Auransa’s Pek Lum on Using Machine Learning to Match New Drugs with the Right Patients

Episode Notes

Pek Lum, co-founder, and CEO of Auransa, believes that a lot fewer drugs would fail in Phase 2 clinical trials if they were tested on patients predisposed to respond. The problem is finding the sub-populations of likely high-responders in advance and matching them up with promising drug compounds. That’s Auransa’s specialty.

The Palo Alto, CA-based drug discovery startup, formerly known as Capella Biosciences, has a pipeline of novel compounds for treating cancer and other conditions identified through machine learning analysis of genomic data and other kinds of data. It’s closest to the clinical trial stage with a gene expression modulator for liver cancer (AU-409) and is also working on drugs for prostate cancer and for protecting the heart against chemotherapy drugs.

The company says it discovered AU-409 as part of a broad evaluation of data sets on a range of close to 30 diseases. The company’s discovery process uses a platform called the SMarTR Engine that uses hypothesis-free machine learning to identify druggable targets and compounds as well as likely high-responder patients. Lum  calls it “interrogating gene expression profiles to identify patient sub-populations.” The company believes this approach can identify unexpected connections between diverse molecular pathways to disease, and that it will lead to progress in drug development for intractable conditions with poorly understood biology, including cancer and autoimmune, metabolic, infectious, and neurological diseases.

Lum co-founded Auransa with Viwat Visuthikraisee in 2014 and is the chief architect behind its technology. Before Auransa, she was VP of Product, VP of Solutions, and Chief Data Scientist at Ayasdi (now SymphonyAyasdiAI), a Stanford spinout known for building hypothesis-free machine learning models to detect patterns in business data. Before that, she spent 10 years as a scientific director at Rosetta Inpharmatics, a microarray and genomics company that was acquired by Merck. She has bachelor’s and master’s of science degrees in biochemistry from Hokkaido University in Japan and a Ph.D. in molecular biology from the University of Washington, where she studied yeast genetics.



Harry Glorikian: I’m Harry Glorikian, and this is MoneyBall Medicine, the interview podcast where we meet researchers, entrepreneurs, and physicians who are using the power of data to improve patient health and make healthcare delivery more efficient. You can think of each episode as a new chapter in the never-ending audio version of my 2017 book, MoneyBall Medicine: Thriving in the New Data-Driven Healthcare Market. If you like the show, please do us a favor and leave a rating and review at Apple Podcasts.

For every drug candidate that makes it all the way through the three phases of clinical trials to win FDA approval, there are about 20 others that fail along the way. Phase 2, where drug makers have to prove that a new drug is safer or more effective than existing treatments, is where a lot of drugs falter.

But often, it’s not because the drugs don’t work. Sometimes it’s just because they weren’t tested on the right patients. Meaning, the people in the treatment group didn’t happen have the right genes or gene expression profiles to respond. If you could find enough patients who were likely high-responders and try your new drug just on them, your chances of approval might go way up. The tough part is identifying those subpopulations in advance and matching them up with promising drug compounds.

That’s where a company like Auransa comes in. It’s a Palo Alto startup that has built an AI platform called the SMarTR Engine. The engine uses public datasets on gene expression to identify subtypes of molecular diseases and predict what kinds of compounds might work against specific subtypes. Auransa used the engine to discover a drug for liver cancer that’s about to enter clinical trials. And it’s licensing out other drugs it discovered for prostate cancer and for protecting the heart against the effects of cancer chemotherapy.

Some of the ideas baked into the SMarTR Engine come from a sub-field of artificial intelligence called hypothesis-free machine learning. And joining us this week to explain exactly what that means is our guest Pek Lum. She’s a biochemist and molecular biologist who worked at the microarray maker Rosetta Inpharmatics and the software company Ayasdi before founding Auransa in 2014. And she says one of the real revolutions in drug development is that almost every disease can be divided up into molecular subtypes that can best be treated using targeted drugs.

Harry Glorikian: Pek, welcome to the show.

Pek Lum: Thank you. Pleasure to be here.

Harry Glorikian: You know, I always try to ask this opening question when I start the show to give the listeners a good idea of what your company does. But you guys are in drug discovery. Tell us how people understand what is the basic approach that you guys have. And I’ll get into the special sauce later. But what do you guys do in the drug discovery space?

