Aug 3, 2023
00;00;00;00 - 00;00;31;24
Hi, everyone, and welcome to the Latest Dose, the podcast that explores the depth of innovation and human compassion in clinical research. I'm your host, Katherine Vandebelt, global vice president of Clinical Innovation at Oracle Health Sciences. Artificial Intelligence, AI, is one of the most popular technologies on the planet, and I find it referenced in most, if not all, industries.
00;00;31;26 - 00;00;59;16
Those of us working in the pharmaceutical industry strive to improve people's lives. Can AI help scientists develop better medicines faster? Human bodies are incredibly complex. Drug development is slow. Since I've been engaged in drug development, many people, teams, organizations, and companies have been working tirelessly to improve the drug development process, the promise, is nothing more than a revolution for the pharmaceutical industry.
00;00;59;19 - 00;01;26;21
The March 8th, 2023 Politico article states “nearly 270 companies are working in AI driven drug discovery”. Let's start learning more about AI driven drug discovery and discuss if or when the promise of AI will be realized. Can AI help speed up the drug development process? Identify new drug molecules that have so far eluded scientists?
00;01;26;23 - 00;02;02;02
Can AI-designed medicines, be safe for people? Have the desire effect on the disease? Meet the rigorous regulatory standards to actually be approved for human use? You know, many of these questions can be answered today with my guest, Andreas Busch, Ph.D. Chief Information Officer at Absci. Andreas brings substantial R&D expertise to Absci’s leadership, a world renowned leader in drug discovery and has led R&D efforts for some of the globe's top pharma companies, including Sanofi, Bayer, and Shire.
00;02;02;05 - 00;02;37;05
Andreas’ leadership has resulted in over ten commercial drugs starting from bench to FDA approval, with several more in late stage clinical development. Andreas holds the title of Extraordinary Professor of Pharmacology at the Johann Wolfgang Goethe University in Frankfurt, Germany, where he also received his Ph.D. in pharmacology. Andreas loves, real football a.k.a soccer, enjoys riding his motorcycle through Alps and playing with his beloved dogs Zorro.
00;02;37;07 - 00;03;04;28
Welcome, Andreas. Thank you for making the time to speak with me today. Hey, it's a pleasure talking to you Katherine. So, Andreas I have been taught that artificial intelligence, referred to as AI, are computer intelligence programs that can handle real-time problems and help organizations and everyday people achieve their goal. And AI is obviously a topic of discussion these days and getting way more attention with the release of the articles around ChatGPT.
00;03;04;28 - 00;03;33;22
Today I'd like to focus our discussion on generative AI, but I thought it would be helpful if you could share with me what's important for me to actually know about this type of AI. I'm glad to talk about it. I guess ChatGPT was certainly a breakthrough in AI and the use of AI for a general population and everybody knows now what AI can do through a GPT.
00;03;33;26 - 00;04;07;07
And if you look at generative AI, what we're trying to accomplish simply is to have artificial intelligence supporting us, creating drugs. And as you know, with ChatGPT, you have to give ChatGPT the right prompt in order to get ChatGPT to do the job for you. And this is similar with our generative AI. We need to give the prompt, which is we need to give our models the target, the mechanism we want to work on.
00;04;07;10 - 00;04;43;12
And then the model produces for us, in our case for Absci, a de novo designed antibody. So that's fascinating. How long have you been developing this approach with these prompts and these programs and actually been using this at your organization? I mean, Absci is actually a company which started as a cell line development company and realized then that for AI to be very productive, you need a ton of data and you need a ton of very consistent, high quality data.
00;04;43;14 - 00;05;14;24
So, these two things have to come together, you know, improvement of AI models, but feeding the AI models with plenty of data. So, the models can get better and better. And we've started really implementing AI for our E.coli expression systems for antibody a bit more than two years ago. And the progress we saw in our generative AI approaches were really very significant, very fast.
00;05;14;26 - 00;05;57;16
Already a year ago we were at a stage that we could optimize existing antibodies, so we basically gave the model the information of … look here is a known antibody, …. can you optimize it for affinity, … can you optimize it for immunogenicity and so forth. And we managed to do that. And just half a year ago, for the first time, give the model the information of the structure of a protein that we wanted to address, to produce for us a binding sequence completely de novo or without any idea of an antibody structure before. I think there was …. really for us …. the breakthrough.
00;05;57;19 - 00;06;28;16
And that is something which we have meanwhile even further progressed in the last half year. We extended this approach to more than one binding regions and we are ready now in a situation to address three of the binding regions of an antibody. And we are very, very optimistic that this progress is going to be extremely meaningful and helpful and what we believe disruptive in biologics research in the future.
