分子预测模型究竟学到了什么：物理规律还是化学家的社会学特征？

Ian Quigley: My co-founder Andrew Blevens and I were at Recursion looking at a small molecule on a screen, and then a medicinal chemist walked by us, his name Mike Karesi, and Mike was like, "Hey, looks like a Stuart Schreiber molecule." And then we were like, "Oh no, you can tell just by walking by and be like who made this? That's horrifying." And if the models can identify the chemist by looking at the molecule alone, then maybe the models can infer chemist intent, [music] which means maybe they don't learn anything at all.

RJ Haniki: Well, they did learn something; yeah, they don't physics. We downloaded all chemicals from all chemists at ChEMBL, which is the biggest repository of these things and their activities. Every chemist [music] who published more than, I think it was 30 papers and 600 molecules. So you get all the molecules and all the stuff. It's 1,815 chemists. And then we trained a classifier to predict a chemist from looking at the molecule alone. And we got a top-five accuracy of [music] 60%. We chose to represent the molecules not with any chemical structures or formulas at all, but just a vector of the probability of creator. So now it's this 1,800-element vector and it's got some probability of who made it. And you can use that representation and then train a model to predict activity from ChEMBL, and it does as well as if you use chemical formulas. We really think that it's learning a lot of sociology. It's not learning ground truth. When you train models on the exhaust, you might be in trouble. Welcome to Latent Space AI for Science podcast. I'm R.J. Haniki, CTO of Mirror Genomics, and I'm joined by my co-host Brandon Anderson. He develops RNA therapeutics at Atomic AI.

Brandon Anderson: Super excited to have Ian Quigley in the studio today. This episode's going to be a treat. Very excited. Ian is the co-founder and CEO of Leash Bio, which in my personal opinion is one of my favorite young tech bios of the last few years. At a high level, Leash is combining epistemic rigor, data collection at a crazy scale, and has a strong commitment to open science and has an incredible series of blog posts, a lot of which we'll cover today, but which I think really tells you how to do machine learning in a very hard domain in kind of the right way. I think I'll have some fun anecdotes hopefully that can trickle in through the episode. But yeah, to start with, thanks for being here, and is there anything about you that you want to tell us before we get started?

Ian Quigley: Um no, thanks, that was very kind. Yeah, we founded Leash historically after being confronted by a particular problem of trying to predict small molecules binding to protein targets, and then finding that the data that was out there was kind of wanting, and so we went out and built this thing at a reasonable scale to kind of try to go after it.

Brandon Anderson: Yeah.

Brandon Anderson: Yeah. So reasonable scale. You're founded upon the idea of doing something called a DNA encoded library. Every person I've ever met who has touched a DEL has come out severely traumatized. [laughter] I don't know any person who has ever said, "DELs, that's a great idea, let's do hundreds or thousands of these things." So why, what? And also please explain what a DEL is before you get started.

Ian Quigley: Yes, yes. So when you're going after small molecule drug discovery, the ultimate goal is to get a small molecule, or a dinky chemical, to bind to the protein that you think is causing the problem in a disease, and then go from there. And some of these proteins are pretty easy to find chemical material that sticks to them, and some are harder. One way to go about this is to have a lot of different chemicals in different tubes, and you then dunk the protein in one tube at a time, and it kind of takes forever. And then a different solution that was proposed some years ago now was maybe if you had a lot of chemicals in the tube at once, in one tube, and then you've got a small DNA barcode that's physically attached to each small molecule. You dunk your protein in the tube kind of like a fish hook, you pull out different ones that bind, and then you sequence the DNA of those. And so the idea is that you could then pool this large number of molecules and then you wouldn't have the logistical overhead of trying to go across large numbers of tubes or individual wells and plates. And so what's happened downstream of that is that you've got chemical libraries that if you do one tube at a time, the biggest ones in the world are in the low single-digit millions. With DELs, the largest one that has been produced is 40 trillion molecules. There's some problems when they get that big, but that is the kind of scale that you could try to attempt with this technology. So the way you create these things is you use this combinatorial chemistry where you have these building blocks and then you build the library piece by piece, but you build it combinatorially.

Ian Quigley: That is correct. Yes, yes. And so to unpack that a little bit, if you've got some number of millions of molecules, what this really means is it is a combinatorial mixture of building blocks, and the number of building blocks is going to be a lot smaller than the final product of all of those. So you might be squared or cubed, right, in some different example. So it's not uncommon for there to be maybe thousands or tens of thousands of building blocks that would then generate a chemical library of many millions of molecules. And this is restrictive in some ways because you're stuck with a smaller number of building blocks. But one thing that's really nice about it for machine learning is that you've got all these different combinations of things that look very similar to each other except for maybe one or two changes. And so that, I'm sure going to get into this later, but that does really enable comparisons so that you can identify, "Oh, it's this part of the small molecule that's driving the binding because I see a lot of variations in this other part that's not affecting the behavior too much." And that turns out to be incredibly useful.

RJ Haniki: The other aspect of this is how you do the readout, which is this DNA sequencing. This has been a big theme on the podcast. We've had many guests who have had some variation of "we do something crazy and then we convert the reading into a DNA sequencing readout." So how does this work and why is this so useful?

Ian Quigley: Right. Yeah. I would definitely say that it is a party line at Leash that if you're going to go off and try to do life science stuff and couple it with machine learning approaches, there's really only one scalable technology in town, and that technology is taking pictures. Sequencing DNA is another way of taking pictures. And so it's basically pictures and DNA sequencing. And if you can shovel your challenging problem into that using one of those approaches as a readout. So, you know, like Recursion takes actual pictures of cells and a bunch of other groups do that, like Eikon does that, and then almost everyone else with the spatial omics stuff or other approaches, they'll do some mixture of pictures and sequencing, which is also pictures. Yeah, there'll be some proteomics people out there who are going to disagree with me, but I [clears throat] I'd still argue the scale is not there yet.

Brandon Anderson: Hot take number one. [laughter]

Brandon Anderson: Yeah, and one of the really cool things about sequencing is that it is dirt cheap. You can, you know, extremely scalable, very well established technique, like we're very good at it.

Ian Quigley: It is, it is very, very cheap. It is getting cheaper all the time. There was a little while where it kind of plateaued in terms of the price dropping, but now a bunch of new companies have started up and they're driving it down again. And so like we very much look at like the trajectory of sequencing cost dropping, and it has implications for how we think about data collection and the scale that we're doing. Yeah, it's really nice.

Brandon Anderson: That's cool. So you have these components that you are mixing...

Interviewer: ...together to form different this combinatorial mixture of molecules and each of those components is barcoded. Is that right?

Leash Representative: That's correct. And so they kind of bind together because they're these protein chain or these amino acid chains that will click onto the ends of each other and fold together and then they either bind or don't bind to your target.

Interviewer: So first of all how do you know that they bound to your target?

Leash Representative: Yeah. So first of all, it is very much like this sort of fish hook analogy where if you were to—so right now the target, you know, is some protein or a protein complex that you're going to try to find chemical material to bind to. It's attached to a solid support. And so some folks do it a variety of different ways, but the way that we usually do it is it's attached to a magnetic bead.

And so you dunk this bead in this complex mixture of all these small molecules. And in a very easy case, let's pretend that the protein has a very distinctively shaped hole in it. You dunk it in there and then the chemicals in that pool, some tiny fraction of them, have the appropriate shape and electronics that they're going to bind into this very distinctly shaped hole.

Then you pull this protein back out and you only pull out the things that are in that hole. You rinse it a couple times in some way. And then you sequence the DNA of those things that come out that are in that hole.

Interviewer: Okay? And then when you say sequence of DNA, you're basically amplifying the barcodes that are in there, is that correct? And so amplify meaning PCR or something like that.

Leash Representative: You do a little bit of PCR and then yeah, then you toss it into the sequencer. And then one thing that you're probably thinking right now is okay, well, we need to compare it to something to get a good measure of relative enrichment.

And so you do have to engineer a bunch of good negative controls to say, "Oh, well, if I just stuck the bead in there by itself, or are there pan protein stickiness signals that are going to pull out certain chemicals from the library?" And so you got to think pretty hard about all those things and make a reasonable comparison to the thing you're actually trying to get binders for.

Interviewer: For the things that bind you sort of have these combinations of barcodes that all sort of okay if this one bound that I have like these—how many barcodes are we talking about for one small molecule?

Leash Representative: Great question. For one small molecule, so there's a couple things. One thing if we want to get into the kind of PCR based weeds, and so for the audience here, if you're making lots and lots of copies of DNA, there's little subtle biases in the DNA that might make them amplify faster than others. And so you want to control for that. So you put in a little bit of unique sequence in there. So when you have lots of copies, you can then collapse it so that they all came from this one event.

And then there's a bunch of other things about barcodes that we worry about at Leash, which is that when you are trying to stick proteins in there, some of those proteins are going to stick to DNA because that's their job, right? There's lots of proteins whose job it is to interact with the genome in cells and turn things on or turn things off or affect them in some other way, and then you've got this assay where you've got a ton of DNA and you're throwing them in there.

Interviewer: Yeah, and that can cause false positives, right?

Leash Representative: Very much so. A lot of people complain about this. This is kind of to the earlier comment about folks having DEL nightmares. This is one of those problems. There's a variety of ways to address it.

One thing that we do is we have large numbers of barcodes that each encode one particular type of small molecule. You say here's one small molecule. That's great. Now we have 32 or 48 different barcodes all encoding this one small molecule.

Interviewer: Oh, I see.

Leash Representative: Right. And so then a single example—one single molecule has got this barcode, and then its neighbor which is the same chemical formula will have a different barcode, and then so on for...

Interviewer: So if there's differential there, then you know that you're binding to the DNA instead of the...

Leash Representative: That's right. And so for the DNA protein aficionados in the audience, one really nice thing about this is that because there's DNA in there, you can see if there's a single DNA encoding barcode that binds way better than everything else, even though it's all the same small molecule. And then you can look at the sequence that you have enriched there and you can recover the described DNA binding preferences of those proteins.

It's like, "Oh, here's a transcription factor. It binds this series of nucleotides that is very well described—you know, it's like six or eight of them." And then you look at the DNA sequences you recover back from this experiment, it's like: "Oh great, we found the consensus binding motif!" So I think our protein probably got made right.

Interviewer: You have this barcode that you've identified and you've kind of gotten rid of the false positives using the technique that you just said. And so now I know that for a given combination, for a given small molecule, I bind to my target. But don't you have all these DNA tags, barcode tags like hanging off your molecule? Doesn't that impact the...

Leash Representative: Yes, they are hefty. [laughter] They are hefty. This is another complaint that people bring up about DELs. There's a couple different answers here, I would say. But like the 10,000 foot philosophical answer is that there are multiple different reasons why you might want to run one of these experiments.

One reason is that you're trying to get small molecule material that behaves a certain way that you're going to progress into drug programs. And so if that is your main goal, then you want to make sure that you're getting some kind of fidelity between having the DNA barcode on there and then having it off there. And that's true. And when we try to pursue interesting chemical material at Leash, you do have to go through the work of manufacturing the chemical off DNA, you've got to run them in these functional assays, and there's a certain level of attrition. For things that are easy, in our hands about 90% of them work; for things that are hard, it's about 50%. So there's some loss; other people have different experiences.

But I would also say that if you're thinking about doing this for machine learning training purposes: all these chemicals with their DNA barcodes are in the tube, you ran the experiment, some of them bound and some of them didn't, and that is real signal that you physically be measuring from the world. And so then you just condition your models on the existence of this barcode. And then you just keep moving on with your model training.

And so in our hands when we rerun experiments—so make fresh protein batches, make a new batch of the chemical library over many many months of time—we get back the same hits in the same order for the most part. And so it's extremely reproducible and robust. And so in terms of ground signal, we feel quite good about it.

Interviewer: Okay.

Leash Representative: You could be blocking the pocket or something with your false negative there. But you don't because you know that the DNA is there. It did not bind. The way that you're thinking about it is like, "Oh, I'm thinking about how the molecule will behave missing the DNA," but we know it's there and we don't pretend that it's not.

Interviewer: Oh, I see. So, does your hit include the DNA then?

Leash Representative: Yeah. Like when we train models, we very much put in a placeholder that marks the position of the DNA.

Interviewer: If you're going to synthesize this for a drug, will you leave that?

Leash Representative: No. Like I said, that's where you get those attrition rates, right? Then you've got to make it and sometimes...