Pek Lum: No, that’s a really great question in the sense that when we first started in about five years ago, we. I’ve always been in the drug discovery field in the sense that I worked for over 20 years ago at that time in a company called Rosetta Inpharmatics, which is really pushing the cutting edge of thinking about using molecular data. Right. And to solve the mysteries of biology. And I was extremely lucky to be one of the core members in when we were very small. And then that really kind of put me in the sense put me in the stage where I could think about more than just one gene. Right. Because the technology was just kind of getting really kind of I would say not rolling forward, like propelling forward, with microarrays.

Harry Glorikian: Yes.

Pek Lum: So I was part of the whole movement and it was really amazing to be kind of like, you know, in the show as it runs, so to speak. And so and then Merck bought us after we went public and worked for Merck and Co. for another eight years, really learning how technology, how we should apply technology, how we can apply technology, molecular data, RNA data, DNA data to a drug discovery pipeline. And really kind of figured out that there are many things that the pharmaceutical world does very well, but there are many things that it also fails in and that how can we do it better? So I’ve always been in the mindset of, when starting Auransa with my co-founder, How do we do it better? And not only just do it better, but do it very differently so that we can address the most, I would say critical problems. So Auransa is really a company started by us to address the problem of why drugs actually fail a lot when we go into a Phase II efficacy trial. Right. Is not like the drug is bad or toxic. And most of the time is you can find enough responders to make your clinical trial a success.

Pek Lum: And that cause, I guess, drugs actually made to maybe against one target. You don’t really think about the biology that much at the beginning or the biology responders. So Auransa was really created to think about first, the heterogeneity of the disease and the heterogeneity of patient response. So we start from looking at molecular data of the disease from the get go. We take RNA, is really the RNA world is coming back with the vaccines.

Harry Glorikian: Right.

Pek Lum: And the RNA has always been fascinating because it tells you about the activity of the cell, of a normal cell versus a disease cell. So we use RNA transcriptomes right, transcriptomics to study the biology and the heterogeneity. So our algorithms, there are many algorithms, one of the first algorithms of the engine is really to look at the biology of heterogeneity, whether we can subdivide a disease into more homogeneous categories before doing anything.

Harry Glorikian: Right. Yeah, I remember when, because when I was at Applied Biosystems, I remember Applied Biosystems, Affymetrix and then Stephen Friend starting this and like, you know, it was all starting back then. And I want to say we sort of had an idea of what we were doing, but compared to now, it’s like, wow, how naïve we were back then compared to how much this whole space has evolved. And it’s interesting you mention, you know, RNA and its activity because in a couple of weeks, I’m actually going to be talking to a spatial genomics company so that you get a better idea from a visual standpoint of which cells are actually activating and which aren’t.

Harry Glorikian: But so, you’ve got an interesting professional career, and I say that because you were working at a big data analytics company for a while that was utilizing an approach that was hypothesis-free machine learning, where the machine was sort of identifying unique or aspects that you should be paying attention to. Maybe that it was seeing that instead of you going in there saying, let’s just look over here, you could see what the machine was seeing for you. How much can you tell us a little bit about that experience? And then how did that influence what you’re doing now? Because I have to believe that they superimpose at some level.

Pek Lum: Right. I think, you know, ever since my first job at Rosetta and then my subsequent jobs really kind of culminated into this into this tech, as you see today. Right. All this experience and certainly experience while being a founding member of a small team at that time of Ayasdi, which is the software company, has been also an eye-opening experience for me because we were trying to create, using a very old mathematical idea called topology, or TDA, really start to figure out whether there’s maybe there’s some things that can’t be learned. Right. And so typical machine learning methods need a training set or a test. But there are just some things where you don’t really know what the ground truth is. So how do you do that? So that’s the idea of like I say, the hypothesis-free approach. And the approach that that that the tech company, the software company that we built is really around the idea that not everything can be learned. But you can actually adapt some very interesting ideas around a hypothesis-free approach and then use it in a machine learning AI framework. So I definitely have been influenced by that thinking, you know, as I as we built the software.

Harry Glorikian: Right.

