Things Have Changed
Things Have Changed
AI in Healthcare: Uncovering Future Breakthroughs with DeepMirror and Dr. Max Jakobs
Imagine compressing a decade of research into mere weeks or turning a failure rate of 90% in clinical drug development on its head. That's exactly what's on the horizon with the incorporation of artificial intelligence (AI) in the realm of drug discovery, and the person leading this charge is none other than Dr. Max Jakobs, CEO and co-founder of DeepMirror. Prepare to be blown away as we navigate the potential AI holds for revolutionizing healthcare, making it possible to sift through massive datasets with unprecedented speed and accuracy.
In our exciting conversation with Dr. Jakobs, we dig into the gritty realities of drug discovery in its current state and the formidable issues researchers face without the aid of AI algorithms. Get ready to discover how AI not only hastens the drug-to-market journey but also boosts the clinical hit rates, saving ample time and a fortune in the process. Plus, we probe into the daunting challenge of FDA approval and how DeepMirror is working to dismantle the bias inherent in traditional drug development.
To conclude, we delve into the groundbreaking advancements in small molecule drug discovery and how cloud-based AI is honing laboratory processes. Dr. Jakobs illustrates how DeepMirror is utilizing AI to unearth the most promising drug candidates more efficiently. We bring to light riveting customer success stories and the promising future that lies ahead with AI-based data analysis in the pharmaceutical industry. Brace yourselves for a riveting exploration of AI's impact on drug discovery and healthcare.
Actually 90% of all drugs fail. So basically you're spending five years and then you have another five years and 90% of the time will fail.
Speaker 2:In a world where 90% of clinical drug development falters, companies invest a decade for the slim chance of success. We dive into this high stakes journey with Dr Max Jacobs, CEO and co-founder of DeepMirror.
Speaker 1:Because, on one end, you can use machine learning, artificial intelligence, to take your past data on drug compounds that you tested against the target and then basically try to predict other compounds that might even be more potent.
Speaker 2:AI is compressing years of research into weeks, tackling vast datasets and promising to accelerate pre-clinical work, but the challenge is steep. Studying clinical outcomes remains a formidable task.
Speaker 1:Individual drug. You have to first make it, test it, and so on. It might take two, three weeks or so. So testing tens of thousands is really not possible.
Speaker 2:The goal, Level the playing field, increase competition and challenge norms. With technologies from the 1960s still prevalent, the industry is ripe for disruption.
Speaker 3:The pacemaker has remained the same from 1960 to now. Largely, it seems like there is a certain level of bias that sticks around that is very hard to think completely out of the box.
Speaker 2:From small molecules to advanced therapies. We're at the cusp of an explosion of ideas and innovation.
Speaker 1:The biggest challenge for any software company in this place has always been to resist digging for gold almost.
Speaker 2:Join us on Things of Change podcast for a deep dive into the future of drug discovery and healthcare innovation. If you had known how important the technology economy was 20 years ago, would you have done things differently? The internet, cell phones, the cloud and data Things have changed. We're here to talk about it, hi.
Speaker 3:I'm Jed, hi, I'm Shikhar. Welcome to Things of Change, your new economics and technology podcast. Healthcare in general is hard. Biotech, healthcare, life sciences industries are one of those industries where there's just a massive amount of data. These are far too long Just because experimentation. It costs a lot and you need to be precise because ultimately, the end users are the patients and consumers are there.
Speaker 3:Having technology come in and improve and influence this level of data that researchers, scientists and companies biotech companies, small and large operate with is a big boon. Your algorithms could be like your buddy that could just help uncover some patterns that you were not seeing before, or even identify potential drug targets. It seems like healthcare in general is one of those industries that could massively be influenced by the wave of AI that we're seeing today. Today, we're super excited to have Dr Max Jacobs, co-founder and CEO of DeepMirror, a health tech company building software, ai software platform that helps researchers accelerate experimentation testing and, by doing so, unlock creativity. Think of chat, gpt or GitHub co-pilot, but co-pilot for your research, like research GPT that names you to be trademarked.