Will Adams: They behave differently. Not always, but some fraction of the time.

Host: Got it. Got it. Got it. So, just need to understand that.

Will Adams: Actually, those are really good questions.

Host: Those are really good. Yeah. Like there are lots of objections and these are many of the reasons why folks are trying to do the thing that we're trying to do.

Will Adams: Yeah. I was going to say going back to my original statement, I again have never met someone who has worked on DELs and said we should do more of these.

Host: So, [laughter] Yeah. Yeah.

Host: So, I think hopefully we will get to why I'm actually excited that this route seems to be working. Hopefully the audience will agree with me before this is over. In fact, actually maybe even a personal anecdote. We have a mutual friend who is I think an angel investor in Leash. And when he told me he was an angel investor in a DEL company, I jokingly said, "Okay, well that's money you'll never see again." [laughter] And I'm like here in front of our entire audience to say I was wrong. I wish I had listened.

I'm curious about first of all, why are DELs so hard? I think we just covered a lot of these. Like what in the historical community of people doing DELs was making it so traumatic as an experience. And then why did you think you could overcome that?

Will Adams: But one thing I do want to say is that when DELs were invented, it was definitely at a time when sequencing was kind of more expensive and it really felt like taking doing sequencing experiments was kind of like a big logistical and financial swing. And then when you built the technology at the time and then pointed it at the problems that it was pointed at, it was very much oh we want to find new chemical material binding to targets that are very challenging.

And so that brings up a couple kind of tactical issues. One of them is that classically folks don't screen large numbers of protein targets. They're usually interested in finding chemical matter for something that's like really hard. And and and they have maybe tried a bunch of other techniques and they haven't worked out. So okay, let's try a DEL screen because we don't have anything. Then they'll go off and make the protein that they're going to try to screen and they'll put a lot of work into it because it feels like kind of a lift to do a DEL screen. And so maybe they might spend two to six months growing big batches of the protein that they care about, doing a bunch of purifications. Maybe they're going to run some functional assays to convince themselves that oh, it's really got to be this good for us to go off and run this screen.

And then when they run the screen, they're going to try to maximize their chances of getting pretty good chemical material. And so then the incentive drives into a place where we want a big DEL, like we want a lot of diversity. And so if you go looking at folks who are offering CRO services or internal DEL groups a lot of times they'll push into these very large DELs in hopes, and that's their motivation and it makes a lot of sense.

What we have done that's very very different is we screen tons and tons and tons of proteins and we are less driven by trying to find individual binders for a particular protein target that might be of high interest for a program. We just want to get tons and tons of active chemical material across large numbers of proteins. And so there we can identify how our DELs behave very very well because every one of our chemical libraries gets screened against many hundreds of targets.

Host: So you're characterizing the chemical library better.

Will Adams: We know very very well and so we start to encounter lots of these motifs that are quirks of the chemistry that we designed or maybe the manufacturing process. A great example is we have one chemical library that we've had made a couple times and even though it's the same chemical library it does behave differently and we did not know that until we screened hundreds of proteins against it. And then you have to stare at those results and say, "Oh well now we're seeing these motifs show up a little bit more." And then as you characterize the library more and more it really helps you evaluate and troubleshoot even your protein production process.

And so the one reason that we felt that this is pretty tractable overall is that I and a couple of the people at the company come from real like genomics methods development backgrounds as compared to chemistry backgrounds. And so we knew what it might take to tweak experimental protocols to really get good signal out of these funky assays.

And I would say it was that plus many years at Recursion. That's where I was last with my co-founder, this guy named Andrew Levens, and a number of other people on our team. And we all made, honestly, like a lot of mistakes in high-dimensional biology, bumping around in the dark figuring out what different failure modes there might have been that would allow models to cheat on the high-dimensional imaging data that Recursion produces. And so we had this real eye for thinking about, "Oh well, you know, there's these sources of noise this way. We're going to have to chase those down. And can you predict the instrument that this was run on? And if you can, that's really bad." And so we incorporated all that thinking into going after this particular assay and we had a feeling that folks hadn't really stared at it in that amount of detail before.

Host: Yeah this goes back to the Ron Alpha Noetic episode where I found...

Will Adams: Mafia.

Host: Yeah, Recursion Mafia. I found there's some real spiritual parallels between both of your companies, this real epistemic rigor checking very carefully everything you're doing, doing things in different ways and sort of semi-overlapping ways. Very data focused like at-scale automation. And I'm curious, was this just core to Recursion's culture that is sort of...

Will Adams: Yes, [snorts] especially in the early years. I would say that there are four ex-Recursion Mafia companies, and they all kind of rhyme. So there's us; there's Noetic with Ron Alpha; there's Enveda with Vish Waluru, August Allen, Hana Gordon, Dave Healey, and all these great people; and then there's also Known Medicine that was run by Katy Rose Scaly that got picked up by Pathos a couple years ago.

And so all these companies are basically looking at different interesting biology problems with extremely noisy measurements, and then they pulled in people from Recursion to figure out like, "Oh, we're pretty good at noisy measurements. Let's try to see if we can get consistent signal out of this particular corner." For us, it's a small molecule chemistry problem; for Enveda, it's mixtures of natural products, and they do a lot of mass spec kind of stuff; and then there's Ron and Katy Rose's group. So yeah, we're all very similar.

Host: So let's pause for just a second. What do you guys actually do? [laughter]

Will Adams: That's great. Sorry, I'm so excited about this, I just jumped in. [laughter] Sorry.

Host: That's great.

Will Adams: Yeah, we are a machine learning small molecule company. And so what that means is we go off and we collect massive numbers of small molecules binding to protein targets. And we use that to drive advancements and machine learning understanding of this kind of stuff. And we also push small molecule material into partners and move it forward towards patients.

Host: Just for the audience that's not familiar, what is a small molecule? Why do we care? How is it different than other therapeutic techniques?

Will Adams: This is great. So, when you're trying to drug something, I talked a little bit earlier about how you've got a protein of...

Speaker A: interest in your body and you're going to try to interact, you affect its behavior in some way, there are a couple different things that are currently on the market that people use. Um, one thing that's common is, uh, antibodies, which are made by cells. They're large molecules. uh and by large uh they're they're big enough that they have a really hard time getting inside cells. They also have a really hard time uh being in a pill. So if you have like antibbody treatment or or one of these protein treatments like GOP like GP1 treatments, uh you're going to have to have kind of an injection. Uh it can affect uh certain proteins that are easy for them to get to. But there's other ones inside the cell that drive a lot of disease that are that are quite challenging for those.

It is kind of more appealing in a lot of ways to instead of have this big thing that's a protein to have a little tiny chemical and tiny chemicals, small molecules here are things that can be in pills. They can go inside your cells and affect behaviors. They do make up most of drugs. Um if you've got any pills in your house, they're almost certainly small molecules.

uh and uh of the different machine learning drug discovery groups that are going after treatments in a general sense. We have made uh we as in all of you know life scientists you know in in the 21st century have made pretty good progress on being able to predict and enable the behaviors of protein type drugs. So you've got a bunch of antibbody comedies like Isomeorphic does this or like Chai or Nava. There's all these groups of very good very smart people pushing on these particular problems and I think part of that is because we understand uh or have enough data for for protein behaviors that that make it reasonably effective you know for these companies to do these designs.

Um our thesis is that uh those capabilities are really not here for small molecule approaches. And so that was one reason why we built this company is we felt that we could go after it in a meaningful way

Speaker B: and small molecules one of the key characterist they go everywhere right in your body. So that's good and bad.

Speaker A: Y [laughter] they they can penetrate into the cell and they can also get through defenses of tumors and cross the bloodb brain barrier more easily but at the same time they do all those things and therefore they can be more prone to toxicity and other side effects.

Speaker B: Very true. So the uh the kind of path uh through the regulatory barriers for small molecules can be a little bit more fraught than those for antibodies. And so antibodies we know pretty well right now about how they behave in the body and what they are likely to cause problems and and what they're not. And for small molecules there's a bunch of um well-escribed problems that some of them might affect. Uh they might cause liver problems for example they might cause heart arhythmias. Uh and as a a medicinal chemist, a person whose job it is to develop these molecules to get them from something that looks okay in a tube to something that you would feel confident enough to give to a human being. Um there are a lot of these pitfalls and they have to do a lot of work uh to avoid these pitfalls. So make sure that my molecule doesn't cause harder arhythmias and then if I change it to stop that's causing those problems does it affect it doing the thing that I want it to do. How do your models help with this larger problem of developing these and then you know maybe with you know sort of these downstream problems.

Speaker A: So one way to think about it is that um a hand some number of these uh problems these sort of offtarget toxicities are going to be driven by your small molecule binding to other proteins in ways that you might be able to predict or identify. And so that's something that we are very much pushing forward to is oh let's also make those proteins. Let's also screen those proteins with our chemistry.

Speaker B: Seems useful if you have a giant assay of

Speaker A: That's right.

Speaker B: of different proteins and a library that you can

Speaker A: Yeah. It is it is not at all an exaggeration to say that it is our aspiration to have screened every protein and then also move on to those uh outside of humans

both for reasons to get variation in the proteins but also to identify maybe you don't want to run a mouse model maybe you should run a rat one

because your chemical material behaves differently in those different species because those proteins might be slightly different.

Speaker B: Interesting,

Speaker A: right? So you want to do that stuff. But yes, going after large numbers of these things allows you to think, all right, well, now I've got my chemical material and then I know um of all these chemicals, which ones and what variations or ones likely to bind these things will cause other toxicity problems.

Speaker B: Got it.

Speaker A: Yep. Yep. And then downstream of that um there's a bunch of other assays that lots of people do that are more kind of emergent behaviors where it's maybe not a single protein, but it's how well it gets across cell membranes,

how it interacts in a more complex way that you require a cell to evaluate. or it might be like a cardiac cell for example for this heart arhythmia stuff. And so those things you know right now there are lots of groups that are pushing on technologies to measure these at the cellular level which is you know one level of biological abstraction away from what we're focusing on right now. But we like to think that models that have a good understanding of the biohysics themselves can be useful and maybe fine-tuned on smaller data sets in the future.

Speaker A: Popping back one more maybe and get maybe getting a little into model. Why why Dells? Dells really are at least in our view um the most scalable technology for measuring small molecule interactions because you can put many millions of them in a tube.

Now one thing that uh that we haven't touched on uh in terms of like these chemical libraries is that when you make them in a pulled process there are probably going to be some that react imperfectly and that the label translation of what you think you have might be might be a little bit off. And so we've designed a lot of controls into constructing our chemical libraries to mitigate this.

When you're using these um you know the word might be array libraries or libraries that are in one chemical at a time in a tube or in a well in a plate um when folks study to figure out how good those labels are like you hope that they're good uh but when folks examine this um they're typically off by 20 to 30%. It's an enormous mislabeling problem. We think it's probably comparable to mislabeling and Dell's maybe worse. Um, and Dells themselves are just enormously enormously scalable.

And maybe I can like I'll tell a quick story um that I'm sure that we'll get into at some point. Anyway, we ran a Kaggle competition a couple years ago uh and we decided to generate the data for the Kag competition and we released it. There's existing uh there's a database that was at the time two years ago called binding DB which was the biggest aggregated collection of small molecules binding to protein targets as compared to cells or something else. And it was about two a little under three million physical interactions. And so we released a data set that was about four billion physical interactions and it was all measured by a single person in the basement of my house. [laughter]

Speaker B: Yeah. The the early blog posts are really funny. You you like literal picture you have an act. It's not quite a garage startup, but it's it's in a basement.

Speaker A: It is a basement apartment with a like a 7 foot ceiling. Yeah. [laughter]

Speaker B: It looks claustrophobic. Yeah. I've never seen a biotech startup started in a basement.

Speaker A: It'll be done for like that, right?

Speaker B: But okay, so the so this Kaggle competition I think is just a fascinating story because you had such high hopes going into this.

Speaker A: We did. [laughter]

Speaker B: You were like and and I I think this is one of those cases of like very nice epistemic rigor where you didn't just provide oh like you know here's a train test with rand random split or something. you you thought carefully about how to organize this Kaggle competition, how to run it and how to kind

Questioner: ...of evaluate fairly. Can you talk about that process, the decisions you made and then also the result?

