AI Discovered Antibiotics: How Small Data & Small GNNs Led to Big Results, w/ MIT Prof. Jim Collins

Hello, and welcome back to the Cognitive Revolution. Today, my guest is Jim Collins, Termir Professor of Medical Engineering at MIT and the leader of an AI-powered project that has created several new antibiotics, which are not only effective against antibiotic-resistant strains, but also work, at least in some cases, via entirely new mechanisms of action. The problem of antibiotic resistance is genuinely staggering in scale. More than one million people are estimated to die globally each year from treatment-resistant infections. And it's getting worse. In 2016, a special commission in the United Kingdom warned that if we don't address the resistance crisis soon, by 2050, we could have 10 million deaths per year, which would put the problem on par with all of cancer. Yet, pharmaceutical companies have largely abandoned antibiotic development because the economics simply haven't worked. It costs just as much to develop an antibiotic as any other drug, but people take them for only a short period of time. And critically, given a new antibiotic that's capable of treating the most drug-resistant strains, the medical system would reserve it to be used as a last line of defense, naturally limiting the size of the market. The good news is that Professor Collins and team seem to have created not just a few breakthrough drug candidates, but a multi-step AI-powered process, which they've refined for the last five plus years and proven by applying it to several different bacterial targets that can select candidate antibiotic molecules from the vast expanse of chemical space in silico with a high enough hit rate that it's now realistic to expect that the antibiotic crisis could be, for practical purposes, solved in just the next few years. That is obviously awesome news, but with so much AI news flying around these days, it's a story that surprisingly few people have heard. Even at the Curve last weekend, where all attendees were super well-informed AI obsessives like me, not many were aware of this development. And I think that's really unfortunate because I'm digging into Professor Collins' work. I found that this is not just a feel-good story, but an example of how even with relatively small data sets and modest compute budgets, modern machine learning techniques, cleverly applied, can drive huge value without anything looking like AGI. In concrete terms, as you'll hear in much more detail. By training small convolutional graph neural networks on datasets consisting of just a few thousand chemical structures and how effective each one was at stopping the growth of a target bacteria, the team was able to create a model that could screen tens of millions of compounds for efficacy in just a few days' time. And then, by using these predictions as part of a pipeline that also scored candidate molecules for novelty as compared to known antibiotics, chemical stability, the ease or difficulty of synthesis, and safety or toxicity for human cells, they were able to identify a small set of very promising candidates, which upon actual synthesis and testing did contain hits that were not only effective against the target strains, but again, at least in some cases, work via previously unknown mechanisms and without harming other types of bacteria. These compounds are now moving toward clinical trials. And while barring some sort of operation warp speed for antibiotics, it will still be years before they're broadly available, I was struck by Professor Collins' estimate that with these techniques at our disposal, the R&D costs to generate a pipeline of 15 or 20 promising new antibiotics could be as low as a few tens of millions of dollars, while the entire process, including the clinical trials required to get them approved, would cost maybe $20 billion. By the standard of recent data center build-out deals which have been dominating the headlines, this is extremely affordable. And the fact that this work remains relatively unknown, even in the AI community, suggests to me that in our haste to create, understand, tame or control, and hopefully live in harmony with fully general or even super intelligent AIs, which we hope will then turn around and cure all the diseases and otherwise benefit all humanity, we risk blinding ourselves to simpler, safer, surer wins, which themselves could still prove positively transformational for the human condition without introducing poorly understood or potentially existential risks. With that in mind, I hope you enjoy this deep dive into how AI, even with small datasets and just a few GPUs, is accelerating the discovery of life-saving drugs with MIT professor Jim Collins. Jim Collins, Termir Professor of Medical Engineering at MIT and creator of Novel Antibiotics. Welcome to The Cognitive Revolution. Yeah, thanks for having me on your show. I'm really excited about this and really excited about the work that you have done. It's an incredible thing. I often reflect on just how many groundbreaking milestone moments are passing us by all the time. And I've been going around telling people literally at cocktail parties and stuff about this work, and nobody's heard of it. I swear when I was a kid, people would have heard about this. I think it would have been like the talk of the town. But these days, there's just so much stuff flying by that people are missing it. So I'm excited to correct that. Thanks. For starters, before we get into all the AI side and ML techniques and all the details, because people who subscribe to this feed are very obsessed with all the AI stuff and probably don't know nearly as much about the biology, can we just do a little primer on the biology of antibiotics and antibiotic resistance? like how do antibiotics work and why do they stop working sometimes? Yeah. So antibiotics are generally small molecules, like an aspirin type that you would take typically orally to treat an infection that you might have in some part of your body, a bacterial infection. The antibiotics act by basically disrupting a protein, typically inside a bacterial cell that will be associated with an important process, be it cell division, protein production, DNA replication. This will disrupt that associated process. And we've shown that that will then lead to downstream stress responses from the bug that will lead to energetic demands that will produce toxic metabolic byproducts that will lead to additional damage inside the bug, damaging DNA, RNA, protein, membranes, lipids, that will trigger additional energetic demands, et cetera, leading to this cycle that will contribute to the disrupted initial process. Why resistance arises is that the bacteria have effectively an intrinsic goal of replicating, surviving and replicating. And so you'll have mutations that occur, so alterations of the DNA or other features, with some frequency every time the bug will divide, as well as in response to a stress directly, such as antibiotics, that will lead to changes in the target of the antibiotic or these downstream features that will make the bug less susceptible to the antibiotic. Those bacteria that have acquired that mutation then now have a fitness disadvantage, meaning they will survive, whereas the other members of the population don't have that, won't survive. So the others will die at that concentration of the antibiotic, but the ones with the appropriate mutation will live and will survive the antibiotic treatment, thus will now be resistant to it. As a result of our overuse of antibiotics, both for human use directly and through animal use, where the agricultural industry will use antibiotics both prophylactically to protect the animals from potential infection, as well as a growth stimulant. We've now seen resistance growing dramatically over the last few decades. And really in the past, the biggest risk would be in hospitals, so-called superbugs. I tell my students, worst place to be when you're sick is a hospital because of these superbugs. Get out as quickly as you can. But these superbugs, these resistant bugs to our frontline antibiotics are now no longer restricted to our hospitals. They're on our playing fields. They're in our childcare centers. They're in our schools. They're in our shopping centers. They're in our communities. And so the problem has escalated due to our overuse and misuse of antibiotics. So to just double click on that a little bit, one of my mantras for AI is that AI defies all binaries. And I think that's like, the more I learn about biology, it seems like that's true for many aspects of biology as well. So one thing I was kind of struck to notice in reading the papers is that I think a lot of times people think of like an antibiotic, you know, it's sort of like a, you know, missile that sort of zooms in and, you know, smashes the bacteria and it's just gone. And as I was reading, I was like, well, it doesn't really kind of look like that so much. The way that the effectiveness of the antibiotic is measured is not like a binary, but rather a scalar, right? It can be like anywhere from sort of not effective at all to like very effective or kind of anywhere in between. Does that like – I'm not sure if this is right, but I was kind of inferring from that that this sort of growth slowing maybe implies that there's actually still like a major role for the immune system. Like what is the actual role of the drug? Is it killing the cells or is it like slowing them down enough that the immune system can rally and destroy them for us? Yeah, it's interesting. So it depends upon the antibiotic and depends upon the type of antibiotic. So they're the ones that we typically think of, you know, the missile that you refer to would be bactericidal antibiotics. So those that have really been developed to kill the bacteria at a safe concentration for human use. The challenge with that is that not every bug in the infection site will see the same concentration. So many will see sublethal concentrations. Briefly, the second class of antibiotics are so-called bacteriostatic, those that are actually selected, developed to inhibit the growth or stop the growth of a bacterial infection without killing the bugs in that infection. Interestingly, as you picked up in the piece, screening done for identifying new