phase of research that we need to do once we can predict what mixtures will smell like, is there a way that we can then recreate that percept without having those exact molecules present again?
0:30:09.5 SC: Yeah. So you’re now looking at the receptors and saying, how do we trick you guys? 0:30:13.3 EM: Exactly. How do we trick you into recreating that same percept, but without having
all 400 of those aroma compounds that were in your cup of coffee?
0:30:22.8 SC: We do that with vision. That’s what TV is. It’s not impossible to trick us. 0:30:27.6 EM: Exactly.
0:30:28.6 SC: In addition to learning about this chemical space and how the structures relate to the precepts they create in people, did you learn anything about how people process odors, what happens out in the real world is mapped onto our biology?
0:30:43.8 EM: Yeah. I think that this gives us some hints. We were really excited to find that this mapping that we created from structure to percept did more than the tasks that we trained it to do. So we were trying to predict what odour labels apply to a given chemical structure, but we found that that same map was really good at predicting the intensity of the odour or how perceptually similar people found two odors to be, the detection threshold of molecules. And so there seems to be some kind of universality in this mapping. We’re extracting the olfactory relevant, parts of the molecule, and they’re applying to lots of different characteristics of odour perception. I’ll touch on one wonk point for the olfactory nerds out there, [laughter] which is that odorants are everywhere.
0:31:40.9 EM: And so it’s really, really hard to collect data on a pure odour compound because everything that we buy from a chemical supplier is 95, 98, 99% pure. But what about that five, two or 1% of other stuff? And so we took an extra step to use a technique called gas chromatography mass spectrometry/olfactometry. So GCMSO is what we call it, where you actually separate all of the chemicals in a sample and then you smell them. We did this process for 50 of the 400 stimuli in our test set. And so we were able then to match the odor that were present in the sample to that being compound that we thought we bought. But then also what is the odor influence of these trace contaminants?
0:32:32.8 SC: Oh, interesting. So you were able to smell a difference between a purified sample and say the mixture that you started with?
0:32:39.8 EM: Well, you smell it coming off the column, so you can smell what’s the odor character of that main peak, that’s the compound that I bought, and then what is the influence of
these trace contaminants in the odor that I smell?
0:32:50.6 SC: And you’re doing this live as it’s transiting. [chuckle]
0:32:53.7 EM: As they’re coming off the column. Yeah, one of the authors and her lab mates, Jane Parker, she did this work and we found that these contaminants are abundant and a lot of them do have odors and that’s what you would expect, right? Because they’re precursors and reagents often and that has a real impact in our data and I think that’s something that’s not accounted for very much in the field.
0:33:16.1 SC: So interesting. Oh, this reminds me of something else that I wanted to talk about. A lot of people don’t realize that the smell space, maybe they’ve never thought of, it is not infinite. You cannot go in every direction and find odors. There are certain things that constrain it. Can you talk a little bit about what we actually can smell in the chemical world?
0:33:33.9 EM: Chemical space is vast, but odor space is a subset of that space. We can only smell things that are volatile or lightweight enough that they leave their source and we can breathe them in and they travel up our airways and reach our olfactory receptors. They need to be relatively lightweight, volatile molecule. And then they need to be relatively hydrophobic to enter the olfactory receptor binding cavities. I have a very shorthand rule of thumb. I call it the rule of three.
0:34:06.1 SC: Oh my goodness.
0:34:07.1 EM: The rule of three is a molecule between 330 grams per mole and with fewer than
three heteroatoms is generally an odorant. So you can impress your friends at dinner [laughter] 0:34:17.0 SC: All right. Well, thank you so much, Emily.
0:34:19.5 EM: It was a pleasure talking to you.
0:34:20.8 SC: Emily Mayhew is a professor of the Department of Food Science and Human Nutrition at Michigan State University. The work was done in collaboration with researchers from Osmo and Monell Chemical senses Center. You can read the paper we discussed at science dot org/podcast. And that concludes this edition of the Science Podcast. If you have any comments or suggestions, write to us at science podcast@AAAS.org. You can listen to this show on our website, science.org/podcast or search for science magazine on any podcasting app. This show was edited by me, Sarah Crespi and Kevin McLean with production help from Podigy. Lots of thanks to our calculus story sharers, Meagan Cantwell, Lizzie Wade, Paul Voosen, and Kevin McLean. Jeffrey Cook composed the music. On behalf of science and its publisher, AAAS, thanks for joining us.
Reading Material 