Pek Lum: And also, when we were Rosetta, we were generating in parallel, data on thousands of genes. And often at that time we were called, “Oh, you’re just going fishing,” you know, but fishing is not a bad idea because you don’t really know which part of the ocean you need to go to catch your Blue Marlin, for example, right?

Harry Glorikian: Yeah, no, no, absolutely.

Pek Lum: Fish a little bit, not the whole ocean, but, you know, to get some, I would say, boundaries. Right. So in that sense, to me, a hypothesis-free approach gives you the boundaries where you can look. So, you know, so the experience, definitely the idea that you can use methods or thinking, algorithms, that could help you in a field where you do not know the ground truth. Like patient heterogeneity, I would say nobody really can pinpoint and say, OK, I can say that, oh, this is THE subtype, these are THE markers. And therefore, I’m going to go after this. And there are many. I guess, for example, you can think of a Herceptin as a great example, right, but when you first started, you know, it was like, wow, OK, you’re going to go after a target. And then the idea of really kind of subtyping breast cancer, you know, I don’t know, 20, 30 years ago. Right. And we’re still learning about, you know, in a patient heterogeneity and we’re just beginning to scratch the surface. So for Auransa, we wanted to use a method very much like the thinking that and the idea that we had, you know, when we were when I was at Ayasdi, is that you could search with some parameters, you know, a very complex space without needing to say, this is my hypothesis. This is that one gene, because we all know that if you have a target, you know…to have to respond you need the target. But if you have the target, it doesn’t mean you’re going to respond. Because things below the target or above the target are much more complex than that.

Harry Glorikian: Correct. And I always feel that there’s, you know, I always call them low hanging fruit. Like the first one is, OK, well, it’s either luck or skill, but I got to one level. But then you start to see people that are not responding. So that means something else is going on and there’s subtypes. Right. So it’s funny how we always also call it “rare diseases” in these smaller population. I’m pretty convinced that at some point everything is going to be a rare disease. Right. Because of the subtypes that we’re going to start to see. I mean, even we’re seeing in a neurological now, or Alzheimer’s. There’s subtypes of Alzheimer’s. No! Really? Shocking. Amazing to me that there’s subtypes. Right. We’ve been dealing with this for ages. And I do believe that these technologies are so good at highlighting something where a human might not have seen it, might not have understood it. You know, I was I was interviewing actually I just posted it today on imaging and agriculture. And they were saying that sometimes the machine sees things that we don’t fully understand how it sees it, but it sees it and points it out, which allows us now to dig into it and be able to sort of identify what that unique feature is that the machine has pulled out. I’m not sure I want drug discovery and drugs being based on something we don’t fully understand, but the machine highlighting something for us that then we can go dig into, I think is an interesting greenfield space that that we need to explore more.

Pek Lum: Right. I think you’re absolutely right. You know, when we first started Auransa, that was the idea that we had. And then my co-founder and I thought, what if we find like hundreds of subtypes? We’re never going to be able to make a drug again a hundred subtypes. So let’s hope we find a small enough number of buckets that we can say this is approximately what it looks like, to be able to be practical to find drugs against those subtypes. So when we talk about subtypes, we are talking about you’re absolutely right, it’s like a leaf on a tree and that we have to cut it off at one point. Enough that things that, OK, this is homogeneous enough that actually makes sense out of it. And that’s where the engine, that’s what the engine does. Basically, it takes data, very, very complex data, things that we could never figure that out ourselves and say this approximately five, six buckets. So we’ve actually not found hundreds of subtypes, otherwise we probably would not have started Auransan, because it would have been impossible. But instead, we find n of one, but maybe a five to seven subtypes at most. That is enough for us to say, the machine says, OK, it is homogeneous enough, go for this. So that’s kind of where we are, where we start at Auransa. And I think that’s an important concept because people often thought about precision medicine as being, oh, I’m going to make a medicine for you and you only. But actually you could learn from, say, breast cancer, and that’s approximately people with estrogen-receptor-positive tumors. And then you will likely respond to a drug like Tamoxifen. And even though we know that the response rate is only about, I think maybe 30, 40 percent. Right. But that’s really good. At least at this point. So that’s where we how we think about the engine as a shining light on a homogeneous enough population that we can actually make a drug against that.

