AI 发现抗生素：小数据与小型 GNN 如何带来大成果，访 MIT 教授 Jim Collins

Hello and welcome back to The Cognitive Revolution. Today, my guest is Jim Collins, Termeer Professor of Medical Engineering at MIT and 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 1 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 his team seem to have created not just a few breakthrough drug candidates, but a multistep, AI-powered process, which they've refined for the last 5-plus years and proven by applying it to several different bacterial targets. It 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 when I was 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 graph convolutional neural networks on data sets consisting of just a few thousand chemical structures, with data on how effective each one was at stopping the growth of a target bacterium, the team was able to create a model that could screen tens of millions of compounds for efficacy in just a few days. Then, by using these predictions as part of a pipeline that also scored candidate molecules for novelty 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.

Upon actual synthesis and testing, that set did contain hits that were not only effective against the target strains but, again, at least in some cases, worked via previously unknown mechanisms and without harming other types of bacteria. These compounds are now moving toward clinical trials.

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. The entire process, including the clinical trials required to get them approved, would cost maybe $20 billion.

By the standard of recent data-center buildout 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 superintelligent 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. Those wins 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 data sets and just a few GPUs, is accelerating the discovery of life-saving drugs with MIT Professor Jim Collins.

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. 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 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 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, could we just do a little primer on the biology of antibiotics and antibiotic resistance? How do antibiotics work, and why do they stop working sometimes?

Yeah, antibiotics are generally small molecules, like an aspirin, 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 would be associated with an important process, be it cell division, protein production, or DNA replication. This will disrupt that associated process, and we've shown that that will then lead to downstream stress responses from the bug.

Those stress responses will lead to energetic demands that will produce toxic metabolic byproducts, which will lead to additional damage inside the bug, damaging DNA, RNA, protein, membranes, and lipids. That will trigger additional energetic demands, and so on, leading to this cycle that will contribute to the disruption of the initial process.

Why resistance arises is that bacteria have, effectively, an intrinsic goal of surviving and replicating. You'll have mutations that occur—alterations to the DNA or other features—with some frequency every time the bug divides, 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 now have a fitness advantage, meaning they will survive, whereas the other members of the population that don't have it won't survive.

The others will die at that concentration of the antibiotic, but the ones with the appropriate mutation will live and survive the antibiotic treatment. Thus, they 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 prophylactically to protect the animals from potential infection, as well as as a growth stimulant, we've now seen resistance growing dramatically over the last few decades.

In the past, the biggest risk would be in hospitals—so-called superbugs. I tell my students, the 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 child-care centers, they're in our schools, they're in our shopping centers, and they're in our communities. The problem has escalated due to our overuse and misuse of antibiotics.

To double-click on that a little bit, one of my mantras for AI is that AI defies all binaries. The more I learn about biology, the more it seems like that's true for many aspects of biology as well.

One thing I was struck to notice in reading the papers is that I think a lot of times people think of an antibiotic as a missile that zooms in, smashes the bacteria, and it's just gone. As I was reading, I thought, it doesn't really look like that. The way that the effectiveness of the antibiotic is measured is not binary, but rather a scalar, right? It can be anywhere from not effective at all to very effective, or anywhere in between.

I'm not sure if this is right, but I was inferring from that that this growth slowing maybe implies that there's still a major role for the immune system. What is the actual role of the drug? Is it killing the cells, or is it slowing them down enough that the immune system can rally and destroy them for us?

Yeah, it's interesting. It depends upon the antibiotic and the type of antibiotic. The ones that we typically think of—the missile that you referred to—would be bactericidal antibiotics, 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 antibiotics, which are selected and developed to inhibit or stop the growth of a bacterial infection without killing the bacteria in that infection.

Interestingly, as you picked up in the piece, screening for antibiotics is almost always done with a growth-inhibition assay and not a killing assay. In a growth-inhibition assay, you apply some library compounds and look to see which of them significantly inhibit the growth of the culture.

Would that indicate that there are some interesting antibacterial properties?

In many cases, those are also associated with killing activity. Some are only associated with inhibition. A killing assay is so much more difficult to get after.

The third level that's also important, and I've already alluded to that, is that you want the antibiotic to largely impact only the bacterium and not our cells. Frankly, if possible, you also want it to impact only the pathogenic bacteria and not the healthy bacteria that make up our gut, live on our skin, or exist on other parts of our body.

Yeah, defying all binaries, indeed. One more primer question: I also noticed in the paper that you ran some experiments on these new antibiotics that you've found to test how quickly bacteria can develop resistance to them. I was struck that, in general, it seems to happen pretty fast.

Should we understand that this is always going on, even in our bodies, on an ongoing basis? I kind of saw this with COVID, too, where it seemed like there were a lot of mutations happening all over the place. Not all of them, of course, break out, but should my mental model be that these resistant strains are popping up all over all the time, mostly not going anywhere, but occasionally getting out of control? Is that the right way to think about it?

Yeah, I'm not sure about “getting out of control,” but they might break out, as you say, from the stress, survive, and then propagate. There are interesting dynamics and challenges in the interactions between different mutations: which ones are beneficial under which situations, which will be retained, which give you an advantage, and which give you a disadvantage.

I think it's fair 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. If you apply it for long enough, eventually resistance will develop.