antibiotics is almost always through a growth inhibition assay and not a killing assay. So a growth inhibition assay where you'll look to see when you apply some library of compounds, which of these inhibit significantly the growth of the culture, would then indicate there is some interesting antibacterial properties. In many cases, those are also associated with killing activity. Some are only associated then with inhibition. Killing assays are much more difficult to get after. The further level that's also in our piece is that you want to, and I have alluded to, you want the antibiotic to really largely impact only the bacterium and not our cells. And frankly, if possible, so only to impact the pathogen bacteria and not the healthy bacteria that make up our gut or on our skin or other parts of our body. Yeah. Defying all binaries indeed. So one more just primer question there is I also noticed in the paper that you run some experiments on these new antibiotics that you've found that test how quickly bacteria can develop resistance to them. And I was struck that in general, like it seems to happen pretty fast. So should we understand that this is like always going on, even like in our bodies on an ongoing basis? I mean, I kind of saw this with COVID too, where it seemed like there was a lot of mutations happening all over the place. And not all of them, of course, break out. But should my mental model be that like these resistant strains are kind of popping up all over all the time, mostly not going anywhere, but occasionally, you know, getting out of control? Is that the right way to think about it? Yeah, I think I'm actually getting out of control, but they might break out, as you say, from the stress and survive and then propagate. There's interesting dynamics, challenges between the interactions of different mutations, which ones are beneficial under which situations, which will be retained, which give you an advantage, which give you a disadvantage. issue. I think it's fair to say, I like to say that if you come across an antibiotic researcher who tells you that they designed or discovered an antibiotic for which there's no resistance, they're either lying to themselves or they're lying to you. And to then speak to it that if you apply it for long enough, eventually resistance will develop. And it becomes interesting to then consider the role of AI as we increasingly utilize AI in this space is that I think AI gives us an advantage in the battle then of our wits against the genes of these superbugs in as follows in the two different ways around resistance. We can continually discover and or design new antibiotics that act in new ways that would not be bedeviled by existence resistance. You still have to get its approval process and now introduce them. But second is also to explore how AI might be able to reduce the probability of resistance over a given time period, meaning extending the runway. And some of the ways that can happen is by discovering and designing molecules, for example, that hit more than one protein target. So that if its efficacy is linked to the action of each individual protein in independent way, now the bug would need to develop mutations in more than one site in order to provide us up with protection, which makes it that much more difficult for the bug to evolve away from the actions of the antibiotic. Yeah. Okay, cool. Well, we'll come back to all that in a second. One more just angle that I want to set this up with is the societal angle. We haven't had a lot of antibiotics recently, And I understand that the pharmaceutical industry broadly has kind of given up looking for them. Given that this is such a big problem with, my understanding is like tens of thousands of Americans and probably a couple orders of magnitude more than that, people dying annually from these antibiotic resistance strains. Why has that happened? Like what is the social failure that has led us to this state? It's interesting. I think it's largely an economic-driven failure that tapped into some interesting aspects of how we handle things as a society. Maybe to just kind of ground the audience, Alexander Fleming discovered penicillin a little less than 100 years ago. So September 1928 is when he serendipitously discovered penicillin. It was not then developed and manufactured as a drug until early in World War II in the 1940s. by a group at Oxford. So we haven't had them for very long, but they have transformed modern medicine, enabling us to have surgeries, deal with any number of injuries, cuts, bruises, ballisters that in the past would have been lethal and no longer lethal. Interestingly, the heyday of antibiotic discovery was in the 1940s, 50s, and 60s. So before the molecular biology revolution, before the biotech revolution, before the genomics revolution, before the AI revolution. What has happened since then is that we've been really in a discovery winter of sorts, that we haven't discovered new antibiotics, and the investment into the field has diminished dramatically. Multiple reasons for that. One is that it costs just as much to develop an antibiotic drug as it does effectively to develop a cancer drug or a blood pressure drug. But an antibiotic, you're only going to sell for a few dollars, whereas a cancer drug or a blood pressure drug you can sell for thousands, if not many more dollars. Antibiotic you'll take maybe over a course of a day or a small number of days. Cancer drug, blood pressure drug you take over many months, years, if not even for the rest of your life. So the economics support the development of non-antibiotic drugs. Further, you have that even companies that stayed in the business and made it all the way through to getting their young molecule approved. Once it was approved, the community of doctors said, oh, we're going to shelve your product, put it on a shelf and keep it for when we really need it. And as a result, many of these companies then went bankrupt after this milestone of getting their compound approved. And so we faced this dire situation. So how do we get out of it? As you allude, we have this underlying epidemic that's been going on for decades. About a million to a million and a half people die each year from bacterial infections around the world. And a UK commission estimated that if we don't address this resistance crisis soon, we'll have upwards of 10 million deaths per year by 2050. So outpacing deaths from cancer. It's a challenge. How do we motivate pharma companies, biotech companies to develop these products they're really not going to generate revenue from? And I think we need to explore public-private partnerships of the type we saw with Operation Warp Speed or Warp Drive, whatever it was called around vaccine development during COVID. Second is I think we need to better engage philanthropists. So we have various months to dedicate different diseases, cancers notably. We have walks, cheddar walks, runs, ribbons, colors. We have none of this for antibiotic resistant infections, though every one of your listeners has lost a friend or family member. I guarantee to an antibiotic resistant infection. Your family member went into the hospital for a or in treatment for us that caught the infection and died. And it somehow does not rise to the prominence in our consciousness of a need to address. And I expect you probably have some fairly prominent, very successful, wealthy listeners on the show. And it's interesting to think of the uber-wealthy and uber-wealthy individual could single-handedly address this challenge for the lifetime of everybody on this planet. So my estimate is for about a $20 billion investment, we could address AMR, specifically antibacterial resistance, over the next many decades. Okay, it's a lot of money for you and me. It's a lot of money for most of your listeners, but it's not a lot of money for an interesting number of individuals. So for individuals who would like to make history but not make a dollar or much money off of it, I think there's opportunities here to really leave an impact on humanity. Yeah. Hey, we'll continue our interview in a moment after a word from our sponsors. Building a voice AI app, its quality depends on the inputs. It's only as good as your transcripts. Missed words, messy text, or misidentified speakers can break your user's experience. With Assembly AI, you get high-quality, low-latency, multilingual transcripts that scale as your app grows. Imagine finishing a meeting and your app has already captured every word correctly, formatted the text automatically, and identified every speaker without any manual cleanup. Teams rely on Assembly AI for industry-leading accuracy, low word error rate, and continuous model improvements, keeping apps cutting-edge and reliable. And with pay-as-you-go pricing, unlimited concurrency, and no minimums, you can scale without worrying about infrastructure, commitments, or surprise bills. Personally, I love how you can provide key terms to guide the model's output, and also that it works with low-quality audio, something I'll definitely keep in mind for our next episode with Zvi. Check out the transcript of this episode, created by Assembly AI in under two minutes, on our website. And start building with $50 in free credits at assemblyai.com slash cognitive. That's assemblyai.com slash cognitive. Today's episode is brought to you by Anthropic, makers of Claude. Claude is the AI for minds that don't stop at good enough. It's the collaborator that actually understands your entire workflow and thinks with you, not for you. Whether you're debugging code at midnight or strategizing your next business move, Claude extends your thinking to tackle the problems that matter. Regular listeners know that Claude plays a critical role in the production of this podcast, saving me hours per week by writing the first draft of my intro essays. For every episode, I give Claude 50 previous intro essays, plus the transcript of the current episode, and ask it to draft a new intro essay following the pattern in my examples. Claude does a uniquely good job at writing in my style. No other model from any other company has come close. And while I do usually edit its output, I did recently read one essay exactly as Claude had drafted it, and as I suspected, nobody really seemed to mind. When it comes to