Harry Glorikian: Yeah. So that sort of leads us into you have this technology that you’ve termed SMarTR, S-M-A-R-T-R engine. Right. What does that stand for?

Pek Lum: You know, that’s my one of my rare occasion where I put my marketing hat on. I don’t like marketing all. And we so and you notice the Mar is big-M, little-a-r. So S is for Subpopulation. Markers. Targets. And Redefining. Because I needed it to be Smartr.

Harry Glorikian: Ok, ok. So and when you like when you’ve described this in the papers that I’ve looked at it, it’s a machine learning mathematical statistical approaches, highly automated and totally runs in the cloud. So can you give us a little more color on the sort of the highly automated, and why is that so important?

Pek Lum: Right. It’s important because it comes from my own experience of working with, like, amazingly talented implementations and data scientist at the at Merck or I know how it goes where biologists will often ask them for something and they would run their magic and they’d give us an Excel sheet or a PowerPoint. Right. It’s always a one-off one of those and one of that because you know, biologists are kind of one-off. So the idea of of us building this engine is not just equipping it with algorithms. So first of all, we don’t have one algorithm, a hammer looking for a nail. We have a problem to solve. The problem is how to find novel drugs, drugs that people have never thought about, for patient populations that will respond.

Pek Lum: So with that in mind, we built a pipeline of algorithms that starting from thinking about heterogeneity, to understanding preclinical models that reflect the biology of human subtypes, to predicting drugs and targets for those, and getting biomarkers for the patients when we go to the clinic. And we have different algorithms for each step of the pathway. So instead of having my team do a one-off thing, we know that if we don’t do good software engineering it’s going to be problematic because first it’s going to take a really long time. This will be kind of higgledy piggledy in Excel sheets and we might be able to solve one thing. But to do this as a platform and as a pipeline builder, it would be impossible without good engineering practices. So we wanted to put this in, like I say, in a framework where everything is connected, so where it gets to run faster and faster through better algorithms, through better software engineering. And this really kind of came from my experience to at Ayasdi, a software engineering, a software firm. And also my co-founder who is a physicist and a software engineer, that we need to have good software practices. So what we did was we built first. We don’t want any servers. Everything is done on AWS and is done in modules. So we create algorithms for each part of the pipeline, of the in silico pipeline. And then we have in such a way that when we take data in, when we ingest data, that we also automate it, and then by the time it ingest data and it spits out, I would say, what subtypes of disease, what biomarkers could be used in the clinic, what targets are interesting to you, what compounds from our digital library of compounds may be effective for that. Everything is more or less connected and could be done up in the cloud and now it finishes in about 24 hours.

Harry Glorikian: When do humans look at it to say hmmm, makes sense. Or maybe we need to tweak the model a little. Right. Because it’s not making sense. When does that happen?

Pek Lum: So we, it happens at several steps. So within our engine we actually have benchmarks in there that we run periodically. You know, for example we have about about eight to ten data sets that we have for breast cancer, thousands of patient tumors. And we know approximately that it should be discovering, and it has discovered ER+ flavored subtypes, ERBB2, HER2+ subtypes, triple negative subtypes. So that is kind of like the rails that we put into our engine as well to make sure that when we actually do tweak an algorithm, it still has its wheels. But what we do is at this point, we generate out all the in-between data, but it’s kept on the cloud. And once it’s up, when it outputs the the list of things, the biologists actually, I would say the biologists with a knack for computation, we look at it and I myself look at it. I love to do data analysis in my spare time when I’m not doing CEO stuff. And we can see that we will look at once it’s done that it also allows you…Ok, so this is an interesting one. The engine on the cloud outputs all of this. And right now, let’s say my CSO, who is not a computational person, or me, or whoever really would be kind of a big pain to kind of go up and install the stuff and look at the things, some things you can’t see. So what we did as a company is to build another kind of software, which is the visualization software on top of that.

