Building Blocks of Life, AlphaFold and AI Drug Development
Stephen Horn and Autria Godfrey look at how AI is being used to understand proteins and to find new drugs. Luke Yates, an assistant professor at the University of Sussex's Genome Damage and Stability Centre, explains what Google DeepMind's AlphaFold 3 does and the debate over access to its code. Kyle Rohde of the University of Central Florida describes his lab's work with the company Atomwise using AI to screen for new drugs against drug-resistant tuberculosis. The episode was recorded before the 2024 Nobel Prize in Chemistry was awarded for AlphaFold.
Transcript
Hello and welcome to Agents of Tech, where we explore the minds of the greatest innovators addressing the world's critical science and technology challenges. I'm Stephen Horn and we're speaking to you from London.
Hi Stephen, great to be with you. I'm Autria Godfrey in Washington, D.C. and today we are looking at understanding the building blocks of life and developing new antibiotics.
The urgency of the need for TB drugs is is often very much underappreciated in countries where you and I probably live because it's not such a big clinical problem, but um until COVID came along, we were the number one infectious disease killer in the world.
Now, the challenge with this whole process with AlphaFold 3 is that element of being able to do drug discovery with AlphaFold 3 to the academic community is um not available.
Coming up in the show today, we'll be talking to Luke Yates, assistant professor in genome stability at the University of Sussex, about using technology to understand the building blocks of life. AlphaFold 3, which is Google's AI model, predicts the structure and interactions of all life's molecules with unprecedented accuracy. We will also speak to Kyle Rohde, professor at Burnett School of Biomedical Sciences, University of Central Florida, about his lab's work to harness the power of AI in drug screening for TB drug discovery and creating new antibiotics.
And we're also absolutely delighted, as always, to be joined by Laila Rizvi, who's director of strategy and research at WebsEdge. Laila, welcome.
Thanks for having me once again. It's a delight to be here.
So Laila, this is a huge topic, right?
Yeah, massive. I mean, uh DeepMind has just released AlphaFold 3. There's a lot of discourse around it and you know, AI to biology is what calculus is to physics. I mean, that's what what's the talk of the town.
Well, we're absolutely delighted, actually, to be joined in the London studio by Luke Yates, a Dr. Luke Yates, assistant professor at the University of Sussex Centre of Genome Stability. Luke, welcome.
Thank you for having me. Pleasure to be here.
Thanks ever so much for for joining in today. We really appreciate it. So, so Luke, can we kick off just take us through what AlphaFold 3 actually is and what it does?
Gosh. So, AlphaFold um is has is an algorithm is is I guess a solution to the folding problem presented by Google DeepMind. Uh and it allows us to calculate uh predict the three-dimensional structure of a protein from its amino acid sequence. And these are strings of amino acids, polypeptides, and they're made of an alphabet of 20 distinct amino acids which have distinct chemical properties. And the sequence of that dictates really how a protein folds into a three-dimensional object which gives it function.
There is a mantra in structural biology which is that structure equals function. And so each protein has a specific shape to do a particular job inside the cell. And there are many of them inside the cell and they don't work in isolation. They often work together and with small biomolecules that float around the cell that are sort of chemical components that are important for signaling and important for um I guess you used and often used as chemical energy to do work. And so proteins are often seen as the workhorse of of inside the cell and they keep us alive.
And they do some remarkable things like replicate our DNA which is an area that I'm very interested in as a researcher. And they do this very accurately and they can replicate these protein machines, nanomachines, can replicate billions of kilometers of DNA in your lifetime all very accurately. And and and if they do come into problems, then that that's where human disease comes in. And so, understanding these nanomachines that are proteins and how they function helps us understand um human uh disease uh states, and we could probably leverage that information in order to develop new therapeutics.
So, now that we have this technology that has the ability to predict the structure of these molecules with this unprecedented accuracy, what do you think that means in terms of accelerating research?
So, in terms of what it's going to do to researchers like me, I've spent many years trying to determine structures experimentally, and I've used techniques like X-ray crystallography and cryogenic electron microscopy. These are very specialized techniques that require expensive machines um and hundreds and thousands of hours of of laborious work, of trial and error, in order to get to a a structure that then, once we've understood, can give us an insight, a deep understanding into the biological process.