Speaker 1:Thank you for having me. We're excited to be here, guys.
Speaker 3:I touched on it a bit, but what are your thoughts as to why this is needed in healthcare in general?
Speaker 1:Yeah, I think it's very important to maybe appreciate that in relationship to other industries in health tech and biotech and pharma these data-driven approaches are still very experimental. Some of the larger companies might have some internal research teams already that do this kind of work, basically providing algorithmic solutions to other parts of the companies, but it's really not yet in the hand of end users really so these kind of data-driven technologies.
Speaker 1:It is really at a stage now where people maybe until a few years ago didn't really believe that there could be some sort of vertical solution in the industry, because everything is, as you said, so complex, so difficult and so specialized to the particular application that nothing can really help everyone in a way.
Speaker 1:But now we're seeing more and more approaches and attempts to actually do this, to bring these kind of technologies that look maybe very difficult into the hands of people. I think you mentioned chatGPT. That has been demonstrated by especially these kind of approaches. Prior to that, people worked with large language models chatGPT is based on, but these were technical people. It didn't really get to any sort of end users, but this kind of leap of design almost made it possible for the first time. This is what we're trying to achieve as well, like having a little leap of design and user experience to finally bring AI-powered algorithms to all these labs that are not really using them at this stage because, de facto, it's in the hand of a select few companies in some way that have successfully applied it. But apart from that, most companies are like basically thinking oh, maybe I should get in on it, but I'm not sure how yet.
Speaker 2:Yeah, one of the things we were really curious about and you talked about it a little bit was how does current state look? What are the kind of issues that researchers are facing today with regards to coming up with drug discovery, and all these areas are really important for the farm and biotech industry space. Can you point out some really high-level stuff that's really difficult for researchers today to achieve without this technology that we're talking about?
Speaker 1:So it is basically around the idea of using your past data to try to somehow make better decisions, right? So at the moment, it takes around 10 years or so, and maybe a billion or two billion dollars to take a drug to market. Wow, and is that the average? That's the average. Yeah, oh my.
Speaker 3:God, it's not a.
Speaker 1:Gaussian distribution, necessarily, but it's a lot of money and a lot of time, and this is because there's so many things you have to do right. Drug discovery goes from initial research of trying to figure out what can I attack to treat cancer, to then trying to find agents like, for example, small drugs that might attack this target and might diminish its activity, then trying to make sure that these agents are very potent, that they are safe for people, that they can be ingested by people, to then finally putting them in the clinic and by that point you're already like five years down, and then you have another five years of clinical work and then this clinical time actually 90% of all drugs fail. So basically you're spending five years and then you have another five years and 90% of the time will fail. But this kind of influences the thinking of a company is quite dramatically right. And then anything that reduces the time to get a gold shot, which basically means shooting into the clinic, right. So actually trying clinic once or improves the outcome in the clinic is super, super valuable.
Speaker 1:And this is what it converges to right, because on one end, you can use machine learning and artificial intelligence to, for example, take your past data on drug compounds that you tested against the target or something like this, and then basically try to predict other compounds that might even be more potent, instead of trying all these compounds. Right, so that obviously speeds you up because you wouldn't have to do so many experiments, because testing sometimes an individual drug, you have to first make it, test it, and so on. It might take two, three weeks or so. So testing tens of thousands is really not possible, whereas you can maybe test a thousand over a few years, and then you really have to be smart which ones you test. That's one way in which these algorithms can help.
Speaker 1:The other way is, of course, once they go into the clinic, you have very little control, right? You do not know what is going to happen. However, many companies already generated some data on clinical outcomes of drug compound because they have tested quite a few over the last hundred years. Right, see, and this is another way you can actually improve this you can try to predict what would happen in the clinic given a certain drug, but there would really have very small data and also, in the other case, these data set are very small. You really have to leverage a lot of tricks from very nice machine learning technologies to make these things work and actually pick up what happens in the clinic is much harder than just trying to predict stuff that happens before the clinic. So I think we're now this day we can try to accelerate the work before the clinic or preclinical drug discovery work. What the stage of trying to improve clinical hit rates.
