Iddo Weiner: Data, moats, and open science: Building scalable AI applications for bio

NicolaHi Iddo, welcome to Models and Molecules.

Iddo WeinerHey, really happy to be here. Thank you for the invite.

NicolaWe're gonna talk about Converge and about AI and all the sort of jazz later, but I wanted to start with a bit of a personal question. And indeed, you started Converge a couple of years back, but you have been at the intersection between data and biology for more than a decade now. So, can you tell us about your story, how you ended up doing that?

Iddo WeinerThis takes us back to academia. During my undergrad years, I started out as a biologist. I still call myself a biologist. Some people call themselves data scientists, ML engineers, bioinformaticians. I call myself a biologist. I'm studying biological systems by any means available. I started out running experiments on the bench, wet lab experiments. I remember I was fascinated by the fact that you can change the way living cells behave while they're living, that was mind-blowing to me. And that's when I said “Oh my god, this is what I want to do”. And then the first time, I remember the first time I sent out a sample for deep sequencing NGS, the data came back, and it was this huge file full of letters and I said, okay, I probably should learn how to program so I can read through this file. That was the moment I started falling in love with the computational side of biology. An interesting story there, by the way, is that professors and some friends of mine and all kinds of consultants told me, listen, you need to choose what you are, you're either an experimental biologist or a computational biologist, or maybe you want to be a data scientist. But pick one thing and be good at it, because otherwise, you're going to end up as a duck. That's the metaphor. A duck can walk a bit, swim a bit, fly a bit, but it's not exceptionally best at any of those things. And my wife actually, she gets all the credit here, she told me, be a duck. Ducks are awesome. And that's what actually happened. I became a duck. And I liked the machine learning side of things. I kept on studying computer science and engineering during my years in grad school. I kept on performing my own experiments in the lab. I went to conferences on computational biology so I can call myself a bioinformatician. I never really chose which one of them I want to do and here I am today, a duck, and I tend to also hire other people who are ducks. I like the interdisciplinary background, I think it's an advantage and that's like that's in a nutshell where I where I started.

NicolaI think you added some more to the duck becoming an entrepreneur, right? Founding Converge, how did that happen? Because you know, I understand you were a scientist, you are a scientist still, but how did you add the entrepreneurship on top of it?

Iddo WeinerAll that was chapter one. That's academia. Then fast forward a bit. I spent many years developing drugs after academia in different companies, mostly from the computational side, but also pipetting in the lab and having things explode and stuff. Always fun. About three years ago, I started talking to two good friends of mine, two colleagues who became co-founders here at Converge with me. And we were thinking, there's all this excitement and innovation going on in the world of NLP. This was a bit before the ChatGPT moment, but everyone was feeling that it's coming. One of my colleagues comes from a business background and the other from a data science background, and I'm the biologist. And we started talking and we said, there's got to be a way to bring all this excitement and all this technology into life sciences. Because many features are very similar, right? You can think of in natural language, we have words, which are bits of information that contain a lot of value in them. And it's kind of similar to if we think of amino acids or nucleotides in molecular biology, they're like single letters, tokens, words, call them what you want. And they contain a lot of information standalone. Unlike, I think, if you think of pixels in images, it's not really the same. Many people today are talking about us needing a different type of modeling for images. But I think natural language and molecular languages are very similar. Information is discrete, it's not continuous, we have interactions. If you think of a protein folding or if you think of different genes in the genome functioning together, it's like long context dependencies like we see sometimes in language. If I'm reading a book about a specific topic, then I need that context to understand what I'm reading on page 150. There are many pieces that seem to fit very nicely between these two worlds. And really, we started from that, from the technology sounding pretty cool. And we didn't really have a business model in mind. So, it's not like we founded Converge to design antibodies to increase protein expression or to find new targets. We founded Converge to bring the excitement of NLP and GenAI into molecular biology. And I think from there, everything just started clicking. I never dreamt of being an entrepreneur. I see myself as a scientist. I just wanted to do cool science. And I felt like this is really cool science, it's important, we need it, people need it, societies need it. Becoming an entrepreneur was like a natural next step from there. It was never “let's just do a startup and figure out the details later”. It was really based on “the technology is cool”.