Leash Representative: Yes, happy to do that. So we were not by a long shot the first group to try to evaluate whether you could use computers to make reasonable predictions of small molecules binding to chemicals and then test those predictions. There are lots of papers doing this, like there's a real classic paper that came out of X-Chem and Google some years before we founded Leash that does this. We had run a lot of our own experiments on public data, and we had suspected that no one was really very good at it. And so we wanted to put together this contest and design the evals in particular in a way that could really test it across a range of eval difficulties. Because we were thinking about, okay, we're going to have a lot of people, hopefully a lot of people compete in this. Some are going to be very good, some are probably going to be beginners. We want to make sure that there are some nice easy wins, then like a bunch of stuff in the middle that's kind of hard, and then some things at the very top end of what would be difficulty, or what we imagine difficulty might be.

Leash Representative: And so to kind of unpack that, we've talked a little bit already about how these libraries are combinatorial. So you're going to have some smaller number of building blocks, but a large number of combinations. We got our hands on a chemical library from a wonderful sponsor, this guy named Xiaoji Liu who runs a shop called AlphaMa in Shanghai. And the chemical library was about 130 million molecules. And the center part of the molecule, you can imagine it kind of like a hub with three spokes, and then the center hub was the same for every molecule. It's a particular chemical motif called a triazine. And so what we did was we said, "Okay, so we got the center hub. Then we got these three spokes for the evaluation data set. Let's have a couple different ways to split this." One way is that we could say, "All right, the chemicals might have one shared building block, one shared spoke, and then the other two spokes are different." Okay, that's one thing we could do. So let's do that for some. And that was the easy stuff.

Leash Representative: Another one we can say is, all right, well maybe we're not going to have any shared building blocks, but we do know that if they're going to come from this chemical library, there's going to be the shared hub. And so we can imagine that some of the spokes might be similar, but they won't be exactly the same, and they won't be shared. So we think this is going to be a little bit harder. Sounds great. And then for the third, most difficult eval, we had a completely different chemical library that we had designed at Leash. There was no common hub, no common building blocks, different attachment chemistries, different everything. And so we thought, okay, we're going to have these three in different proportions. We're not going to tell anyone that's what we did. We're just going to say here are the evals, and let's go see how it goes.

Leash Representative: And so we launched the competition. It was two years ago, and you guys can look it up if you want, we've written a lot about it. And then after 2,000 teams competed, the winner did, you know, okay on the stuff where there were shared building blocks. Depending on the target, we saw some boosts in performance for where there were no shared building blocks, but the core hub was still the same—the triazine was still there. And then when you go to the most challenging eval, no one did better than chance.

Questioner: Yeah. The winner was like the 1,200th just random person who submitted a Jupyter notebook, right? Or something.

Leash Representative: Actually, yeah. So the winner was this guy named Victor Shlepov who put together a slightly different approach, and his was good, but again, like everyone else, no better than chance. Three of the top five were the noisy tutorial notebook. But just to be clear, usually with a Kaggle competition, there's like a public eval for you to test your algorithm, I guess, and then there's the private one that's held out. That's right. And the competition runs that, and all you know is the score.

Questioner: So, was your library included in the public? Okay.

Leash Representative: It was hidden back to make it spicy.

Questioner: Right? Okay. Yeah.

Leash Representative: And it was spicy.

Questioner: It [laughter] was very spicy. So, what did you take from this? What was the lesson? And I mean, first of all, what did you hope to get from this? You kind of touched about that, but then how did that sort of affect your thinking about like how do you proceed forward there?

Leash Representative: Yeah. Yeah. We had a couple hopes and some were fears. One thing I will be candid about is, you know, we had released this data set that was a thousand times bigger than anything anyone had ever seen. And we gave it about a 5% chance that recent advancements in kind of like co-folding approaches would mean that it would obviate our company entirely. Like, didn't happen.

Questioner: No, it didn't happen.

Leash Representative: No, but we were like, you know, there's like a small chance like we are really messing up here, and that didn't happen. We suspected more likely that folks weren't very good at it and that the data might help, and so we were hoping to see that. Um, one thing that—

Questioner: And anyway, you'd want to know that. You'd want to know that before you [give] 10 years of your life.

Leash Representative: Absolutely. Absolutely. And another thing that we wanted to know is there's lots of different ways to represent molecules and give them to computers. You could use these things called SMILES that do it some way. You can do 3D graphs. There's like weird spherical convolutional nets. Like people have tried kind of goofy things to try to represent molecules in a useful way. And there'll be lots of people in the field who will argue that like, "Oh, I think this way is better than this way" or whatever. And we did really encourage competitors, especially those who had a chemistry background, like let's try some of these different representations and let's see how it goes. And we were grateful to find that the thing that we were doing, which are just basic SMILES, kind of beat everything else. It wasn't like a resounding thumping, but it was better enough for us to continue using that approach.

Questioner: And SMILES just, if I recall correctly, is basically a sequence, a textual sequence of strings. It doesn't encode any 3D structure into it, it's all emergent from—

Leash Representative: It is a graph traversal, yeah, translated into a text string. And so you can start the graph at different points, and so there's what they call canonical and non-canonical SMILES depending on where you start. But yeah, it's an extremely simple representation. And so yeah, we were thinking that or wondered if some of these fancier ways of representing molecules might beat at least this particular problem, and they super didn't.

Questioner: So was your takeaway that you should go full force into this? Was this encouraging to you, or was this discouraging to you? What was the collective emotional reaction at Leash after that? I've always wondered that. I mean, you kind of wrote about—you had several follow-ups, but I'm just curious to hear.

Leash Representative: Yeah. Yeah. This is a great question. Um, so like we are extremely bitter lesson pill over here. And so our thought was—

Questioner: Some of it was kind of delight, which is like, "Wow, this is much harder than we thought. That's kind of great."

Leash Representative: [laughter] That's kind of great because if this problem is like all the other ones that we're aware of: if you can get pretty good signal out of a scalable measurement regime, then you should just try to go get that.

Questioner: Yeah.

Leash Representative: Yeah. No, I mean for us it was pretty validating, like, "Oh man, like this is an opportunity here."

Questioner: Especially if your scalable measurement regime is hard and you have specialized knowledge to do that.

Leash Representative: Absolutely. Absolutely. It was like, "Oh man, like..."

Speaker A: This is the time to like pour gas on this. [laughter] This is great.

Speaker B: Wow. Yeah. I mean that that's awesome. So then you were just like, okay, let's full force. We're we're going to just start collecting data, collecting data, collecting data.

Speaker A: Correct.

Speaker B: Yeah. And so you did that for a while, and then what came next? And maybe first also, what were the lessons you learned in the process, and then building to kind of your first useful outcome?

Speaker A: Yeah. Yeah. There are a couple things from the process that we thought was very very useful. One thing that we realized is that the competition was based on three protein targets that were fairly well studied. It became clear to us that we really wanted to, you know, for lots of reasons, but including that getting performance across novel proteins probably meant that we'd have to do a whole lot more. So we were always motivated to do that, but that really underscored like, oh man, we're going to have to make a lot of proteins, aren't we, here? And so that was part of it.

Part of it was thinking that we had this very large number of small molecules, like this 100 million compound library, that didn't help as much to us as better characterizing a smaller one. And so that was also like a strategic choice that we made. This is actually quite important because you have a fixed budget, and so you can either let's say spend more money on sequencing getting higher resolution for what you already have for a smaller set, or you can get more molecules and then hope that the noise sort of averages out or something.

Speaker B: That's correct. Like having a fixed budget really does affect like your strategic choices.

Speaker A: For us, we wanted to make sure that we're getting a lot of variation across both axes as compared to like one axis and like a handful of proteins.

Speaker B: Sorry, which two axes?

Speaker A: So that would be like the number of chemicals against the number of proteins. Okay. Right. And so, you know, you can imagine like, oh, I'm going to do three proteins in a billion chemicals, or maybe I should do like thousands and single digit millions. You know, or some other combination of those.

Speaker B: Single digit millions being probably one of the tiniest DELs ever constructed.

Speaker A: It's pretty small. The smallest one I've seen is 50,000 molecules.

Speaker B: Oh, okay. Yeah. I guess you can make a custom DEL if you want to do like...

Speaker A: You can do whatever you want.

Speaker B: Yeah. Yeah. [laughter]

Speaker A: Yeah. When you have the technology, just do it. Two building blocks. Yeah.

Speaker B: Yeah. Right. You can do whatever you want.

Speaker A: And so that's the nice thing is you have enormous control over both the library size and its characteristics and like this barcode issue that I talked about. You can expand that kind of stuff. And we put a lot of thought into kind of bells and whistles in our own DELs to really boost performance and signal over noise and all this stuff.

Speaker A: But getting back to the kind of the competition lessons, you know, it was also clear to us that models could think about memorizing individual building blocks and that would boost their performance. That seemed pretty clear. It seemed pretty clear that they might be able to get an idea through probably memorization that if you have a very similar building block, it might behave similarly to some of those performance metrics. And then also that when we have something that's pretty far away, at that current regime, they simply could not do it.

And when we released these results basically saying, "Hey, you know, no one did better than chance," we did get some blowback where folks were like, "Well, did anyone good compete in your competition? Like maybe everyone's just bad, maybe it's all first year grad students." And so as a result, we just released all of the data in its entirety. And so if anyone out there thinks they're awesome, then please show us, cuz like we want this problem solved. But we haven't seen anyone do great or perform great on that stuff yet.

Speaker B: And when you say all the data, you have including all like the raw read counts and everything. So you can go through and say, "Oh, it's not just a processing error," like you can go from the raw data and use it.

Speaker A: You could take our binary labels, or you could do it yourself. All the evals, everything, [sighs] it is all on Polaris, which is like a nonprofit group that was run by Recursion. It's there for anyone to use. Please use it, you know, like we intentionally released it hoping the community would make good use of it.

Speaker B: Do you think though the structure of this data set was fundamentally limited in a way that maybe without some external augmentation, you're not going to be able to do any better than what people have pretty much already done?

Speaker A: I think if you had really perfect structural understanding or like really good structural models, then you ought to be able to do okay on this, right? Because if we all have models that truly understand like the physics of the interactions between the amino acids of the small molecules, then it doesn't really matter what the distribution of the small molecule training set would look like if they're really good at this.

And I think when we see structural model predictions and folks say, "Oh, well, you know, I've got this time split on the PDB, for example, and our model's never seen these molecules or this protein before, and then hopefully we can kind of use this understanding to make reasonable folds." That's very much in the spirit of being able to make the predictions on this stuff cuz no one has seen these, especially our internal molecules—no one's seen those before.

Speaker B: That's cool. So Kaggle competition, you went ahead and you've learned some lessons. You went full force data collecting, regrouped, committed to scaling.

Speaker A: Yep.

Speaker B: All right. So then I think something like six months later you released or you had a blog post about your model Hermes.

Speaker A: Yep.

Speaker B: Can you tell us about that? And [laughter] some of the learnings there, and especially how you evaluated that and all the care which went into that model?

Speaker A: Hermes is a model that is the original brainchild of a data scientist at the company named Max Kleiman-Weiner, and also Sean Francis-Lyon, and Andrew Blevins of course is the CTO there. What the thinking was originally was that we had a bunch of data and we were thinking about how to split it and different training regimes, and so we wanted to come up with a model that was really really dead simple so that we could do lots of potential experimental iterations. It wasn't the purpose of it wasn't to be a state-of-the-art model. It was like, "Ah, you know, we want to try stuff fast," and so we built Hermes.

And then the basic architecture of Hermes is two different embeddings. One of them is a Evolutionary Scale Modeling (ESM) embedding, that's a protein embedding that comes from a group that was at Meta and then went off to start their own company, and then they got picked up by Chan Zuckerberg more recently...

Speaker B: We had Alex Rives on the podcast so we can throw back to that episode if you want to hear more, go back a few years.

Speaker A: Yeah, thank you Alex if you're listening. So we did that for the protein side. And then we had a different embedding called KPGT, which is basically like a token predictor chemistry kind of the other side...

Speaker B: Something which predicts SMILES or...

Speaker A: That's right. Yeah. Just predicts SMILES tokens. And so that's it. These very very two simple embeddings, and then we just put them together with cross-attention, and that was it. And then we used it to predict a binary label.

Speaker B: Simplest possible thing.

Speaker A: It's incredibly simple. Uh, very very small and then we train some data on it. Um,

Speaker B: actually you know not the simplest. you could have done a MLP after those right instead of cross engine. So why why attention?

