Building AI Foundation Models for Molecular Design

Our models have been downloaded more than a million times from more than a hundred thousand unique sources. We know our team is seeing every large pharmaceutical company using our models hundreds or thousands of biotech companies. Back in 2020-2021 was when we started to see a lot of progress on this very fundamental task of predicting the structure of proteins. The big moment there was with AlphaFold, we realized that AI was there to stay and could really have dramatic impact on a fundamental problem of biology. We have seen in the past few years the development of autonomous robotics labs, a lot of automation and experimental validation as well. And so if you think about combining those autonomous labs with platforms and tools, if you want scientists to directly use your models and integrate them into your workflows, you need to really build products that do that. AI is beginning to change how biology is understood, modeled and engineered, and the implications for drug discovery are profound. For decades, computational biology improved incrementally, constrained by limited data, slow experiments, and tools that struggle to generalize. Recent breakthroughs, starting with AlphaFold, marked a turning point. AI models began solving core biological problems at a level that surprised even experts. But structure prediction was only the beginning. As these models evolved to capture bio-molecular interactions, binding strength and molecular design, a new question emerges. How did these capabilities move from research papers into everyday scientific workflows? In this episode, Boltz co-founders Jeremy Woolwind and Gabriela Corso joined E-16Z's Jorge Condé in Zactoric to discuss the launch of Boltz, a public benefit company building AI infrastructure for molecular biology. They displayed how open-source models helped accelerate adoption across academia and industry, while Boltz is being built as an infrastructure company rather than a therapeutic company. And what changes when frontier molecular AI is productized? The conversation explores how AI could reduce early drug discovery bottlenecks, improve molecular design, and enable faster iteration between computation in the lab. We hope you enjoy. So welcome to the A-16Z podcast. We are thrilled today to be hosting Jeremy Woolwind and Gabriela Corso of Boltz. Why don't we get started with quick introductions, but for the folks that don't know, I think it's safe to assume, Gabriela and Jeremy, that the vast majority of our listeners aren't in the business of discovering drugs or aren't in the business of regularly trying to discover novel biology or chemistry. So why don't we start with the brief introduction of what Boltz is and what you're working on? So Boltz is a public benefit company, and our mission is to advance AI for biology and make it accessible to every scientist building towards a healthier future. Our models basically try to model how proteins and other molecules interact with one another, and through that help scientists develop new drugs or other biological tools. All right, so let's take that a little bit up a level, and in layman's terms, why is that a big deal? So why is it a big deal to be able to use AI for molecular modeling or any of the number of things you just numerate? And maybe it's worth even like defining what molecular modeling is, which is that it's the science of trying to predict how molecules behave, so that can be structurally, and it can also be from the perspective of their properties. Like does this molecule bind to this particular molecule with what strength with what physical chemical properties, and if we can predict those things, then there's a lot of problems that you can dramatically accelerate the progress for whether that's drug discovery or even areas outside just like agriculture by manufacturing and all sorts of applications that touch on molecules in one shape or the other. Okay, but scientists have been using computers to design molecules for a very, very long time. What's new this time? What's different this time? And probably one way to respond to the questions to help it put it in context for the non-scientist. When it comes to the use of AI, what is sort of the chat GPT moment for the use of AI in the life sciences? I think computational tools have been used for a long time. I think the power of machine learning, I think as we've seen in other areas as well, has been the ability to really sort of transcend the performance of classical tools, and in many cases also do that considerably quicker. And in particular back in 2020, 2021 was when we started to see a lot of progress on this very fundamental task of predicting the structure of proteins, proteins being maybe the most important modality of molecule out there largely orchestrates a lot of how about functions. And that was the big moment that was awful, which largely solved that problem for some large categories of proteins. And I think the performance and the quality of the prediction that surprised everyone. I think it was not necessarily just a constant progress, but really like an enormous inflection point. And I think that's sort of the first time where people like myself, I think Gabriela as well, but I think many others in the field sort of realized that AI was there to stay and could really have dramatic impact on a fundamental problem of biology in a way that also I think opened a lot of question as to what else is possible. So you guys have released a series of models during your time at MIT, starting from both one then bolts to and then more recently bolts Jen, can you please walk us through sort of what each model accomplishes and why was important and impactful to the field. So about a year ago, we announced bolts one, which at the time was the first fully open source models to approach alpha volt 3 level of accuracy. Then we follow up with bolts to bolts to went beyond alpha volt 3 and bolts one that were just predicting the structures of this biomolecular interactions, but also predicting the strength binding affinity with which is more molecules binds to protein. And so we use that to also showcase first applications of these models to design new molecules that could bind particular proteins and finally bolts Jen that we released a couple of months ago is a model that is able to design arbitrary proteins to bind any biomolecular target and we have validated it when experiments of validation across very different modalities across different targets with many collaborators across the academic and industry. Gabriella one thing that struck us about bolts Jen was just how generalizable that model is you guys have wet lab validation across several modalities which so far has been rare for these models and even in the platform you're announcing you're making compatible for both small molecules and biologics. Can you talk a little bit about the movement towards more generalizable models what unlocked that and does that mean that we're moving away from more specialized models in the