It becomes interesting to then consider the role of AI as we increasingly utilize AI in this space. I think AI gives us an advantage in the battle of our wits against the genes of these superbugs. This follows in 2 different ways around resistance.

We can continually discover or design new antibiotics that act in new ways and would not be bedeviled by existing resistance. You still have to get them through the approval process and then introduce them. But the second is to explore how AI might be able to reduce the probability of resistance over a given time period, meaning extending the runway.

Some of the ways that can happen include discovering and designing molecules, for example, that hit more than 1 protein target. If their efficacy is linked to the action of each individual protein in an independent way, the bug would need to develop mutations in more than 1 site in order to provide us with protection. That makes it much more difficult for the bug to evolve away from the actions of the antibiotic.

Yeah, okay, cool. We'll come back to all that in a second. One more 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 given up looking for them. Given that this is such a big problem, with tens of thousands of Americans—and probably a couple of orders of magnitude more than that worldwide—dying annually from these antibiotic-resistant strains, why has that happened? What is this social failure that has led us to this state?

Interesting. I think it's largely an economically driven failure that tapped into some interesting aspects of how we handle things in society.

Maybe just to ground the audience, Alexander Fleming discovered penicillin a little less than 100 years ago. September 1928 is when he serendipitously discovered penicillin. It was not developed and manufactured as a drug until early in World War II, in the 1940s, by a group at Oxford.

We haven't had them for very long, but they have transformed modern medicine, enabling us to have surgeries and deal with any number of injuries—cuts, bruises, and blisters—that in the past would have been lethal and are no longer lethal. Interestingly, the heyday of antibiotic discovery was in the 1940s, 1950s, and 1960s, before the microbiology revolution and before the biotech revolution.

What has happened since then is that we've been in a discovery winter of sorts. We haven't discovered new antibiotics, and investment in the field has diminished dramatically. There are 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 is only going to sell for a few dollars, whereas a cancer drug or blood-pressure drug can sell for thousands, if not much more.

An antibiotic you'll take over the course of a day or a small number of days. A cancer drug or blood-pressure drug you take over many months or years, if not for the rest of your life. The economics support the development of non-antibiotic drugs.

Further, even companies that stayed in the business and made it all the way through to getting their new molecule approved found that, 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.” As a result, many of these companies then went bankrupt after reaching the milestone of getting their compound approved.

We face this dire situation. As you allude, we have this underlying epidemic that's been going on for decades: about 1 million to 2.5 million people die each year from bacterial infections around the world. 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, outpacing deaths from cancer.

It's a challenge. How do we motivate pharmaceutical and biotech companies to develop products that they're really not going to generate revenue from? I think we need to explore public-private partnerships of the type we saw with Operation Warp Speed, or whatever it was called, around vaccine development for COVID.

Second, I think we need to better engage philanthropists. We have various months dedicated to different diseases, cancers, and so on. We have walks, charity walks, runs, ribbons, and colors. We have none of this for antibiotic-resistant infections.

Every one of your listeners has lost a friend or family member, I guarantee, to an antibiotic-resistant infection. Your family member went to the hospital for a certain treatment, caught the infection, and died. Somehow, it does not rise to the prominence in our consciousness of a need to address it.

I expect you probably have some fairly prominent, very successful, wealthy listeners on the show. It's interesting to think that an uber-wealthy individual could single-handedly address this challenge for the lifetime of everybody on this planet.

My estimate is that, for about a $20 investment, we could address AMR, and specifically antibiotic resistance, over the next many decades. That'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.

For individuals who would like to make history, but not make a dollar or much money off of it, I think there are opportunities here to really leave an impact on humanity.

Well, let’s get into how you would put that money to work. I know you’ve been working on this for a number of years, and the first reported antibiotic that you found goes back to 2020. Just give a high-level overview of the trajectory of the work for the last half dozen years, and then from there we can really dig in, especially to the most recent work, the data sets, the techniques, and all of the nitty-gritty details.

Maybe I’ll go back even a little earlier. Our lab’s been working on antibiotics now for a little over 20 years, and we’ve used machine learning, a sub-branch of AI, in that context from the very beginning. Our initial efforts were really using machine learning to infer, or reverse-engineer, biomolecular networks inside bacteria in order to better understand how antibiotics act. Our goal there was to better understand mechanisms of resistance, as well as to identify molecules and come up with ways that we could boost existing antibiotics.

Here at MIT, in 2018, the institute launched a campus-wide initiative in AI, recognizing that the institute had been asleep at the wheel on this third wave of AI. The first wave was really the very early days, from the late 1950s to the 1960s, when folks like Marvin Minsky and Seymour Papert here at MIT led the way with their interest in early neural nets and perceptrons. Then, in the 1980s, there was the second wave, with interest in things such as Lisp and other language-based programming languages, and executive programs, with folks like Patrick Henry Winston.

This third wave focused on big data and deep learning, eventually leading to more advanced language models. The institute realized that we really hadn’t stepped up, so in March 2018, we launched a campus-wide initiative. I had the opportunity to sit next to Regina Barzilay, one of our AI stars here on the faculty, who has done a lot of work applying AI to problems in biology and medicine. We realized we both had an interest in drug discovery and thought, “Wouldn’t it be interesting if we could apply our interests to get after antibiotics?”

Then we brought on Tommy Jaakkola, who’s another faculty member and AI expert. We really didn’t have money to do this, so we bootlegged the project. We looked around to see what we could pull together and assembled 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.