Speaker 3:That's still very difficult and the companies that were successful also were more towards the preclinical work there's a direct link between them spending all this time, all this money and still having just At best ten percent chance of getting a home run. All that is still linked to higher drug prices because they need to recoup those cost that they put for ten years and on top of actually reducing costs for farmer.
Speaker 1:That itself wouldn't really help because it could still charge whatever they want. But if you increase competition between companies as well by actually enabling everyone to get to the same level of capability, that's also key for this and in a way, that is also the thing we're trying to achieve a reliving the plane of these kind of technologies.
Speaker 3:A large part of what's in the medical device industry, farm industry and biotech are Improvements or iterations from what was created in the nineteen sixties, of the seventies. And you smile over there. I was so surprised because I remember my first day at walking with an abit and they were like, okay, this is not like rocket science. The pacemaker has remained the same from nineteen sixty to now, largely. But I'd like to get your take on that statement that I made. It might not be completely true, but it seems like there is a certain level of bias that sticks around that is very hard to think Completely out of the box, because this is tried and tested and has been used with the public. You have public health data for forty, fifty years. So you're like, okay, you don't want to deviate from that normal.
Speaker 1:It's difficult to deviate from the normal if you have to, in the end, get fda approval, and it's easier Get approval for something that is similar to something that is known. But that's, of course, only one Issue. I think this ties very well into what we're mainly doing right now, which is so, first way people build drugs. So what's more? Molecules, which is like tiny molecules that easily go into cells. They have been used.
Speaker 1:There were the first drug that have been used for the years and farm company is developed quite a few of them over the last century or so, and initially these were found mainly from natural analog. So people are just looking at plants trying to figure out what the ingredient was that you, just you are the certain disease of cured headache. Then people start deviating from that to be to try to find analogs of these like slight changes. So you can think of this almost. As this is huge desert of molecules, most of them are toxic to people. There's a few islands in there that have promising molecules, but the desert in between is big and vast, so finding these islands is extremely difficult. That's why you Tend to often go from natural analogs right then that was the status quo.
Speaker 1:Then in the 80s and 90s people started doing more with high content screening, so they would just generate millions of different compounds and just screen them over some sort of target that they wanted to attack. That of course gave many more results, but often these compounds in the end tend not to be great in the clinic Because the experiments that people did on these compounds were actually not very predictive of what happened later on in the clinic. So it was a bit more difficult. And then, especially now after covid, we see quite a lot of development, in particular people called advanced therapies, which are like therapies based not on small molecules but based on RNA, cell therapy, antibody therapy, peptide therapy and so on. So suddenly this whole zoo of new options that pops up and the whole space became way more complicated.
Speaker 1:And also people are going back to the small molecules again and realize, oh shit, the ones that we actually work with we're very similar with each other, like all the different ones that pass through the FDA, but often very slight modifications just of the same molecule. And then a company such as in silico, for example, come around the corner and start designing molecules that To see. If this weird because they haven't really been done before, but clearly they work. So there is some, in some ways, you can harness, like these generative approaches to come up with new ideas, new solutions to all the problems, and it's a bit like a An explosion of ideas that's happening right now, of different viewpoints you can apply to all solutions again.
Speaker 3:So it's extremely exciting, can you provide customer success stories or something where you are able to just take some sample data and was able to Improve an influence there.
Speaker 1:research we did a lot of consulting work prior to actually embarking on product development, and some of the consulting work, for example, was image analysis, where we essentially help people interpret microscopy data In early drug R&D. And there it's straightforward because typically it takes people a few minutes to analyze an image where it takes an AI working a few seconds. So you're really talking about For the fifties speed up here, so that's very straightforward. In in other ways, we did some work on optimizing molecules, as I mentioned, and there, as you can see, with some of the bigger companies, you can speed these things up by what five to ten X like you can find the most optimal drug candidate about five to ten times faster if you use some sort of AI powers decision making and In turn and this is also where we see our initial product now, which we have been developing last year, because people have been working with small molecule in the dawn of time, right, and this is also the modality that most people have experience with.
