Lantern Pharma Inc. (LTRN) Earnings Call Transcript & Summary

As I was saying before, we often work with a treatment-naive condition, which is basically the molecular data that's taken before treatment starts and -- that's obviously the rational way to go about this. A lot of people will use various types of molecular data after treatment for a different analysis. But if we have it before treatment, we can actually know and make the decision at the right time in an actual clinical setting and that goes for whether we're using the model or whether we're just using a biomarker drive from the model. So as far as our approach goes, we usually -- we do some standard things, but the most important thing that we usually do is future engineering. When we work with biological data, it is basically different than every type of data set that you will be trending like in the machine learning course somewhere. They usually have what's called like a long data set, do you have like a few types of features, so you don't really need to worry about selecting features too much or future engineering in many cases. And you have maybe millions of samples or tens of thousands of samples, but in biology, that's very expensive and you have very many features. So it's called like a wide data set. So we had to do feature engineering, Otherwise, we have a problem of overfitting, which means we get excited because we think we have like 100% accuracy, but it doesn't translate outside of whatever we're training on. So to evaluate this, we usually -- we'll just make a training and test it. So we do -- in the training set, you can have a separate validation set to sort of tweak the parameters of the model and also decide like how to do your future engineering. And we use machine learning metrics to determine which features to include. We use something called the [indiscernible] method, which is very common. It was developed in economics -- basically, you can think of it as a way to determine like what is the contribution of a feature in the context of an uncontrolled comparison? Like you could think of this like in an analogy to support like what is the -- for like player A, like what is his contribution. And so we do this with all of the features and then we start to sort of drop them. There's 2 reasons to do it. One is to avoid overfitting another is -- it makes the model more interpretable. So if we have to like go through like 50 features that we're involved in, why a model came to some conclusion, it's going to be very convoluted and nobody can really explain that. So part of our approach is feature engineering. And I think that's -- I would say this is one of our specialties. This is something that is not done in machine learning generally. And -- we basically can do 2 things from that. One is we extract biomarkers, as I have been saying, and also, we just make the models better when we give it a better pool of features to work with. So when I first started to work on this data, there was initially less accuracy on the clinical data that I was getting for past models that we're using RNA and I mentioned we did a future engineering approach to sort of resolve that problem. But whatever is going through that, I was testing a lot of different things to try to come to a solution. So I ended up making a tool that I would call a brute-force tool. So basically, it ties all of the different parameters that randomly selects like which features to include. And this basically gives an unbiased way to go through the training data and come up with the features and final models. And it also has an interesting property, which was just came about by chance. It wasn't part of our approach before, but -- and that is basically when you make thousands of models and you have a lot of different feature evaluations, you can get an average future evaluation, which turns out to be very important because -- it's basically just an estimation, which can have some variance, if you look at a single model. So sometimes, we may overestimate the importance of various features if we're just looking at a single model and there's also a lot that are sort of missing because in the process of trying to only use a few features in a single model, you end up with a small set, and you can't do much pattern analysis in that situation. So it turned out to be a very helpful thing that was useful for a general machine learning process in biological data. And yes, with that, we make models after we get them, we try to interpret them see if it makes sense and do things like predictions on public data. As I was saying, I was making this process that would randomly sort of select things and find the optimal sort of way to get to the result by brute-force. But in practice, if you have something that is negatively associated, it may be an exclusion criteria, but it's not as useful as an inclusion criteria, Like, if you have something that's saying like you're going to respond, you will be much better off than saying like you probably won't respond. It still has like a lot more uncertainty, even though we can get good accuracy with the model if we want to sort of find some signature of drug response, it's much more valuable to look at positive ones. So following some suggestions by Andrew, what we were doing was, first, getting all of the positive features and then using those as the future pool to do -- and then we would do feature selection and model building. And one of the interesting properties from that was -- well, first, it worked. It was not quite as good because there was like less features available but it was still very effective. And like you had over 80% accuracy in our asset. And you had the interesting property of not having false positive. So that means if it is predicting positive, you have a lot more certainty, if this is real. So you may have like less accuracy overall, but you have more confidence in your positive predictions. And that's a good thing in many cases.

So there's -- I think the important thing that Joseph said is a drug signature versus a tumor signature. And a lot of the negative features are often tumor signatures. So when you have mutations in RAS or amplifications in [indiscernible] HER2 or [indiscernible] mutations, those are negative features that are tumor signatures in general, any tumor that as those is going to [indiscernible] I think -- nothing to do with our drug, whether they benefit from it or not, those patients would not do well. And pretty much any drug you test for patients with those mutations will generally not do well. That's changing now that people are taking drugs specifically for those mutations. But before it was just a general poor prognostic marker. You look at it the other way and you find mutations where now you give your drug, you take a patient that would have done poorly and flip them and not do well, that's a drug signature. And that's really what we are interested in trying to find out if that allows us to then enrich for patients that will do well. And [indiscernible] patients do poorly without knowing anything about your drug, but you can't include patients that will do well unless you actually [indiscernible] drug.