NicolaI can imagine, and you said that of course the idea came to you guys before this so called ChatGPT moment, right? So, did how you experience that? You know, founding a company at the same time in which that was happening, I can imagine that might have been quite a wave to ride.

Iddo WeinerWe and many other people in the world were kind of seeing it coming. Basically, before the application was out, we saw the models and their performance on benchmarks. And then the application, the ChatGPT API that hit the world and everyone was like, oh my god, was kind of like the application layer on a really, really cool technology. We loved it from the first moment and we try as a culture here in the company to always be early adapters of any novel models coming out and we're still operating in that way today. Basically, what we saw in that, what we saw in like GenAI, in the natural language space, we see it as complimentary to our offering, because we are building molecular models. If you can use a language model, it can swap out things like customer service or explaining your technology to a customer. That's how we positioned ourselves and built our product. We have these agents like many others have that talk to the customer, explain results, take input from the customer, wrap around the molecular outputs, but then the core technology that we are building is the molecular side of it. And we were super excited about ChatGPT at the time, about Claude Code around five months, what is it, five and a half months ago.

NicolaI can totally imagine that, and I think you said something indeed interesting, right? You were talking about a comparison between this large frontier language model and more like this molecular model. I also believe there are many similarities you mentioned, for instance, the fact that they are kind of discrete, they operate on discrete elements. I'm sure there are also differences, right? So, for instance, the things that come to mind to me, well, while a language model is maybe more limited towards the scalability of parameters, so you want to scale parameters just because you can throw it so much data. Maybe when you are on more let's say molecular type of generative models, perhaps, I don't know, curious to hear what you think about, just the amount of parameters not the limiting factor. Maybe more the data or the diversity that you throw at it, maybe a limiting factor. What's your thoughts there?

Iddo WeinerThe size of data is extremely tricky. I think it's actually one of the motivations for "let's build LLMs in molecular biology". I'll maybe start with a short story. In my days being the computational team lead in a company that's developing drugs, people would come to me with all kinds of requests. Like, here's a molecule we're working with, and here are the outputs on 30 or 40 different assays. Can you look at all these assays, do some hocus-pocus with your machine learning and improve my molecule. Or we just finished Phase I, we dozed 17 patients, 7 responded and 10 did not respond. We don't know why, can you run a machine learning classifier and tell us how we need to select patients for Phase II. So, like all these big questions in biology and biotech and drug development were and still are constrained on the amount of labeled data you have. You can't build a machine learning model on 17 data points. The tools we had five years ago were things like decision trees or regressions. The only thing you can do there really is start with those 17 or 35 data points that the teams bring you from the lab. So, in biology we're always constrained on labeled data. What we do have is unlabeled data. It's like, how many protein sequences are there? How many sequence genomes are there? The answer is surprising. There's one to two orders of magnitude more molecular data out there than there is text online. The bank of unlabeled biological data is enormous. We're not constrained there. We couldn't do anything with that before LLMs and self-supervised training. So, we didn't really care about that so much. But one of the big inventions that led to the GPT moment or whatever historians decide to call it is the self-supervised training, meaning that I don't need labeled data, I just need data that represents the world I'm interested in. And then I do whatever with it, I hide some of it, and I try to complete it based everything else, you can do that with proteins and you can do that genomes. So that piece I think is okay for the pre-training. Where it gets complicated is how do you then post-train your models? And for that you do need labeled data and there the problem kicks in. Let's talk about antibody design, we're very active in that field and many others are, it's a very hot topic today in AI for bio. How many well-labeled antibody, antigen sequences do you have with structural data, with binding affinity measurements, et cetera. That's where we start hitting a data wall. And you're right, we can't just increase the size of the model and throw more compute at the problem and just hope everything will be solved. So yes, you are right, and that's why we need alternative methods. I think eventually we will get there. But in the meanwhile, you need to be, and this is our approach, you need to be very clever in the way that you go and collect your own data. Instead of, for instance, burning your entire experimental budget on a huge screen that will give you a lot of data, but in one dimension, our approach at Converge is let's use that same budget. And it's a significant part of our R&D budget goes for experiments. Let's break it into baby steps and run a lot of smaller experiments, where each time we carefully choose which sequences we want to go and experimentally label, bring them back into the model, retrain, make new predictions, see where it works, where it doesn't, where is the confidence low, go and label those and vice versa. And we've been doing this for two years. We now have about 10,000 data points that we've collected this way. And it's an ever-progressing kind of effort.