Speaker 1:So we now pushing out or initial product exclusively for small molecule drug discovery, where we essentially take customer data. Let's say somebody tested twenty, fifty compounds in the lab, then we would they upload the data to our app and then they can predict the same properties that they measured for these compounds for ten, twenty hundred thousand other compounds. Then prioritize the next step, take that back to the lab, get the results back into the app, then close the design, may, test and analyze cycle and this is really what we're doing. So it's this idea of laboratory optimization, by reducing the amount of work you have to do and suggesting the work you should do next, and there we did a few case studies, which you can also find the no web page, where we think it can speed up by about four, x and it's current state. But we're still working on some benefits there as well.
Speaker 1:Yes, really exciting, because you go, you design a compound in the app which is similar to your compounds. You potentially make it, then you test it, then you go back again and design the next one based on the information you have before. Yeah, instead of like sometimes stabbing a bit in the dark, you can really use this to guide your thinking and also, sometimes something might come up which you would not have thought about. We have molecule that maybe you have never really synthesized before, but make sense and then you might find something which is maybe a new way, says this past desert.
Speaker 3:Very interesting. So the way you're describing it, it's like an application that researchers can basically log onto on a browser, get into the are, into your portal and input their data set. And is it something that they get right away? Or is it one of those things where, okay, they input it and then they head back and then you're doing the post processing for a day or two and then they come back, see all okay, what these are the potential outcomes, and then test it out. Is that the workflow that you're seeing?
Speaker 1:It has to be faster than that because otherwise, okay, becomes more of a consulting business again and we really want to make these inside available like almost instantly, because in the end people do a lot right. They might try this and they might try this. They want to play a bit around on the app. So, yes, it's a cloud, simple clouds, vertical SAS app you would call it, and we initially when an interface with which people log in the upload a bit of data that they click a button, predicts or other data that they might have provided themselves, what generated using generate a guy on our platform and then it should take the pain. On the data set size, anything between four, what's a two minutes to a few hours for the really big ones. But then nobody in early drug discovery has big data sets because you would start with Maybe tens of compounds, that a couple hundred compounds. Once you have a thousand compounds, you either failed or one in the clinic already, but that you don't need a I am anymore because you already had your shot.
Speaker 3:That's really cool the people who are ultimately their researchers.
Speaker 1:And on top of the game, they come into your platform knowing maybe one of these hundred might work, so they've already narrowed the problem space and then finding the oasis in that Problem space is a lot quicker and probably gets more signal yes, so basically, initially it's just about finding something that might what we just call the hit, which is like maybe one, two, three compounds, so, and then, once you have a hit, try to generate a few compounds that are like in this molecule space, to just have a look, and then at that point we, like already have narrowed into the space a little bit. I really want to look into that region, which could still be massive right, and then generate a few in that area. Then it's really where I was, so where would come in?
Speaker 1:because it would basically take the initial Results and then go from that to generate new suggestions. That might make a lot of sense, because going really from zero to something here Is it would be really interesting. But in terms of what ML AI is capable these days it's really not that straightforward because there's basically infinite molecules you can make right and then trying to do that. We have a grant now pending. That tries to do something like this is a bit more, let's say, high risk, so let's see if it works out.
Speaker 3:But yeah, maybe we'll know next year's like a baseline, and then you walk through it and you paper it. Now you can only do a.
Speaker 1:I am that's awesome once you know something and where you get that something from is really key.
Speaker 2:Does it get better as you onboard more customers onto the platform, or are these companies like super touchy about their data sets that you can never use everybody else's data sets to establish baselines for new customers?
Speaker 1:Let's depend.