We applied them to E. coli. E. coli is both a model organism that we use in molecular biology to understand different biological processes, but it’s also a pathogen, and many of these listeners may have unfortunately experienced it, whether in a urinary tract infection or food poisoning. We applied each of those compounds to E. coli to see which exhibited antibacterial activity, as evidenced by growth inhibition.

We took those data and discretized them to say yes or no. If you achieved at least 80% growth inhibition, you were considered antibacterial. If you didn’t achieve that, you were considered not antibacterial. We took the structure of each compound and trained a deep neural net, specifically a graph neural net, that could learn, bond by bond and substructure by substructure, those features associated with the feature of interest—in this case, antibacterial versus 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, which consisted of just 6,100 compounds. We 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.

Interestingly, only one molecule fit all 3 criteria. It was the molecule we called halicin, in homage to HAL, the killing AI system from 2001: A Space Odyssey. HAL killed humans in the movie; halicin, our molecule, killed bacteria. It turned out to be a remarkably potent new antibiotic.

It’s really striking to me how small that data set is. We’re so used to—today, when I think of large language models, I think of something like 1 to 15 trillion tokens as the range of the data set just on the pretraining, right? Then they do lots of post-training and reinforcement learning and all that kind of stuff on top as well.

If you had proposed to me, 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 was probably at least 2 orders of magnitude too small. How do you think about the fact that this works at all?

It’s interesting, Nathan. I think your response is very similar and consistent with the response we got from our colleagues in the AI space. When we presented what we had and what we were doing, they dismissed us and said, “Don’t even start to explain. You have far too little data to do anything meaningful.”

A few points of note: One is that, with the growth-inhibition data that we collected, we could have discretized it differently. When you look at those data, you could say, “This compound achieved 90% inhibition; this one was 10%; this one was 30%.” But we didn’t. Recognizing that we didn’t have a lot of data, we just discretized it to 0 or 1. We reduced the coarse-grained feature of the data and binarized it.

The second point is that, again, it’s interesting that to this day people still say these models are very data-hungry—the more data, the better. We’re not challenging that. But it’s interesting that here we had a good number of hits, having known antibiotics—around a couple hundred—in that data set. We weren’t looking to get a 100% true-positive rate. We would have loved it, but in the end, when we tested it, we had about a 51% to 52% true-positive rate.

That might sound small if you’re trying to differentiate a picture of a cat from a dog on the internet, but it’s really good for looking at a prediction of whether you came up with a new antibiotic. Usually, for a random screen or a large screen, it’s well less than 1%. So, I was surprised by how well the model performed—pleasantly so.

I think it speaks both to the value of positive data in these compound structures and to the fact that they were really rich and enriched for antibacterial activity in that case.

On the question of discretizing the data, again, I think if you had said to me, “I’ve got this relatively small data set, and I’ve got these measures that range from 0 to 1 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?” my intuition would have been that you should try to predict the scalar quantity and then maybe apply the threshold at the end.

I imagine you probably tried both. These models weren’t huge, right? It wasn’t as if you only had the compute for one run.

I suspect we did, and I can tell you that if we did, the data weren’t good, meaning the results weren’t good. Doing a correlative model, there just really wasn’t sufficient data to get predictive capacity now from a completely new structure, or to understand where we sit on that line.

But given that, I think the difference is that getting after the discretization of a molecule that really would inhibit growth means you’re now getting after specific structural features of the compound that make for a good antibiotic versus one that doesn’t. If you got enough data and enough compounds, I think you could predict where you’d sit on that line from 0 to 1. But I think in that case you’d need many hundreds of thousands, if not millions, of compounds to fill that out.

Here, we were rich with a lot of compounds with no effect and a good number with an effect, 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? If you wanted to set out to do tens of thousands or hundreds of thousands of compounds against bacteria, what would that cost?

There are 2 or 3 levels of cost. One is actually curating and buying the compounds themselves. In this case, for the initial one we had, I think we had the library available. We then subsequently, as part of the Antibiotics-AI Project, 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, what, on the order of $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 bid to get after them, especially ones that could be up to on the order of $100 per compound. It goes up very, very quickly. For some of these larger public health challenges, there are compound libraries available in pharma that include about 1 million molecules or so, and maybe larger, but they don't make them publicly available. I wish they would in some cases. Maybe they already mine them for the features they like, but it would be great to make them available for things like antibiotics.

The second level of cost is: how do you screen? When you start getting to these tens of thousands, that's a lot of work for a grad student. So, it's one of the few spots where we'll use robotics, liquid-handling robots, but they're costly. It probably runs us on the order of $20,000 in robot time to screen a 40,000-compound library. I wouldn't say that linearly scales when we go to 1 million, but it's a decent cost.

For example, our 40,000-compound library, we've now applied to 7 different bacterial pathogens and 3 different human cell lines. We've done this 10 different times.

Honestly, that's astoundingly little money in the grand scheme of things, right? In the world of AI that I'm following on a daily basis, we've got, in the space of the last couple of weeks, like $100 billion from Nvidia to OpenAI and Oracle. It's like tens of billions—it has to be tens of billions or you're not even making the news. So, we're talking like a full 3 orders of magnitude less to do the biggest-scale versions of the experiments you're running, and probably 4 orders of magnitude less to do some of the ones that you actually ran.