NicolaWould you say some type of active learning in which you basically kind of understand which type of next piece of data you need to generate to solve the problem that the previous piece of data surfaced?

Iddo WeinerActive learning is exactly the term. It's a word that gets thrown around a lot. And sometimes people mean different things when they say that. Our approach is definitely inspired by active learning. And again, the road until we are at a GPT-3 moment, we're not there yet. In many ways, we're in GPT-1 days. Like the amount of tokens we have to train on labeled data is about at the moment we had Converge have about 1 billion tokens of data. And when you tokenize like one amino acid it equals one token. So, the amount of data is much smaller. That's roughly the training data size for GPT-1. So that's roughly where we are. And that just means that the models hallucinate quite a lot. But the good news is that you can still build pretty good products off it. Imagine that you're a customer and I generate some antibodies for you now. And let's say I generate, we usually generate 24 for our customers for experimental testing. Imagine that our hallucination rate is 90%. That still means that you get two or three antibodies in this campaign that hit your spec. And that's pretty good because our process takes a few hours and voila, you got your antibodies. So, models are going to get better and better, but we do think that in the meanwhile there's already a valuable product to put on the table which is the good news.

NicolaAnd do you have a feeling for how much this is converging towards the ChatGPT-3 moment? Meaning, how far are we from something like that? Or what do you think we're really missing to get there? Is it just the amount of data, is it something else?

Iddo WeinerWe see ourselves as a player and a community, and the community is going to succeed together. And of course, we want to build, we want to push forward, we want to lead. But we also think that data sets for the community are important, and sharing progress, sharing ideas, open sourcing some data sets, open sourcing models is something very important. So, every time I say "we", it's like "we" Converge, but also "we" as the community. I think we will get there and any predictions today in the GenAI era of how fast things will take. You always go back to them a few years later and say, wow, that was really off. So, I'm tempted to say faster than some people will think, but we’re definitely not there yet. In terms of the constraints, so definitely data. And there are many efforts to collect more data, to build more data, more data sets. I think that's like the trivial answer, right? I think there's another answer. Of course, we need more data. And anyone who's been working a lot on trying to build models for molecular biology probably is aware of this. But for those who are not, then maybe this can be a shortcut. The first thing you find when you start modeling like real biological challenges is that the homology signal in the data set is extremely strong. So, for instance, any interesting challenge that you might take, you can very easily solve naively with just homology inference. I'll give you an example. Let's say you're trying to model antibody-antigen binding, we're with the antibody examples today strongly. This is considered a very, very hard challenge. But if you build a random data set of here are many antibodies and antigens and let's say I take them to the lab and some bind and some don't. And now I ask you to predict which ones will bind and which won't. A very simple algorithm that will go to whatever databases is out there, a public database, and just like do a kind of a nearest neighbor say, okay, this antibody, its closest hit in the database is this other antibody that does or does not bind. That predictor is going to work unbelievably well and that is of course completely cheating, right? We're not learning any biochemistry, any biophysics. That type of model cannot help me engineer because then if I make a mutation that type of model is still going to think that I have binding or not because I'm still similar to that same hit.

NicolaThis is a very interesting point that you're making here about the homology, right? Because I totally see the part of cheating, but do you see a possibility perhaps to use that information in the opposite way? So, if I find something which actually binds but is not homologous to something else, use that information actually to derive something which goes beyond the homology.