coding and agentic use cases, Claude frequently tops leaderboards and has consistently been the default model choice in both coding and email assistant products, including our past guests, Replit and Shortwave. And meanwhile, of course, Claude Code continues to take the world by storm. Anthropic has delivered this elite level of performance while also pioneering safety techniques like constitutional alignment and investing heavily in mechanistic interpretability techniques like sparse autoencoders, both internally and as an investor in our past guest, Goodfire. By any measure, they are one of the few live players shaping the international AI landscape today. Ready to tackle bigger problems? Sign up for Claude today and get 50% off Claude Pro, which includes access to Claude code when you use my link, claude.ai slash TCR. That's claude.ai slash TCR right now for 50% off your first three months of Claude Pro. That includes access to all of the features mentioned in today's episode. Once more, that's Claude.ai slash TCR. Well let get into then how you would put that money to work I know you been working on this for a number of years and the first reported antibiotic that you found goes back to 2020 Just give kind of a high-level overview of the trajectory of the work for the last half dozen years. And then from there, we can kind of really dig in, especially to the most recent work and the data sets, the techniques, and all the nitty-gritty details. Yeah. So maybe I'll maybe go back even a little later. So our lab is working on antibiotics now for a little over 20 years. And we've used machine learning, so a sub-branch of AI, in that context from the very beginning. And our initial efforts were really using machine learning to infer, reverse engineer, biomolecular networks inside bacteria in order to better understand how antibiotics act. And our goal there was to better understand mechanisms of resistance, as well as to identify molecules, come up with ways that we could boost existing antibiotics. Here at MIT in 2018, the Institute launched a campus-wide initiative in AI, kind of recognizing that the Institute had been asleep at the wheel on this third wave of AI, first wave really in the very early days, late 50s to 60s, where folks like Marvin Minsky, Seymour Peppett here at MIT led the way with their interest in early neural nets and perceptrons. And then in the 1980s, in the second wave, where there was interest in things such as LISP, other language-based programming languages and executive programs with folks like Patrick Henry Winston. This third wave focused on big data, deep learning, eventually, the large language models. The Institute realized we really hadn't stepped up. And so in March of 2018, we launched a campus-wide initiative. I had the opportunity to sit next to Regina Barsley, one of our AI stars here on the faculty, who has done a lot of work in applying AI and problems in biology and medicine. We realized we both had interest in drug discovery and thought, would it be interesting if we could apply our interests to get after antibiotics? And we brought on Tommy Jockala, who's another faculty member and AI expert. And we really didn't have money to do this, so we bootlegged the project, looking around to see what could we pull together. And pulled together a small training library of 2,500 compounds, which was remarkably small. This consisted of 1,700 FDA-approved drugs, including the known universe of antibiotics, plus 800 natural compounds. Applied them to E. coli. So E. coli is both a model organism that we use in molecular biology to understand different biological processes, but it's also a pathogen. Many of you may unfortunately experience it, whether in a urinary tract infection or food poisoning. Applied each of those compounds to E. coli to see which exhibited antibacterial activity, as evidenced by growth inhibition. took those data, discretized it to say yes, no, if you achieve at least 80% growth inhibition, you would consider it antibacterial. If you didn't achieve that, you would consider it not antibacterial. Took the structure of each compound and trained a deep neural net, specifically a graphical neural net, that could learn bond by bond, substructure by substructure, those that were associated with the feature of interest, in this case antibacterial and not antibacterial. We then applied it to an internal library at the Broad Institute, where I also run a lab, that was the drug repurposing library at the Broad that consisted of just 6,100 compounds, but asked which of the molecules there were predicted to be antibacterial, which were predicted not to be toxic against human cells, and which did not look like existing antibiotics. And it's the only one molecule that fit all those three criteria, which is the molecule we call halicin in an homage to HAL, which was the killing AI system from 2001 Space Odyssey. HAL in the movie killed humans. Halicin, our molecule, killed bacteria and turned out to be a remarkably potent new antibiotic. So it's really striking to me how small that data set is. we're so used to, you know, today I, when I think of large language models, I'm like, you know, it's sort of one to 15 trillion tokens is the range of data set just on the pre-training, right? And then they do lots of post-training and reinforcement learning and all that kind of stuff on top as well. I guess if you had proposed to me, you know, in my ignorance before seeing all these results, that something like this could work with just a 2,500 compound library to learn from, I would have guessed that that was like probably at least two orders of magnitude too small. How do you think about the fact that this works at all? You know, it's interesting, Nathan. So I think your response is very similar, very consistent with the response we got from our colleagues in the AI space. So we presented what we had and what we were doing, they dismissed us and said, don't even start to explore. You have far too little data to do anything meaningful. So a few points of note. One is that the growth inhibition data that we collected, we could have discretized. So when you look at those data, you could see that we could have said, okay, this compound achieved 90% inhibition, this one was 10%, 30%. But we didn't recognize we didn't have a lot of data, so we just discretized. zero one. Okay. So we reduced the coarse grain now, the feature of the data, you know, to your point, binarize in this, making it binary. Two is that, you know, it is, again, interesting, you know, to this table, you know, that these models are very data hungry, the more data, the better. And the latter, we're not challenging, but it's interesting that here we had a good number of hits having known antibiotics of a couple hundred in that data set. And we weren't looking to get 100% true positive rate. We would have loved it. In the end, when we tested it, we had about a 51, 52% true positive rate, which might sound small if you're trying to differentiate a picture of a cat from a dog on the interweb. But really good for looking at prediction. Did you come up with a new antibiotic? where usually for a random screen or large screen, it's well less than 1%. So yeah, I was surprised how well the model performed, pleasantly so. And I think it speaks both to the value for positive data in these compound structures and the fact that it was really rich and enriched for antibacterial in that case. So the question of discretizing the data or not. Yeah. Again, I think if you had said to me, hey, I've got this relatively small data set and I've got these measures which range from zero to one for how much a given compound inhibits the growth of the bacteria. Should I discretize it or should I try to train the network to predict the scalar quantity? Yeah. I think my intuition would have been you should try to predict the scalar quantity and then maybe like apply the threshold at the end. I imagine because these models are not huge, right? It wasn't like you only had the compute for one run. So I imagine you probably tried both. I suspect we did. And I can tell you that if we did, the data weren't good. I mean, the results weren't good. So that, you know, doing a correlative model, there just really wasn't sufficient data to get predictive capacity now from a completely new structure of where we sit on that line. But given that I think with differences is that getting after now that discretization of a molecule that really would inhibit growth, you're now, I think, getting after specific structural features of the compound that really make for a good antibiotic. Versus, and then I think if you got enough data, if you had enough compounds, I think you could predict where you'd sit on that line from zero to one. but I think you probably in that case, we'd need many, many hundreds of thousands, if not millions of compounds to fill that out. Here, we enrich a lot with no effect and then a good number with effect. And so the training set itself was kind of binarized. Yeah. It's really interesting. How much does it cost to collect this sort of panel data in the first place? Like if you wanted to set out to do tens of thousands, hundreds of thousands of compound to bacteria? So there's two or three levels of cost. One is actually curating, buying the compounds themselves. So in this case, the initial one we had, I think we had the library available. We then subsequently, as part of the antibiotics I-Priot, put together a library of 37,000 additional compounds. And I think that cost us about $150,000 to put together. So it's still not a large amount when you think there's a board of order, you know, $5 per compound. If we go to any of the vendors, you're going to be anywhere from $10 to $20 per compound in a larger bit to get out to then specialty ones that could be able to order $100 per compound. So it goes up very, very quickly. Now, for some of these larger public health challenges, there are compound libraries available in pharma that include about a million molecules or so and maybe larger, but they don't make them publicly available. I wish they would in some cases. Maybe they've already mined them for the features they like, but to make them available for things like Interbox. The second level of charge is now how do you screen it? So when you start