Pek Lum: So we have on our other end a visualization software that we call Polo because it’s exploring that basically connects everything the SMarTR engine has done into something that’s visualizable. It has a URL, we go to it and let’s say, for example, my CSO wants to know, OK, the last one you did on head and neck cancer, you know, how many subtypes did you find? What is the biology, what’s the pathway? And it could do all of that by him just going then looking at things. Or he can actually type in his favorite gene and then see what the favorite gene actually is predicted for how it behaves across over 30 diseases, and you can do that all at his fingertips, so we have that part of the engine as well, which is not the engine. We call it Polo, which is our visualization platform.

Harry Glorikian: Right. It’s funny because one of the first times I interviewed Berg Pharma and they were talking about their system, I was like, if you put on a pair of VR glasses, could you see the interconnectivity and be able to look in a spatial…. I was on another planet at the time, but it was a lot of fun sort of thinking about how you could visualize how these things interact to make it easy. Because human beings I mean, you see a picture. Somehow we’re able to process a picture a lot faster than all this individual data. I think it… I just slow down. I rather look at a visual if it’s possible.

Pek Lum: It is so important because, you know, even though the engine is extremely powerful now, takes it 24 hours to finish from data input to kind of spitting out this information that we need. Visualization and also like the interpretation and just kind of making sure kind of like the human intelligence. Can I keep an eye on things. The visualization platform is so, so important. That’s why I feel like that we did the right thing in making and taking time, putting a bit of resources to make this visualization platform for our preclinical team who actually then needs to look at it and go, OK, these are the drugs that are that are predicted by the engine. Can we actually have an analog of it or does it have development legs? Does it make sense? Does the biology makes sense. And so now we’re basically connected everything. So you can click on a, you can find a drug in a database and it will pop up, you know, the structure and then it will tell you, hey, this one has a furan ring. So maybe you might want to be careful about that. This one has a reactive oxygen moiety. You might want to be careful about that. As we grew the visualization platform, we got feedback from the users. So we put more and more things in there, such that now it has a little visualization module that you can go to. And if you ever want to know something, I can just, I don’t have to email my data scientist at 1:00 am in the morning saying, hey, can you send me that Excel sheet that has that that particular thing on it that I want to know from two weeks ago? I can just go to Auransa’s Polo, right? As long as I have wi-fi. Right. And be able to be self-sufficient and look at things and then ask them questions if things look weird or, you know, talk to my CEO and say, hey, look at this. This is actually pretty interesting. And this one gets accessed by anybody in Auransa as long as you have Wi-Fi.

Harry Glorikian: So so it’s software development and drug development at the same time. Right. It’s interesting because I always think to myself, if we ever, like, went back and thought about how to redo pharma, you’d probably tear apart the existing big pharma. Other than maybe the marketing group, right, marketing and sales group, you tear apart the rest of it and build it completely differently from the ground up? It was funny, I was talking to someone yesterday at a financial firm, a good friend of mine, and it’s her new job and she’s like, my job is to fully automate the back to the back end and the middle and go from 200 people down to 30 people because we’re fully automating it. I’m like, well, that sounds really cool. I’m not really thrilled about losing the other 170 people. But with today’s technology, you can make some of these processes much more automated and efficient. So where do you get your data sets that you feed your programs?

Pek Lum: Yeah, let me tell you this. We are asked this a lot of times. And just kind of coming back again for my background as an RNA person. Right. One thing that I think NIH and CBI did really well over 20 years ago is to say, guys, now we no longer doing a one gene thing. We have microarrays and we’re going to have sequencing. There’s going to be a ton of data. We need to start a national database. Right. And it will enable, for anybody that publishes, to put the data into a coherent place. And even with big projects like TCGA, they need things that could be accessed. Right. So I think it is really cool that we have this kind of, I would say, repository. That unfortunately is not used by a lot of people because, you know, everything goes in. That’s a ton of heterogeneity. So when we first started the company, before we even started the company, we thought about, OK, where is it that we can get data? We could spend billions of dollars generating data on cells, pristine data, but then it would never represent what’s in the clinical trials without what’s out there in the human the human world, which is the wild, wild west. Right. Heterogeneity is abundant. So we thought, aha, a repository like, you know, like GEO, the Gene Expression Omnibus, right, and ANBO or TCGA allows this kind of heterogeneity to come in and allows us the opportunity to actually use the algorithms which actually have algorithms that we look for. We actually use to look for heterogeneity and put them into homogeneity. These kind of data sets. So we love the public data sets. So because it’s free, is generated by a ton of money. It is just sitting there and it’s got heterogeneity like nobody’s business. Like you could find a cohort of patients that came from India, a cohort of patients that came from North Carolina, and group of patients that came from Singapore and from different places in the US and different platforms. So because the algorithms at first that studied heterogeneity is actually, I would say, platform independent, platform agnostic, we don’t use things that are done 20 years ago. They were done yesterday. And what we do is we look at each one of them individually and then we look for recurrent biological signals. So that’s the idea behind looking for true signals, because people always say, you go fishing, you may be getting junk out. Right?