Iddo WeinerYeah, exactly. It's both a challenge and a big opportunity. Just to bring home this thought, I sometimes talk to people who are building models in NLP, and they just don't have this issue. People who build models in NLP, there's something called Common Crawl. You just bring in all the information that's available on the worldwide web for everyone, you bring it in, you sort it into documents, and you start your training. And there are, you know, millions, billions, trillions of these documents online. And you can just start training. A good metaphor is: imagine that all these millions of documents online actually have like 1,000 or 20,000, if we think of genes in the human genome, original documents that were just copied with minor changes into millions and billions of documents online and then you just start modeling all that, that's going to collapse very, very soon. So just throwing a lot of biological data into these very powerful models always collapses into the model just learns homology. This is both a problem and an opportunity. So, if you're aware of this and you stratify for homology very properly, then you start seeing the signals you actually care about. You can even harness further and say, okay, here I have homology, but my label actually flipped. Those are the most interesting cases, and I learned the most out of them. So, by the way, when I described earlier the way that we try to build our data set, those are exactly the type of data points we're looking for. Very minor changes, but a flip in or a big difference in the experimental label. I actually think that too many people in our community today are just brute force throwing a lot of biological data into LLM classic models and reporting these awesome results that are actually just model collapse into homology. That’s unfortunate because it's in a way a bit a waste of resources and the more we're aware of this I think the better the models will get and ideally, we're going to learn what the best way is to control for this homology and kind of come up with new pre- and post-training techniques for biological data.

NicolaI follow your reasoning here and it makes a lot of sense. At the same time, I wonder how much data we do need to generate to find those pearls, right? So, or perhaps maybe active learning is a way to get there and reduce the amount of data and just get there. But that was my question, right? So, do we need to generate a lot of data to get to those pearls, or?

Iddo WeinerSo ChatGPT basically, and this is like I didn't do my homework well enough, so this is off the top of my mind and sorry if the numbers are a bit off, but ChatGPT is basically GPT-3, which is immense, but that's only in the pre-training, plus post-training with about 10,000 data points, like 10,000 well-labeled sentences that got like prompts and questions and answers that got a score from a human in the loop. 10,000 was roughly 10,000. Yeah, I might be off a bit, but 10,000 data points well labeled for post-training of a good foundation model equals ChatGPT. So, I don't think we need like, you know, millions and billions of labeled data points, maybe 10,000 or 20,000 well labeled data points on top of a properly trained foundation model can do the trick. We might see that like we're building these types of things. I know others are. Again, I say "we" as a community and also "we" as Converge. I think we might see that model out there for everyone to use sooner rather than later.

NicolaAgain, so we're talking about data, we're talking about way to get to the data or to generate the right data. There are people that say the moat is not the model, the moat is the data for AI companies in this field. What's your take there?

Iddo WeinerThe question about moats is always interesting. I just read the other day someone who said “the only real moat is brand”. So, like, I think that's maybe a bit too extreme.

NicolaMy marketing friends will be very happy to hear that.

Iddo WeinerYeah. Definitely if you have a good brand and that's a great moat and you know, us scientists, we don't usually think about the world that way. But yeah, that's one moat. Another moat obviously is regulation. So, if you have let's say a drug that passed, you know, that has approval from the FDA to be dosed to patients, that's a great moat because it's not very easy to replicate that. If you have a patent on an asset, that's a good moat. And then people who are developing software or AI models in bio, we don't typically have any of those moats. And then the question becomes, is it the data, is it the model, is it the services that come with the model and data together? I'm not sure I have the answer. I definitely don't think that having a good model is a good moat over time. And I can tell you how we think about this at Converge. We have a mixture of using open-source models, using models that we licensed or tools that we license. For instance, we have a license for tools from the Baker Lab and then the University of Washington. We think they put a lot of work into that, and we don't want to try and compete with what they did, we’re just using their tools. So, we do a lot of that. We'll take an open-source model, we'll continue its training. We also build our own foundation models from scratch in cases where we think there's not something good enough out there. So, really a mixture of everything. And our focus is on value. We're trying to always think, who's the customer? What do they need? How can we support them? Sometimes it’s with a better foundation model, sometimes it's with a better application layer. How this will play out over time is, I don't have a good verdict for this. I think it's probably going to be a combination of a lot of things. If you have a good model or a good infrastructure that allows you to always stay with the state-of-the-art model under your hood, together with a proprietary data set that you use to finetune or post-train it in any way. And on top of that, you have like an application layer and services that over time build brand and users just like using your products. That's probably my best advice for anyone thinking about this. At the same time, I realize it's like saying everything together is almost like not saying anything. Because this is, this is a tough question I admit.

