Can Open Data Become the New Infrastructure for AI Drug Discovery?

Hello, Harry. Harry, can you hear me? Okay, okay. Okay, let's get started. Hello, everyone, and welcome to Join Care Space. Thank you for joining us today. While we are getting started, we are free to share the space with anyone who interested in AI scientific research on these science. This topic is, can open data become the new infrastructure for AI drug discovery? The open elements found in the announcement highlighted in our agenda gives us a few useful starting points. The initiative brings together laboratory experiments, open data sets, AI models, and community challenges, where predictions are tested against experimental results that participants haven't seen. That's right, that's raises some practical questions. What is AI already changing in drug discovery? How much difference can open data make? And how can communities help to share the resources into useful research? We are highlighted to have Natasha, City of Join Care, Harry, co-contributor of MeridianX, and a higher CMO of Freedom is joined us today. We are beginning with brief introductions, walk through four questions, and also the room for community questions if time allows, so let's get started. Before we get into discussion, each of our guests briefly introduced itself, also your project, and what's interesting to you about today's topic. First, let's invite Natasha from Join Care. Hello, Natasha, can you hear me? Yes, I can hear you. Yeah, hi, I'm Natasha, and I'm the CTO of Join Care. I'm a machine learning engineer. I'm also an award-winning founder and entrepreneur, and I have a very interdisciplinary background that includes mathematics, philosophy, and AI. My experience spans the AI and ML pipeline from data acquisition and analysis to custom model building and the post-training and evaluation of larger systems. And at Join Care, I focus on the development and deployment of our care agents, and on how AI can be applied rigorously and effectively across project evaluation and research and development. And of course, the question of data is crucial for these things. Okay, thank you, Natasha. Next one, let me invite Ohio from Freedom Assets. Hello, everyone. I'm Ohio, the theme of Freedom Assets. Freedom Assets is a multi-train tax aggregator focused on token swaps, liquidity access, and NFT trading, or non-castodial and executed directly from the US wallet. Personally, I spend a lot of time on AI and the emerging technologies. So I'm very interested in how AI is changing scientific research and the Jack discovery. I'm very glad to join this discussion today. Okay, thank you, host. Let's turn to the next speaker. Thank you, Ohio. The last but not least, let me invite Harry from Matt Berean, X, Harry. Good morning, everyone. I'm Harry, the core contributor from MeridianX. And let me introduce MeridianX briefly. So MeridianX is a stablecoin based metal settlement layer for the global aviation system. So in simpler terms, we buy wholesale, sell retail by utilizing USDC from online travel agencies and connect to the upstream airlines. And so we are like a vendor between them. Like in a more simple term, we strive to make your airplane ticket cheaper. And when it comes to the topic, we are going to discuss today. Can open data become the new infrastructure for AI drug discovery? I found it very interesting because I heard that andthropic has their own biological research labs. And even the RCO thinks that biology could be the next break through point that AI is able to help human scientists to do. And so I'm very excited to join from MeridianX today. Thank you. Thank you, also thank you all. Let's begin with what is actually beginning actually happening in this discovery today. There's a lot of excitement around AI research and also in AI drug discovery. So let's start with the practical picture. How is AI changing in drug discovery and what is already happening? And what still needs more time to prove itself? The first one, I would like to invite Natasha from joining care. Yeah, thank you. So first of all, I think it's probably worth just talking about drug discovery and what the main kind of stages of drug discovery on how it works. So drug development overall is a very expensive and lengthy process. It typically takes 10 to 15 years. And drug discovery is where it all begins. Now the starting point of any drug discovery is an unmet medical deed. So it will be a disease that we want to cure, or we maybe just even want to find a better treatment for. Biologists develop a hypothesis about what mechanism might cure or slow the disease. And then they will select something called a target, which is usually a protein, like an enzyme or receptor that plays a key role in the disease. This target then usually needs to be validated because we want to know that it definitely does drive the disease. And then it's screened. And that means that you monitor the effect of exposing this target molecule to thousands or even millions through them. And the compounds that show promising activity against the target are called hits. And then you have this hip to lead phase, where chemists refine the best hits into lead compounds and then optimize them, including for something called admet properties. So that's how the body absorbs, distributes, metabolizes and excretes the compound, and also whether it's toxic. And the best compound or molecule is then chosen as the pre-tinnical candidate. So that's the one, if everything goes well, that will all may eventually become a drug. Now AI has already changed a lot of this process. We actually talked about this a bit in our AMA a couple of weeks ago. In terms of the very start of the process, models like alpha-fault-3 have