Speaker 1:some customers are really interested in that, because particularly academic customers are very interested in making data more widely available so that sometimes even ask Whether this is an, and of course it can be done because our database can distinguish between, say, hi Lee, it's the tiny secret, just part of the database where very happy, happy and I could never be used for anything but serving to a customer. But other customers can also. We call centers Multiplayer. They can engage in multiplayer mode, which basically means that they provide some data and then that data might be used For other people as well, which is not so interesting when you try to make drugs against the particular target, because it really tends to work on different things anyways.
Speaker 1:But it becomes really interesting for the things I mentioned before the idea of trying to predict what happens in the clinic, because if you have, if you have a few compounds and you know whether these metabolize well or whether these are toxic, group that with many other drugs of which you know how they metabolize or how toxic they are, that's really cool and there's not that much data on that right now. If people are willing to share that, that's something people often willing to engage with. It also helps the problem, because typically people do not want to share the structure of a compound because that's IP, because as soon as that becomes public knowledge, it can be patented anymore. But if you just say, all we can use that model to build better models for everyone, but your structure will never be retrieval from that model, that's really something some people are interested in so, max, the biggest news, at least in the US, is ozampic and those weight loss drugs.
Speaker 3:It's like the biggest news. Everyone's always hang on, so now I can. It's a really cool thing because you lose the weight and then it's easier to control it then. But it's interesting that it stems from diabetes drugs that we've used all this one. When I was reading through your blogs and the papers, it just feels hey, hang on. Something like this where you could expand what a drug could do or just think of it as a new way, like a new paradigm through an AI, would be a lot faster than 60 years of trial over the world's population. So that seems like something that we might just start uncovering, where we use traditional drugs and running the co-pilot and we're like hang on, this could also do this and go from there.
Speaker 1:Yeah, Things like we are doing is, let's say, you have a bit of information, initially like again 20, 50 compounds and so which are completely new and have never been done before, but they're not really that potent. They're quite toxic, so they're quite annoying, but you can use the information from them to basically go and screen public databases. There's millions of compounds out there that you can just buy, so you can then basically use this information to try to repurpose one of them. For example, I don't know, these 50 compounds are really good against killing this particular cancer, but they are also doing all kinds of other bad stuff. So maybe I use this information from these compounds to then screen other compounds which then flag up and maybe these compounds have never been used against this disease, and then you can start just repurposing them because they're already FDA proof. So that's pretty cool and history is littered with these kind of, let's say, super versatile compounds, right?
Speaker 1:Doxycycline is one of them. I tend to discover a new application of that every month by chance. Then Viagra is actually one of the most famous ones, because Viagra was initially. I can't remember what the clinical trial was about Hard, hard.
Speaker 2:It was the hard.
Speaker 3:Yeah.
Speaker 1:There's also so many drugs that kind of failed clinical trials, but they were already shown to be safe for patients and we don't even know if they can do anything yet. So that's also then really sad and often. These then sometimes become available and you can maybe ask a farmer company if you can look into them on top of this vast chemical space and say we haven't yet explored. We still don't really fully understand how, first of all, some of the drugs we use actually work and then, second of all, how we can actually repurpose them to do even completely different things, right, because we already know that they're safe. Right, so they can do so many things potentially. It was really exciting. We have medicine. We only scratch the surface until now, and now we like starting developing all these new therapies, all these different approaches. Cell and gene therapies are really awesome, especially for like cancer.
Speaker 3:Yeah, yeah, it's super exciting.
Speaker 2:One of the things we wanted to ask was as you envision the future of deep mirror and the problems that you're solving today. You look five, 10 years down the line. How do you see that? What are the challenges that you foresee right now that you're solving, and how you'll get to the future state that you want?
Speaker 1:The biggest challenge for any, let's say, software company in this space has always been to resist digging for gold, almost. Because once you have something that kind of works, everybody will try to tell you oh, now you have to go all the end to the patient. But then you tend to put all your money onto a single component at some point and then again 90% failed and then you might have your cool platform software which sped up your preclinical work. Let's say it's five, x or something like that. But that still doesn't necessarily mean that afterwards you will have an higher success rate necessarily.