00;06;28;18 - 00;06;49;01
So, this is exciting and extremely fascinating. So, I'm going to go to a statement you made about the data. So, can we talk a little bit about that? So where do these sources of data come from? What types of volume are you talking about? And I guess more importantly, as somebody who has worked with data for many, many years,
00;06;49;01 - 00;07;11;00
and one of the things that people will often ask about is ….should you use that data? Is that data appropriate? Is it reliable? Some people use the word quality. So, in order to achieve these impressive results, can you tell us a little bit about, more about, the data that's being used? Where does it come from and all those things?
00;07;11;03 - 00;07;36;13
Sure. To make it clear, what we're doing is, once we know the structure of a mechanism we want to address, let's assume whatever a membrane protein like a G protein coupled receptor, whatever you name it, we identify the region to which we want our antibody to bind and we give this information in the structure of this region to the model.
00;07;36;14 - 00;08;08;25
The model then delivers to us a number of model hits. Artificial intelligence generated hits. Information about what the model thinks the binder should look like. And what we do then, and that's the very straightforward answer to your question of the quality, is we generate those hits in the laboratory, we express the genes relevant for those binding regions in our expression system.
00;08;08;27 - 00;08;42;06
That's a microbial expression system, E coli. And then we simply have a test available called the Ace assay, in which we then validate what is indeed the binding affinity of those calculated binder. So that gives us then immediately an experimental validation of the AI suggestions and of the AI results. And therefore, we feel very, very comfortable that of course the quality of our predictions is very high as we validate them right afterwards.
00;08;42;08 - 00;09;25;10
Not only that, we validate them, but we can then again also use the information of those data to further improve the model. You ask, how many data do we generate? Well, the nice thing about E coli is that it replicates very, very fast and we can express huge libraries. The libraries again are the genes suggested by the model, and we can express easily your libraries of 500,000 or 1 million binding regions and as a consequence can measure 2-3 million of individual binders in a week or two.
00;09;25;10 - 00;09;57;08
And we can of course, also then see how well those binders are expressed in the cells and can measure up to a billion data points and protein interactions per week. Okay. So, I have to ask, if you didn't have the generative AI and the capabilities that you've just talked about, how long would it take for a human to do this without these additional tools and capabilities?
00;09;57;10 - 00;10;28;20
I think the really exciting piece about what I'm describing to you is that the model not only spits out a binder of a certain quality, but it spits out, already something which we can in a multidimensional way, optimize. So, if you go back to a traditional way of how to generate an antibody, which would be through mouse immunization or rapid immunization or what is called a phage display, you also can get a binder.
00;10;28;20 - 00;11;08;21
However, that binder comes without any potential optimization you would want to see. For example, you know, you get a binder. But you cannot influence in this traditional way the affinity, you cannot influence the solubility, the immunogenicity, and so forth. All of those parameters are very, very important for an antibody. Our model can spit that out, and I think that is a breakthrough, especially if you consider this is indeed a rounded up, optimized candidate.
00;11;08;23 - 00;11;33;05
This is not just, you know, a first antibody, which then can take over years, years really to get finely optimized. So again, going back to revolutionizing this and actually making it very different. So, but this is so different than what some people are familiar with or what they've been educated. Absolutely. They've done in the past. Are you familiar,
00;11;33;10 - 00;12;02;11
you probably are, but I'll just check; are you familiar with the technology adoption curve where they use the terms innovator, early adopter, early majority, late majority and lagger. Sure. Yeah, that's what's kind of coming to my mind. This is so different than what scientists have been doing in the past. I guess how broadly is this currently being used or where do you see the industry right now with this way of working?
00;12;02;11 - 00;12;25;02
Are we in still the innovator stage? early adopter? or am I a bit behind and we're actually in the majority stage? so can you talk us through that, please? That would be great. Yeah, I think we certainly consider our approach at the forefront of biologics research right now. And our focus is, of course, entirely on generation of antibodies.
00;12;25;04 - 00;12;55;11
That is our focus and I think it really needs this focus to make the progress which we are having right now. But how in the context of in general, R&D of biopharmaceuticals, there are many, many more aspects which AI can address what we are doing with large molecules, with antibodies other companies are doing with small molecules, with the chemicals.