Speaker A: Um the thinking with the attention was uh had to do with some other experiences that we'd had um kind of with uh graph convolutional like graph models for protein structure. And one thing that we noticed is uh if you've got these sort of graphs and there's a couple models that do this at other work that we had done, we had scrambled the positional coordinates of those and performance for some of those models more or less stayed the same. And that made us think that like attention was super critical. And so we ended up just making sure that the model was at least at least at least what we hoped is that it would be looking at the different parts of the proteins to small molecule structures that could be useful here. So this is I think oftenimes known as the Picasso problem in like image generation where um let's say CNN's for example would were famous for a while where they that you could scramble just the features of a picture like Picasso like flip the nose around or just like rotate the eyes or something and the CNN would still depict oh this is a Picasso because it just picked up on individual local features rather than some sort of global structure and this has been a problem in small molecule drug discovery for for a while where often times it picks up on local motifs and then it doesn't have any global understanding of the topology. So especially for something like Adele where you have this combinatorial structure of individual building blocks you could imagine the model learning each building block independently and then essentially just decomposing into the sum of three building blocks or something or I mean is a simple concept. Yeah. Yeah. So cross attention

Speaker B: in some sense forces certain molecules to interact with like certain residues in a protein. Yep. or at least maybe not forces but provides the offer. Think about it. It's an inductive bias. Yeah.

Speaker A: Yeah. Kind of what you're getting at is um you know we were thinking about how to evaluate this and so we had done some of the similar tricks that we did uh over with the cow competition where we had you know training sets and data that we generated internally um and then we'd looked at chemical libraries that were further away. Uh but then also we wanted to think about splits um to kind of probe Hermes behavior. And so because Dell's have this combinatorial approach, you can do the same thing where you might share certain building blocks. And so we do that and of course performance is better which we suspect is for memorization. Um but the other thing is you can imagine that there are building blocks that are very similar to one another that if you're just trying to split on the identity of those then uh then you might have things that like are close enough that that memoration memorization or some other you know cheat mode might you know convert predictive power. And so we came up with a a scheme where we then did uh um basically clustering of these different building blocks. We had building block clusters on similarity. Uh and that way we would just avoid entire clusters from being overlap between test and train. And so when you do that um then you know the performance does drop. Um and so uh one thing that you might be alluding to is uh we have a you know we use lots of lots of reasonable baselines here that we're really worried about. And so for almost everything that we do, we'll use like XG Boost, right? You should use something like that just to see how it's going. Right. Yeah. Every Yeah. The uh classic young machine learner practitioner, they spend six months optimizing the favorite model and then someone walks by and does like ex you know [laughter] XG boost random forest and they're like, "Oh yeah, yeah, that just beat like six months of work." Oh [laughter] yeah. Yeah.

Speaker A: Yeah. And and so we uh and so we did these different like split regimes uh to try to evaluate how well things would do relative to like actually boost performance. And so one thing that was really really gratifying to see is that you see XG boost drop steadily as you get further and further away with these building block cluster approaches. It's like if you share two building blocks, it's a certain level. If you only share one, it gets worse. If you share one one, you know, only none, it's worse. And if you do build block, it's even worse than that. Uh and then we compared that to the Hermes performance, which was pretty steady across all these different regimes.

Speaker B: I see.

Speaker A: Yeah. So it made us think like, oh yeah, you know, you've really got to get pretty far out of distribution for XG Boost to really tank. Yeah. And that's something that we haven't, you know, we haven't, it's in a couple plots and some work that we published. We haven't yelled about it too much.

Speaker B: Wait, do I thought do you don't you have a blog post like XG Boost is all you do? That's why we did that. Yeah. Yeah.

Speaker B: Running those splits is something that we've put a lot of thought into into trying to figure out like, okay, how do you stop your models from really, you know, memorizing and remembering things? This might be a good time for a little bit of a detour, but referencing another one of your blog posts which I found absolutely hilarious from my own personal experience. I have seen machine learning models fail like pick up on so many weird detractors and your one your blog post about molecule ownership or authorship was a really fun story.

Speaker A: The clever Hans post. Yeah. So many years ago, uh my co-founder Andrew Blevens and I were at Recursion looking at um a small molecule on a screen and then a medicinal chemist walked by us. This guy uh this guy named Mike Karesi. Uh and and Mike was like, "Hey, looks like a Steuart Triber molecule." And then we were like, "Oh no, you can tell just by walking by like who made this? That's horrifying." And then that haunted Andrew in particular for like many years. Like it was like six or seven years. And we joked uh you know I don't know every couple of months we're out for beers like man we really should figure out like how bad this problem is.

Speaker B: Why is that a problem?

Speaker A: Okay so it's a problem where it's very easy to imagine you're going to try to train uh models to understand chemistry. So you're going to get the data from lots of groups because it's very tempting to get data from lots of groups. And if the models can identify the chemist by looking at the molecule alone, then maybe the models can infer chemist intent, which means maybe they don't learn anything at all. And so,

Speaker B: well, they did learn something. They learn something. Yeah. Yeah. They don't learn physics, [laughter] right? They're not learning the thing you're trying to get them to learn. Uh, and so we were really afraid of this.

Speaker A: They learned that certain chemists are better than other woods.

Speaker B: Yeah. We learned all kinds of stuff, but maybe to set to set this up for the audience. Um, so we went out and like we, you know, we published this, you know, over Christmas. You can look it up. We have a leaderboard. We have a goofy online game you can play. Um, but uh we downloaded all chemicals from all chemists at Kemble, which is the biggest repository of these things and their activities. Every chemist who published more than I think it was 30 papers and 600 molecules. So you get all molecules and all the stuff. It's 1,815 chemists and then we train a classifier to predict a chemist from looking at the molecule alone and we got a top five accuracy of 60%. It's a very very easy problem.

Speaker A: Yeah.

Speaker B: Yeah. It's super easy for the models to do. Where they get confused is where you have two chemists who were on the same publications and were at the same institution. And so we think it's very likely is that there is a distinctive chemical signature style that individuals have when they make molecules. And then when those individuals have students or or trainees, then they adopt some fraction of that, you know, and so that was that seemed pretty pretty clear to us. Uh and so that was horrifying enough of on its own.

Speaker A: And this is a data set that many academic projects, many PhDs have trained on, tried to predict on, find patterns, and this is really like a major data set for the field.

Speaker B: Yeah. Like if you are not generating your own data, you're using this. And then if you are generating your own data, then you are probably victim to this because your organization probably has multiple chemists who have their own distinctive chemical styles and that they are using their styles to go after different types of projects.

Speaker A: But the key here I think is that those chemists also tend to solve similar problems. Like I might have a career here and I have a style, but I also am working on a class of problems, right? And so you're learning the style instead of the problem. You're learning the sociology of the problem. And the other thing that you're learning is that when you're a chemist at these organizations, ordering a molecule to be manufactured and tested is kind of a pain, right? So what will happen is you will make many choices of things you will not pursue, and you do not write those down. "Oh yeah, it would be dumb to do that." And it's like, yeah, that's their motivation. Like our team couldn't, just wouldn't make it. Like they could say like, "We could do this, but then they would cost too much." Or the lab team, the chemistry team will say, "No, we can't make it now." Or, you know, like we don't have the building blocks on...

Speaker B: "I don't think it's going to work." Like there's all these reasons why you might not make a choice. And then not making that choice is a fingerprint that is going to land on what does get made. And then, you know, when you talk to some medicinal chemists, some of them are, you know, bless them, they're superstitious people, and some of them will say like, "Oh, this is..."

Speaker A: Yeah, I bring this up like every single time. Medicinal chemists are some of the most superstitious people I've ever met.

Speaker B: Right. And so it's like, "Oh, well, this building block worked in this program that made it to the clinic. I'm really fond of it. Maybe it helps with solubility. So I'm going to put it in things that I like." And so then your model will learn like, "Oh, the chemist likes this because they put so that means it's probably working." And so it learns all this weird stuff. And so like we were very afraid of this. And so we took, you know, we ran this model to identify who the chemist was. And then we chose to represent the molecules not with any chemical structures or formulas at all, but just a vector of the probability of creator.

Speaker B: So now it's this 1,800 element, you know, vector and it's got some probability of who made it and you can use that representation and then train a model to predict activity from ChEMBL and it does as well as if you use chemical formulas and if you combine them...

Speaker A: According to sociology only.

Speaker B: Yes, [laughter] if you combine them, it doesn't get you meaningfully better. And so yeah, we really think that it's learning a lot of sociology. It's not learning ground truth. It is just learning like the choices and the fingerprints of the people who are trying to develop these molecules. And it's not a knock on them, right? I mean, that's not what they're trying to do. They're trying to, you know, get two drugs by making the fewest number of chemical series possible. Of course, they're going to do it like that. But when you train models on the exhaust, you might be in trouble.

Speaker A: So, this means that this is not a great machine learning training set.

Speaker B: We do not believe so. [laughter] It's funny. In my own life, you know, I've had many of these similar lessons. It's well known that things like molecular weight will just, you know, all you have to do is sort by molecular weight. That's like one of the single [laughter] single biggest predictors you can possibly have. Like wasn't it CASP 16, right?

Speaker A: CASP 16 where it was tied for first place.

Speaker B: Yeah. Yeah.

Speaker A: Second place, like it was really close.

Speaker B: It was really close.

Speaker A: Yeah. So all these people pulling out their fancy machine learning models, you just sort the molecules by heaviest molecules better find binders. Yeah.

Speaker B: Yeah. So...

Speaker A: Yeah, but this is not an intrinsic property to chemistry. This is a property of the data set.

Speaker B: Yes. Yes. This is a property of the sociology of what gets made, what gets tested...

Speaker A: And like the way medicinal chemists work.

Speaker B: Heavier molecules more likely to bind.

Speaker A: A lot of times it's because they're seeing improvements to affect other properties. That's part of it. And you probably got some parts too you want to talk about.

Speaker B: It's like basically you have a chemist who they find a hit and then they start optimizing it and they're like, "Oh, I have a hypothesis that this is the binding site. So, let me add this functional group to bind here." And then they're like, "Okay, well that that was a stronger binder." And then so they're like, "Oh, well, what if we add this other functional group to bind to this thing?" And what ends up happening is they just keep adding things until it tastes really good.

Speaker A: And then yeah, until it's sufficiently good. And then, you know, then instead of using combinatorial approach...

Speaker B: Say Izzian said also then you have these other properties. You're like, "Okay, well now we have this really good binder, but it's not soluble. So, let me throw on, you know, some hydrophilic functional or some functional groups on the other side of the molecule, which is not in the binding pocket to like make it..." and you just sort of you just keep tacking these things on and next thing you know, you get this big good binder. And so, you just sort of by molecular weight and it's like the single biggest predictor [laughter] that you can have in most public data sets.

Speaker B: [clears throat] But that has nothing to do with chemistry.

Speaker A: No, like I mean I think a good way of intuition is like if a thing's working, they're going to do more stuff to it...

Speaker B: Because they like it.

Speaker A: Yeah. [laughter]

Speaker B: So you just do more stuff to it. It gets more baroque or you have more like complicated expensive reactions or, you know, you get like weird building blocks that are hard to get.

Speaker A: Yeah.

Speaker B: You know, things that you wouldn't do if you didn't think it was working.

Speaker A: And so it's all these signatures that the models would just kind of pick up.

Speaker B: And so yeah, it's quite terrifying. So...

Speaker A: Yeah, this is like the horse something...

Speaker B: Clever Hans. Can you tell the story about Clever Hans?

Speaker A: Yes. Yes. So, a little history lesson here is there was a horse trainer in Germany, I think it was like the late 1800s. I mean, I wrote about this. I should remember it. And there's a horse named Clever Hans. And what Clever Hans seemed to be able to do was complicated math. So, you would show the horse, you know, like a chalkboard that's got a math problem. And then the horse would stamp the horse's foot on the ground and then stop when the correct answer was arrived at. And everyone was amazed that horses are brilliant. And then someone had the good idea to go get a questioner who did not know the answer to the math problem. Yeah. And then the horse completely fails. And then it turns out that the horse is just looking at like the posture and the micro expressions mostly of the owner. So that when it got to, you know, let's say the answer seven, then the owner gets like real giddy when it's at six, right? to stop. And there's this like wonderful paper put out by a German psychologist who actually evaluates the actual micro expressions and postural changes that he had observed on the horse's owner is quite extensive. So yeah, like if you're dying for a silly read, you can go look it up.

Speaker B: So then the analogy is that you're reading the face and not the...