NicolaYou don't think that we will see a sort of commoditization of those models such that at some point it will be kind of interchangeable, which one people will use?

Iddo WeinerWe have Anthropic and OpenAI. Everyone knows this, so I won't go into the details. But basically, the foundation models seem to be kind of interchangeable and similar scores on similar benchmarks. And we can probably throw 5 or 10 more names into this list. It's not like you will have, and that's a huge field and potentially bio is a bit smaller in its reach and usage. So, I assume that there will probably be a few kind of leading companies having state of the art models, but it's volatile. It's going to change often. A model alone is probably not a good moat. Datasets, probably a bit better than just models but I guess the point I'm trying to get at is that you I think the best moat is having a good way of putting all these components together and building value that users actually enjoy communicating with and that actually helps drug developers develop better drugs. That's really what's going to be the moat over time and I don't think anyone is close at the moment to solving that.

NicolaYeah, I 100% agree on that. So, we operate also as in an adjacent field and we see the same. So that was actually my next question. What do you think about service or, let's say, being able to translate all these different components and all these different models into something, an outcome, something which is valuable for the customer, for the company that you're talking with. Because that I think at the end of the day, that's what the customer needs. It doesn't need a model, it needs an outcome, right? And I think that at Converge also I think you phrase yourself as AI partner or something like that. So, I think that's also something that you believe in?

Iddo WeinerYeah, totally. Basically, different companies decide on their distance compared to the customer. You can put models out there and charge per API call or charge for the compute of using your model. Once upon a time, we had marketplaces for models, today, everyone's talking about connectors. That's one business model that, if this is the customer, it puts you at some distance from it. Our approach is a bit different. We want to be very close to the customer. We have our own environment, our own web application where our users log in, they have accessibility to all the services we provide, they have a solution scientist that they can get in the call with anytime they want. It's very common that our users want to get on calls, show us results they got from the lab, think with us or with tools that we put online or on the app, think about what's the next step, help them plan, they have some success here, now they want to do something else. So, our approach is a more intimate approach with our customers. There's a variety there, and different companies are positioning themselves in different places. But that's definitely a differentiator between different offerings. How good are you at actually being able to understand what your customer wants, and can you scale with that? And that's by the way, one of the places where our agent, like the NLP agents, the classic agents kick in because today it's very easy to kind of download the knowledge that we have and our scientists have into this agent and offer that as kind of a first line of customer service. And if that's not enough, then of course that human is going to come in and listen.

NicolaI think that part of being close to the customer is definitely a very actual, very in this moment important business model, or at least that's how I also personally feel it is. Also, because in the past there have been some comments regarding various, let's call it AI proof of concepts project, things like that that did not really show the value of those initiatives. So, there's been a bit of negative press around those. But I think partly it is because each organization, especially enterprise organization, is very different as its own internal structure, his own internal need, and there is a need as you expressed, probably to be close to that to understand and deliver that value that eventually is what this company needs, which also shapes a bit different business models when it comes to companies, especially maybe software companies that approach biotech and how they call them today, tech bio type of companies. So, what's your view about there? Because I think there are many thoughts around different business models. There are people that say, well, historically software companies did not have much traction in the biotech field. There are people that says, well, this is going to change. There are other people that say, well, actually if you have a great model, why you don't make drugs yourselves rather than selling it to other companies. What what's your thought about business models out there and maybe what's going to succeed?