And that is affordable, as you said. Basically, let me frame it this way. I saw an announcement in the news this morning that xAI, Elon Musk's company, was in the midst of raising $20 billion in its most recent round. That's the number I just quoted that could solve the antibiotic resistance crisis for the coming many decades.

You have young kids, and I have some little kids. For the lifetime of your kids, that could be solved, which is stunning for a single instance of a private company. If you break it down, it's anywhere from $500 million to $1 billion to $2 billion per drug to be developed. Depending upon how one sets things up, I think you might get that even down to $100 million in some cases for certain drug indications with orphan status.

So, again, I think if there are some wealthy individuals who are publicly spirited and recognize the need for public good, here's a great example of AI for good. It's a great example where, with additional capital, I think we could take these compounds into patients and actually begin to expand our armamentarium—our portfolio—to go after these superbugs.

When you talk about solving the whole thing, big picture, for decades at $20 billion, how many drugs are we talking about? Is that 20 drugs at $1 billion each, including all the clinical trials and all that kind of late-stage stuff?

It's on the order of 15 to 20 drugs—that would be the real pitch here. 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 certain infrastructure and get things in place. But when you look at what we could do, I think it's, again, an interesting oversight that, for some reason, hasn't risen to public consciousness at the level it should.

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 are 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, or climate change, those are multi-trillion-dollar problems. This is on the order of tens of billions—low tens of billions. So, we have some work to do to convince folks, but 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 the interesting way to make it interesting again. It's raising the attention of your community and my community. Again, it's a true, beautiful example of AI for good.

Yeah, it's an area I would also love to see the U.S. and China decide to race. I proposed maybe even teaming up, but racing—I'm a very competitive guy, so racing would be marvelous as well.

Yeah. I mean, I would love to see that. This is a whole other digression, but I'm what passes for a China dove these days, in the sense that my outlook is: I think this AI stuff is going to be a really big deal, and we might need to work together across the U.S. and China to end up in a good place.

The alternative is that we race to weaponize and create a whole new Sword of Damocles, and that all sounds terrible to me. But obviously, the spirit of competition is high and rising. So, I was kind of thinking, maybe we can make a medal tracker for new drug discoveries or something like that. I know that the CCP loves to collect gold medals, so if we create more gold medals for antibiotics and similar efforts, 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, a very small amount compared to all the trials and downstream stuff. What are we talking about?

That's a fair point. I think it's on the order of millions per compound to develop it preclinically before trial. Low millions would probably be the fair number. Really, the cost comes in when you start queuing up your phase 1, phase 2, and phase 3 trials for antibiotics. But on the order of low millions from early discovery, hit-to-lead optimization is, I think, a decent estimate.

For example, we're working with Fair Bio. Fair Bio is a non-profit we helped launch as part of the Antibiotics-AI Project. With Fair Bio, 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—to establish a very robust pipeline driven by generative AI.

Looking at that, you're at a little under $2 million per compound to get it through to being IND-ready.

Yeah, so just a couple of percentage points as a down payment could take that to the scale you'd need to stock the shelves indefinitely.

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.

A couple of things jumped out at me about your techniques, and I'll give this to you as a prompt, 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 interested in whether you tried both, and how you think about why you chose 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, I assume they're all randomly initialized differently as a way to avoid any one of them overfitting or going weird, and then you ensemble all of those to make the predictions. What else do we need to know about the actual architecture of the networks?

There, you're spot-on. I think the reason we chose the graph neural net with a convolutional net as the overlay was that this was the platform that Regina and Tommy's team had developed under the banner of Chemprop. This preceded our work on Inner Bodyx. This was work done by Yang in their lab and, I think, by Kyle Swanson and Wengong Jin. It was a marvelous platform.

We started this probably in late 2018. Transformers were just beginning to appear at the time and weren't really that popular yet. Our strategy is generally to create an ensemble of models trained similarly—I'd say, in fairness, trained identically—but with different initial conditions. We then average across the ensemble, a wisdom-of-crowds approach, of sorts.

More recently, we've explored large language models that have been developed. These are not graph neural nets, per se, but language models that can look at a string of symbols from a compound structure. They've done okay. Many of these are pretrained on large libraries of compounds, but they've not yet outperformed our graph neural net. We've seen that they seem to be learning slightly different schemes, and they make predictions that are a bit different.

We've explored whether we can do a multimodal or hybrid model with a little bit of success, giving us a little bit of a bump, but not a considerably high one. More recently, we've actually implemented Uni-Mol, which is another graph neural net that is pretrained using quantum-mechanical calculations. Work that we'll be submitting soon shows that it significantly outperforms Chemprop, our earlier graph neural net.

It's interesting. Here, we're really, I think, taking advantage of the fact that these models were set up in part to look at graphical representations, and it fits beautifully for the compounds that we're looking at. You can think back to the chemistry class you had in high school or college. It's those types of structures we're feeding in. Now, in many cases, the model is considering a 2D representation of a 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 Evo, Evo 2, ESM, Fold.

Yeah, Evo and Evo 2 are basically genomic models trained on DNA. ESM is protein-based. We've been looking at language models that were set up specifically for small molecules. We looked at ChemBERTa, which was a dominant one. NVIDIA has its own, and IBM had a version, so we've looked at the leading cases.

Evo wasn't well set up for what we're doing, because we're doing small molecules, and neither was ESM. But in each case, we are intrigued by each of those models and how we can apply them to some other things we have going in the lab.