Speaker A: That's right. Yes. You do not understand math. The models do not understand chemistry at all. They are just, you know, remembering the sociology of the creators.

Speaker B: Got it. Going back to...

Questioner: Hermes and your epistemic rigor, and trying to build something useful out of all the data you trained. You find public datasets are untrustworthy and you have a bunch of data which you trust internally, but now you need to understand how do we accurately and carefully use this to do something useful in the real world. So how did you bridge that gap for us to think about this?

Respondent: Like we did a bunch of these splits like I kind of described earlier. But then the next thing that seemed very obvious for us to do is like, well, you know we've trained models on internal data and we kind of don't like to train models on external data. So maybe we should try to predict on external data instead, but without ever training on external—

Questioner: Without ever seeing it.

Respondent: And so that does put us in a nice position because I haven't seen any other group train models for this task exclusively on internally generated data. I think it's very tempting to go get these other sources especially because there's so many possible protein targets out there because people don't screen as many as we do typically.

And so we took large collections of external molecules and then we predicted them to see how well we would do across, you know, many different proteins, lots of different positive negatives that we chose very carefully. It was, I want to say the total number of tests in the Hermes paper, I think it's like 13 million or something across like 500 protein targets. It's some huge number.

And then we benchmarked it against what was at the time the state of the art structural model which is done by the Boltz team. Which is a very good model, and ours cannot do a lot of the things that that model can do. But what we were able to do was do about as well as those guys on those public datasets in terms of like this binding problem.

Questioner: When you compared to Boltz, it required some sort of tricks to kind of have a fairish comparison. I think Boltz also trained on this data. Was that right?

Respondent: Sort of. So there are a couple large collections that we weren't totally clear on the different subsets that they had trained on. And so I could think of one example in particular where it was very, very clear and I'll talk about that in a minute. But a lot of the eval that we used were ones that were also serviced by the Boltz team, only we put a lot more molecules in there to go see how it would go.

So yeah, if you look at the different performances between Hermes and Boltz, you know, we're better on some proteins and protein families that Boltz is worse at and vice versa. And we suspect this is because of different distributions of those in the training sets, so that they understand a little bit more or, you know, honestly maybe memorize a little bit more and they're able to extend that a little bit.

Respondent: One thing that was very validating for us is one of our test sets was another orthogonal chemical library that our models had not seen that we had used internally, but it was a copy of a chemical library that had been out in the literature. And so there was one target in particular, it's a protein called BRD4, where there's an identical molecule from our screen that was also in the PDB (Protein Data Bank) training set that Boltz had almost certainly seen.

Our models did kind of bad on that molecule. It was a distant chemical library, they weren't very good at it. And Boltz was perfect at it. And it's because we suspect that Boltz had seen this protein and this individual molecule in a crystal structure already as part of the training set. So they were awesome at like, "Oh yeah, that checks out." Feels pretty good.

Questioner: So does this mean that in your opinion this validates your scaling hypothesis? Add more proteins, more diverse chemical libraries and now you can basically fill in the space more broadly or—

Respondent: We hope so. Yes, we hope so. Yeah, we think that going after large amounts of diversity and in particular going after intentionally chosen what you might call activity cliffs is the way to do this.

And kind of to unpack that a little bit: it's either on the protein side or the small molecule side. Do you have two things that look very, very similar except for a tiny change, and that tiny change drives like outsized changes in behavior?

A real classic example of this that medicinal chemists will talk about is this idea—and it's kind of mythical, it's not exactly real—but this thing called a magic methyl. And the idea is that you've got a molecule and then maybe you've got this one little thorn that is or is not there. And the presence or absence of that thorn means either it works great or it doesn't work at all.

When you have those situations, then maybe your model will start to learn like, man, having a thorn could be really meaningful, and so maybe I need to think about these very small changes. And so one thing that we do a lot, and it's baked into DELs, because you've got two out of three building blocks might be similar across large numbers of examples and have variations of behaviors. And then on the protein side, we very intentionally engineer mutations in those proteins to do the same kind of thing.

Questioner: I'm a little confused here because if you have these building blocks and the size of your variation is multiple atoms, right? Many at—how big is like a building block in terms of atoms?

Respondent: Right, right. It might be like six to like 30 or something if we're ignoring the hydrogens. But you can imagine that there are building blocks that themselves are very similar with like a single atom substitution.

Questioner: Oh, I see. Okay. So, maybe there's two building blocks that have one a thorn and not a thorn and then you can see that in the sequencing.

Respondent: That's right. That's right. All right. And maybe that thorn just changes position. And then as you're thinking about this, you're like, "Okay, well, I might have tension between my goals. I might want lots of rougher, more coarse diversity because I'm trying to find binders across large numbers of protein targets that may be very different, or I may want super fine-grained changes that are very, very subtle to try to really teach the models these very, very subtle differences."

And so as we think about designing our own chemical material, we have to span this kind of tension. So yeah, we think about it a lot.

Questioner: And both of those matter.

Respondent: Both of those matter.

Questioner: Both of those matter.

Questioner: And it also matters on the protein side too, right?

Respondent: It matters very much on the protein side. And so as you look at current like sort of co-folding models, a couple different groups have shown this. You can take a protein and it's got this amino acid sequence and it codes some protein that's got a big pocket in it, and maybe that protein is represented in the training set. So the model has a pretty good idea about what it looks like when there's a small molecule chemical in there. So you got this pocket, a thing that goes in the pocket.

You can just change the amino acid sequence that you give to the model to consider and you can use that to break the pocket. You could fill the pocket in like you're filling it with cement, or you could remove it entirely so that there's nothing there or it's gone or it's much bigger. You do any of these things and you get—at least what we've seen so far, if you use current state-of-the-art models to go see what the model will do, it will shove the small molecule in where the pocket used to be.

Questioner: Okay. [laughter]

Respondent: Because it just remembers where it is.

Questioner: Yeah.

Respondent: And that's bad.

Questioner: So it's like looking at the large scale structural things, not so—or like the hypothesis anyway is—

Speaker A: ...that it's doing this.

Speaker B: Yeah. The hypothesis is that it doesn't quite understand, you know, when you have a pocket in a protein, there's going to be a small number of amino acids that will have hydrogen bond positions that are sticking out. There's going to be maybe a little bit of pi stack and these other things that, you know, sort of drive this chemistry, and those things really ought to be very, very important—and they are very important in the physical world when you're measuring them. And so if you break them and the model understands the physics, then it probably should know that it can't shove a molecule in there. And that's the thing: a lot of these structural training sets don't have great negatives, because what makes it into there is—

Speaker A: Yeah, that's the thing, right? You know, they don't have a lot of these variations that really disrupt these behaviors.

Speaker B: And no one's going to make crystal structures with a bunch of these new—like we're currently doing first.

Speaker A: But how do you get negatives?

Speaker B: The way that we get negatives is that we have a really good understanding of the distribution of the number of those molecules in a baseline condition.

Speaker A: Okay.

Speaker B: So like when we do the screen we know, "Oh, okay, these bound a lot because now we're enriching for them when we pull the protein out."

Speaker A: And now let's break the pocket and they are no longer enriched. Okay. They are no longer going. That one amino acid substitution that we made is making a bunch of these chemical series drop out.

Speaker B: But this is a one-off experiment. This is not a DEL.

Speaker A: No, this is a DEL.

Speaker B: Oh, no. So you're not saying you broke it; you're saying, "I find the case where it was broken."

Speaker A: No, we break it. We break it. So let's say we screen a protein and the pocket is intact.

Speaker B: Yeah.

Speaker A: Sounds great. Okay. We get a bunch of hits.

Speaker B: Yeah.

Speaker A: Let's go even further here. Let's say that we know someone else has already identified a chemical that goes in that pocket.

Speaker B: Let's buy it. Let's stick it in there. Run the screen again.

Speaker A: Yeah.

Speaker B: That means all the chemicals that normally go in the pocket will be displaced by a high concentration of the one that—Oh, you have a competitive assay.

Speaker A: Oh, we do all the time.

Speaker B: Okay. Okay.

Speaker A: We do this all the time. Okay. And then you say, "Okay, well that looks pretty good. I believe that." And then you say, "All right, now we know that this chemical that someone else has characterized goes in this pocket because maybe they made a crystal or some other thing." Yeah. Great. Let's look at the amino acids in that pocket and then we will change the DNA construct that makes the protein for us. We will now mutate physically this one residue. So now, you know, this particular amino acid—now we put a big bulky one in there and the pocket's full.

Speaker B: Like the different side chain or something.

Speaker A: Different—all these things you can do.

Speaker B: Yeah. And this is really easy now because we have AlphaFold and, you know, Boltz or all these structure models. You can use this to design your proteins as you—

Speaker A: You can use this. And for things that aren't as well characterized, we use conservation. We have a structural biologist on staff who kind of looks at these things—like, "Oh yeah," you know, this guy Edward Craft is amazing. And so, all right, now we're going to make this one mutation. Let's do yet another screen. And look: all the chemicals that got booted out when we hit it with the one thing that's blocking the hole, they're also gone when we broke the pocket. So now we have really good evidence that they're all going into that pocket.

Speaker B: Right.

Speaker A: Right. And then here's the point where the protein chemists come in and they say, "Well, if you made a mutation, how do you know that you didn't just break the whole thing?" Right? Reasonable question. You say, "Okay, well, we only lose the chemical series that got kicked out by that competitor. All the other ones that are binding everywhere else in the protein are intact." And they come back—

Speaker B: Intact. Yeah. So you didn't just break the protein.

Speaker A: You didn't just break it, right? And so, and then you think, okay, well, if you're going to be out screening 100 protein constructs a week, which is what we do on big weeks... like, sorry, so you've broken the pocket somehow.

Speaker B: Oh, yeah.

Speaker A: Then how do you also find that there are other things that do that start to connect to bind to the protein?

Speaker B: So, when you change things around, you might change the shape in a way that does encourage other things to bind. We have definitely seen—

Speaker A: And is that an information source for you?

Speaker B: It is. It is. And so another thing that's really nice about this, and the nice part about DELs, is you can imagine—if you're not thinking in large numbers—that, "Oh, I've got this one chemical that was binding and now I broke the pocket and now it's gone."

Speaker A: Yeah.

Speaker B: But because of the combinatorial nature of the DELs, maybe I have five different chemical series. It's hundreds of molecules, and there's another group that's the same hundreds of molecules with a single change in the building block. So you have these repeated examples where it's like, "Oh, this one tiny thing shifted the population behavior enormously," and it's a big population. And so it's like, "Oh, I am now really convinced." I can show you guys a plot that we—I don't know if you put pictures on podcast or whatever—

Speaker B: ...but it's very obvious like, "Oh, this one side group on this building block is clearly interacting with this one amino acid residue," because we have like five independent examples of population behavior shifting because of it. It's beautiful. It's very, very clear.

Speaker A: Yeah. Go ahead, tell us about this plot.

Speaker B: We pulled this thing up. And the reason that we pulled it up is we think it's really important for there to be a lot of variation in the proteins as well as the small molecules. And a reason that we think this is you want to be able to get a model trying to learn medicinal chemistry to really start to put together that you're having these hydrogen bond interactions between, like, this part of this small molecule and this part of the protein. We've seen other work that I talked about earlier about how, if you digitally break a pocket and you use the current state-of-the-art models, they have a really hard time understanding the physics and they'll just shove the molecule back in that thing anyway. So it suggests to us that they're memorizing. And so we thought one way to go get around this is to experimentally engineer mutations in binding regions of these proteins and then rescreening. And so what that looks like—

Speaker A: Physically doing that—

Speaker B: Physically doing that, you take the DNA of the wild-type protein that you've already screened, and then you use microbiology techniques to change one or more than one amino acid in that protein so that the functional part where the small molecules are binding is no longer the same—it's different in some way. We have different methods to try to identify which amino acids are the smart ones to break. Sometimes you break ones that you're confident will do things; sometimes you don't know. But we can do this at a pretty big scale, so it's not a big problem to us to take a couple wingers at each one. And so then you run the screen, and now when the pocket is broken, what do the molecules do? Do they still go in there? Do some of them go in there? Is there something common among the ones that stick around? Is there something common about the ones that get dislodged or pitched out? And then you have lots of different examples of these where you might have three to five-ish amino acids in a given pocket that are driving the binding behaviors. Maybe you try to break all of them independently, and then you redo the screens each time and get a good feel for what part of the small molecule is—

Researcher: ...interacting with the part that you have changed. And then when you do this with a large DEL, then you'll have lots and lots of repeated examples of similar molecules. They're all following the same behavioral trajectories. And that way you get a really good feel for, oh, I've got lots of examples. They're all pretty similar, but the one thing that is super common is this one little bit on the small molecule. So I have a suspicion that that is interacting with the pocket that I just broke.