And so, um each structure roughly costs about $100,000 to solve a single structure. And so, to date, we have solved uh as scientific community about 200,000 of them. And so, and we've deposited these structures in a public database that's freely accessible to allow other researchers to leverage the information we've gotten from these studies. And of course, to replicate that today would cost 20 billion US dollars. And so, that's a lot of money and a lot of time and effort. And so, if you can go from a amino acid sequence uh and a use a computer instead of hours of lab work to get to a pretty good approximation with reasonable accuracy of a three-dimensional structure of a protein, that would save you billions of dollars in research funding.
Wow. So, do we already have some specific cases or examples where AlphaFold has provided key insights?
So I in my field in I guess genome instability, we're looking at ways in which the cell and it the protein machines recognize and repair DNA damage. And the DNA damage if it repaired incorrectly can lead to human disease, developmental diseases, or it can lead to cancer. And so there are there have been a few labs in the world who've really leveraged AlphaFold as a discovery tool. They know a factor is involved somehow in the process of DNA replication. They're not quite sure how it works. And so by in essence screening on a computer across all the different variables of proteins they can think about that may be important for this process, they can narrow it down to handful of candidates that they can go along and then test in the laboratory. And then that reduces the trial and error process. And so there have been a few breakthroughs in this area where labs at Harvard Med School, Johannes Walter's lab at Harvard Med School, has leveraged AlphaFold to find a new protein function, which explains why when mutated it causes or gives rise to a developmental disorder in humans. But at a molecular level it helps understand how it orchestrates the beginning of DNA replication, which is called initiation, and that's a really key process and has to be timed very well in order for the cell to divide properly without error. And so I guess this that that to me is a really good example of leveraging AlphaFold for discovery based science, but also helps explain specific human diseases.
I use it in my own research to fill in a lot of the missing gaps. So we determine experimental structures, but sometimes tech technical reasons pieces are missing. And so, we might use AlphaFold to fill in the gaps, um which then may lead to new and interesting avenues in which to tread.
That's incredibly interesting. And I think I might be um jumping the gun a bit because you're obviously talking about discovery. Um but, what does AlphaFold mean for more kind of therapeutic development and therapy development? Um and particularly because you mentioned cost-effectiveness. I mean, one of the most expensive of things is drug development. Um and one of the most profitable things is drug development. So, um yeah, what does AlphaFold mean for drug development and therapy therapeutic development?
Okay, so AlphaFold 3 is the is is the latest and and and most sophisticated um machine learning model for protein folding. What's amazing about AlphaFold is that you can now include small molecules in your folding experiments and to see how these small molecules, which could be lead compounds in a drug screening exercise, how they interact specifically with your uh target your therapeutic target, which is ultimately a protein of a specific fold. Um and in doing that, um we we can leverage, I guess, AlphaFold 3's capability to um to really springboard um a sort of what we call a structure-based drug design program. Where you take the accurate structure with a potential lead compound bound, and you use a chemistry department to to really modify that um that scaffold, that chemical entity into something that is that that looks more like a drug-like molecule that could probably get through um like clinical trials, for example. But, that that early discovery phase, if you've got a target for a disease, and you want to find chemical entity that could be a good drug compound, that that costs maybe up to half a billion dollars, maybe up to a billion dollars for research companies, for pharma companies in particular. So, any way you can reduce that bill is only going to make drug discovery faster.
Now, the challenge with this whole process with AlphaFold 3 is that element of being able to do drug discovery with AlphaFold 3 to the academic community is not available because AlphaFold or DeepMind and Google have shielded that aspect and I presume because they have commercial interests themselves through Isomorphic, for example, who are collaborating with pharma proper. And so, we can only test commonly seen biomolecules found inside the cell, found in biology, which are available to us, but there are there is a limited number of small molecules we can make use of in our predictions. And I guess that's that's a bit of a challenge for the academic community who want to do drug discovery and might not get access to the the very cutting-edge AI tools to do it.
Luke, I want to get into a little bit of controversy that AlphaFold has has had recently. And that is DeepMind's decision to release AlphaFold's predictions freely to the scientific community. We did reach out to Google DeepMind. This is what a spokesperson said to us, quote, "We've been excited to see the impact our free-to-use server has had so far on academic research and we look forward to expanding the reach of this breakthrough even further when we release the AlphaFold 3 model. We are publicly committed to releasing AlphaFold 3 for academic use, which will include inference code and weights within 6 months of our methods paper release in May. This will allow researchers to benefit from AlphaFold 3 full capabilities for non-commercial use, including the ability to model novel ligands." What are your thoughts on the implications of this open access to this type of technology?