No, okay, cool. There's maybe another question around bridging these modalities, but I'll save it for a little later.

But keep going through the process. We've got this ensemble of convolutional graph neural networks trained. Then the big computational step, which I understand is still not all that big, is taking increasingly 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 are the ones that are predicted to kill this particular bug.”

There's also an interesting Monte Carlo tree search algorithm that seems to, if I understand correctly, cluster molecules in—not functional space, but physical space, right? Those that have similar structures are clustered together, so that you can not only get the prediction, but also 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.

With the trained model, we'll feed it structures from in silico libraries that are either curated from compounds available for purchase, in some cases, or that can be synthesized. In other cases, they might be arbitrary—we think we can synthesize them, but we're not really sure.

In that initial phase, we screened computationally about 110 million compounds. That's an enormous library from a real-world standpoint; you would never empirically screen it in the lab, but we did it over the course of 3 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, they are 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 silico libraries. Their REAL Space is on the order of 65 billion to 70 billion, and we've been screening those now. It's getting up there, right? It's several orders of magnitude larger than that initial 110 million, which itself was very large.

The second space you were speaking to was an effort led by Felix Wahl. It centered around the idea: Could we get at explainable AI to better understand common structures among the best-scoring molecules, enabling us to better identify novel structural classes? Could we set up—in this case, we did—a Monte Carlo tree search to look at rationales or substructures across the top-predicted compounds, to see whether a substructure rationale is overrepresented in a way that would suggest that maybe we're onto a new class that goes after a similar mechanism, but has chemical diversity of some sort within that class? That would increase the chance that we really did come up with something new and meaningful.

We posted 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. It gave us insight into the chemical structures that had 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 dominate my consciousness most of the time.

And that is a really interesting alternative way to try to make the AI process explainable.

Yeah. So, okay, the move from tens or hundreds of millions to tens of billions of molecules is definitely a huge leap. Can you give me a little bit of an intuition for chemical space? I actually was a chemistry undergrad, but I'm embarrassed to say that I don't have much intuition for this. At that time, we were working on natural product synthesis, and I understand there's been a pretty big shift in the field from picking one molecule and doing whatever it takes to synthesize it, to a more sane approach that's more like a building-block, almost Lego-style approach to putting these things together.

So, a few levels. An academic lab like ours will typically have compound libraries on the order of 10,000—tens of thousands. Large research centers will have libraries on the order of hundreds of thousands, maybe a million. Pharma will have libraries on the order of millions, low millions.

And when you say libraries, these are in hand?

Yes, they are compounds that you have in little vials. For example, the Broad Institute has a center for C-dots center for discovery of therapeutics that we work with, and they have this beautiful robotic system that has on the order of 800,000 to 1 million compounds on site. You can program it via barcode to have your liquid-handling robot go grab them and couple it to another liquid-handling robot that can apply them to our bacterial cells.

So there is the physical-world compound, typically on the order of numbers that are in hand. I think Edamine, for example, has maybe 4 or 4.5 million molecules stored, ready to send you right away. Now, you go into the in silico space, and you touched nicely on your building blocks. Edamine's 65 billion is based on a set of building blocks and a set of recipes—synthesis steps to get there. 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 make those. So you're at 10^10 compounds.

Going back again to your chemistry days, the estimate is that there should be on the order of 10^60 compounds. I'm not sure they all fall into the drug-like space, but let's just say 10^60 compounds, which is more than astronomical. We're literally just scratching that surface with, say, Enamine. I mentioned the 65 billion; I may be off by a little bit there. Their REAL Space, I think, is now 220 billion.

Okay, so you're now at 10^11. Still not even close to 10^60.

Yeah, 10^60 is really hard to wrap one's head around.

What's out there? How real should we think of that as being? Are those actual possible compounds?

I don't know. I was just in a meeting with my team, and they claimed that there's a group—BiosolveIT, I think is the team—that can get to trillions of molecules from their building blocks.

Okay, that doesn't sound too crazy, given that Edamine is confident with the 220 billion. So, okay, an order of 5 more to get to a trillion.

Yeah, I get it. I've never seen the calculation for 10^60. I've heard the number and 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 an order of trillions 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 is. One of the challenges we had as we moved from discovery to design is whether you can actually synthesize what the model made. There is a big challenge, and one of the best in the business is my colleague here at MIT, Regina Coley, who's developed AI models to predict the synthesizability of a compound.

You want to get after the answer to, can you synthesize it? Then you want to ask, can you do it in a reasonable number of steps, and can you do it for a reasonable amount of money?

Yeah, so I guess regardless of how vast the chemical space ultimately is, it's safe to say it's pretty big, and you're turning through an increasing fraction of it with basically a pipeline of steps, right? 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, filters or classifiers or scores. I'm not sure if these are applied in sequence or all at once, but you're scoring all the molecules on how novel they are—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 putting them into a pill and shipping them around the world, and, as you said, how easy they are to synthesize. How is that happening? Is that a sequential thing, or is there one kind of heuristic that weights all those things and gives you an order?

In the past, we've done a multistep application of the models and/or filters. We've begun to explore multitask modeling. Are there advantages to doing multitask modeling? I think in the case where there's overlap in features being common across, say, the pathogens, we've seen some advantages. But for the most part, it's a matter of applying these different models in sequence, really to cut down on the space—to move from this interesting thing in the computer to something you've got to pay to get and actually test.