Iddo WeinerFirst of all, I think the elephant in the room is indeed that, like if we look 30-40 years back at software companies for bio, we won't find many huge companies with huge success stories. We will find companies that started out as software, built something interesting, and then decided they're just going to go and become a regular biotech and either license it to some pharma, the asset that they found with their computational tools, or become a clinical states company themselves. That's the elephant in the room. I am pretty bullish about the idea that this is changing because of the technology. I think we're in a place today, it's still early, we're only two or three years into this revolution, where offering out scientific services as a software is for the first time something real, something reasonable, something that we can build. And it has two components to it. The first is why today we can solve some of the more challenging problems with machine learning, which machine learning has been around for a while, but not this type of machine learning. And that's what we covered in the previous questions, so I'm not going to go into all that, but I do think the science itself is in an exciting new world. And the other side, which this is what investors always look at, and it's the right question, is how do you scale? Because if I can serve 5, 10, 20 customers and they're very happy, but I need to have 5, 10, 20 people who are actually serving these customers, I'm not going to become viral, I'm not going to grow, I'm not going to be able to scale. And that's, I think, where the other part of AI, like the NLP and agents come in because we can understand all these communication events with customers and we can load that. And here we don't have to develop anything. This is already there. Any type of customer support, you go online today, you probably meet an agent, so we don't have to invent that. That piece is already there. We just need to fine tune it a bit with our know-how. So, we have good science. We can build good science with these new models. And we can scale with these new models. So, the idea of having a SaaS business model in the world of biotech, I think, for the first time in history, is actually achievable. And we're going to start seeing big companies who do exactly that.

NicolaWhich I think is an interesting element because it's my understanding, or at least that's how I see that, that yes, there are SaaS companies that are indeed scaling and becoming very valuable for biotech. There is also, let's say, ambition I think from pharma companies to be closer to tech and becoming maybe more tech companies. So, there is that trend that I see. There are some pharma companies out there that say we want to be the technology company of today to be the medicine company of tomorrow. You know, that type of vision. And, there are tech companies that want to be very much pharma companies. Perhaps also what recently happened with Anthropic. They really are starting their own program. So, I see a lot of, let's say, crossovers. What are your takes there?

Iddo WeinerLines and boundaries are getting blurry, which I think is cool. It's exciting. If we don't need a boundary, then let's get rid of it. We have models for virtual cells and one of the use cases there is to look at analyze some type of omics data and try to search for novel targets. So, let's say we now built a new model and we want to validate it. So how do you do that? You go and search for a novel target, right? Because if you want to convince customers that this is a good product, then you need to show that it actually works. So, we do these kinds of things and we run this type of model on some cohort and we get a hit. Then what do you with this hit? You go to the lab, and you test it out and you run knockdown or some CRISPR assays on it and you see what happens to the cells. Do you get the phenotype you thought you would? In many cases it works. And then you say like, wait, oh my God, I may have just discovered a new target for some disease. So, I can like put this in my slide deck and tell potential customers like here look it works and that's our main business. That's what we do. But on the other hand, I also found an interesting target like maybe this on its own is commercially viable. So, we have these thoughts all the time and we're not actively developing drugs, but we are like also getting this long tail of like ideas on how one might go about that and I assume other kind of AI for bio companies are as well. So that's one side of it and how does that play out? It's still early, I'm not sure. The other side is... And I think actually, like the announcement by Anthropic from was it last week that they're going to develop some drugs for orphan disease. I don't know if it's like an early sign of more tech companies becoming drug developers or if it's a one-off kind of thing. I was excited about it. At the end of the day, there's so much unmet need and patients need drugs, doctors need drugs to treat patients, that there's people eventually that the reason we're in this field is that there's a lot of suffering and a lot of, again, unmet need in this clinical world. And I think we need more resources here. So, to me, it's a great thing that a big tech company is taking direct interest here. I laughed a bit because some friends of mine who come from the kind of AI side of things said, well, look, they're like going after orphan disease. So, they're probably, you know, they don't want to compete with pharma who might be their customers. And that's funny because orphan disease actually, like a lot of pharma companies are developing drugs for orphan disease. Orphan disease like the FDA and other regulatory bodies have a lot of upsides for companies developing drugs for orphan disease. The regulatory path is a bit easier. Usually, if you are able to develop something that actually works, the price tag is much higher because there's no competition. So, I'm aware of many biotech companies who are going after orphan disease as a business strategy. It's definitely not like a philanthropic effort. It's a commercially viable effort and Anthropic of course is allowed to do this. I'm excited about it. I have no insight into how real it is and what will come of that. All in all, I think that the blurring of boundaries is a good thing. If companies have capacity to have more than one business model, I don't see anything negative in that.