Yeah, this has been a big controversy in the field. A lot of structural biologists were very disappointed that the full AlphaFold code was not made available. Earlier iterations were made available. AlphaFold 2 Multimer was made available to the community. And what was really nice is that the structural biology community could tweak it and tinker with it in ways which helped to address their biological issues. Again, there are some examples where there are a group of proteins which which change shape as a as like a switch mechanism and they have two distinct, very different structures, but AlphaFold couldn't decide which structure to predict. And so a lot of researchers found ways in which to almost hack AlphaFold to to do that kind of work. And this is the this is where the structural biology community really shines. It's everything we do is very much open source and we offer any softwares, any developments back to the community. And so I guess it it jarred with the scientific community that this was not available.
But their response has has whilst been quite vocal, their response has also been to try and mimic or replicate AlphaFold 3 like machine learning algorithms for themselves without the without the code available to them. We've got a few contenders here that have just been released that are as good as AlphaFold 3 and they're completely open source, which means that the community can then go away and tinker with the code, optimize it for their own needs. We might see a bootstrapping, if you like, of discovery. And of course, when AlphaFold 3 finally releases their code, they'll be able to exchange notes and be able to find the best way, fastest way possible to to get to viable drug-like compounds, possibly as one avenue, or or improve, discover research significantly when we compare all these different methods.
So, Luke, obviously you've spoken about this issue of open sourcing, but what some people argue is that open source code means that this powerful technology might be in the hands of people who might misuse it. So, I can't really imagine what misusing AlphaFold might look like, but could you elaborate on that?
Some of the concerns about these types of AI models is that when they become too powerful, it's the they're ultimately tools and bad actors or people with bad intentions can make use of them for nefarious means. So, you could think about bioterrorism as a real threat when it comes to thinking about developing weapons that are that are biological, for example. Now, the challenge with that kind of argument is is that this is all kind of predictive and theoretical. You still need a good foundation of good quality science to validate any model that is predicted. And there is a there is obviously a barrier for people. People can't go to their garage and set up a dangerous experiment based on a prediction from AlphaFold.
But there there is a a growing thought about, you know, generative AI as AlphaFold has just released its new AlphaProteo, which is a generative AI program which produces protein binding molecules that are not seen in nature, that are targeted usually to sort of antiviral or anti-cancer ends, a bit like antibodies. Yeah. Um but they're much more effective. But again, you could imagine someone might want to make a venom. I mean, venoms are made of proteins and toxins are that could cause harm. But in in order to kind of achieve that, it still requires a lot of infrastructure, scientific infrastructure, to realize that that concern. And so, I think there's a there is a real worry on some level, but I think it can be well mitigated. And I think we as examples as researchers, we're not allowed to buy DNA from pathogenic organisms. We're not allowed to get that synthesized and shipped to us. The companies that provide this synthetic DNA will say, "We don't like the fact that you may be trying to work on bubonic plague or the causative agent of the bubonic plague." And they will ban you from trying to acquire that. And so, I think there's there's definitely scope to mitigate this risk quite significantly.
So, what we're hoping here, Luke, aren't we, that the good outweighs the bad when it comes to this. Well, thanks ever so much indeed for joining us and talking us through that. We really appreciate it, Luke. So, thank you.
My pleasure.
And staying with biotech and AI, we now have Kyle Rohde, Professor at Burnett School of Biomedical Sciences at the University of Central Florida. Professor Rohde's lab is partnering with biotech company Atomwise to see how AI can speed up testing of potential drugs to treat tuberculosis. Dr. Rohde, thank you so much for joining us today. I want to kick off with multi-drug resistant TB. It's a growing public health threat, and could you explain to us why this resistance develops and how it complicates treatment strategies?
Every year there's actually about a half a million cases. Half a million of the 10 million total cases of TB are drug resistant. And so, resistance to these antibiotics usually emerges spontaneously just due to mutations in the bacterial DNA that may change the target that the antibiotic is trying to bind to, and thus making it unable to do its job. And so, once you have a mutation like that, those mutants kind of gain the upper hand and they can continue to to grow even while you're still taking those antibiotics. And so to become multi-drug resistant is if you one bacteria accumulates these kind of mutations in multiple drug targets, now you can have multi-resistance. And so that can make um the treatment for TB, which even if it's normal drug-susceptible TB, that takes four drugs for four to six months for treatment. So if you have a drug-resistant strain, that then becomes four drugs that may be IV drugs or injectable drugs that have to be taken for up to a year to hopefully um resolve the disease. And so the cost of that treatment, the side effects uh becomes um very difficult clinically to manage. And so that that raises the urgent need for for new TB drugs to that can kill these resistant strains.