Yes. So this has been an exciting collaboration. I think it was really exciting to be able to work with clinical data then we actually had a good team. They had a lot of collaborators that were like very intelligent and informed and some people that have like Dan Bilodeau at the Mayo Clinic and his group, and there's several groups at Mayo Clinic at [ Brown ] and various collaborators. So it was a really interesting group that like personally was nice to get information from. And yes, getting feedback on the models like we were talking about the positive models, like that was just something I never thought of before. And when we were working with genomics data, this is definitely a data type that we want to work with. We went to AACR and ASCO together. And when I was at ASCO, I actually saw a lot of posters that are like how do we use mutation data to like sort of predict results of the clinical trial. And you can see some of them are saying the same thing that I was talking about mutation data is sparse. So there's no solution for this yet. So just kind of put us at the cutting edge of this problem. And I think by finding a solution that worked we're able to work with a lot of other data we could work with, like this will come up in many areas. It could work in liquid biopsy data and you can get -- instead of just like looking at 1 or 2 mutations of predefined interest, you can just look at this in an objective way that's unbiased and high throughput. So it turned into something valuable for tool development. And when I was trying to sort of find something that worked better for this, that was when I sort of made a platform for automated machine learning and all you have to do is put in like your outcome data and then whatever your features are. And then it will work automatically overnight. And this became basically my go-to tool that I've been using ever since. So it wasn't planned that way, but it was just a side benefit to and from the perspective of RADR, it's good to have like some more clinical experience with an ongoing trial instead of just like something that has been -- maybe some data from some common public sources like TCGA. There is some good public data out there, but it's obviously very valuable to like work with something that's currently in the clinic in Phase II. Yes. So when we were working with this mutation data, like we had a panel of about 400 genes, and this is sort of a common thing you can find something that has like 1,500 and so we can work with that type of data, if you have that and you have the more samples, the better. But if you have a reasonable amount of samples, and we can make a good model in I think we can always find the best position for a drug. It does help if the drug is good. So yes, I should mention that one of the good things working with Actuate is when we were able to start to divide this into a training data, and we wanted to say like this group is responders, this group is nonresponders. We could roughly balance the data set. But for many drugs that we're going through a Phase I, you probably have very imbalanced data set that's mostly nonresponders, and that will be very troublesome to work with. But for many drugs that have -- that aren't in that situation, whether you're working on a preclinical or clinical stage we can basically work with any type of data. And sometimes this can be data that is publicly available to use some well-known cell lines, we can pull that from some coming data sets. So as long as you have some sort of sensitivity or outcome measurement, we can make a model for you and we can identify biomarkers and then try to see where is the best place to put the drug in what would be the best combinations and things like that.

So our main focus now has been 2 programs. The main trial, the one that's most advanced is the development of elraglusib patients with metastatic pancreatic cancer. I refer to that as the 1801 3B study. That study is now completely approved. And we've publicly disclosed an analysis of interim data back in April. We went public, as you said, and that was in our S-1 and in our public information. It's also available on our website in our corporate deck. And we have promised another analysis of the interim data, that will be the November data, but it will be available in December. And that will occur at approximately 70% of the patients in the control group, which [indiscernible] because the endpoint of the study is survival. Early next year, we anticipate having top line data. And so the trial will read out and we have outcomes data on that full patient population, which has a control arm of just the standard of care for first-line treatment is gemcitabine and Abraxane and experimental [ elraglusib ] plus gemcitabine and Abraxane. So very excited to see that data preliminarily that was very exciting and the data was in favor of the experimental arm, but it remains to be seen what we see [indiscernible]. We also observed that in a pediatric indication in Ewing sarcoma, we had a number of surprising complete responses. So our drug was able to eliminate tumors in refractory resistant patients. So these are kids that essentially had failed conventional treatment. The tumors that come back and at least few of those patients, we completely put them into remission. And some other patients had partial responses. A number of these are durable and these patients are still alive. Ewing is a very rare -- fortunately, it's a pediatric and it's rare type of cancer indication, but it's something we're not pursuing using the various regulatory mechanisms that help develop drugs for rare diseases and rare pediatric indications. Now going forward, that's where we really see an opportunity to apply the RADR data. We are interested both based on our Phase I and based on the modeling in pursuing different indications, including metastatic melanoma, non-small cell lung cancer, possibly colorectal cancer. And in those indications, we have the ability to potentially apply some of the learnings from the RADR analysis, the RADR experimental approaches to that trial. How would we do that? So it's too early right now to just jump in and patient select based on that data has to be validated. But what we can do is we can actually put patients in like we normally would, but then stratify them for analysis based on how RADR would have stratified them. And we'll see for the analysis whether using RADR stratification will give us a better outcome than just using the normal type of analysis where we just include all comers. And so that will be a way how we envision actually be able to validate the data. If it showed promise, we would then apply that to our pivotal trials and then talk with FDA to talk about how we actually use that commercial space. I think that will be no problem for all of us to have [indiscernible] division. I should mention that new indications are dependent on additional funding. We are not currently pursuing those. We're really focused on pancreatic and Ewing but hopefully, with future financings, we'll have the money to [indiscernible]. So we have website actuatetherapeutics.com. It's a public company. So public company website where we post our posters, all our trials, all our manuscripts and abstracts on the website. It also has contact information. So just any questions that anybody has or interest in current or future trials, please reach out, we'll always respond. And we have a lot of people reaching out to have interest in the program, and we're happy to talk to you about it.