future. One of the trends that we've seen in many other applications of machine learning but we've also seen by writing with alpha volt 3 in our field is that the more data about biomolecular interactions and other things you put into these models the better understanding the develop and then the better that they generalize. And so this is also something that we have seen with both Jen that we have trained to generate binders across any biomolecular target across proteins of all sorts of different modalities. And obtaining performances at the state of the art across each of the different modalities and then we were very lucky to be able to collaborate with many collaborators across the industry to validate each of these different modalities and different targets. And it should feel like largely intuitive right it's like if these models like understand anything about physics then what it understands should be quite transferable across modalities and I think we're I think we're going to continue to see this like more. Not to overuse that term too much like these like foundation models you know that can do a lot of things and that's not to say that they won't be like specialized models for some things like for example small molecule properties I think is something that feels a little bit more specialized still. But we've definitely seen that in other areas of mission learning and we're seeing it here as well. So just like I said at the outset that the vast majority of these listeners probably aren't in the business of regularly discovering drugs. I would bet that the vast majority of listeners to this podcast probably have heard of alpha fold. So that's like the one tool in sort of the scientist toolkit that is probably jumped into the ether into the zeitgeist for the average person. Can you talk a little bit about why you chose to work on bolts given that alpha fold was such a big advance so I think as Jeremy mentioned both him and I were working on machine learning kind of before alpha fold was published and alpha fold was one of the reason that convinced us to actually come into the field. And so what alpha fold does is predict the structure of individual proteins now proteins are critical for a lot of things in biology by the way the proteins operate is by interacting with other molecules are the proteins. And so when it comes to designing a new drugs understanding a biological system where you really often care about is predicting the structuring the strength of interaction. And so both Jeremy and I started our PhD right after alpha fold to was published and what we worked on in our PhDs trying to model this by molecular interactions. We did a lot of progress in the field and that has led to bolts being created in 2024. And one thing that really struck us when we started learning about bolts what was on the academic side is just how many collaborators you guys were able to assemble around the bolts project. We all know that convincing talented grad students to work on your project instead of their own is an incredibly difficult task. And so how are you able to sort of set the vision that so many other exceptional people were excited to to follow. We really started working on the bolts model soon after alpha fold free was published alpha fold free was particularly important moment for the field on the one end on the positive side because it's set a new standard for in terms of the performance of you know our ability to model by molecular interactions. So on the other end it was really the first time that really the state of the art was no longer accessible by most people in the industry. So Jeremy and Sarah started working together around this time on bolts and you know bolts one we came together to create bolts one which was the first fully accessible open source model to approach alpha fold free level of accuracy. So I created a lot of excitement in the field a lot of people started using it and also at MIT and beyond we started you know working with a lot of people that you know join the project. Or you know directly as well as you know indirect for the community and started contributing to the models and to the tools. Because of a few things you know that we did that I think we're critical and sort of building out the momentum around the models and around the community. You know I think one of those was you know the open source aspect I'm sure we'll talk about that a bunch. The second one was I think you know our effort towards like making engineering quality and sort of like ease of use really central. I think you know it's it's one thing to put code out there it's another to like do it in a way that is going to be really easy for people to like hop on and do interesting work and I think we put a lot of thoughts in that. And I think you know the third one was sort of like community building and we created the Slack channel we like you know gave multiple like you know keynotes and talks I think like that sort of contributed to you know the community really building up up to a point where like you know to be honest like even this like Slack community of like few thousand people is like largely self sustained now where like I you know we don't necessarily answer a lot of questions on there anymore it's like very much like people. You know answering each other's question I think like there's there's really kind of a really powerful momentum effect that gets created from that. And so yeah I think that's like really comforted us in the idea that you know like open sourcing and like investing in a community gives a lot of we work back. You guys have been very intentional from the very beginning to do science in the open you've continued that commitment even now that you've created a company based off of that work and this is going to be a continued commitment for you guys wise open source so important. One of the things that we keep like telling each other I think often I think Gabri came up with it the first time that statement you know it was like like open source is like we think is not only optimal from like a societal standpoint but also really powerful from a commercial standpoint and so we've always seen it as like a powerful motion for us to leverage. And so you know on the one end yeah we come from an academic world I think you know we have received a lot from the academic world and I think it's important to us to continue to do a lot of open science we believe in open science. And so I think that we're very much committed to continuing to you know make the most advanced open source models in the world and then on the on the flip side like I think what it really allows you to do is put your stuff to the test in a way that you know a commercial you know close product cannot when things are open people can dissect it in and out and you you know you really allow. The model to shine or to or and to show its you know limitations and failures and so it's really powerful way to learn and to iterate and I think the last thing that we found really great about is that. You know by the time people