Big Pharma steps up race for AI-discovered drugs

Are you Swiss and planning to move abroad? Or have you already taken the leap? I'm really excited to tell you that Swissinfo has a new podcast coming out just for you. You can listen and watch it from the 25th of November. It's called Ade Mercy Shweitz and is available in Swiss German and French. It covers everything you need to know about setting up your new life abroad. We speak to Swiss around the world who've already made the move, and we ask experts to share their experiences. You'll find Ade Mercy Shweitz, wherever you get your podcasts or in our sweet plus app. Hello, I'm Jo Fey and this is a Swissinfo podcast. Swissinfo is the multilingual and international public media company of Switzerland. Artificial intelligence is already revolutionising many areas of life, and this is especially true in medicine, pharmaceutical giants such as Swiss firms, Russia and the Vartis are betting big on artificial intelligence to discover new drugs, to treat a range of diseases. But there's a long road ahead to bring AI-discovered drugs to patients. In some of our previous episodes this season, we've talked about the decades and massive amounts of money spent trying to develop drugs that extend lifespan, or slow Alzheimer's down, or to improve women's health. But what if we could dramatically reduce the time and money it took to develop these drugs? What if we could develop even better medicine? And that's the promise of artificial intelligence. I'm with our healthcare reporter Jessica Davis-plus to talk about how AI is revolutionising the development of drugs for a range of diseases and whether it's living up to its promise. Hi, Jo. Thanks. This is a really exciting area. I spoke to a number of people about how artificial intelligence is revolutionising medicine. One person in particular I spoke to is Mothius Stegard. He discovered a drug candidate, EA-2353, for a rare degenerative eye disease. But it started out very low-tech, with a notebook and a pencil. I hope you're not in San Francisco right now. Stegard is a trained medicinal chemist and he worked in drug discovery at the Swiss From Roche for years. For nearly a decade he jotted down in a notebook chemical structures that researchers found had an impact on stem and progenitor cells. These are cells that can regenerate and damage tissue. But to arrive at a drug candidate Stegard needed to find the pattern in the chemical structures. This would take years and a lot of money testing in a lab and even then a lot would be left to chance. He decided to send the chemical structures to Gisbert Schneider, a former Roche colleague. Schneider now teaches computer assisted drug design at the Swiss Federal Institute of Technology, ETH Zurich. Here's what Stegard told me. I went to Gisbert and asked him to apply his machine learning algorithms to design new molecules that come out of these structures. Out of that initial work then came the two clinical candidates that we have now. It still took five years of work by medicinal chemists to improve the properties of the molecules. But Stegard, who founded the company Endogena in 2016 to develop the drug candidates, still gives credit to AI for these initial discoveries. It's actually not about speed. It's really finding the needle in the haystack if you like. How would I have done it without AI machine learning algorithms? I'm not so sure it would have been possible at everyone. That's the key thing. How could I have come up with novel molecules? It's very difficult even for a very trained medicinal chemist to design on a paper new molecules that we believe are active. I could have done it, but it would be more like a guess. The algorithms can find these particular patterns, patterns of small molecules, which is not necessarily almost often not visible to the human eye, not even to a trained medicinal chemist. It's estimated to take around $2.5 billion to bring a new drug to market. It also still takes about a decade for a drug to make it to market, and the failure rate is miserable. Only around 10% of drugs actually make it to the market. AI's potential to dramatically reduce the time and cost it takes to discover and develop drugs has fueled an AI investment boom. Over the past decade, investors have pumped more than $18 billion into biotech firms and startups that put AI at the center of their drug discovery workflow. Big pharma companies are now jostling to develop and buy up the latest AI platforms and technologies. Roch has been investing billions to overhaul the company's digital infrastructure and make AI a more integrated part of its research and development process. It's hired top computational biologists from MIT and Cambridge University. The plan is to build out a team at its San Francisco subsidiary Genentech and other sites including in Basel. It also signed a multi-year research collaboration with US chipmaker and video. Roch isn't alone. There have been hundreds of deals between pharma and AI drug discovery startups in the last decade. Novartis in Basel signed a major partnership with Google DeepMind's offshoot isomorphic labs in early 2024 to develop drug candidates. When I talk to experts, they say that pharmaceutical labs have been using machine learning to assist drug discovery for years. But there's something different now. I asked Gisbert Schneider from the ETH about this. We see a third wave of AI in pharmaceutical discovery, likely facilitated by a number of seemingly disconnected developments. One is the speed of computers and the availability of cheap computer time and data storage has certainly made possible certain applications, large-scale applications. These are not miniature baby models anymore. We're dealing with today. These are really valuable useful models trained on a lot of data. The latest generation of AI models can analyze and find patterns in vast and disparate data sets and even images, making it extremely useful for drug discovery. Scientists are usually dealing with trillions of cells and around 20,000 protein-coding genes in any one person. A major breakthrough came in 2020 when Google's DeepMind launched AlphaFold. It is an AI algorithm which could predict the three-dimensional structures and interactions of proteins. Protein folding is one of the hardest problems in biology. AlphaFold was key in determining the protein structures of the virus that causes COVID, helping scientists develop vaccines in record time. There are now a host of AI tools to search medical journals for relevant data, screen molecule libraries for promising drug candidates and identify disease targets. Some