getting to these tens of thousands, that's a lot of work for a grad student. And so it's one of the few spots where we'll use robotics, liquid handling robots, but they're costly. And so it probably runs us of order $20,000 in robot time to screen a 40,000 compound library. I wouldn't say that linearly scales when we go to a million. And so, but it's decent cost. So, for example, our 40,000 compound library, we've now applied to seven different bacterial pathogens and three different human cell lines. So we've done this 10 different times. I mean, honestly, that's like astoundingly little money in the grand scheme of things, right? I mean, in the world of AI that I'm following on a daily basis, we've got, you know, in the space of the last couple of weeks, like $100 billion from NVIDIA to OpenAI and Oracle. And, you know, it's like tens of, it has to be tens of billions or you're not even making the news. So, I mean, we're talking like a full three orders of magnitude less to do sort of the biggest scale versions of the experiments you're running and probably four orders of magnitude less to do some of the ones that you actually ran. And that is, you know, affordable, as you said, for. It's affordable, yeah. So I'll frame it along those lines. I saw an announcement on the news this morning that XAI, Elon Musk's company, was in the midst of raising $20 billion on their most recent run. That's the number I just quoted, could solve the antibiotic resistance crisis for the coming many decades. So you have young kids, I have little kids. For the lifetime of your kids, that could be solved, which is stunning for a single, in this case, private company. If you now break it down, it's anywhere from 500 million to a billion to 2 billion per drug to be developed. I think depending upon how one says it, you might get that even down to 100 million in some cases for certain drug cases with orphan status. So, again, I think if if there are some wealthy individuals there that are publicly spirited and and recognize the need of public good, I do think here's a great example of AI for good. And it's a great example where with additional capital, I think we can take these compounds into patients and actually begin to expand our armament, our portfolio to go after these superbugs. When you talk about solving the whole thing big picture for decades at $20 billion, I guess like how many, which by the way is under a tenth of a percent of US GDP, another way to think about it. How many drugs are we talking there? Is that like 20 drugs at a billion each, including all the clinical trials and all that kind of late stage stuff? You know, it's a border 15 to 20 drugs would be the real pitch here. And, you know, in fairness, it's not that everything's ready to go if Elon Musk wrote a check for 20 billion. I think we need to put in a certain infrastructure, get things in place. But, you know, when you look out of what we could do, I think it's it's again, it's an interesting, for some reason, oversight that hasn't raised to public consciousness levels at the level that should. And the term existential has kind of gone out. It's been set up as being overused. But if you look at individuals' lists of existential threats to humanity, antibiotic-resistant infections is on that list for most individuals. And it's the cheapest risk that could be solved on that list. Whether it's global poverty, hunger, climate change. Those are multi-trillion dollar problems. This is, you know, avoid tens of billions, vote tens of billions. So it's, you know, we have some work to do to convince folks. But again, it's, you know, for those who really want to make history without worrying about making money, I think it's a good one to go after. And I think AI is becoming, again, an interesting way where it's making it interesting. I think raising the attention to your community, to my community, where, again, it's a true, beautiful example of AI for good. Yeah, it's an area I would also love to see the US and China decide to race. And, you know, I propose like... Maybe even team up, but race, I'm a very competitive guy, so race would be marvelous as well. Yeah. I mean, I would love to see, this is a whole other digression. I'm what passes for a China dove these days in the sense that my outlook is like, I think this AI stuff is going to be a really big deal. And we might need to work together across the US and China to end up in a good place. The alternative being we raced to weaponize and created a whole new sort of Gamocles. And that all sounds terrible to me. But obviously, the spirit of competition is high and rising. So it was kind of like, maybe we could make a metal tracker for new drug discoveries or something like that. I know that the CCP loves to collect gold medals. So if we kind of create more gold medals for antibiotics and similar, that could be a good thing. I guess one more point on the money before going deeper into the techniques. The money to actually develop the drugs is, again, like a very small amount, right, compared to all the trials and downstream stuff. That's a fair point. Yeah. So I think that it's of order of millions per compound to develop it preclinically before trial. Low millions. would probably be a fair number. And really the cost comes in when you start queuing up your phase one, phase two, phase three trials for antibiotics. But of order, you know, low millions from early discovery hit to lead optimization is, I think, a decent estimate. And so, for example, we're working with Fairbio. Fairbio is a nonprofit we helped launch as part of the Antibiotics AI project. And with Fairbio, we have fantastic support from ARPA-H, a federal agency. Together, we've received a $27 million grant to develop 15 antibiotics through preclinical development. So to establish a very robust pipeline driven by generative AI. And so, you know, looking at that, you're avoiding a little under $2 million per compound to get it through to being IND ready. Yeah. So just a couple percentage points down payment could take that to the scale that you'd need to stock the shelves indefinitely. Hey, we'll continue our interview in a moment after a word from our sponsors. AI's impact on product development feels very piecemeal right now. AI coding assistants and agents, including a number of our past guests, provide incredible productivity boosts. But that's just one aspect of building products. What about all the coordination work like planning, customer feedback, and project management? There's nothing that really brings it all together. Well, our sponsor of this episode, Linear, is doing just that. Linear started as an issue tracker for engineers, but has evolved into a platform that manages your entire product development lifecycle. And now they're taking it to the next level with AI capabilities that provide massive leverage. Linear's AI handles the coordination busywork, routing bugs, generating updates, grooming backlogs. You can even deploy agents within Linear to write code, debug, and draft PRs. Plus, with MCP, Linear connects to your favorite AI tools, Clawed, Cursor, ChatGPT, and more. So what does it all mean? Small teams can operate with the resources of much larger ones, and large teams can move as fast as startups. There's never been a more exciting time to build products, and Linear just has to be the platform to do it on. Nearly every AI company you've heard of is using Linear, so why aren't you? To find out more and get six months of Linear business for free, head to Linear.app.tcr. That's Linear.app.tcr for six months free of Linear business. Build the future of multi-agent software with Agency. A-G-N-T-C-Y. Now an open source Linux foundation project, Agency is building the Internet of Agents, a collaboration layer where AI agents can discover, connect, and work across any framework. All the pieces engineers need to deploy multi-agent systems now belong to everyone who builds on Agency, including robust identity and access management that ensures every agent is authenticated and trusted before interacting. Agency also provides open, standardized tools for agent discovery. seamless protocols for agent-to-agent communication, and modular components for scalable workflows. Collaborate with developers from Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and 75 more supporting companies to build next-gen AI infrastructure together. Agency is dropping code, specs, and services, no strings attached. Visit agency.org to contribute. That's A-G-N-T-C-Y dot O-R-G. Being an entrepreneur, I can say from personal experience, can be an intimidating and at times lonely experience. There are so many jobs to be done and often nobody to turn to when things go wrong. That's just one of many reasons that founders absolutely must choose their technology platforms carefully. Pick the right one and the technology can play important roles for you. Pick the wrong one and you might find yourself fighting fires alone. In the e-commerce space, of course, there's never been a better platform than Shopify. Shopify is the commerce platform behind millions of businesses around the world and 10% of all e-commerce in the United States. From household names like Mattel and Gymshark to brands just getting started. With hundreds of ready-to-use templates, Shopify helps you build a beautiful online store to match your brand's style, just as if you had your own design studio. With helpful AI tools that write product descriptions, page headlines, and even enhance your product photography, it's like you have your own content team. And with the ability to easily create email and social media campaigns, you can reach your customers wherever they're scrolling or strolling, just as if you had a full marketing department behind you. Best yet, Shopify is your commerce expert, with world-class expertise in everything from managing inventory to international shipping to processing returns and beyond. If you're ready to sell, you're ready for Shopify. Turn your big business idea into cha-ching with Shopify on your side. Sign up for your $1 per month trial and start selling today at shopify.com slash cognitive. Visit shopify.com slash cognitive. Once more, that's shopify.com slash cognitive. All right. Well, we can do another call to philanthropists at the end, but let's go deeper into the