transformed protein structure prediction, which helps researchers understand their target and where a drug might bind to it. And then when it comes to finding hits, AI is now widely used in two ways, firstly to virtually screen huge libraries of existing compounds and also to use generative models to design brand new molecules. And then during lead optimization, AI is used to predict admet properties. And beyond conventional small molecule drugs, AI is also used to design peptides, optimize antibodies and engineer proteins. Another thing we discussed a couple of weeks ago in our AMA was how AI agents have changed the game because they're increasingly able to automate entire parts of the scientific webflow. So in the beginning, they're able to generate, compare and refine hypotheses. We're seeing that with multi-agent systems like Google's co-scientist and Robin. AI agents are also designing therapies. For example, Stanford's virtual lab, which used an AI principal investigator and specialist scientist agents to generate 92 new SARS-CoV-2 nano bodies. And actually just last month, Stanford medicine researchers published a virtual biotech that consisted of 37,000 AI agents. And those agents actually proposed a lung cancer treatment strategy that turned out to be the same as one that a major drug maker was already independently developing. And that is since received FDA breakthrough therapy designation. So AI has already been hugely successful in these ways I've described. But of course, as I mentioned at the beginning, drug development is a long, expensive process. And it extends far beyond that initial drug discovery. All AI successes so far have really been in these very early stages. So after the initial drug discovery, you have pre-clinical testing. The point of that is to get permission from regulators to begin human trials. And then you have phases one, two and three. And in those phases, you move from a first small trial in healthy volunteers that happens in phase one to the first tests in actual patients with a disease in phase two, to then large scale trials as multiple centers, conducted on several hundreds to several thousand patients. And if you make it through all that, you move into regulatory review, which is the sort of approval stage. Now, as I was saying, AI design drugs do extremely well in the early stages, sort of up to and including phase one. A study published last month had a look at 129 of these, AI design drugs. And in phase one, they had a success rate of nearly 80%. That's compared to historically somewhere between 40 and 65%. And a good example has been in silicone's rentals certab. That went from target to phase two in under 30 months. Normally, we would expect that process to take 68 years. So that's fantastic. But then in phase two, which is where most drugs fail, it's a very different story. So that same study found that the success rate for AI design drugs was under 20% in phase two, which is actually below the historical average of roughly 30 to 40%. Now, to be fair, more than half of the drugs in that study were cancer drugs, they tend to fail more in phase two anyway. So that's probably part of that statistic. But either way, the early advantage isn't showing up quite as much as we'd want to at the next stage where it really matters. And I think the biggest reality check at the moment is that only one drug that can arguably be considered AI design has ever successfully cleared phase three. And that's Sassositinib. It's an oral psoriasis candidate by Decada. By the way, that's actually likely to be approved by the FDA in early 2027. So watch this space. That will definitely be a milestone for AI and drug development. Now, some issues are already visible in the early stages. For example, AI research tends to concentrate on problems where large, clean data sets already exist, even when those problems are not the most important bottlenecks. And so we still need to see how helpful AI can be in rare diseases or other fields where new data needs to be generated. And a recent study shows this is a problem even for these very well-known models like AlphaFault III. So AlphaFault and similar models can now actually predict how a drug molecule sits in its target spiding site. But when researchers tested them on thousands of new structures, the models only got the placement of the molecule right if they've seen something very similar in training. So these models may be good at tweaking known compounds, but they may not actually be that good at finding brand new hits. Yeah, so to sum up, AI has already transformed the various stages of drug discovery. It's getting drugs into the clinic faster and with better phase one success rates. But there is still quite a bit left for AI to prove over time in terms of what its impact actually is at the later stages of drug development. So can these promising early results actually translate into success in the later trials where most drugs fail? And can AI deliver when data is sparse? And I would say these are still quite open questions. OK, thank you Natasha, very brilliant. And so let me invite the next speaker, Haya from Freedom Assets. Hello everyone. And thanks Natasha, very impressive. And for me, I think that AI is already very useful in drug discovery, but we should be realistic about where the value comes from. Drug discovery has a huge surge problem. As we know, there are too many possible molecules, targets and combinations to test experimentally. AI can have researchers narrow that search space very faster. It can have identified the promising molecules predict properties, pre-ority high candidates and decide which experience I was running first. This is also what I find interesting about platforms such as PetsNet.