We're expanding that. So, as you alluded to, we've got really good models for antibacterials. We've got good models for toxicity. It's easy to calculate whether it's different from an existing antibiotic. We need more data to get after drug-like properties beyond toxicity: solubility, bioavailability, and lack of metabolic liabilities. So we're working hard with Fair Bion to get those data to train what we're probably going to call DrugProp 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.

I'm sure your listeners have heard many times the comment that we've cured cancer 900 times in a mouse or rat, but very few times in humans. In my world of antibiotics, if the molecule works in a mouse model effectively from a systemic-delivery standpoint, the estimate is about a 90% chance that it would also work well in humans. Our AI models today 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 at also predicting whether they're good at killing the bug in a mouse. That's where we need these additional data and 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. It's a really good question. I think the medicinal chemist remains pretty skeptical about the value of AI. In part, they feel that their decades of work and expertise in identifying molecules, modifying the molecules, and tweaking the molecules can still outpace AI. I think in some cases they can. Now, in fairness, there's no medicinal chemist I know who can review 110 million molecules.

I do think there are a few points to this. We ran an interesting case where we had a really talented medicinal chemist as a postdoc in my lab, and we pitted him against the synthesizability model that Connor Coley's group came up with. We gave both the model and the postdoc 30,000 compounds to see how they were ranked, and compared that with what Ed Ekins would say. The human outperformed the model.

The next level is, how do we better capture this intuition? I do think there are opportunities to recruit panels of medicinal chemists and develop the medicinal chemist in the loop through reinforcement learning with human feedback, to improve the models for these different features that they're able to capture and maybe not necessarily describe or quantify.

There's also the interesting anecdote, which I'm sure we're aware of, that when Google DeepMind was developing AlphaGo, they had trained the model on synthetic games and even human games. But it really began to make a big leap when they brought in an expert—I think he was maybe the 1,000th-ranked player—who could give additional human insight into strategy and moves. It elevated their system to a new level, and I think we need to figure out how to do that better in drug discovery.

So, that head-to-head—when you say 30,000, does that mean the human expert actually went through and scored 30,000 things?

Yeah, it was just the human expert right-swiping and left-swiping 30,000 compounds.

Holy moly.

I've been happily married for 35 years, so I don't know what left-swiping and right-swiping means. I know what I like and what I don't like, so I don't know which is which. Holy moly is right. It was so impressive that this guy did this.

It was interesting when I asked Andreas Bouton—he's now a talented young professor at Karolinska Institutet in Stockholm—what was guiding him. He was really looking for liabilities. He would say, “Bad, bad, bad.” If he couldn't find a liability, he said, “Good.” That was an interesting take, and it's not how I would have set up my model.

Again, I think examples such as this show that we need to do a better job. There remains some dismay, significant dismissiveness—maybe rightly so in some cases—and some hostility from medicinal chemistry toward these types of approaches. I think we need to turn to medicinal chemists for help to make them that much better and do a better job of capturing the medicinal chemist in our code and 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 not going to be 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: Would this make for a good drug? Will it be stable? Will it be available at the site of the infection? Will it not be cleared too quickly? Will it not be broken down by the liver too quickly?

These are things where I have many medicinal chemistry friends, but they can look at a molecule and say, “I like that molecule,” or, “That's an ugly molecule for XYZ.” We should 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. I was with someone who was either working for Baidu or had insight into Baidu. Baidu, at the time, had hired 15,000 individuals to label data. This goes back to about 2018.

Supervised learning was the big thing, and so labeling feature-rich data was key. They had hired 2,000 medical students who were labeling medical images, like pathology. I thought that was brilliant: that a country, or a company within a country, could have the resources to 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've learned could make a big difference.

Yeah. Honestly, it's really striking how well this seems to work, given how little data you start with and just how coarse that original signal is. The whole thing is—it's worth repeating, too—if I understand correctly, there's no encoding of the target bacteria at all, right? It could literally be anything. It's totally abstracted away into “it works” or “it doesn't work.”

Yeah, 0 or 1.

That's amazing. So, when things work better, what does that mean? I'd like to understand a little bit better the tradeoffs between compute costs and synthesis costs. I guess those are the 2 big costs in getting to the end of something where you're like, “Okay, this actually kills a bug.” Now we have to go into the medicinal chemistry phase. How do you think about those tradeoffs? If the model works a little bit better but maybe it's also a little bit bigger—maybe we do 10 times as many molecules—how do you think about that?

The compute cost isn't very large. You can treat it as a fixed cost in the background. I do think getting after larger training libraries is expensive. Compute, not so much.

Synthesizing is the big challenge, right? Because you're making a commitment in the outlay of money and time, and there's the question of the probability that you've got something that's good. The next cost is animal models. Animal models are not inexpensive. You now have to decide which ones I'm going to advance, and whether it's good enough to go.

If that looks good, now you're back at, “Okay, I've got to maybe do an analog generation.” Which of those will I synthesize? It becomes interesting within a little academic lab like mine how much of it you really commit to. So, it's the tradeoffs.

Given our experience on Gemini and the challenges of synthesis, I'm much more inclined toward looking at these libraries of molecules, like Edamine, where I'm guaranteed I can synthesize. I still might need to pay a bit more for those that are either difficult to synthesize or not readily synthesized. 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 the probability is that it's going to work.