techniques because this is really where the AI obsessives, I think, want to understand what's going on. So a couple of things that jumped out at me about your techniques, and I'll just kind of give this to you as a prompt, so to speak, and then you can elaborate on what you think people should better understand. One point was that you're using graph neural networks with a convolutional approach as opposed to some of the new graph transformers. I was kind of interested in if you tried both or how you think about that, why that particular architecture. I also noticed that you train a number of them. I think it's 20, if I understand correctly, identical convolutional graph neural networks basically as a way I think to sort of because I assume they like all differently randomly initialized as a way to sort of avoid any one of them kind of overfitting or going weird. And then you sort of ensemble all of those to actually make the predictions. What else do we need to know about the actual architecture of the networks? So, you know, you're spot on, right? So, you know, I think the reason we chose the GraphenolNet with ConvolutionalNet as the only, was that this was the platform that Regina and Tommy's team had developed under the banner of ChemProp. So this preceded our work on antibiotics. So this was efforts done by Yang in their lab, and I think Kyle Swanson and Wing Gong Jin, and really a marvelous platform. We started this probably, it was late 2018. Transformers were just beginning to appear at the time weren't really yet that popular. Subsequently, and maybe just to speak up, so our strategy is generally we will create an ensemble of models trained similarly, I would say in fairness trained identically but with different initial conditions that we then will average across the ensemble, kind of wisdom of crowds of sorts. we subsequently in more recent work we've explored large language models that have been developed so these are not graphical neural net per se but language models where you'll look at a string of symbols from a compound structure and they've done okay so many of these are pre-trained on large libraries of compounds but they've not yet outperformed our graph neural net we have seen they seem to be learning a slightly little different schemes and they make predictions that are a bit different. We have explored can we do a multimodal or hybrid model with a little bit of success giving us a little bit of a bump but not considerably high. More recently we've actually implemented Minimole which is another graphical neural net that is pre-trained using quantum mechanical calculations considerations and work that we'll be submitting soon shows that It significantly outperforms chem-prop, this earlier graph neural net. So it's interesting. Here we really are, I think, accounting for that these models were set up in part that they could look at graphical representations. And it fits beautifully for the compounds that we're looking at. You can think of back to your chemistry class you had in high school or college. Those types of structures were feeding it. Now, in many cases, the model is considering a 2D representation of the molecule. And we are also beginning to think about how we can better take advantage of 3D representations for improving and extending the predictions. I can't resist the follow-up question. What are the foundation models that you have looked at? The ones I've studied are like EVO, EVO2, ESM fold? Yeah, so EVO and EVO2 are basically genomic ones on DNA. They ESM are protein-based. So these, we'd be looking at language models that were set up specifically for small molecules. So we looked at Kembert was a dominant one. We looked at, oh gosh, I'll think of the name, but we looked at NVIDIA has their own. IBM had a version. So we've looked at the leading cases. Evo is not well set up for what ours, because we're doing small molecules, as well as ESM. No, but in each case, we are intrigued by each of those models for how we can apply for some other things we have going in the lab. Oh, okay. Cool. There's maybe another question around, but I'll save it a little later, around kind of bridging these modalities. Yeah. But, okay, so we'll keep going through the process. So we've got these ensemble of convolutional graph neural networks trained. And then the big computational step, which I understand is still not all that big, is taking increasingly larger and larger libraries of, again, if I understand correctly, both real and hypothetical molecules out of these big libraries of molecules and just crunching through literally millions of them, tens of millions, I think even over 100 million in one case, to get all these scores and say, okay. here's the ones that are predicted to kill this particular bug. There's also an interesting Monte Carlo tree search kind of algorithm that seems to, if I understand correctly, like cluster molecules in sort of not functional space, but sort of physical space, right? Those that have kind of similar structures are clustered together so that you can not only get the prediction, but also kind of look at, oh, there's a cluster of things here that are all predicted to work well, and they have something in common. Therefore, that thing that's in common seems to be the key that's actually driving. And that is really interesting, too, in terms of just an alternative approach to explaining what's going on in the AI system. Yeah. So first, yeah, with the trained model, then we'll feed it structures that are then from in silico libraries that are either been curated of ones that are available for purchase in some cases and or can be synthesized. In other cases, might be kind of arbitrary. They think we can synthesize, but we're not really sure. In that initial Halison piece, we screen computationally about 110 million compounds, So enormous library from a real world standpoint that you would never empirically screen the lab. But we did it over a course of three days on the computing platform we had at the time. Since then, just as an aside, we have been doing a lot of work with Edamine, which is a small chemical synthesis company in Kiev, Ukraine. Obviously quite occupied with the war going on right outside their synthesis company. But they've been great partners and notably they have much larger in silica libraries. They're real spaces of order 65 billion to 70 billion. And we've been screening those now, which now, you know, it's getting up there, right? You're several orders of magnitude larger than that initial 110 million, which itself was very large. The second space she was speaking to was an effort that was led by Felix Wong, which is around the idea, could we get after explainable AI to better understand structures, common structures amongst the best scoring molecules, enabling us to better identify novel structural classes? Meaning, could we set up in this case, which we did, was a monoculture research to look at rationales or substructures across the top predicted compounds to see is there a substructure rationale overrepresented that would suggest that maybe we are on to a new class that goes after a similar mechanism but has chemical diversity of some sort within that class, increasing the chance that we really did come up with something new and meaningful. And we published that in Nature a little over a year ago to a lot of interest, And it really turned out to be quite a powerful approach that gave us insight into the chemical structures that have been learned by these models that could matter in the antibiotic space. Yeah, I find that really interesting. I find all of this really interesting because it is quite a bit less black boxy at a few different points than what I'm used to when I just look at the big foundation models that kind of dominate my consciousness most of the time. Um, and that is a really interesting alternative way to try to make the AI process explainable. Yeah. So, okay. The, the, the move from, you know, tens or hundreds of millions to tens of billions of molecules definitely is a huge leap. Can you give me a little bit of an intuition for chemical space? I actually was a chemistry undergrad, but I am embarrassed to say that I don't have much intuition for this. At that time, we were like working on natural product synthesis. And I understand there's been like a pretty big shift in the field from pick one molecule and, you know, do whatever it takes to synthesize it to a more kind of sane approach that is more like building block, kind of almost Lego style approach to putting these things together. So, you know, a few levels. So, you know, an academic lab like ours will typically have compound libraries in order of 10,000, tens of thousands. Large research centers will have libraries of orders of hundreds of thousands, maybe a million. Pharma will have libraries of order of millions, low millions. And when you said libraries, these are on hand. These would be compounds that you have in little bios. So, for example, the Broad Institute, they have a center for CDOT, Center for Discovery of Therapeutics, that we work with. And they have this beautiful robotic system that has a voider 800,000 to a million compounds on site that you can program via barcode to go have your liquid-hailing robot go grab, come it out, and then couple it to another liquid-hailing robot that can apply to our bacterial cells. So there is kind of the physical world compound, typical order number. So I think, for example, Eden has maybe four and a half million molecules stored, ready to send you right away. Now you go into the in silico space and maybe you touched nicely on your building blocks. At Edamine, their 65 billion is based on a building block, set of building blocks and a set of recipes, synthesis steps to get there. So I don't know what the number of building blocks is, but it gets you to the 65 billion and they're confident they can do those. So that's the order you have there. So you're in 10 to the 10 compounds. Going back again to your chemistry days, the estimate is that there should be avoided 10 to the 60 compounds that I'm not sure fall into the drug-like space, but let's just say 10 to the 60 compounds, which is more than astronomical. And we're literally just scratching that surface with, say, I mentioned the 65 billion, I may be off by a little bit