Pek Lum: So let’s say, for example, we go to, the engine points to a spot in the sea, in the ocean, and five people go, then you’re always fishing out the same thing, the Blue Marlin, then you know that there is something there. So what we do is we take each data set, runs it through an engine and say these are the subtypes that I find. It does the same thing again in another data set and say these are the things that I find. And then it looks for recurrence signals, which is if you are a artifact that came from this one lab over here, or some kind of something that is unique to this other code over there, you can never find it to be recurrent. And that’s a very weird, systematic bias, you know, so so because of that, we are able to then very quickly, I would say, get the wheat and throw away the chaff. Right. And basically by just looking by the engine, looking at looking for recurring signals. So public data sets is like a a treasure trove for Auransa because we can use it.

Harry Glorikian: So you guys use your engine to I think you identified something unexpected, a correlation between plant-derived flavonoid compound and the heart. I think it was, you found that it helps mitigate toxic effects in a chemotherapy drug, you know. Can you say more about how the system figured that out, because that sounds not necessarily like a brand-new opportunity, but identifying something that works in a different way than what we thought originally.

Pek Lum: Right, exactly. So in our digital library, let me explain a little bit about that. We have collected probably close to half a million gene expression profiles. So it’s all RNA gene expression based, representing about 22,000 unique compounds. And these are things that we might generate ourselves or they are in the public domain. So any compound that has seen a live cell is fair game to our algorithms. So basically you put a compound, could be Merck’s compound, could be a tool compound, could be a natural compound, could be a compound from somewhere. And it’s put on a cell and gene expression was captured. And those are the profiles or the signatures that we gather. And then the idea is that, because remember, we have this part of the engine where we say we’re going to take the biology and study it and then we’re going to match it or we’re going to look for compounds or targets. When you knock it down, who’s gene expression actually goes the opposite way of the the disease. Now, this is a concept that is not new, right. In the sense that over 20 years ago, I think Rosetta probably was one of the first companies that say, look, if you have a compound that affects the living cell and it affects biology in a way that is the opposite of your disease, it’s a good thing. Right thing. So that’s the concept. But, you know, the idea then is to do this in such a way that you don’t have to test thousands of compounds.

Harry Glorikian: Right.

Pek Lum: That is accurate enough for you to test a handful. And that’s what we do. And by putting the heterogeneity concept together with this is something extremely novel and extremely important for the engine. And so with this kind of toxicity is actually an interesting story. We have a bunch of friends who are spun off a company from Stanford and they were building cardiomyocytes from IPS cells to print stem cells. And they wanted to do work with us, saying that why do we work together on a cool project? We were just starting out together and we thought about this project where it is a highly unmet medical need, even though chemotherapy works extremely well. Anthracyclines, it actually takes heart, takes a toll. There is toxicity and is it’s a known fact. And there’s only one drug in the market and a very old drug in the market today. And there is not much attention paid to this very critical aspect. So we thought we can marry the engine. At that time were starting up with oncology. We still we still are in oncology, and they were in cardiomyocytes. So we decided to tackle this extremely difficult biology where we say, what is a how does chemotherapy affect heart cells and what does the toxicity look like? So the engine took all kinds of data sets, heart failure data sets, its key stroke and cells that’s been treated with anthracyclines. So a ton of data and look for homogeneity and signals of the of the toxicity.