Right. And so how is uh is AI being utilized to find these new TB drugs um and particularly for these resistant strains? Um what is AI helping us do what we maybe couldn't do before?
Basically, when you're discovering new antibiotics, there's two major strategies. One we call phenotypic screening. So you may take hundreds of thousands of maybe drugs and have to treat the bacteria hoping to find some new ones that can kill it. Um but that means you have to have access to hundreds of thousands of drugs and you probably have to have a robotic setup to do hundreds of thousands of tiny experiments that a graduate student would want to do by hand. And in the case of TB, you would have to have that kind of sophisticated robotic setup in a specialized biosafety level three laboratory. And there's very few places that are equipped um to that level. And so that method is been what has worked, but it's very laborious and expensive. And then you find a a killing compound, but we have no idea how it works.
So the other strategy is kind of the opposite where you take a target protein that you have reason to believe is essential for the bacteria to stay alive, and then we call it target-based screening. So, you have to find an inhibitor of that target that you assume will kill the bacteria. But, TB is very well known to be highly impermeable, so you may find the best inhibitor of the best enzyme in the world, and it can't reach its target. It can't get through the cell wall of the bacteria, and so that becomes sort of the second bottleneck.
So, where AI fits in would be for this target-based type screening where if you know a target, now you can sort of short-circuit the the need to physically test hundreds of thousands or millions of compounds, and let the AI platforms predict the likely hits. And so, you may only have to screen hundreds instead of hundreds of thousands to find starting points for for new drugs. And so, that is at that early stage of the drug discovery where we think these AI platforms can can hopefully save us a lot of time and and effort. And also, make it possible for smaller labs that don't have robotics setups and capability to screen hundreds of thousands um to actually be a part of the solution.
Absolutely. So, how are you how are you working specifically uh on TB with with this technology?
Uh our sort of involvement in in using AI for our drug discovery efforts, um we know it's a difficult task, and so we're actually open to trying any strategy. And so, in the in our case, uh Atomwise was the the company that is the the source of the the platform that we made use of. They had started an industry-academic partnership called AIMS, where they had involved, I believe it was 400 different labs that they partnered with to utilize their platform to help us find possible drugs that then we tested and then shared the data back with them. And so that that is how we began to work with AI and our drug discovery efforts.
And um for those that I know more about the actual AI side than I do, their platform is known as AtomNet. It's a deep convolutional neural network. What what they're able to do is drawing on a huge databases of known protein structures and chemical structures and drug target interactions, they can model a protein of interest in all the different shapes and conformations it might, you know, bend into in real life and then virtually screen millions of compounds without me or my staff having to to lift a finger. And so then try to predict ones that would have a a good binding affinity with with the target that you're aiming for. And so that's kind of how we became involved and Atomwise actually provided the compounds. Not they didn't just send us a list of things I could go buy and test, they actually provided those compounds for us to to test on TB.
Sure. And so Dr. Rohde, I'm going to come back round on the microbiology side things. So you said you focus on specific proteins which are critical for TB survival. What are those proteins and are those protein targets that you're targeting in the search for new antibiotics? What do they do and why have you specifically chosen them?
When we began this project, Atomwise actually asked us to submit sort of a wish list of targets that then they would choose the ones that best fit their platform that had the most potential to work. And so two of the ones that that we move forward with, the criteria I guess were ones that we knew from genetic studies were essential. So, if you tried to delete that gene, the bacteria would die. So, if we inhibited that, we had reason to believe it would should kill the bacteria. And the the other was that there were not any current antibiotics that acted on those particular targets. So, if we found something, it would be truly novel and have the hope of being able to kill bacteria resistant to the drugs that are already out there.
So, the first one is actually part of the ribosome. It's sort of an accessory subunit on the ribosome, which bacteria ribosomes are one of the most targeted processes in a bacteria. So, we know it's a it's a well-validated process. We were just targeting a new component of that machinery. Um and there are there no current drugs that target it. There's one natural product that targets the protein we were going after, but it's not really a usable drug. It it works in the lab, but it's not safe to use in people. So, it that that kind of proves that it's possible. We just need to find a a new molecule that hopefully has more drug-like properties.