NicolaFrom a business perspective that you mentioned, I have, however, a few questions, right? Some that you also hinted at yourself. How is it that you are, basically you being Anthropic, competing basically with your customers, as you also pointed out, right? Because it's not a small segment. You are, it's a totally viable one. So, you're competing with your customers. So, what's your take there? Second question I had: what's your take regarding Anthropic is going to eat it all? Meaning, is Anthropic going to be the model then and the resource? And what happens to all the other AI companies? So, there are a few questions that I think still raise.

Iddo WeinerYeah. So, the competition with your customers is a tricky point. I'm not sure how to think about that. I'm sure they have given it some thought and weighed it and said, we're going to go for it anyway. It is a risk. Let's kind of dial it back down to an early startup like ours. We definitely don't want to compete with our customers. And this is something that we say. We're not developing drugs. We are giving them services and we are more conscious about not competing with our customers. As you grow and you have such a big portion of the market, you can maybe allow yourselves to take more risks in that sense. So I think for that, the verdict's still out. Regarding the second question about this paranoia that one company, specifically at this point probably Anthropic, is going to eat everything, it's hard for me to believe that that's what will happen. First of all, I'm like an admirer of their work and we use their applications, and we use their models at Converge. So, I'm a fan, not a hater. Looking at the strategy there, basically what Anthropic is building is one type of models. They're building models for natural language. As far as I'm aware, Anthropic doesn't even have a model for image. If you ask Claude to generate an image for you, it will probably do it, but it's going to call some other model. And of course, in the bio world, if you ask Claude to generate a molecule for you, it's going to call some other model. And so I think it sometimes seems like, you know, Anthropic is doing so many things at the same time and going after so many different fields, but actually behind the scenes like they have this one type of model which works very nicely and they have a lot of harness and like the product thinking there is should we should probably give kudos to. But all these other types of models I think are, you know, at the moment very much reliant on other players building all these other types of models. And eventually, there's going to also be a lot of competition on the natural language model, of course. OpenAI is the player in this game, Google's products are now gearing up. So, I'm not too worried about that. But yeah, I mean, I'm probably more excited than I am worried about recent events.

NicolaAnd I think it's a great attitude to have more one of abundance than anything else. So that's really nice. And you said also another thing that I think is very nice. When you say we as a field rather than all on top of we as a company, which I think is very nice. Which brings me to maybe one last question or one last thought regarding how you see open science, open source and all of that in this specific segment or market.

Iddo WeinerWe see ourselves as scientists. And like to me, the definition of being a scientist is or the definition of science is I have some claim, I share that claim with the world, but I don't just share my claim, I also share the evidence that I've seen and I let people scrutinize it and try to replicate it and test it and find flaws in it and that's why science works and that's kind of what in my mind makes the world go round, right? And in academia this works beautifully because you're incentivized basically by peer reviewed publications, which are supposed to be exactly that. I'll skip all the parts of people trying to hide things in papers. I think like 99 % this pretty much works. And then when you get to industry, it becomes a bit more complicated. Now, once upon a time, industry was kind of, okay, someone in academia found something like a whatever, a drug, a target, a method, a purification method, I'd like license that from the academia and now I scale it, I manufacture it, I do the operations and I do the marketing and sales. And it's okay because that's not directly science and there's no reason to be super open about that. It's changing though. We see more and more scientists today choosing to practice science in industry. We see more and more like model building and hypothesis generation in the world of drug development, you know, finding targets, developing molecules, we see all that, developing tools for these things. We see more and more of that happening in the industry. So, this is like, you can almost think of it as basic science being conducted in industry. And here, I think as a field, I think we need to be careful. Because on the one hand, of course, I believe in the free market. And of course, companies are incentivized by making money. And it's OK. We don't need to be shy about that. That's how the game works. At the same time, I think we do have some obligation as scientists to share with the world our methods, our results, to share enough for people to be critical about it, to offer improvements, to try and replicate our claims to see that they actually work. And the moment we stop doing that is, to me, it's a bit scary for science, which is why… So again, here at Converge, I can give some examples. Recently, we published an open-source model for Virtual Cell alongside with a preprint that's also under review now, like peer review. We don't have to, but we wanted to do it because we want people to see it. Every time we talk about like, you know, we generated an antibody that does XYZ. We also give the sequence and we give all the data. An antibody sequence is usually enough for someone to go ahead and replicate that and I would like to advocate for more of that in our field and the one last piece I'll add here: what I've learned is that from a sales perspective, it turns out that this is actually beneficial as well because if you're trying to build a business where your customers are scientists, they actually appreciate this openness. So, if you have negative results, you report them. If sometimes you try and fail, you talk about that. Whatever you can open source, open source. If you can't open source, give me enough information so I can replicate it. And all these things resonate with your clients. So, it just happens that the moral imperative of conducting science and the sales motion of trying to like talk to customers and explain to them why to work with you and how your products actually work. They end up converging to the same kind of set of principles that each company can color as they want. But I believe very strongly in kind of being open, answering questions and explaining how the technology really works and where it doesn't. And it's good for science. It happens to be that it's also good for sales. So, it's a general recommendation I have.