00;12;55;13 - 00;13;24;12
Then you can of course, beside the generation of drugs, discuss options of AI to identify the right mechanisms. Because of course you always need to start in a disease with the right mechanism to address. Otherwise wonderful antibodies or wonderful small molecules are not really worth a lot if you're working on the wrong mechanism or target.
00;13;24;15 - 00;14;05;27
So, I think when it comes to generation of antibodies, we are at the forefront. We certainly want to extend our knowledge in the future to other biologics beyond antibodies. But there are other approaches of AI which of course are also very productive and they all really did grow over the last couple of years based on the existence and availability of vast amount of data.
00;14;05;29 - 00;14;39;26
So how much, how expensive is this? So, we've talked about how it works. We talked about how it's going to save significant time, what it needs to run. Totally appreciate your focus area in antibodies and so forth and other companies are doing other things but how expensive it this is? Is it really cheap? Is it moderate? And I'm not necessarily asking you to tell us the price, but what sort of investment, I guess, or what sort of expense should companies think about as they get engaged in this type of work?
00;14;39;29 - 00;15;03;00
Yeah, I think that we should try probably to look at the end game. What is the end game, really. I mean, our goal clearly is , once we know a target, at the click of a button we will have the information of how the optimized antibody looks like. The consequence, and of course, the click of a button does not cost a lot of money.
00;15;03;00 - 00;15;45;18
As you can imagine, you're doing that every day yourself. But as you can imagine, the traditional path is a very, very different one. The path it takes from a target to a traditional antibody really means tons of lab work… it means tons of iterative processes… it involves many, many people, consumables and so forth. Until you indeed have an antibody in hands which you then start producing first in vitro and in vivo data later on, those data will still be needed.
00;15;45;18 - 00;16;18;06
So, you will need of course, once you have the antibody spit out of the model to characterize the antibody in the relevant disease models. But until then, of course, I would say the cost saving and the time saving are enormous. My assumption is right now, if you look at benchmark and the industry, the cost to come from a target to a candidate antibody is somewhere in the range of $5 - $10 million.
00;16;18;08 - 00;16;46;13
And you can imagine that a click of a button is certainly going to be faster and cheaper. I think McKinsey actually coined this phrase, pilot purgatory, which means that organizations are hesitant to take on new ways of working. They see better ways, they see exciting ways, but because they don't necessarily understand them or they're not that familiar, they require a lot of change in their organization, they’re hesitant.
00;16;46;17 - 00;17;14;10
And so often, we pilot things or I have piloted things or my company has piloted things in my past. And then what I notice across the industry, this slow adoption can kill very valuable innovation because we're constantly piloting them. Do you see those concerns with what you're talking about, or how do you recommend that we prevent that or escape it in this particular situation?
00;17;14;10 - 00;17;54;29
Because it looks so, so promising. I think like every breakthrough technology, there will be the winners and fast adopters and there will be the slow adopters. Listen, I've been R&D head in pharmaceutical industry for over 20 years. I was R&D at Bayer and I was R&D at Shire and I've certainly dealt with a lot of associates, you know, which were skeptic of new technologies and like you heard from the McKinsey reports, not readily available always in a situation to adopt technological breakthroughs.
00;17;55;02 - 00;18;42;06
Having said that, once the breakthrough is obvious, that's the latest moment. Then you can get on board and everybody knows that at the end there is going to be, if really the promise comes through, which I just described to you, that there is no way that you could say, okay, let's wait. And I think this is going to go much, much faster than a number of other breakthroughs in the past, I think, not just the entire world's population got prepared to apply AI to ChatGPT, but the industry is really eager to apply AI along the entire value chain of R&D and even beyond that, onto marketing aspects of drugs.
00;18;42;06 - 00;19;25;26
So I have to say, yes, there always is a chance of resistance, of adoption of technologies in R&D organizations, but I am completely convinced that once our approach has been validated on a couple of targets, that will be the case in my assumption is within the next year, it is going to be a must without very little alternatives for industries to adopt it because it brings them into the situation to be faster, to come up with molecules which have a higher probability of success based on a multi parameter optimized profile.
00;19;25;28 - 00;19;57;13
And the two things together; being faster & being better optimized, gives you a competitive advantage, which you cannot, cannot give up. Do AI designed medicine, meet the rigorous regulatory standards that are being used to get drugs approved to humans? So, it sounds like this might be changing the data package, it might be changing how we actually might need to talk to regulators.
00;19;57;16 - 00;20;37;06
Am I understanding this correctly or what is it I need to understand with regards to regulatory requirements? Actually, we should distinguish between what I expect over the next 5 to 10 years versus in the more distant future. What we will deliver will undergo exactly the same regulatory processes as all drugs, no matter how they are delivered, no matter whether they come from traditional small molecule approaches or traditional biologic approaches, the regulatory process will be exactly the same.