The second protein is actually an enzyme that is needed to for the bacteria to assemble and crosslink its cell wall, which is critical for it to survive. And it's related to proteins that are the targets of penicillin and penicillin-related drugs. But tuberculosis actually uses a very weird form of that enzyme that's very different from the ones in E. coli and other bacteria. So, a lot of penicillin drugs don't work. And so, we wanted to aim for this target because it's very TB-specific. So, if you're taking antibiotics for 6 months, you don't kill all the healthy helpful bacteria in your gut and other places. Um and again, it's novel. And so, if we could if we could find find a new drug to replace penicillin um for TB.
Professor, when you compare traditional drug discovery methods with these AI-driven approaches, what are the most significant advantages in terms of speed and cost?
I think for us it's been pretty uh sort of a remarkable advantages on the both of those the the speed and even just feasibility. We we don't have access to hundreds of thousands of compounds. Um so we're not even able to carry out a screen like that unless it's part of some larger government-funded program. And so for one it it's not just speed, it's just feasibility to even take on the project. But if it can eliminate weeks to months and thousands of dollars physically screening, um that really accelerates the the front end of our of our project. And I can only estimate the cost because for us these drugs were provided to us for free from Atomwise. And so our cost for a few pipette tips and our our poor man's robot were um were very minimal. And so it I think it really could save sort of a year or two if not more at the at the early stages. And I can't even calculate the amount of money it saved us.
Well, we know that AI runs on data for it to mine um so that we can get these outputs. So how do researchers like you contribute to this process?
So for for this type of of drug discovery, typically people had relied on uh need a 3D crystal structure of your protein of interest. And so computer modeling could do a very low-scale approach of docking and giving molecules a score, but you have to have a 3D structure. And so what AtomNet and other AI platforms are are trying to do is to say you don't even have to have the 3D structure of your protein. We can take advantage of structures of similar proteins and fill that gap sort of computationally. And so, structural data is very important, but what the AtomNet and similar tools are trying to make is is make it not essential. It's helpful, but you might still be able to move forward without the actual structure.
And moving more onto kind of the global side of things, um how could AI-driven drug discovery transform, you know, the fight against TB on a global scale, especially in regions which are most badly affected by the disease?
Right. I think the you know, the the urgency of the need for TB drugs is is often very much underappreciated in countries where you and I probably live because it's not such a big clinical problem. But, um until COVID came along, we were the number one infectious disease killer in the world. And then and the COVID pandemic actually reversed progress in TB um control. So, our numbers had been gradually decreasing. Now, they went back up again because all of the resources that were were channeled to COVID. The urgency is there, but the drug companies have basically left the left the building. The big pharma is basically not involved in antibiotic development. There's very few players still left because there's no money in it. And so, as these efforts fall to academic scientists like myself, we don't have a robotic setup that can do it the old-school brute-force way. And so, I think these tools become even more important enabling little guys like me to to be part of this effort.
So, I guess multi-drug resistant bacteria generally outside of TB causes quite a lot of deaths worldwide. And are you hoping that this kind of approach might be used for other diseases if successful that is? And then what I guess if you find a collection of really um good drug-like candidates, what's the next steps if you're, you know to moving towards a clinic that is?
Yes, I think one of the the big benefits of of the Atomwise strategy and approaches that they were intentionally testing it across multiple disease platforms to sort of prove this is broadly applicable everywhere from cancer to malaria to TB. Really it's just a protein and a chemical interaction that should work as long as you know enough about the target that you want to inhibit. So I think the the promise of it as it develops is broad across all from infectious disease to other other fields. And so I think it could be very well become a very integral part of sort of all drug discovery as as it is validated and and improved upon.
As as you may know and others that the road from a an exciting hit to anything that even looks like a preclinical drug candidate is a very long and expensive road. And so we're sort of now at the stage of trying to prove that the molecules that were predicted are actually acting and killing in the way that they were intended to and not just accidentally killing by another mechanism.
Kyle, thank you so much indeed for taking the time to talk to us today. We've we've learned a lot and and really appreciated it. So so thank you very much indeed for that.
Thank you. It's been a privilege.
Thank you. And Luke, thank you too very much for joining us today Agents of Tech. We've really appreciated that too. Until next time on Agents of Tech when we'll be looking at AI in education and bridging the gap between computer science and society, it's goodbye.
Goodbye. Please make sure to like and subscribe wherever you get your podcast from. Goodbye.
Agents of Tech is brought to you by WebsEdge.