there, their unreal space at Edomine, I think is now 220 billion. Okay, so you're now 10 to 11, still not even close to 10 to the 60. Yeah, 10 to the 60 is really hard to wrap one's head around. Yeah. What's out there? Is that like, is that, how real should we think of that as being? You know, I don't know. You know, I was just in a meeting with my team and they claim that there is a group, Biosoft, I think is the team, that they can get to trillions of molecules from their building blocks. Okay, that doesn't sound too crazy given either me and as confident with the 220 billion. It's okay. Order five more to get to a trillion. Yeah, I get it. 10 of the 60, I've never seen the calculation. I've heard the number. I've passed along the number. I've seen it in many different spots. I don't know how pragmatic or real world that number is. But a border trillion seems okay to me for now. It will be interesting to see to what extent AI can help us better define what the real space of possible compounds are. one of the challenges we had as we moved from discovery to design is can you actually synthesize what the model made and there is a big challenge and one of the best in the business is one of my colleagues here at MIT at Connor Coley who's developed AI models to predict the synthesizability of a compound and there you want to get after the answer to can you synthesize and then you want to ask can you do it in a reasonable number of steps can you do it for a reasonable amount of money Yeah. So I guess regardless of how vast the chemical space ultimately is, safe to say it's pretty big and you're turning through an increasing fraction of it with basically a pipeline of steps. You've got the prediction as to whether or not it's going to kill the bug. And then you're also applying, I guess it would be filters or classifiers or scores. I'm not sure if these are applied like in sequence or all at once, but you're scoring all the molecules on how novel they are, like how different they are from existing antibiotics. Their predicted level of toxicity to human cells, which is obviously important. How stable they are, which is obviously important for like putting them into a pill and shipping around the world. as you said, how easy they are to synthesize. How is that happening? Is that a sort of sequential thing or is there one kind of heuristic that weights all those things and gives you an order? So we, in the past, have done a kind of multi-step application of the models and our filters. We've begun to explore multi-task modeling and are there advantages to do multi-task modeling? I I think in the case where there's overlap in features being common across, say, the pathogen, or we've seen some advantages. But for the most part, it's a matter of coming in with these different models in sequence, really to cut down on the space, cut down on how can you move from this interesting thing in the computer to something you really got to pay to get and actually test. We're expanding that. So as you alluded, we've got, I think, really good models on antibacterial. We've got good models for toxicity. Easy to calculate. Is it different from an existing antibiotic? We need more data to get after drug-like properties beyond toxicity, solubility, bioavailability, lack of metabolic liabilities. And we're working hard with FairBio to get those data to train what we're going to probably call drug prop AI. So you can get after drug properties and use this to really increase the chance that the molecule would make for a good drug. That would be effective not only in a mouse but in a human. And I'm sure your listeners have heard many times the comment that we've cured cancer 900 times in a mouse or a rat, but very few times in a human. In my world of antibiotics, if the molecule works in a mouse model effectively from a systemic delivery standpoint, the estimate is about 90% chance that it would also work well in a human. Our AI models to date are really good at coming up with molecules that will kill the bug in a dish. they're not as good as they need to be yet of also predicting are they good at killing the bug in the mouse and that's where we need these additional data additional steps to move the power of ai further down the antibiotic development pipeline and i guess one other question there is is there any role for human taste or judgment in this process absolutely Yeah, no, absolutely. It's a really nice question. I think the medicinal chemists remain still pretty skeptical about the value of AI. In part, they feel that there are decades of work and expertise that identifying molecules, modifying the molecules, tweaking the molecules, can still outpace AI. And I think in some cases they can. Now, in fairness, there's no medicinal chemist I know that can review 110 million molecules. but I do think a few points to this we ran an interesting case where we just had a really talented medicinal chemist as a postdoc in my lab and we pitted him against the synthesizability model the kind of college group account but so we gave both the model and postdoc 30,000 compounds to see how they were ranked and compared it to then what edamine would say and the human outperform the model. So then the next level is, you know, how do we better capture this intuition? I do think there's opportunities to recruit panels of medicinal chemists and develop kind of the medicinal chemists in the loop with the reinforcement learning with human feedback to improve the models for these different features that they are able to capture and maybe not necessarily describe or quantify. There's the interesting antidote, which I'm sure, where it was Google D Mime was developing AlphaGo. They had trained the model on synthetic systems and even human systems, but really began to make a big leap when they actually brought in an expert, I think he was maybe the thousandth ranked player, who could give additional human insight of strategy and moves and it elevated their system to a new level. And I think we need to figure out how to do that better at drug discovery. So that head-to-head, when you say 30,000, does that mean that the human expert actually went through and scored 30,000 things? So it was discrete. So the human expert right-swiped, left-swiped, 30,000. Holy moly. And so I've been happily married for 35 years, So I don't know what left swipe, right swipe means. I know one is, I like one I don't like, so I don't know which is. But holy moly is right. That was so impressive that this guy did this. And it was interesting when I asked, Andreas Luton is his name. He's a talented now young professor at the Karolinska Institute in Stockholm. And I asked him really what was guiding him. And he was really looking for liabilities. So he would say, bad, bad, bad. And if he couldn't find a liability, he said, good. And so that was an interesting take that is not how I would have set up my model to do. And so, again, I think it's examples such as this that I think we need to do a better job. Now, there remains some significant dismissiveness, but maybe some rather so, and some hostility from the medicinal chemistry towards these types of approaches. And I think we need to turn to them for help to make them that much better and do a better job of capturing the medicinal chemist in our code, in our models. And just to make sure I understand where that is in the pipeline, these liabilities are for a molecule that has been identified as likely to kill the bacteria. It's like, oh, but I can spot that this is going to be hard to synthesize or it's going to be not available. I think in this test, it was really only around synthesizability. but you can imagine it'd be great for me to also be able to do something similar of you know would this make for a good drug will it will it be stable will it be available at the side of the infection will it not be cleared too quickly will it not be broken down by the liver too quickly um and you know these are things where you know you have many medicinal chemistry friends but they can kind of look at a molecule yeah i like that molecule or oh that's an ugly molecule for xyz We need to sit these guys down and get that. I'll give another related example, which you don't hear as much about anymore. But pre-pandemic, I remember I was meeting with a number of different folks from China coming through the MIT area. And I was with someone who was either working for Baidu, I had insight into Baidu. And Baidu at the time had hired 15,000 individuals to label data. So this goes back to about 2018. Supervised learning was the big thing. And so labeling featureizing data was key. And in that they had 2000 medical students they had hired that were labeling medical images, might be pathology. And I thought that was brilliant that a country could or a company within a country could have the resources that could take advantage of that. I think we don't need something of that scale, but I think figuring out how I could rally a group of medicinal chemists to commit to letting me peer into what they've learned and how they act on what they learned could make a big difference. yeah honestly it's really striking how well this seems to work given how the little data you start with and you know just how how coarse that original signal is i mean the whole thing is because it is worth repeating too like again if i understand correctly there's no encoding of the target bacteria at all right it's just like literally could be anything it's it's totally abstracted away into a, it works or it doesn't work. Yeah. Zero one. Yeah. That's amazing. Yeah. So when things work better, what does that mean? Like, I guess I'm, I'd like to understand a little bit better the trade-offs between you could do, you know, you've got like compute costs and then you've got synthesis costs, right? I guess those are kind of the two big costs in getting to the end of something that you're like, okay, this actually confirmed kills a bug. Now we got to go into the sort of medicinal chemistry phase. Yeah. How do you think about those trade-offs? Like if the model works a little bit better, but maybe it's also a little bit bigger, you know, maybe we do like 10 times as many molecules. Yeah. The compute cost isn't very large. You know, you can kind of, as a fixed cost in the background, you know, I do think