Pek Lum: So this is a little bit different from the disease biology, but it is studying toxicity. And we then ask the engine to find compounds that we have in our digital library, that says that what is the, I would say the biology of these compounds when they hit a living cell that goes the opposite way of the toxicity. And that’s how we found, actually we gave the company probably about seven, I forget, maybe seven to 10 compounds to test. The one thing that’s really great about our engine is that you don’t have to test thousands of compounds and it’s not a screen because you screened it in silico. And then it would choose a small number of compounds, usually not usually fewer than 30. And then we able to test and get at least a handful of those that are worth looking into and have what they call development legs. So this I would say this IPSC cardiomyocyte system is actually quite complex. You can imagine that to screen a drug that protects against, say, doxorubicin is going to be a pretty complicated screen that can probably very, very hard to do in a high throughput screen because you have to hit it with docs and then you have to hit it with the compounds you want to test and see whether it protects against a readout that is quite complex, like the beating heart.

Pek Lum: And so we give them about, I think, seven to 10 and actually four of them came out to be positive. Pretty amazing. Out of the four, one of them, the engine, noticed that it belonged to a family of other compounds that looked like it. So so that was really another hint for the the developers to say, oh, the developers I mean, drug developers to say, this is interesting. So we tested then a whole bunch of compounds that look like it. And then one of them became the lead compound that we actually licensed to a a pharma company in China to develop it for the Chinese market first. We still have the worldwide rights to that. So that’s how we tackled toxicity. And I think you might have read about another project with Genentech, actually, Roche. We have a poster together. And that is also the same idea, that if you can do that for cardio tox, perhaps you can do it for other kinds of toxicity. And one of them is actually GI tox, which is a very common toxicity. Some of them are rate limiting, you might have to pull a drug from clinical trials because there’s too much GI tox or it could be rate limiting to that. So we are tackling the idea that you can use to use machine, our engine, to create drugs for an adjuvant for a disease, a life-saving drug that otherwise could not be used properly, for example. So that’s kind of one way that we have to use the engine just starting from this little project that we did with the spin out, basically.

Harry Glorikian: So basically, you’re sort of, the engine is going in two directions. One is to identify new things, but one is to, I dare say, repurpose something for something that wasn’t expected or wasn’t known.

Pek Lum: That is right. Because it doesn’t really know. It doesn’t read papers and know is it’s a repurposed drug or something. You just put in it basically, you know, the gene expression profiles or patterns of all kinds of drugs. And then from there, as a company, we decided on two things. We want to be practical, right. And then we want to find novel things, things that, and it doesn’t matter where that comes from, as long as the drug could be used to do something novel or something that nobody has ever thought of or it could help save lives, we go for it. However, you know, we could find something. We were lucky to find something like this flavanol that has never been in humans before. So it still qualifies as an NCE, actually, and because it’s just a natural compound. So so in that sense, I would say maybe is not repurposing, but it’s repositioning. I don’t know from it being a natural compound to being something maybe useful for heart protection.

Now for our liver cancer compound, it is a total, totally brand-new compound. The initial compound that the engine found is actually a very, very old drug. But it was just a completely different thing and definitely not suitable for cancer patients the way it is delivered.

Harry Glorikian: This is the AU 409?

Pek Lum: Correct? Entirely new entity. New composition of matter. But the engine gave us the first lead, the first hit, and told us that we analyzed over a thousand liver tumors and probably over a thousand normal controls, found actually three subtypes, two of them the main subtypes and very interesting biology. And the engine predicted this compound that it thinks will work on both big subtypes. We thought this is interesting. But we look at the compound. You know, it’s been in humans. It’s been used. It’s an old drug. But it could never be given to a cancer patient. And so and so our team, our preclinical development team basically took that and say, can we actually make this into a cancer drug? So we evaluated that and thought, yes, we can. So we can basically, we analogged it. It becomes a new chemical. Now it’s water-soluble. We want to be given as a pill once a day for liver cancer patients. So so that’s how we kind of, as each of the drug programs move forward, we make a decision, the humans make a decision, after the lead us to that and say can we make it into a drug that can be given to patients?

Harry Glorikian: So where does that program stand now? I mean, where is it in its process or its in its lifecycle?

Pek Lum: Yeah, it’s actually we are GMP manufacturing right now. It’s already gone through a pre-IND meeting, so it’s very exciting for us and it’s got a superior toxicity profile. We think it’s very well tolerated, let’s put it that way. It could be very well tolerated. And it’s it’s at the the stage where we are in the GMP manufacturing phase, thinking about how to make that product and so on.