Speaker 1:But people still think in this old school biotech way where you basically go and okay, so you play around a bit, you find something, and then you put all your hosts and then you become a proper biotech company that then gets bought by a farmer or your partner with some farmer companies and so on, and staying clear of that like trying to resist this digging and really trying to stay a service provider and delivering value to everyone. That will be quite tricky, but we see other companies in this space that kind of manage to become a fully vertical Sastau solution. Benchling is maybe one of the most famous one. Nobody thought that something like Benchling would work when they came around because effectively it was like a workflow solution for researchers, and why would researchers pay for that? But then over time people realized researchers go to farmer companies, have big pockets and suddenly Benchling has contracts that are more than a million and annual revenue.
Speaker 3:You said something that is so interesting. I see the promise of this technology. I'm thinking, oh, massive amounts of data. You put AI there. We will be able to uncover certain patterns. But you mentioned something so interesting where sometimes it might just be helpful to get a cleaner set of data, a smaller set of data, and then expand from there. So that is really interesting I have not thought of that before and how AI can help with.
Speaker 1:Sometimes the less is more You're also defining the question very well for that particular data set. Like in our early days, we once got contacted by a company that essentially asked us oh, we have these millions of images here. Can we pipe those through an algorithm and learn something? And then again, initially you think, OK, maybe, but then you look into this and then, as long as you don't really ask a proper question about the data, you just end up with nothing, Like if you don't ask a question you don't know.
Speaker 1:you won't find an answer because everything will be statistically significant, and then you might be amplifying noise as well, because it's a big data set and you can't just have a look at it, and you might not learn anything.
Speaker 1:It's like this whole idea of looking at each grain of sand on a beach and then trying to understand what a beach is. It doesn't really help. You have to have some sort of question about the data you might analyze. There's something similar happening in academia now, where people were thinking a lot about something called spatial transcriptomics, which is basically the idea of taking images of tissues and then also looking at genetic information in these images and basically generating not even terabytes but petabytes of data on this. And still they don't really know how to apply this, because just by looking and collecting data without narrowing the question, we don't know. This is where we come in, right that people still are very much required to define a question, define the hypothesis and then help get a machine to help with the number crunching. Right Can't replace it yet.
Speaker 3:Yeah, max this was really great. Before you leave us, we always like to give our guests the mic. You already have a mic, but we want to pass the virtual mic to you so that you can give a shout out to your team, the work that you're doing and where people can reach you, because we have a few OVCs and founders who listen to us and they're always interested to hear what people are building and how they can reach them.
Speaker 1:Yeah, people can reach me straight on the max at deepmirrorai, and without the team, nothing of this would have been possible. I'm quite fortunate because I have two co-founders One is more on the technical side, one is more on the product side. I think the synergy between the two of them is pretty much amazing. We also over the last year hired two more people who help us more with customer service and things, like Cecilia and Jacob, who's like basically taking charge of all the machine learning, and potentially in the next year we'll grow the team to 10 people. So I think we might be looking to hire soon. That's very exciting and at the moment we're very much bootstrapped for our consultancy contract. We have a bit of investment. We might be looking for some later this year to re-scale up the product Nice.
Speaker 1:But, yeah, that's all. Very. Sometimes I would say we're the pre-seed just looking to find the product with which we'll go on the seed rocket. But you really have to. In such a complex space, as we mentioned, you really have to watch out that the thing you're building is actually something useful.
Speaker 2:I wanted to say thank you for coming on the show. It was really nice to meet you and hopefully we have you on again.
Speaker 2:Check in a couple of years' time. See how deep mirrors do we leave you with thought. The future of the pharmaceutical industry is on the brink of transformation. From AI-powered drug discovery to the emergence of novel therapies, the landscape is shifting rapidly. These innovations usher in a new era of health care. Only time will tell. Stay tuned, stay informed and join us again to explore how things are changing around the world. Until next time, stay curious.