the other, you know, getting after larger training libraries is expensive Compute not so much Synthesizing is the big challenge right Because you making a commitment both on just outlaying the money the time and what the probability you got something that good The next is that animal models. Animal models are not inexpensive. And you've now got to decide which of the ones I'm going to advance and is it good enough to go? If that looks good, now you're back at, okay, I've got to now maybe do an analog generation. Which of those will I synthesize? And it becomes is interesting within a little academic lab like mine, how much of it are you really to commit to? So it's the trade-offs. Given our experience on generative AI and the challenges of synthesis, I'm much more inclined towards looking at these libraries and molecules, like getamines, where I'm guaranteed I can synthesize. Still might need to pay a bit more for those, either difficult to synthesize or not readily synthesize, but I'm more comfortable there than coming up with an exotic new molecule that I'm not sure I'm going to synthesize. And what's the probability that it's going to work? So can you give a sense of where we are today in terms of like, okay, you start with millions, tens of millions of molecules out in chemical space. You do the predictions, you do all these filters for the novelty, for the non-toxicity humans, for the stability, for the synthesizability. It seems like we get out of that process like dozens and then you synthesize dozens and then like we get down to the end, like a couple actually work, at least in a mouse. Is that basically the right? Yeah, I'd say that we probably get down to many hundreds before we try to synthesize would be probably the fair set of filters we've set up. And then after that, you were correct. And so I think that we can expand the starting point of libraries, of course, but I think it's where we have not yet gone is that really from the early hit to a so-called lead development. And that's where can we get after more drug-like properties, so-called ADMETs, if absorption, excretion, metabolism, toxicity, distribution, as well as PKPD, so the dynamics of the drug. it's where I think AI, again, most of your listeners talking in the AI space, I think AI has done a really nice job in early discovery efforts, certainly antibiotics, but I think in other drug spaces. AI has not yet really been utilized much further downstream, I think because of lack of data. And it becomes interesting of where are you going to get that data? Who makes the commitment? And, you know, it's expensive. And I think the companies that make those commitments of the companies are going to have advantages going forward. So in terms of improving the models, the big way that that – because you're already seemingly quite successful. I don't know if there's any negative results or any strains you've tried this on where it didn't work. But to read the papers, it seems like the general workflow of identify a target bacteria you want to be able to kill, run the panel of all the test molecules against them, train the network, do the pipeline, apply it to the huge swath of chemical space, get candidates out, synthesize those. It seems like that is pretty consistently working. Is that right? I think that's fair. I think with an acceptable success rate, it's decently working. I think our goal always is, can you do things even better, even a higher success rate? Could you really identify those compounds that, boy, that's really a great starting point? And And I think we have a little work to do there because we still need to get after those other drug-like properties is kind of the key thing. And figure out how do you accommodate for multi-objective optimization across these compounds. And I'm hopeful we'll get there, but we're not there yet. Yeah. But it sounds like from the standpoint of society, to the degree that you can improve the models and get even more confident predictions, that would effectively reduce the cost of the development process from the end of – from where you are like, okay, it works in a mouse to the point where we actually get into clinical trials. Yeah. Yeah. That is a cost that should be great if it was cheaper. And of course, you have like intellectual interest in finding new techniques to make it better. But from society standpoint, it seems like it already works well enough that we should just clone your lab 10 times and like apply this to as many targets as we can basically immediately. Yeah, I think that's fair. I think that our group and other groups have shown that AI is a valuable tool. And I think it needs to be utilize more. I do think for sure in the early discovery, it's reducing cost, increasing hit rate, success rate, and thus increasing our chance of getting to new molecules that can make a difference for human health. There's a couple of other aspects of the results that we've barely touched on too. So again, it's worth just repeating. These are drugs that are working against drug resistant strains, right? This is not just that they work, but they work on things that other drugs don't work on. Key point, worth emphasizing again. At least some of what you found has been shown to work with a new mechanism, meaning it is working by disrupting the bacteria in a different way than other molecules. And it's even, oh, there's also specificity. So again, at least some of what you've done here has shown to work against the target and not disrupt the other quote-unquote good bacteria. Maybe I'll just comment on that briefly. So that's worked in several cases, and that was surprising. And why it was surprising is that the models were not designed to yield a so-called narrow spectrum. So narrow spectrum antibiotic would be one that goes after the pathogen of interest but spares the commensals, or good guys, as well as other pathogens. In fairness, the models, we've only trained against a particular pathogen. So they were only trained against the pathogen, but they weren't then counter trained to avoid the other ones. And yet, in the case of a bouset, which was a molecule we discovered to be effective against this interfective amani, in the case of also a molecule we discovered that was designed effective against gonorrhea, in each of these cases, they were narrow spectrum. And so it was intriguing that I think we got it more or less by luck, but that the model pointed us to these molecules that were sparing most of the good guys in our gut. Yeah. I guess if I had to attribute that to something, it would be the novelty filter. The theory would be like, if it's very different from other antibiotics, maybe it's more kind of particular to the target, even if that wasn't like explicitly. Yeah, maybe that's an issue. We haven't, you know, it's an interesting notion. It is possible. Um, you know, I, I'll, I'll, I'll put an esoteric spin on it that we're doing, um, phenotypic screens. And in many cases, I think we're getting after membrane acting, uh, antibiotics. So they're targeting targets in the kind of outer layer and inside layer. And it appears that many of them are getting after lipoproteins and lipoprotein transport, Again, kind of an esoteric point. But I think that the narrow spectrum aspect of compound may be that these lipoproteins are very specific to the given pathogens. So that while we're not targeting a target, we're targeting the pathogen and looking at the phenotypic screen. I think we're selecting for compounds that are getting out these lipoproteins that are specific to the pathogen and not to other pathogens. And I think that's what's happening. And so we're also now beginning to, can we use AI specifically now to start with the lipoprotein as a target? And then can we find small molecules that would interact with those targets, those protein targets of interest? Can you speak a little bit also to resistance resistance? Like the, that's another notable finding that, again, at least some of the drugs you've discovered. So yeah, that's a fair point, right? So halicin, for example, we compared it to Cipro. So Cipro is a very commonly used quinolone antibiotic and applied each for over 30 days inside a lab to E. coli. Within a few days, there was significant resistance to Cipro. Several fold, then after 30 days, there were many hundred fold levels of resistance to Cipro. When we applied halicin, after a few days, we didn't see any resistance. In 30 days out, we didn't see any resistance. They said earlier, we'll eventually see resistance if we look out long enough. I think that the resistance to resistance of halicin was likely that it's targeting multiple molecular targets, so multiple proteins, probably at the membrane level. And again, to my earlier points where I think we can also use AI with this as an intent, intent to goal, intent to goal, is that because it's hitting multiple targets, the bug can develop mutation against one or two. but because maybe three or four and any one of those can be a killer to the bug. I think it just puts off the development of resistance. Fascinating. Okay. So kind of zooming out for a second and then, you know, talking about the future and how these methods might evolve. I mean, I guess my, I want to emphasize my general sense that like this should just be scaled up even as it is. We don't need to worry about getting too much more clever. You know, give me the warp speed. for new antibiotics. Um, but that said, cause it's interesting if nothing else, I'm interested in, I guess like, when do you think this could have first worked is, is one interesting question. I'm not sure, you know, I, I'm always reading like new AI and ML literature and it's always like some new technique, but I'm not sure here, like what was the limiting technique or were all these techniques kind of sitting out there potentially for a few years before you came along and figured out how to integrate them? I think they probably were sitting there for a few years. I do think that these graphical neural nets were the real trick. I think the ability of these models to learn chemical structures and break them down and associate them with a feature of interest, not because antibacterial are not toxic. That was a differentiating tech. And we're now expanding that to other features and other schemes. And that, you