Harry Glorikian: So that that begs the question of do you see the company as a standalone pharma company? Do you see it as a drug discovery partner that that works with somebody else? I’m you know, it’s interesting because I’ve talked to other groups and they start out one place and then they they migrate someplace else. Right. Because they want the bigger opportunities. And so I’m wondering where you guys are.

Pek Lum: Yeah, we’ve always wanted to be, I say we describe ourselves as a technology company, deep tech company with the killer app. And the killer app is drug discovery and development especially. And we’ve always thought about our company as a platform company, and we were never shy about partnering with others from the get go. So with our O18 our team, which is a cardioprotection drug, we out-licensed that really early, and it’s found a home and now is being developed. And then we moved on to our liver cancer product, which we brought a little bit further. Now it’s in GMP manufacturing. And we’re actually looking for partners for that. And we have a prostate cancer compound in lead optimization that will probably pan out as well. So we see ourselves as being partners. Either we co-develop, or we out-license it and maybe one day, hopefully not too far in the future, we might bring one or two of our favorite ones into later stage clinical trials. But we are not shy about partnering at different stages. So we are going to be opportunistic because we really have a lot to offer. And also one thing that we’ve been talking to other partners, entrepreneurs, is that using our engine to form actually other companies, to really make sure the engine gets used and properly leveraged for other things that Auransa may not do because we just can’t do everything.

Harry Glorikian: No, that’s impossible. And the conversation I have with entrepreneurs all the time, yes, I know you can do it all, but can we just pick one thing and get it across the finish line? And it also dramatically changes valuation, being able to get what I have people that tell me, you know, one of these days I have to see one of these A.I. systems get something out. And I always tell them, like, if you wait that long, you’ll be too late.

Harry Glorikian: So here’s an interesting question, though. And jumping back to almost the beginning. The company was named Capella. And you change the name to Auransa.

Pek Lum: That’s right.

Harry Glorikian: And so what’s the story behind that? Gosh, you know.

Harry Glorikian: When somebody woke up one morning and said, I don’t like that name.

Pek Lum: It’s actually pretty funny. So we so we like to go to the Palo Alto foothills and watch the stars with the kids. And then one day we saw Capella. From afar, you look at it, it’s actually one star. You look at closer, it’s two stars. Then closer, it’s four stars. It’s pretty remarkable. And I thought, OK, we should name it Capella Biosciences. Thinking we are the only ones on the planet that are named. So we got Capella Biosciences and then probably, we never actually had a website yet. So we were just kind of chugging along early days and then we realized that there was a Capella Bioscience across the pond in the U.K. We said what? How can somebody be named Capella Bioscience without an S? So I actually called up the company and said, “Hey, we are like your twin across the pond. We’re doing something a little different, actually completely different. But you are Capella Bioscience and I am Capella Biosciences. What should we do?” And they’re like, “Well, we like the name.” We’re like, “Well, we like it too.” So we kind of waited for a while. And but in the meantime, I started to think about a new name in case we need to change it. And then we realized that one day we were trying to buy a table, one of those cool tables that you can use as a ping

Please rate and review MoneyBall Medicine on Apple Podcasts! Here’s how to do that from an iPhone, iPad, or iPod touch:

• Launch the “Podcasts” app on your device. If you can’t find this app, swipe all the way to the left on your home screen until you’re on the Search page. Tap the search field at the top and type in “Podcasts.” Apple’s Podcasts app should show up in the search results.

• Tap the Podcasts app icon, and after it opens, tap the Search field at the top, or the little magnifying glass icon in the lower right corner.

• Type MoneyBall Medicine into the search field and press the Search button.

• In the search results, click on the MoneyBall Medicine logo.

• On the next page, scroll down until you see the Ratings & Reviews section. Below that, you’ll see five purple stars.

• Tap the stars to rate the show.

• Scroll down a little farther. You’ll see a purple link saying “Write a Review.”

• On the next screen, you’ll see the stars again. You can tap them to leave a rating if you haven’t already.

• In the Title field, type a summary for your review.

• In the Review field, type your review.

• When you’re finished, click Send.

• That’s it, you’re done. Thanks!