Interviewer: What are we looking at on this diagram?

Researcher: So what we're looking at on this plot is, um, all those little dots are small molecules from a chemical library that we have designed and made. Along the axes are how well the small molecules bound. So if it's like really down the axis, it's binding well. If it's more towards the origin, it's not really binding at all. Um, and then we've colored a couple different blobs, and those blobs are different chemical series. The chemicals look very similar to each other, but the difference is where we've pointed that kind of chemical looking structure. And so if you're squinting and you look at these, um, they're all pretty similar, but there's a couple subtle differences, and that whole population we believe is driven by interactions of those differences to the one mutant that we made.

Interviewer: So in this plot there's on the bottom the wild type target, and then on the horizontal or the vertical axis you have the mutant type target. And so there's this red one, and the red one is pointing basically up.

Researcher: It is binding better.

Interviewer: It is binding better. So that particular one is like the mutation made that particular one bind better, which is speaking to the question that I asked a few minutes ago.

Researcher: Yep.

Interviewer: And then maybe, uh, some of the—most of these seem to be above the axis. So most of this mutation caused most of these to bind better.

Researcher: That's correct. Yeah. There's a second mutation in that pocket that I can also share with you guys, um, where all of the—all of those population, all those blobs get kicked out entirely and they do not bind. So then the pocket was broken in a meaningful way, and then here the binding is increased. Um,

Interviewer: So, and so that basically everything will be along the bottom.

Researcher: That's correct.

Interviewer: Okay. That's great.

Researcher: So you can give us postproduction.

Interviewer: Yep.

Researcher: Awesome.

Interviewer: So doing this you can map out both—well, you can—you can treat the activity cliff issue both in compound space and in protein space.

Researcher: That is correct. And this is particularly useful if you're, let's say, working at kinases where there's like a hundred kinases which differ only by, like, in the important binding site you might differ by one residue at a time or something. So this can be incredibly powerful for certain classes like that, in addition to your own sanity.

Interviewer: Yep. That's right. That's right. Yeah. And we do this—you know, as I mentioned, we do this, uh, we do this a lot. Yeah. Many, many hundreds of these types of screens.

Researcher: One of the things you can do with this is, uh, Hermes sees this cross attention. So you can actually interpret where Hermes thinks the binding pocket is. So it's learning the binding pocket through these type of maps. Yes. So, uh, one thing that we've sort of written about is using these attention mapping approaches to have a look at where, uh, Hermes is, you know, looking in amino acid space. And when we first did this work, we were expecting to—for it to either look at, uh, the binding pockets. That was like the first intuition. And then maybe like a more measured intuition was like, well, maybe what we'd hope it for it to do is maybe look at all those residues because it probably doesn't really know what's going on yet. And then when we looked, uh, it was not looking at either of those things. It was looking at really short sequences of amino acids that were away from the functional part of the protein and also away from part of the protein that looks similar across many proteins.

And so for maybe the non-biology people in the room, the way a lot of, uh, molecular evolution has worked is that there's a stretch of a protein that does a pretty good job and that often gets shuffled among other proteins as well. So that you have this conserved region. It's called a domain or, you know, conserve protein domain and so that part looks pretty similar across many proteins. It's something that biologists will use to, uh, infer protein families and relationships and functions. So if they all the same kind of piece, they're probably doing something kind of similar. And so we expected the models to look at that, but they don't. They look outside that, and they look at these little stretches that people don't think typically do a whole lot. They're like a little bit different. They're non-conserved. There's a lot more variation in those.

And we realized that Hermes was probably coming up with a nickname for the protein. It's not thinking about, oh, this is the functional part. It's like, I don't need to learn that. I can remember, you know, 10 or 12 amino acids that are different here than everyone else, and I can just memorize it. And we thought that was bad [laughter] because maybe many of the—like if you have a variations in the pocket or whatever, but they all have the—these other parts that are similar, then maybe and the variations in the pocket mostly don't matter, but some of them do. Then you might—the model might learn, oh, I can just pay attention to these.

Interviewer: Yeah. The concern was that the model again was not really learning any physics. It's like, oh, I can just memorize this stuff. It's very much in keeping with, uh, when you try to break the pocket with the co-folding models and they shove it in there anyway. It's like, oh, it doesn't know what's going on. It's just made this nickname.

Researcher: Yeah.

Interviewer: It's like, oh man, so how do you fix that? And then that's when we hit upon trying to do these, uh, these mutations at scale. And so what happens is if you have a protein that looks almost exactly the same,

Researcher: ...but there's one tiny difference. And then the behaviors change a lot as a consequence of that one tiny difference. You're kind of encouraging the models to look at all the amino acids. And

Interviewer: It can't do that for the section that's non-conserved because if it does that with these two things of different behavior, it's going to do a bad job.

Researcher: Yeah.

Interviewer: And so we went back and did model, you know, train models with these additional mutations like the one on this plot that you guys are seeing and then, you know, looked back at the attention maps and now Hermes will look at all of these things much more evenly. So you're basically using the attention map to find these cases where it's memorizing nicknames and then enrich the mutations in the pocket to kind of encourage it to pay attention to the pocket instead.

Researcher: Yep. Yep. And it's not necessarily just to be across any one gene. It's more like the general problem of like, oh, I can't use this nickname strategy effectively anymore.

Interviewer: You want to teach the model that: No, don't do that.

Researcher: You don't do that. No, no, that's—that's lame. Don't do— Uh, so I love this because we've seen—we're— I think that the most pragmatic use of mech-interp has been in biology and in probably other related sciences where—and the reason why I think it's qualitatively different is because you go backwards from—you're learning about biology from the model rather than just trying to interpret what—like, what is—like, what is it thinking. In this case, maybe it is actually that forward case, but I've just seen so much mech-interp being applied to biology because you're really trying to understand—you don't have the language in these machine learning models. There is no language to lean on. So, you really have no clue what it's thinking unless you're inspecting the model.

Interviewer: Well, it's really interesting.

Researcher: So, yes, and one thing that we have found very fruitful in the very recent past is actually transitioning to language models.

Interviewer: Okay.

Researcher: And what we are currently doing is, um, we're using, um, readable interpretable English to represent the small molecules and the proteins and their behaviors.

Interviewer: Wow. Okay.

Researcher: And that turns out to work pretty good.

Interviewer: Yeah. I [laughter] mean this is—no. So this is because...

Host: I want to just frame this a little bit before you tell us about this because there's actually a really interesting paper in the single cell domain called GenePT in which people—I can't remember the authors, they are from Stanford and it was somebody who was a statistician, not even really in bio. Maybe they are now, but that person, they took the textual embeddings of the descriptions of genes.

Researcher: Yep.

Host: And then they use those. And so like you're using the ESM embeddings—

Researcher: For Hermes, yeah.

Host: So you were using the ESM embeddings. They were just like, let's just use the textual embeddings. And the beauty of this is because now you have like a gazillion years of human time put into understanding each of these genes, and that's all captured to some extent by these textual embeddings, that knowledge of the embedding. So yeah, please tell us about what you're doing.

Researcher: Yeah. So we started this project a little while ago cuz one thing that we realized is that we had this growing collection of activity cliffs and profiles of the same chemical material across large numbers of proteins. And we wanted to know if we could actually use reasoning agents to go explore it and use it in a lab-in-the-loop kind of situation. And so that demanded that we started working with reasoning models and language models in order to kind of let them go play.

And so what that looks like is we took some open source language models, like we did some Qwens and GLM and the stuff that other people are using, and we taught them medicinal chemistry in ways where you say: here is a structure, describe it in natural English; give me the diff between two molecules in natural English. And then we got it to the point of where we say: okay, here are some molecules that are binders, here are some that are not binders, can you tell us why you think the binders are binding?

Researcher: And then you can also give it hard negatives, so you have these two collections that think, okay, well I think this chemical series has this, and it will write these things down in readable English, which we think is great. And then we said, okay, now that you sort of know these basic concepts, we're going to give you 10 binders and 10 non-binders that are close, ask you to decide why you think the binders are binding, and now we're going to give you the opportunity to run your own experiments. Here is a catalog of millions of molecules. Why don't you pick five or 10 and test your hypothesis about why they're binding?

And so then the model will go into the data that we have already collected. It will pull out what it thinks are the most informative choices. It will get back the answer instantly like it's a coding sandbox, and then it will update its hypothesis about the binding. And then the ultimate goal is to get the best binder in as few steps as possible, very much like a medicinal chemist would do.

Host: You're doing this based on your a priori run experiments, but it's looking up in a database.

Researcher: That's right.

Host: Right. So you've basically created a reinforcement rollout.

Researcher: It is a reinforcement rollout of a large scale that's growing all the time, that's full of these activity cliffs and full of different protein variations. And since we know that when you change the molecules in these small activity cliff ways it is a causal change—like we know that when you put this thing on there, the behavior changes in this reproducible, measurable way.

And so you can also use these trajectories to come up with reasoning traces kind of at a pretty good scale. And so right now we have these agents running through the dataset. We just give them 10 binders and 10 non-binders, you give them a budget of 100 molecules, and they'll go through probably like 100 to 200 calls where they're just going through this long, many-step reasoning problem and then come up with what they think are the answers to the best at the end.

Host: So why do you think this is superior to something like Hermes, or maybe like sort of scaling up the Hermes approach where you're sort of building some attention thing?

Researcher: Yeah, we think so. We think it's much more resistant to the memorization problem. And so like while we are very happy with where the attention goes with Hermes, when you throw in all these mutants, as you add more parameters to a model, they're going to memorize the mutants too, right? And we want this thing to really understand these causal changes, and it seems like using natural language is like a pretty good regularizer for the problem.

And like we are unaware of other problems where you have a reinforcement learning loop that is physically grounded that is this scalable, this fast. Like you can imagine when you're doing lab-in-the-loop stuff where you're going off and making robots and going after these problems, they will also necessarily require limited resources. They will still make some choices about what to make in the first place, what not to make in the first place, and then doing those loops will take a little bit longer. And if there's a scalable way to measure that landscape and then continually make that bigger and bigger and bigger, and to allow the agents to play in it, we think it's probably going to go a lot faster for this particular problem.

Host: That's interesting. Yeah. I mean, I would say X-Cell from Zephyr is getting there, but they're not actually language; they are encoding as a sequence. So in some ways the perturbation is similar, but this is different. And then Lila also rhymes with this, but—

Researcher: It does. Like all these guys are really smart and pushing these automation things.

Host: Yeah, yeah. But they're not doing these pooled screens—

Host: —and then using that as, they haven't, as far as I know, figured out how to make that into a just a database lookup. So you can do huge number of rollouts.

Researcher: Yes. Yes. So right now the way that we like to think about it is if you're a medicinal chemist and you're working at a pharmaceutical company, you do this thing called a structure-activity relationship (SAR), and what that means is: oh, when I change the molecule in this way, it changes the behavior, and so I'm trying to optimize for these different properties. And you go through what's called an SAR campaign. That's early drug development: let's make the molecule better. You might see a couple dozen of these, maybe 50 of these in your career.

Researcher: And so right now, you know, last weekend we pulled out—

Researcher: —over 10,000 of them. And we're [laughter] not even trying that hard,

Host: Right?

Researcher: So yeah, and so now is the opportunity. Now we'll be restricted to the space of things that we have screened. And then at the end, you can ask the models to make additional choices from say catalogs. And then you can, in the future, go order those and retest them like you might for like a traditional lab-in-the-loop approach.

But right now we give these agents 10 binders, give them a budget of 100, and they'll get the best binders out of six and a half million in that budget. And so it seems like it's working.

Host: Yeah. So tell us more about why you think this is—I get the argument about it being a better regularizer, but also why do you believe that this is effective? Right? Because you can imagine, well, it doesn't get things wrong so badly, but it also doesn't get things that right. But—

Researcher: I think it's because you give it the opportunity to make all those choices.

Researcher: I mean, it's like running these coding sandboxes. You just have like really rapid reinforcement reward signal. And the thing is, you also have to include the choices that a human being would not have made, because we're measuring all the molecules across all the proteins.