know, the deep neural nets, you know, were introduced, you know, of order many years ago, but really became quite popular about 10 years ago, largely on image analysis out of groups like Yashiro Bengio and Jeff Hinton, Jan LeCun. So I think that was a differentiator that we could. I think it can be scaled up, I think, quite nicely. I think we need more data. I think we need more talent. I think the models can be extended in clever ways. I think they need to look at bigger chemical space, associate and bring in more biological data, more chemical data to get after mechanism, to get after features, using more and more generative AI to get after design properties. So can we design the molecules specifically to go after multiple targets from the get-go? I think all of this is possible. And I think we'll see developments along these lines in the coming few years. How far could this go in terms of other things? Like you mentioned, we've cured cancer, you know, in mice lots of times. I mean, obviously a human cancer cell is presumably a lot closer to other healthy human cell as compared to bacteria versus the human cell. But could we imagine a similar technique working for human cancer cells? Yeah, most definitely. This and I mean, human cells and for other conditions other than cancer as well. So Felix Wong, a postdoc in my lab, who's also leading integrated biosciences in the Bay Area, is focusing among many things, but it was also age-related conditions, aging, and using this platform, identify molecules that It could act as senolytic, so they could eliminate so-called zombie cells amongst our cells. So these are cells that have stopped dividing that are thought to underlie neurologic conditions, scarring, skin conditions. Worked beautifully. He's exploring the approach now to many other schemes and other conditions. Within the infectious disease space, we work with folks to do it for antifungals, antivirals, antiparasitics. the more complex, I think there's potential for cancers, potential for neurological conditions, potential for metabolic conditions. So this and related, both phenotypic and target-based screens, I think has tremendous potential. I guess if you imagine applying the same kind of core approach to all these different things, and for some of the more challenging ones, it maybe doesn't work at first or it doesn't work as well, what kind of enhancements do you think you would need? Earlier, we touched on like Evo and Evo2. And I know those are sort of, well, first of all, Evo1 was like only bacterial. But there's an interesting possibility there where these foundation models can potentially be used to identify targets. There's like, of course, lots of different models now that do like, you know, all kinds of different molecule binding, small molecule to protein binding. you know, if you sort of had to imagine a next generation pipeline that kind of brings in either some foundation models or other specialist models, like what kind of elaboration would it be? Yeah, I'd say it's multiple levels. So I do think, you know, AlphaFold for predicting 3D protein structure certainly was a major advance. It's been widely used, but it's not really good enough yet for drug development and that the 3D predicted structures are not at a fine enough level that you could use it in a target screen. We need work there. Our binding predictions for affinity of small molecules against an identified binding pocket are not there yet. They're not as good as they should be. So I think if we could advance each of those, now we could do so much more in silica. From a target ID, from understanding how the drugs act, we need, I think, to better use AI to embrace the complexity of underlying biology. We've been fixing single targets, but these targets operate in very complicated networks that vary depending upon cell type, that vary depending upon context. I think we need to develop tools that can embrace that complexity to identify the meaningful targets and understand how the drug would interact with them. I think interfacing those layers, and then it's also how do we get to phenotype? Can we make predictions? So right now, we're not very good at predicting functional phenotype from our interactions. Now, I'll give this a case. There's a lot of interest, I'm sure you've seen in where you are in AI scientists and AI-driven kind of automated labs. And you were saying they never replace all scientists. And I think it's way premature to make these claims. But they're being set up in my world. I should do things where we know exactly what needs to be done. And so here's your recipe. But so, for example, in E. coli, you know, this very simple organism, model organism, it's been studied for decades. When I first moved into molecular biology 25 years ago, I was told don't work on it. by big, big people in molecular biology, because everything's known. Well, it's a bug with 4,000 genes. You look at the genome, 1,500 of those genes, we still don't know what they do. And so if I now give AI the challenge, okay, functionally annotate each of those 1,500 genes and run the experiments to validate it, we can't come anywhere close to that. So I think that AI has tremendous things to offer. And we've got these other types of models I alluded, but we also have a ways to go in order really to take advantage where I really still very much believe in AI as a thought partner for us. And it's still keeping the human in the loop is critical for many of these advances. Cool. I don't know if I have any additional followups there. One question I always want to touch on, at least for a second, is possibility of dual use or safety concerns. And one thing I do like about this approach and the sort of not just ensemble, but like pipeline nature and sort of many specialist models is it seems like it's not really prone to a dual use problem in the same way that a lot of other much more general purpose techniques might be. Maybe I'm missing something there, but what would you say is the sort of risk, if any? I think for the most part of what we're working on, it's kind of single use. Can we help humans from these problematic pathogens? There is one interesting case of dual use that's worrisome, and that is for these tox models that we developed. So here we're developing models to predict toxicity of a compound against a human cell or set of human cells. And we're mainly interested in identifying those that are not toxic. Well, those same models can be used to identify molecules that are toxic. And so now you can imagine bad actors using them to identify molecules in natural product space or design them in such a way that they're highly toxic. And in particular, it would be problematic as if they develop ones that would get after mechanism toxicity for which we don't have countermeasures. So that's a, unfortunately, dark but possible dual use that we hadn't thought about until we actually published our toxicity models. that I alluded to back in 2024 in a nature piece. And some of my friends in the federal government came and said, Jim, you know, are you worried about this? And I said, boy, and I had to admit that we hadn't really thought about it, but we have since thought about it. Is there anything that can be done to create versions of the model that don't have that sort of reverse the sign? No. No, right, because you're going to get a score. I mean, I guess it could be that you only score those that are not toxic and then you don't know how toxic. I guess you could, but anybody could easily pull apart the model and just shift the threshold on that. Yeah. That's kind of the, yeah, I could imagine interpretability techniques there could be sort of applied for bad, even if you, and I'm generally a big fan of interpretability. But if you did have something that sort of only, you know, I put a score of 0.8 or higher or whatever, but I asked everything below. Presumably the model inside still does have a representation that you could access if you were determined enough to do it. Yeah, yeah, tough point. Cool. This is fascinating. And I really appreciate all the time and remedial education that I've got from you. Which I agree with you. Thanks for the great, great conversation. Is there anything more that we should say about just, you know, call to philanthropists? Like, where are we on warp speed? You know, I think I encourage young people to think about this problem. And from the AI standpoint, from the microbiology standpoint, from the drug discovery standpoint, these are exciting, practical problems that can make a big difference. And we need more young people to take them on. I think it's a great time to be a young person in science. It's okay. Half the world seems to be on fire, but it's still a great time to be in the young, in size young person because so many cool tech technology being developed and we need young folks to take on this problem. So that would be my call in addition to the philanthropist to at least think about, you know, can you make a difference here? Cool. Well, I hope that a couple of listeners are inspired and might go a little bit in that direction based on your example. Again, this has been fantastic. I really appreciate it. Professor Jim Collins, thank you for being part of the Cognitive Revolution. Great. Thanks for having me. I really enjoyed our discussion. If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, CognitiveRevolution.ai, or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts where experts talk technology, business, economics, geopolitics, culture, and more, which is now a part of A16Z. We're produced by AI Podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at AIpodcast.ing. And finally, I encourage you to take a moment to check out our new and improved show notes, which were created automatically by Notion's AI Meeting Notes. AI Meeting Notes captures every detail and breaks down complex concepts so no idea gets lost. And because AI Meeting Notes lives right in Notion, everything you capture, whether that's meetings, podcasts, interviews, or conversations, lives exactly where you plan, build, and get things done. No switching, no slowdown. Check out Notion's AI Meeting Notes if you want perfect notes that write themselves. And head to the link in our show notes to try Notion's AI Meeting Notes free for 30 days.