Can you give a sense of where we are today? You start with millions, tens of millions, of molecules out in chemical space. You do the predictions, and you do all these filters for novelty, nontoxicity to humans, stability, and synthesizability. It seems like we get out of that process with dozens. Then you synthesize dozens, and we get down to the end with a couple that actually work, at least in a mouse. Is that basically right?

I'd say we probably get down to many hundreds before we try to synthesize. That would probably be a fair result after the set of filters we'd set up. After that, you were correct.

We can expand the starting point of libraries, of course, but I think where we have not yet gone is really from the early hit to so-called lead development. That's where we can get after more drug-like properties, so-called ADMETs. You have absorption, excretion, metabolism, toxicity, distribution, as well as PK/PD—the dynamics of the drug.

It's where I think AI, given that most of your listeners are probably in the AI space, has done a really nice job in early discovery efforts, certainly in antibiotics, but I think in other drug spaces as well. It hasn't yet really been utilized much further downstream, I think because of a lack of data. It becomes interesting: Where are you going to get that data? Who makes the commitment? It's expensive. I think the companies that make those commitments are the companies that are going to have advantages going forward.

The big way to improve the models is that they're already seemingly quite successful. I don't know if there are 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 identifying a target bacterium you want to be able to kill, running the panel of all the test molecules against it, training the network, doing the pipeline, applying it to a huge swath of chemical space, getting candidates out, and synthesizing those—it seems like that is pretty consistently working. Is that right?

I think that's fair. I think it's working with an acceptable success rate. Our goal is always: Can you do things even better, at an even higher success rate? Could you really identify those compounds that—boy, that's a really great starting point?

I think we have a little work to do there because we still need to get after those other drug-like properties. That's kind of the key thing, and we need to figure out how to accommodate multi-objective optimization across these compounds. I'm hopeful we'll get there, but we're not there yet.

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 where you go, “Okay, it works in a mouse,” to the point where we actually get into clinical trials.

Yeah, that is a cost. Sure, it'd be great if it was cheaper.

And of course, you have an intellectual interest in finding new techniques to make it better. But from society's standpoint, it seems like it already works well enough that we should just clone your lab 10 times and apply this to as many targets as we can, basically immediately.

Yeah, I think that's fair. I think our group and other groups have shown that AI is a viable tool, and I think it needs to be utilized. I do think, for sure, in early discovery, it's reducing cost, increasing hit rate and success rate, and thus increasing our chance of getting to new molecules that can make a difference for human health.

A couple of other aspects of the results that we've barely touched on, too: 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. There's also specificity. 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. That's worked in several cases, and that was surprising. The reason it was surprising is that the models were not designed to yield a so-called narrow-spectrum antibiotic.

A narrow-spectrum antibiotic would be one that goes after the pathogen of interest but spares the commensals—our good guys—as well as other pathogens. In fairness, we've only trained the models 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 abaucin, which was a molecule we discovered to be effective against Acinetobacter baumannii, and in the case of a molecule we discovered that was designed to be effective against gonorrhea, in each of these cases they were narrow-spectrum. It was intriguing. I think we got it more or less by luck, but 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 that if it's very different from other antibiotics, maybe it's more particular to the target, even if that wasn't explicitly—

Yeah, maybe.

That's an issue. It's an interesting notion. It is possible.

I'll put an esoteric spin on it. We're dealing with phenotypic screens, and in many cases I think we're getting after membrane-acting antibiotics. So they're targeting structures in the outer membrane and/or inner membrane. 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 could be because these lipoproteins are very specific to the given pathogens. So while we're not targeting a specific target, we're targeting the pathogen using a phenotypic screen, I think we're selecting for compounds that are getting at these lipoproteins that are specific to the pathogen and not to other pathogens. And I think that's what's happening.

We're also now beginning to think: Can we use AI specifically to start with the protein as a target, and then can we find small molecules that would interact with those targets, those protein targets of interest?

And can you speak a little bit also to resistance? That's another notable finding, at least with some of the drugs you've discovered.

Yeah, that's a fair point. So, halicin, for example, we compared it to Cipro. Cipro is a very commonly used quinolone antibiotic, and we applied each for over 30 days to E. coli in a lab.

Within a few days, there was significant resistance to Cipro—several-fold—and 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, and 30 days out we didn't see any resistance. As I said earlier, we'll eventually see resistance if we look long enough.

I think halicin's resistance to the development of resistance was likely because it's targeting multiple molecular targets. So, multiple proteins, probably at the membrane level. And again, to my earlier point, this is where I think we can also use AI with this as an intent going forward: Because it's hitting multiple targets, the bug can develop a mutation against 1 or even 2, but if there are maybe 3 or 4, any 1 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 talking about the future and how these methods might evolve, I guess I want to emphasize my general sense that this should just be scaled up even as it is. We don't need to worry about getting too much more clever. Give me warp speed for new antibiotics.

But that said, it's interesting, if nothing else. I'm interested in, I guess, when you think this could have first worked. That's 1 interesting question. I'm always reading new AI and ML literature, and there's always some new technique, but I'm not sure here what was the limiting technique, or whether all these techniques were sitting out there potentially for a few years before you came along and figured out how to integrate them.