Host: Yeah. So it has to be able to make the counterintuitive decision and get a feedback on that.

Researcher: It's got to have the counterfactuals. It's got to make mistakes, and you need to make sure that the mistakes are present in the data or future potential—

Ian: mistakes are present in the data in ways that are not really biased for it to, you know, learn chemist sociology fingerprints.

Host: Yeah, that's that's really fascinating.

Host: So through all of the, you know, language pre-training corpus, it it kind of has an intuition for some basic chemistry that I think most medicinal chemists might have an idea of. And now you're giving it sort of like an infinitely fast lab in the loop.

Ian: That's correct. Sorry.

Ian: Maybe reasoning too. I would say that the reasoning skills might even be more important, right? The ability to say, "Oh, in my reinforcement campaign, I've learning campaign. I've learned things and now now that allows me to reason about the next harder question

Ian: and then just keep making choices, you know, and and you know, right now we like we're really excited about the early signal." And also, you know, like we're a tiny shop, right? There's only 10 of us. Like I'm really looking forward to when uh you know our our experimental corpus grows at you know like 2x 10x you know like 100x the rate that it currently does because then like the amount of of space for these agents to explore like expands dramatically.

Host: Let's talk about that space because so my understanding and this is from me like you know talking to Claude and googling and whatever but my understanding is that Dells are like limited in that they are they tend to be create flat molecules or flatter molecules.

Ian: It depends on the building blocks. Okay. It depends on the building blocks. Yeah. So there are you can imagine that um you will be somewhat constrained by designs in terms of like oh I'm going to get some number of building blocks usually two to three and then stitch them together in some way. And so like when we do uh chemical library design choices typically it's like like one one that we run a lot right now is a forked design. It's like a Y-shaped molecule. And then there's other ones that are linear shaped molecules. There's yet other ones that are kind of like circular kind of macrocycle shaped things.

Ian: Um, and then you have to think about where is the DNA going to go and like how is that going to cause me problems and then also the the commercial availability of building blocks and the chemistries required to stitch those together. You can go for custom stuff takes longer. It's kind of more appealing to us to go faster. Yeah. And there's certainly going to be corners of chemical space that we're probably bad at. And then you have to think about, oh well, future libraries should have this kind of geometry. And then I got to think about the different proteins are going to throw at it.

Ian: And there'll be like a pretty uneven distribution of like a whole chemical library made this way might be really good for these types of targets. It might be really crappy for these other types of targets. And so then you get to think about all right, what am I what what what corners of protein space am I optimizing for? And those kinds of tensions.

Host: Yeah. So, so it sounds like the argument is well, we know that we're not optimizing the whole space, but it's a lot more than it was before. So, that's better. And then we are as we get better at this and understand or we have bigger budget or whatever it is, then we'll expand to these corners that we know that we're lacking in, but or maybe we don't even know we're lacking in, but we will find that out.

Ian: Right. Exactly. Exactly.

Host: Do we want to move on to the uh last phase?

Ian: Yeah, you can. Go for it. Let's see how it goes.

Host: All right. So, uh, for the audience context, um, Ian, uh, requested, I want to confirm that this was his idea, not mine. A call it a a don't ask me or come at me, bro section where we we ask the hardest problems we can we can come up with,

Host: which is, you can probably tell is not very hard from me. [laughter]

Ian: Yeah,

Ian: I think I've asked my hard questions already.

Host: We only had about 24 hours to prepare, so I think most of these will be Claude's hardest questions, but anyway, we'll see how this goes. Yeah, the the context for this is I was at um I was at a cancer and micro therapeutics meeting last week and uh and the the format of the meeting is half of the time is for content presentation and the other half is for very aggressive grilling. Uh and it was uh it was it was horrifying and inspirational and so uh here I am willing to do it in public.

Ian: Let's do this.

Host: Dells as uh R.J. was just alluding to often times have very specific properties um and you can get very good at optimizing indel but ultimately they frequently fail at certain things which are hard to translate to actual medic medicine. Y

Host: so certain failure modes being such as anything which is easy to assemble is easy to dis disassemble bad for you know admit properties you only have so much room to kind of build on because you've already sort of used some of the comatorial building blocks and so on. Anyway, how is this limiting in terms of like long-term out of library generalization?

Ian: I think a lot of it is going to be somewhat dependent on like building block availability and like one's willingness to go into custom synthesis. So, if we're especially for the fine grain mapping stuff, you know, being able to say move a nitrogen all the way around the ring at all one of those positions is something that's very very desirable, but it's also something that's pretty unlikely that you're going to get like commercially for all sorts of building blocks you're currently generating. So, like is that going to be a failing mode? For sure. Yes. Yes, I would I would love to be able to get more of those and like beef up our custom synthesis. That's something that we're looking at. Uh but right now trying to get those is challenging.

Ian: One critique that Claude might be thinking about in terms of molecules easy to put together and disassemble. For many years it was uh the most common bu way to attach uh building blocks between del chemistries is amide bonds or protein bonds. And so those are quite digestible uh you know in in the body. Uh but in the past few years uh there have really past like fourish years um there have been like a really nice um fluoresence of chemistries that are friendly to DNA that don't destroy it during the attachment process because that's very important that will give you carbon carbon bonds, carbon nitrogen bonds, these kinds of things. So it's actually pretty good for work a day simple chemistry. Yeah. So like I'm actually not too worried about that part. I'm I'm more worried about building block diversity uh than there sources for that. Yeah.

Host: What about adnet properties in general? like you know you you've been focusing very much on binding.

Ian: We have. Yeah. So uh of projects that we've done uh trying to look at reasonable adap and like PKPD properties. For what it's worth, we've gone directly from uh from chemical library hits to good PKPD in a mouse in 6 months with a very small number of medcam cycles. So at least the chemical material that we currently have or at least that you know those few series that we did that for uh you know they turned out pretty good in our hands. I guess you know we will continue to see like how that modeling looks and also of course we want to do more of these like admit that target models and then think about how to incorporate other types of you know more emergent models along with our designs.

Host: Was that speed driven by your ability to create custom like local cominatorial libraries or was that speed driven like what allowed you to get there in six months like

Ian: um the so we when we designed those chemical libraries at the beginning we very much optimized that space uh for drug likeness and so we think that's part of it is that was like a really good starting point you know when we do our chemical library designs uh we have this uh kind of like tsson sampling level approach where we come up with an optimization function you say okay well I want the libraries to have these properties And so some of those properties are going to be I want to bind these things. I want to avoid these things. And then some of these properties are going to be oh there's other admit model approaches that you can use to make predictions about how your

Speaker A: chemical material might turn out. And so you put that as part of the optimization function. And then what you do is you say okay of all the chemicals that I could make given my geometry. I don't want to make that many. I want to make a smaller number. Uh and so you you get you basically set up and generate a bunch of synthetic libraries and then you score them on this optimization function. And so we did that with this library. Uh, and the starting material look pretty good. So, I'm sure that we'll continue to do that.

Speaker B: Along those same lines, has anything that you've done like made it to so it made it to animals and

Speaker A: so PKPD in a mouse is that's Yeah. So, what that what that means is uh can you dose the mouse with the thing uh and then see how it gets into the mouse's body and tissues.

Speaker B: Yeah. So PKPD is pharmaccoinetics pharmacodnamics meaning how does the body move the the drug through the or how does the metabolism move the body the drug through the body.

Speaker A: Yep. And then are there biomarkers of the drug's uh intended activity. And so what that means here is this is like a drug this is a compound we're working on the inflammation. And so we were following uh cytoine levels in mouse blood and they and they did what we wanted them to do and expected them to do based on the targets of the drug.

Speaker B: So, uh, Sean, uh, likes to ask when money. I want to ask when people [laughter]

Speaker A: Yeah. I mean, for those we talked to, we have a couple of these things, uh, that we're talking to folks about partnering up to spin up new entities to go chase them down.

Speaker B: Okay.

Speaker A: And so, yeah, and of course, like all these things take like sometime you've got to do like there are these traditional things forward once you uh declare what's called like a development candidate, then you've got to do a bunch of other work about manufacturing and additional safety work before you can get into people. So, all these things take a little longer.

Speaker B: Yeah. And so you're going to spin out the idea is to spin out a so people maybe even internal they will say like I really want to take this this thing to market or try to get it like take it to the next step.

Speaker A: Yeah. Yeah, I mean one way that we kind of think about it is um you know we built this really great engine for like data generation and and model training and and now as we've talked about like these reinforcement learning reasoning approaches um and most other domains when you build an engine to go do that stuff to go like help on those problems um the subset that you're generating along the way is not itself intrinsically valuable. No one wants to sell reasoning traces about like political theory,

Speaker B: but like there are going to be little bits of gold that we have generated as a consequence of doing all the screening

Speaker A: because you're actually reasoning about not only that this model is good or sorry, this molecule is good, but that this here is why that's good.

Speaker A: Here is why that's good and here is chemical material that's proprietary to us that has, you know, activities that are in, you know, protein target classes that others might find hard to to drug, right? And so you've got starting material for you know right now we have starting material that looks pretty selective to us for uh I think like 700 targets like it's enormous number uh you don't want to chase all those down uh but you know as as opportunities they are all opportunities with different levels of you know kind of economic value capture and desiraability and so you think all right well we got all this stuff like let's see what we can get out let's see what we can get out of it.

Speaker B: Discovery is like a small percentage of the total cost of developing a drug and you can argue that well you have a lot any given discovery is cheap maybe but that you know because of the failures you have a lot of them. So but e even then you know

Speaker A: to get one drug one drug like all the steps of a single drug candidate

Speaker B: y

Speaker A: that is successful is maybe you know on the low side 200 million on the high side a billion um depending on who [clears throat] you have to test on and lots of other factors but and then you know the discovery is a couple million so so are you tackling the right problem here

Speaker A: um and so I think there are a number of early discovery problems that are worth that much, you know, like like reasonable amounts. And so what does that mean? There's a bunch of targets right now with like really good genetic evidence that are very challenging to drug with small molecules. Um, you know, a pretty straightforward example is, uh, for a lot of these, um, uh, kind of cytoine inflammatory indications, folks are using antibodies to drug them and going after a small molecule therapeutic that you could take as a pill instead of an infusion, one that doesn't require uh, like cold chain, you know, transport uh, to get to different patient populations. Like having those things is enormously desirable,

Speaker B: meaning you have to have refrigerated trucks,

Speaker A: right? Yeah. Like and like if you want to get these if you want to get these antibbody based drugs to some of these folks like this is like a very hard problem.

Speaker B: Yeah. And very expensive

Speaker A: and very expensive and so and so folks are like really quite excited about can you find a small molecule therapeutic that does the same thing uh that antibbody against these

Speaker B: so there's some antibbody it's it works.

Speaker A: Yep. And so we know it works

Speaker B: and and so you're like well why can't we just go after the same target with a small molecule?

Speaker A: And it turns out that those proteins are very hard to drug for a variety of reasons but we like to think that you know the technology that we're developing can can get good material. And so we do in fact have selective material for a lot of these cytoines. And so that's one example of something that's like a very desirable target where of course like there's going to be much more uh capital requirements to like get it across the line to actually get it to market but the interest is sufficiently high enough.

Speaker B: So you're going to ruin the antibbody market basically. I mean if [laughter]

Speaker A: like if you can like then it's good for some stuff like it's great for some stuff and like there's a lot of energy being thrown you know at those groups because we're good at protein folding right and so like they're making a lot of progress and maybe more broadly I would say is like going after the right problem. It is a little bit of a lampost problem where the things where we can make enough scalable measurements in the life sciences are these early problems. we can't really run a million mice experiment, right? Like we can't run many millions of phase one trials. And so we're going to get better at that stuff. We're going to start to like go up these different layers of biological abstraction with like better data and better reasoning and physics models. But like right now like the the technologies that we have to really make the number of measurements that we need are going to be like these early approaches I think.

Speaker B: Yeah.

Speaker B: Yeah. And I would argue also that you can if you have a oracle some AI oracle that says oh this like all the admmit properties will be good and this generalizes to exactly this population. So you have this oracle filter then all of your failures in clinical trials go away.

Speaker A: Yeah you can imagine a world where you can come up with an or you have some new target that you know some new evidence that covers uh and then the Oracle could give you small molecule material instantly that you would feel comfortable running straight to like a mouse study.

Speaker B: Right. Right. Right. So, but so that requires training on data from mouse studies, clinical trials, whatever.