studies suggest AI could reduce the time and cost of drug discovery by 25-50%. Beyond making drug discovery more efficient, AI has the potential to identify and even generate molecules that chemists haven't even dreamed of. Some algorithms are even generating molecules from scratch, like the one Schneider developed with Stegard. To convert or transfer the traditional approach of molecular screening, one after the other, high throughput screening, high throughput experimental screening, to convert the whole drug discovery process to design, to generate molecules instead of searching for them. We are no longer looking for new drugs by testing them one after the other, although this can be done by virtual screening as well, quite nicely today. But the trick now is to invert the process and say, "computer, please generate." You can say, "Please or not." You can omit the please, of course. "Computer, you generate a new molecule that is IP-3 and has desired properties." So it kills pain by interacting with a certain desired set of receptors, for example, and has potentially fewer off-target effects, fewer potential side effects by design. And that is the conceptual breakthrough. I would say these generative AI tools we have today have become so prominent and no one wants to be left out anymore. And Big Pharma is back on track now to develop and apply these tools. As an example, Roche's subsidiary Genentech has developed its own AI model to identify new antibiotics. It has been found to be 90 times more effective at identifying compounds with antibacterial activity compared to using traditional methods alone. Some of the molecules identified by the algorithm have completely different structures than those used to train it. Despite the massive investment and excitement over novel discoveries, there's still some reticence to boast about what AI has actually achieved. No AI discovered drug has been approved and launched on the market. But some are getting close. Yes, it does speed up drug discovery. I'm not sure 30 percent or 40 percent or 15 percent is the right number here, but it speeds up drug discovery in particular in the early drug discovery phase when it comes to finding that the first hit, the first lead compound, the first ideas, which new kind of chemical matter should we synthesize and test. Here, I see an advantage. Second, it will save and it does save cost and material, precious material, because we can reduce the cost limit and materials consuming high throughput screening exercises. We can trim them down towards focused screens to limit their breadth by using AI tools that select the compounds to be screened. So early drug discovery can be sped up and we can come from an idea of a new drug, so I don't know, we want to have a stem cell module later or whatever, to success and pre-clinical development to ready for clinical trial entry, and here the 30 percent certainly is the right margin, 30 to even 50 percent faster than before. Success on a computer screen or even in a lab doesn't always predict success in patience. Here's Mattias Steger again. AI delivered what we wanted, and if it's not going to be successful, then it's not about AI, then it's about we made a mistake. Well, mistake, we didn't optimize to drug in the best way or the biology doesn't. I mean, it's not always like that the drug is wrong. When I come back to the clinical trial failures, it's often not that there's something wrong with the molecule. In most cases, the problem is that the molecule does exactly what it should be doing, but the biology with the link between the biology the molecular biology and the disease doesn't pan out as the hypothesis was. That is the most responsible factor for drug failures. Even if AI generated drugs fail in clinical trials, researchers hope that this information is fed back into the models to generate better drug candidates the next time around. According to Schneider, the hope is that companies fail less and faster, avoiding huge cost outlays and unnecessary testing of drugs and animals and humans. Given how much we don't know about human biology and the lack of data, experts warn that we still have to be careful about overhyping AI's impact on drug development. As soon as we enter the human body or biological system, we have to concede our imperfect understanding of human pathology, human psychology, and the effects a small molecule can have in a human being and there is this element of unpredictability. AI will never, and I say never deliberately, never be able to perfectly predict the drug effect on a human body. Take for example, you take an aspirin and some people react well to aspirin, others don't. To some degree, we can try to make predictions, whether a certain drug will work in a clinical trial, have a certain desired efficacy in a certain cohort of patients, and these are ongoing efforts. And I think we will come so far, say 80%, 85% correct, but then there is the individual human being, which can be an outlier, a perceived error. So there are certain limitations. This was a complex answer to a complex question. So when it comes to predicting clinical outcomes, clinical trials, the late stage drug development process, the questions we have to answer there are two complex for today's AI systems. Thanks so much Jessica for that report. I think it's really incredible how AI is being used in this way, and just how fast it's moving to in developing new drugs. But as you said, there are still no drugs on the market right now that have been discovered by AI. So how far off do you think we are really, and are there any inherent risks that could be involved in developing drugs like this? Those are really good questions. I mean AI is already helping speed up drug research in development in many ways, but experts say that we'll likely see the first fully AI discovered drug on the market by 2035. One big question though is how regulators will actually view drugs discovered by AI, especially given so much of what AI does as a black box, meaning that how AI comes up with an answer is not fully transparent. That's a super interesting point. I can't wait to see how this field develops. Thank you for bringing us that report Jessica. Coming up in the next episode of the Swiss Connection Science podcast, we're traveling to Basel, where we meet Michael Hall. He made a ground breaking discovery more than 30 years ago that has many people in the aging field very excited. Today's episode was recorded and edited by our science and video journalist, Michaela Andina. For more content, visit our website, Swissinfo.ch. I'm Jo Fey. Thanks for listening. [Music]