Yeah, I think they probably were sitting there for a few years. I do think these graph neural nets were the real trick. I think the ability of these models to learn chemical structures, break them down, and associate them with a feature of interest—in our case, antibacterial or nontoxic—was the differentiating technology.

We're now expanding that to other features and other schemes. The deep neural nets were introduced a number of years ago, but really became quite popular about 10 years ago, largely on image analysis, out of groups like Yoshua Bengio, Jeff Hinton, and Yann LeCun. So I think that was the differentiating technology that worked.

I think it can be scaled up 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 we can look at a bigger chemical space, associate and bring in more biological data and more chemical data to get after mechanism, to get after features, using more and more generative AI to get after design properties.

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? You mentioned we've cured cancer in mice lots of times. Obviously, a human cancer cell is presumably a lot closer to another healthy human cell compared to a bacterial cell versus a human cell, but could we imagine a similar technique working for human cancer cells?

Yeah, most definitely. This applies to human cells and to other conditions besides cancer as well. Felix Wong, a postdoc in my lab who's also leading Integrated Biosciences in the Bay Area, is focusing, among many things, on age-related conditions and aging. He's using this platform to identify molecules that could act as senolytics, so they could eliminate so-called zombie cells among our cells.

These are cells that have stopped dividing and are thought to underlie neurological conditions, scarring, and skin conditions. It worked beautifully. He's expanding the approach now to many other schemes and other conditions.

Within infectious disease, we've worked with folks to do it for antifungals, antivirals, and antiparasitics. For more complex diseases, I think there's potential for cancers, neurological conditions, and metabolic conditions. So this and related phenotypic and target-based screens, I think, have 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 maybe it doesn't work at first or doesn't work as well, what kind of enhancements do you think you would need?

Earlier we touched on Evo and Evo 2, and I know those are sort of—well, first of all, Evo 1 was only bacterial. There's an interesting possibility there where these foundation models can potentially be used to identify targets. There are, of course, lots of different models now that do all kinds of different molecule binding, small-molecule-to-protein binding.

If you had to imagine a next-generation pipeline that brings in either some foundation models or other specialist models, what kind of elaboration would it be?

I'd say it's multiple levels. I do think 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. The 3D-predicted structures are not at a fine enough level that you could use them in a target screen.

We need work there. Our binding predictions for the affinity of small molecules against an identified binding pocket aren't there yet. They're not as good as they should be. So I think if we could advance each of those, we could do so much more in silico, from target ID to understanding how the drugs act.

We need, I think, to better use AI to embrace the complexity of underlying biology. We may fixate on single targets, but these targets operate in very complicated networks that vary depending on cell type and context. I think we need to develop tools that can embrace that complexity to identify the means to try to understand how the drug would interact with them.

I think interfacing those layers is important, and then there's also the question: How do we get to phenotype? Can we make predictions? Right now, we're not very good at predicting functional phenotype from our interactions.

I'm sure you've seen, where you are, a lot of interest in AI scientists and AI-driven, automated labs, and you were saying they never replace all scientists.

Yeah, I think it's way premature to make these claims, but they're being set up in my world largely to do things where we know exactly what needs to be done, and so you can just do the rest of it.

For example, in E. coli, this very simple model organism we've studied for decades, when I first went into molecular biology 25 years ago, I was told, "Don't work on it," by big, big people in molecular biology because everything was known. Well, it's a bug with 4,000 genes. You look at the genome, and 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. So I think AI has tremendous things to offer, and we've got these other types of models I alluded to, but we also have a ways to go in order really to take advantage.

I really still very much believe in AI as a thought partner for us, and keeping the human in the loop is still critical for many of these advances.

Cool. I don't know if I have any additional follow-ups there. One question I always want to touch on, at least for a second, is the possibility of dual use or safety concerns.

One thing I do like about this approach—and not just the ensemble, but the pipeline nature and the many specialist models—is that 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 risk, if any?

I think for the most part, what we're working on is kind of single-use: Can we help humans with these problematic conditions? There is 1 interesting case of dual use that's worrisome, and that is these tox models that we developed.

So here we're developing models to predict the toxicity of a compound against a human cell or set of human cells. We're mainly interested in identifying those that are not toxic. Well, those same models can be used to identify molecules that aren't 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. In particular, it would be problematic if they developed ones that caused on-mechanism toxicity for which we don't have countermeasures.

That's an 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. Some of my friends in the federal government came to us and said, “You know, are you worried about this?” And I said, “Well, no.” I had to admit that we hadn't really thought about it, but we have since talked 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 capability?

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. I could imagine interpretability techniques there could be applied for bad. And I'm generally a big fan of interpretability, but if you did have something that only put out a score of 0.8 or higher or whatever, but masked everything below, presumably the model inside still does have a representation that you could access if you were determined enough to do it.

Yeah, good point.

Cool. This is fascinating, and I really appreciate all the time and remedial education that I've got from you. Much obliged. Thanks for the great conversation. Is there anything more that we should say about just a call to philanthropists? Where are we on warp speed?

I encourage young people to think about this problem from the AI standpoint, from the microbiology standpoint, and from the drug discovery standpoint. These are exciting, tractable 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. Okay, half the world seems to be on fire, but it's still a great time to be a young person in science because so much cool technology is being developed, and we need young folks to take on this problem. So that would be my call, in addition to the philanthropists: at least think about whether you can 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.

Well, thanks for having me. I really enjoyed our discussion.