NicolaIf you practice open science, you'd even maybe have an edge when it comes to being able to sell or at least to provide solutions to customers that probably as you mentioned are looking actually for that type of openness to be able to also evaluate and compare, right? Because otherwise, how do people make a proper assessment if they're lacking the possibility to test it on a fair and equal playground, right? And perhaps also what's your take on benchmarks, right? Because I think, at least that's my impression, we are lacking a bit of, let's say, community-based benchmarks in this field, because it evolves so rapidly that the benchmarks that are there are maybe too few or too sparse with respect to what we need today.

Iddo WeinerAgree 100%. Couldn't agree more. We definitely need like having baselines and like, you know, leader boards and performance on these benchmarks is like… people coming in. We constantly hire ML engineers coming in from other fields. And the first thing they ask is “okay, like what are the benchmarks here for bio?” And I kind of move uncomfortably in my chair and say, like, we don't really have establishment. In some cases we have, right, for like for 3D structure prediction of proteins, okay, but like for many fields we really don't have anything and I completely agree that that's a glove that should be picked up at some point.

NicolaAnd it's very interesting, right? Because the real bank, the initial bank in the whole bio and ML bio happened because those benchmarks were hit and were shown. Somehow the attention was driven because these benchmarks were there, and someone managed to really strike some really high score. If those benchmarks wouldn't have been there, I would I maybe we would not have that kind of attention to what was happening, maybe. Unfortunately, we're running a bit short on time, so I think we need to get to get to a closure and as we always do here at the Models & Molecules, we have one last question, a bit of a tricky one, the contrarian one that I will would like still to ask you, which is what are your beliefs in the current market or in the current field of AI in bio that you hold but most of your peers would disagree with?

Iddo WeinerIt is a challenging question. I think maybe we can talk about wet lab assays. And also, maybe I'm a bit biased because that's in some way where I started. So, I think many people have this belief that as our modeling capabilities become better, we'll need the lab less and less because we'll be able to predict things and we won't need to run the assays anymore. I completely disagree with that. I definitely think that the way that we're conducting experiments is going to change. Probably we want more high throughput essays and less tedious characterization of specific molecules. And maybe for some assays, we'll have good enough predictors and we won't need them. But then the only thing that will happen is, OK, the frontier is going to move and now we need a new type of essay for the next type of feature that we need. We're always going to need data. The bottleneck is on the data, as we discussed. And the place where we generate data for molecular biology is in the lab. So, I think wet lab testing and wet lab essays are with us for the long run.

NicolaSo, it's not going to be less lab, it's going to be more lab in the future.

Iddo WeinerI think so.

NicolaMaybe in a different way. Maybe in a different way what's been done so far.

Iddo WeinerYes, yes.

NicolaAlright. Iddo, I thank you very much for this conversation. I really enjoyed it. I hope you enjoyed it too. And, well, I wish you a good continuation, a lot of success with your company.

Iddo WeinerThank you very much and thanks again for the invitation. I enjoyed it too.

NicolaThank you, Iddo.