Speaker A: I mean, maybe I mean I could say definitely like AdMet properties if you're confident in the target and you need to make sure that it's not hitting all these other off targets. Uh but even then like that's

Speaker B: you probably don't need clinical results to feel the confidence to run off and run a mouse study right away.

Speaker A: All right. Well, we're on, you know, everyone's favorite pet question. My favorite pet question is what I call the koseri paradox of machine learning and discovery.

Host: Yes. Yes. Um, okay. So Shri Kosuri, great tweet. I love this tweet. It was, yeah, he was on with us a few weeks ago. We hopefully will release his episode in the next few weeks. Sadly, I actually didn't even bring this up on the episode. I'm so disappointed. Anyway, but yeah, the tweet is basically: what is the business model in machine learning for drug discovery? Because in order to build a good machine learning model, you need lots of data and good high quality data at scale. You clearly have that. But then if you need that to train the model, like what do you need the model for? Because you just go collect the stuff, and you clearly can collect this stuff quite well. So yeah, we talked a lot about data collection. We talked a lot about machine learning, and I'm actually kind of curious in some ways: where does the actual machine learning—is it necessary for your business model?

Guest (Leash): So, I mean of course it would depend on the business model. I think if your only rationale is to go off and get good chemical material for a particular target, then of course Sri is right, right? You would have it. You could get started if you had enough chemical material broadly cast across all these other targets in a way that can kind of drive a deeper machine understanding of the chemical behaviors, then the way by which you optimize those early materials have the potential to go much faster, especially if you can do it at the scale that we've done it, especially if you can do it at like the low capital infusions that we've been able to do this. We have not raised a lot of money at Leash, and we've screened an enormous amount. And then I think getting even, kind of even stepping back, one of the reasons that I brought up this language model stuff is because we are starting to wonder if it would enable a much larger sort of scope of capabilities than just identifying novel chemical material. Like if we get really good at chemical reasoning, then we can get really good at lots of different industries and their applications without having to physically go and screen in all those different domains.

Host: Like what's an example of that?

Guest (Leash): Uh, so let's say we get really good at making these predictions, then we can run off and go work in agriculture like immediately, right? You could go work in veterinary science, you go work in material science, and you don't have to—if you're really good at it—you don't have to screen all those proteins too, because you have like a good fundamental understanding of this, right? And the same thing could be argued for like all the protein folding companies, right? It's like now that these models are good, there are teams that are going after protein targets that have not been screened yet or are very challenging to screen, and getting like really good starting material, like the Novel guys did it with this basically like a receptor on cells that has like a little tiny piece of a mutated cancer gene that's like really hard to do, and they managed to pull it off. And so you can imagine, oh, there could be a world where we do this with small molecule material as well. And so like I don't think Sri is wrong in terms of finding like start—you know, if you're just trying to drug this one thing and you screen a lot and then you already have the thing, you don't need a model for the thing. But that's not exactly the same as generalizing across all things. And you know, we are hopeful that if you have enough of this variation of engineered activity cliffs, that you will actually be at a point where you can do all things, in which case at which point you don't need to screen at all ever again.

Host: Yeah. Well, and arguably you guys, if I don't—I don't know exactly how much you raise. So I mean you're for a campaign—you're talking that's like a couple campaigns that would cost the same amount.

Guest (Leash): Uh yeah. Yeah. I mean maybe to put some real perspective on it. Um, so we have developed and screened about 2,000 different protein constructs, um a total of probably about 5,000 screens in the history of the company across at least 6 and a half million molecules each.

Host: Which I don't want to get into like a huge numbers game, but it's certainly almost as much as anyone else on Earth. Maybe more than everyone else.

Guest (Leash): Probably not. There's probably a couple groups that like if you know... but like the number of mutations, I think it's probably more than everyone else. That seems likely to me. Uh and so yeah, like we just put a huge amount of work into making this like very cheap and very scalable. And if I was doing one at a time...

Host: Yeah. If you went to a CRO and you did one at a time, it would probably cost you 100K each depending on, you know, how much handholding you got, right?

Guest (Leash): Right?

Host: Yeah. You just said five million six, what is it? 6 and a half million molecules. That's not a lot of molecules, especially since it's combinatorial space. You know, everyone say 10 to the 40 or something. Big numbers. I mean, obviously, you don't need to understand all the chemical space.

Guest (Leash): You just need to make drugs consistently, repeatedly. That's your goal.

Host: But is that actually enough chemical space to meaningfully do anything outside of your chemical space?

Guest (Leash): Great question. Um, so the pleasant answer is, you know, I said earlier that we have trained models on internal data only and that we're able to extend and make generalized predictions into ChEMBL, which is very different corners of chemical space. So yes, do we think it's enough? Not remotely, you know. And so part of our road map forward is making, you know, larger numbers of chemical libraries in different corners of chemical space, and kind of spanning kind of what I talked about earlier where some are like these very fine-grained changes to more appreciate activity cliffs, and then the other side it's like these big coarse changes to try to capture new proteins and sort of protein shapes.

Host: This is—sorry, not a hard question, just I'm genuinely curious: are you able to predict like potency, like relative potency meaningfully? Like can you do KD curves, like where you know titrate like more and more larger libraries, or?

Guest (Leash): So people have wrestled with this a little bit in terms of trying to get the KDs directly out of DELs, doing things like maybe varying protein concentration is something that folks do. One thing that we see is when we do replicates of screens, we get back the same hits in roughly the same order, which makes us, you know, kind of optimistic that you could get these kinds of things. We started to move more into SPR, biophysics-related measurements, try to figure out if we can maybe come up with like a potency ladder that we could then be as part of the chemical material that we're screening to try to capture those things. But I won't say that like we're good at it yet.

Host: Like you can't just use reads as a proxy for KD or something?

Guest (Leash): Um, I think the relative ranking, yes, but I think reads for KD—reads for KD, not really. And then you start running into weird stuff. So like there's lots of proteins that we've encountered where it turns out that like the DNA barcode itself... there are proteins that are not described as DNA binders that we have discovered are. And so you can imagine that that might affect the KDs of things, and you have to figure out how to mitigate those problems and that kind of stuff. So like, yeah, I think the relative ranking is something that's quite doable. I think translating it into KDs is something that we're actively working on, and we think we might be able to get like some reasonable, you know, kind of like sloppy translation, but I wouldn't say that we're there yet.

Host: Okay, this one, this one was pretty brutal: The Bitter Lesson. Yeah, I'm just going to read this. This one is pure Claude. I had no thought in this: "The bitter lesson is really the law. What stops a better funded lab from outscaling your DEL engine?"

Guest (Leash): Oh yeah, of course people ask us this. Um, so and the reason like, you know, here's what I...

Guest: would really say is that uh both at leash and at recursion uh my team has made so many mistakes in accidentally bumping into the cheating modes that just kind of come out of the consequence of doing laboratory work. There are so many of these and and so like you have to chase these down and a lot of it has to do with tracking batch effects with metadata for example and so then you have to build a culture at your organization where it's like all right we have to track all these things we have to build tools track all these things we have to make those tools frictionless that the people doing the work actually will record these things uh and then you have to go back and look at all the data back into the founding history of the company to make sure you're getting really consistent results and then when you see big changes you have to identify what drove them then you have to mitigate those and it's just like it's like a huge huge amount of work. And we also rebuilt kind of from scratch the selection process. We rebuilt from scratch a lot of the protein expression process. And then we have all these other proxies that we use to evaluate how well the protein production went and if it's good or if it's bad and these other filters. And it's just um it takes kind of a unique culture and patience and willing to make these mistakes and then have the team like really think very deeply about how to address them. And I think unless you've got folks who, you know, really

Host: process mode.

Guest: Yeah, I really think so. There's just so many things where it can go sideways and you have to be paying enough attention all the time to identify them when they're going sideways and they might be things that you can write automated filters for and maybe not early on. Um, and so like you know, like one pretty good example I'm happy to talk about is um people ask us a lot, how do we know our proteins are any good? We have lots of ways to figure this out. I think of one of my favorite ones right now is um we have noticed that when proteins misfold and we're able to identify misfolding in very crude cases like if you put the protein you know run on a gel which is this protein trick where you try to figure out how big it is um and it misfolds then you can kind of detect that but if it misfolds a little bit it could be hard. There are drugs whose job it is to refold misfolded proteins called molecular

Host: chaperoons. Okay.

Guest: You can call crers too. That's what we would call them. Yeah.

Guest: Uh there's a very famous one that treats cystic fibrosis patients called lumicaptor. We have a bunch of chemicals in our chemical library that looks like lumacaptor. This thing whose job it is to refold misfolded proteins. Guess what kind of proteins it hits when we screen them?

Host: Cy fibrosis.

Guest: Uh misfolded proteins folded proteins in there. [laughter]

Guest: That's what it's for. And so like and it's in our libraries. And so you know like and we discovered this by accident like we saw this you know these chemical series keep popping up across all these different protein types it's like man like what is that is that captor and then you start looking at the protein expression and then like how it went it's like oh it it goes when when things aren't going well and so now we can use it as one filter for if we think if if we get a bunch of lumac heads that protein probably didn't fall fold right. We probably don't want to use that as training data. we probably want to rerun it with slightly different conditions and make it more friendly to the protein. Then we'll see what those hits look like. And so like unless you're paying that kind of attention like it's really I mean it's really hard. [laughter] Yeah. And you have to build tools so that you can just look at these things and have people who know to look for these things and recognize them. It's uh Yeah. It's it's a hard problem.

Host: Yeah. Going back to uh everyone who touches Dells comes out traumatized. Somehow you've uh you figured out supported this.

Guest: Yeah. [laughter] You're also traumatized but in a good way. Yeah. No, it's really hard. That's, you know, and that's why it's it's and and and you know, when you have the motivation to just finding like one or two really awesome hits to start off a drug program, like then you're not going to chase down like all of these different aspects of the assay. And so I get it. But like if you're motivated for it to be machine learning fodder, then like yeah, you have to pay attention to all this stuff.

Host: One of the questions that we always ask at the end is if you could by fiat remove a bottleneck, what would that be?

Guest: No. But when I think about Bif FIA and it's one of these things that is actually currently happening which is watching the sequencing costs drop like a rock like I'm thinking about okay what do we really want to do we want more chemicals and we want more measurements those are the things that I want to make my life a little bit easier like at the problems that are immediately on the ground

Host: so you want sorry I don't understand cheaper sequencing more molecules

Guest: okay yes

Host: in sequencing like can you give some context for how much it's dropped

Guest: oh my god like many many many orders of magnitude It is uh I think

Host: the first genome cost a billion dollars and now we can do a human genome for a thousand, right?

Guest: I think it might be 100.

Host: Really cheap. Yeah. There were stretches where it was dropping faster than Mo's law.

Guest: Yeah.

Host: Was so wait why then why do you want even cheaper sequencing?

Guest: Oh, I just I'm just I'm just happy with the trend.

Host: Okay. What would you remove by fiat if you like that hasn't been removed?

Guest: Yeah, here's a good thing. What I would want is there are different gizmos that you can use to kind of purify small molecules. Okay. So when you uh right now when you make them you you put some stuff in a pot and you you heat it up or do some other conditions and you get the output and and that will work at some level of efficiency. Might be great, might be crappy. And so in either case, especially if it's crappy, there are these other gizmos that you can use to kind of clean it up. So people use like HPLC or some of these other tricks. None of them are super parallelizable that are really really good. And so being able to like clean up stuff at a massive scale would be really really nice because like we you know what do we want to do? We run all these molecules in this funky assay that we've already described has given people lots of trauma. It would be great to then get you know large numbers of molecules in a different assay that would be really clean and purified and easy to work with uh and then go do that next because then you could really orthogonally validate all those measurements that you're making. Yeah, that would be really really nice. And of course, everyone complains about like GPUs and stuff like

Host: GPU that's what everyone says like besides [clears throat] GPU. [laughter]

Host: All right. So,

Guest: and then the other question is um what would you like people to take away from the episode? I would like people to take away from the episode that like you know at least for us we we believe that we're building something that's really really exciting and you know like what feels like physically grounded substrate for all sorts of interesting reasoning and biomedical applications and so if that's of interest to them like please reach out to us like we're you know we're hiring machine learning people we want to talk to partners and so you know and we're building this thing that we think will be transformative so come talk to me.

Host A: Amazing.

Host B: Cool.

Host A: That was great. So much fun. Thank you for coming.

Guest: Thank you. Yeah.

Host: Yeah.