On our website, you can see links to posters for our collaboration with Actuate or ASCO and AACR posters. And I think soon you will be able to look forward to a manuscript that we'll be looking to publish. And if you have interest in using RADR, please reach out to us.

All right, thanks for watching everybody sticking through the whole time, and thanks also to anybody watching in the future. We're going to pull up some questions. Just give us a moment. So the first question, do you anticipate that tumors will develop resistance mechanisms against GSK-3 inhibitors? And if so, what strategies do you think you think could be effective in identifying and overcoming these challenges to maintain long-term efficacy and treatment? So, so far, we only have data from the beginning of treatment. So if we do have longitudinal data, we'll be able to look at things like resistance mechanisms and we can apply our model directly on that type of data. If those are coming from different mechanisms than the resistance that we see for just baseline resistance -- so longitudinal data would help that as for whether we will anticipate that happening probably will happen to every drug to some degree, hopefully, at a lesser degree but it's inevitable to happen sometimes. And there should be some evolution of mechanisms that are different. So hopefully, they're not too far off from what we have identified now. But once that does happen, I actually think identify and resistance mechanisms is one of our specialties. So when we do the machine learning modeling to find what is like a good biomarker, we also find things that are negatively opposed to the drug. And we can do this for the drug that we're looking at as well as some other drugs that are sort of publicly known. And we can see like what drug is going to attack the resistance mechanisms that we identify. So once we do see that occurring, if we've collected data beforehand, we can actually see what is sort of -- we can build a model for that specifically and then identify the right drug combination for that situation. Next question. Okay. So what insights can we gain from the current collaboration that could inform positioning of other similar compounds? So if you're talking about other like similar compounds like another small molecule GSK-3 inhibitor, one thing I would say is we have actually seen like the Lilly compound has done an NCI 60 screen in public data. So we've actually done this directly, and it's actually very different. So you can actually get some insights that are generalizable by type like you can make a lot of generalizations, I'd say, like an EGFR inhibitor, comparing one to another one for these drugs where there's a lot of variety. But each drug is usually unique to a substantial degree. So we will probably see quite a lot of unique patterns that would be coming out if we did modeling. So to address those, we would actually just want to have data for those different molecules. And if we have data, I think we can work with almost any molecule out there. Okay. Does Lantern have other radar collaborators. Can you tell us more about the possible areas that RADR can be helpful? Okay. So yes, at the moment, we're collaborating with interesting company called Oregon Therapeutics, they have a very interesting drug, which we have never seen before, called a PDI inhibitor. You can see more about that at our website or at Oregon Therapeutics' website. So they have given us a lot of preclinical data to work with. And in this case, we're trying to find like we'll make a model from that, we'll get biomarkers and look for good indications and possible drug combinations. So we can do the same thing for anything where there is data. It doesn't necessarily have to be in drug development, but we can work in those areas very well, and we can work with just about any type of data that we -- we've tried about every type of molecular data and all of them were able to work well in. So other areas where RADR may be helpful? So I'll just mention we have some -- we have some other areas that are sort of not covered in this talk, but we do things like blood-brain barrier permeability test, and those are based off of like the chemical structure of small molecules. So this is one of the most well-known areas of computer-assisted drug development where you use something like a smile string to predict. Will this react in some way? We did that to predict blood-brain barrier permeability, and we have a very good result in that, but it could also be applied to other problems that could be addressed such as target binding inhibition, things like that. So we can work with most areas of drug development. Any other questions. Well, thank you. And please reach out if you have any interests.