00;20;37;06 - 00;21;08;04
The regulatory process will be… you need to show in phase one, phase two and phase three clinical trials that the compounds are safe and efficacious in patients. You will go through before you go to the clinic through extensive pre IND activities to get to that stage. Those regulatory aspects will not be different from generative AI generated drugs versus the drugs coming from traditional pathways.
00;21;08;06 - 00;21;42;27
The only difference I can see immediately versus the future , I can see that based on the chance that we should be able to predict with AI a much better profile. And already also if we go into systems biology, get more information about potential side effects, mechanism based and so forth, the probability of success to get through those regulatory processes is going to be increased.
00;21;43;01 - 00;22;18;13
That is the one aspect in the long term off course, I do see that regulators want to understand what really is the productivity also of AI methods in clinical development. They want to see how valid my predictions were of, you know, development aspects based on AI information and I can see that AI will have a significant impact in the future also on regulatory processes.
00;22;18;16 - 00;22;43;20
Again, I know this sounds repetitive, but that's so exciting to me. Being working in this industry so long to see these types of changes is it's just very, very inspirational. As well as I'm getting older, hopefully some of the targets we're looking at, well, hopefully bring some solutions for things that have eluded us for many, many, many years.
00;22;43;23 - 00;23;14;23
So, I guess as a place right now, I thought it might be helpful to our listeners for me to sort of go back to our introduction, where I asked a number of questions. You shared such great information and I wanted to just sort of make it a little more simpler. So, if you're able to sort of answer these questions as yes, no or to be determined or that type of thing, I will pose the questions. Can AI help speed up drug development process?
00;23;14;25 - 00;23;48;00
Yes, it was a simple one. Yeah, that was a simple one. Can AI identify new drug molecules that have eluded us? Yes, very clearly! We can get into a space where traditional methods may not be able to get to. There are a number of, if I look at antibody research, a number of targets, which right now with the traditional methods are not really addressable, particularly in membrane proteins, ion channels, and so forth.
00;23;48;02 - 00;24;15;16
We believe that we will be available immediately, more, or less, to address all of those difficult drug targets and make them amenable for treatment. Fantastic! And then you just address this with the regulatory question, but can AI design medicines, be safe for people? Again, you know, my assumption is and that's not a yes or no, it's a yes.
00;24;15;19 - 00;24;47;23
My assumption is we can deliver multidimensional optimized compounds which will be tested. My assumption always is we can predict higher safety or the safety in general. However, we will test it. There's at this point no way around, in the regulatory process, to avoid testing for safety. And then can AI design medicines, have the desired effect on the disease?
00;24;48;00 - 00;25;19;07
Absolutely. But I think we should be very clear that in the first place you have to work on the right target. The antibody has to be, in our case, directed against a highly validated target. If the target is, if the mechanism, if the target is not the right one, then the very best antibodies with highest affinity with wonderful other parameters is not helpful, is not useful.
00;25;19;08 - 00;26;04;17
So, you have to start off with the right mechanism, but then absolutely and generative AI generated antibody will be of gigantic use. Fantastic! So, as we close, where do I go to learn more about the work you're doing in Generative AI? Where would you suggest I expand my knowledge? I think reading always helps and I think if you if you really look there, there is, especially in a number of nature magazines a very frequent report on updates of news around generative AI.
00;26;04;19 - 00;26;34;18
We have published our papers and our news and bio archive, which is a place where we can publish the results without giving away all of the details of our technologies, which is of course important for us at our present stage of the company. But there are also, meanwhile, a number of meetings and platforms for AI which are worthwhile to attend.
00;26;34;20 - 00;27;01;27
Well, thanks for that. And those that are listening and they wish to connect with you, would they go to Absci.com, is that a good location? That's a good one. Wonderful. Well, thank you so much, Andreas, for sharing this exciting, innovative way to really bring new antibodies and align them with targets and really help bring some new solutions to the community, really appreciate your efforts.
00;27;02;02 - 00;27;30;09
Thank you for all that you do to making our lives better. And thank you for spending the time with me today. Thanks, Katherine, for having the possibility to talk to you. Thank you for listening to the Latest Dose, the podcast that explores the depths of innovation and human compassion in clinical research. Before you go, show us some love by subscribing and make sure to look for us next month.
00;27;30;14 - 00;27;39;00
Goodbye. [Music]