are interested in sort of talking to us and working with us you know that they've had the chance to use the model to have the chance to see what works for them what doesn't work for them and I think it's almost like automating like the pilot phase in a way you know and and so for all of these different reasons it's felt quite central to our thesis and you know we'll continue to be I love this idea of automating the pilot phase. If you had to estimate how many people have touched a bolts model have played with the bolts model and the reason why I asked that question is you talk about community you talk about you know automating the pilot phase one of the things that struck us and get into know you all is that when you when you announce a new model it gets it generates a lot of enthusiasm when you present. You feel auditoriums now I'm not going to call you guys rock stars but that's a pretty remarkable thing to have that kind of support from the from the community from scientific community. When you sort of look at that do you guys have a rough estimate of how many folks are using bolts models and of course how it's being used and what's been surprising around the usage and how that's the community has taken up the work you have put out there. Yeah we keep some look at statistics and it's always you know very. Humbling to see kind of all the support and our models have been downloaded more than a million times from I think more than a hundred thousand unique. So our sources we know teams in every large from studio company using our models hundreds or thousands of biotech companies. And in general this is a momentum that grows over time and we've seen it starting from both one two bolts to two bolts Shen it's a community that grows and so over time as we. Without a new model the excitement kind of grows with it and obviously all the user base also kind of shifts from kind of the older model to the new model and also it also becomes a bit of a standard in the community that people compare to the people build upon and that was it helps with this flywheel. You know given that that momentum that you just described what is sort of your view on I'll call it probably unfairly I'll call it what it's I'll call it an arms race the sort of an arms race that you know bolt puts out a model and someone else puts out a model and then bolts puts out another model and so on and so forth. How do you view that sort of competitive dynamic is it a positive is it something that will ask them to over time let's talk a little bit about sort of the arms race when it comes to AI models. So I think on one side it's amazing from the scientific side you know if you see the progress of the past few years it's amazing and especially when the is you know releases of models are open in the sense of you know either open source or at least you know. Open kind of description of kind of the methods this really help advance you know science and and so on the one end it's amazing on the other hand you know from a company stand point you have into put things open source always being kind of the benchmarks that everybody uses to. Compared admitted against obviously puts has a lot of you know motivation on you know always shipping kind of the best models and you know keeping a very fast pace and so. You've had an incredible amount of success built as building up the community and driving that enthusiasm and really bringing. You know cutting edge models to bear on some very important problems one of the things you said at the outset was you created a company bolts pbc. Why start a company what is that company going to do that is different than what you've done to date with the bolts project and walk us through some of the announcements that you're making today as part of that launch. So one of the things that Germany and Saro soon realize after starting the bolts project is that on one end for the research side we were going to need a lot more resources that we could have possibly gotten on the academic communities to be able to stay at the frontier. So on the one end we need the investment on the other hand we also realize it's not enough to just put a model on GitHub just open source in models. It doesn't really get all the impact that we want to have in the field and in particular if you want scientists directly use your models and integrate them into your workflows you need to really build products that do that. And so the big announcement we're making today is the release of bolts lab which are platform that integrates protein and small molecule design agents into scientists workflows in a way that it's very easy and intuitive for them to use. It's kind of like switching the audience a bit in some ways you know I think we've gone from a very computational audience to a much broader audience with the product. And I think that that was like one of the critical like sort of problems that we wanted to solve. And it's also about like you know understanding like where is it that we can also work with our commercial partners to you know come up with new solutions and solve problems that you know maybe the base models are not necessarily well suited for. So yeah this it's sort of like expanding the scope really of the of the audience of the tools. So when we're thinking about building this products especially for all these enterprises that use them to create new drugs is really critical to build a frontier technology company that can support them that can develop them with a level of engineering that is required for such enterprise context. And so that's part of the goal of why we need to be a company. Yeah and to pick it back on on that point like you know we've seen it obviously a lot in the other lamb space where the amount of infrastructure that you need to run the model at scale is is tremendous. We're not necessarily in a space where the models are as heavy but the amount of times that you need to run the models to be able to like search the molecular space. It can also be extremely large and so you know part of our company is actually also very much like infrastructure like how do we put in place enough you know raw computational power to be able to like executes you know these molecular discovery workflows as quickly as possible. And yeah that that takes a ton of engineering for us and we've built really the team also around around that problem. I love the emphasis that you guys put on product in our experience as investors betting on the best teams and best models often isn't enough because models when they're predominantly trained on public data as you guys know can get rapidly commoditized. So for us it's really important that companies have to build a great product and so maybe let's double click on that just a little bit. You've mentioned this great emphasis that you guys put on product but obviously building a great product here isn't trivial because there's an infrastructure component there's a security component. There's an intuitive interface component there's a feedback loop with the with the wet lab. How have you guys been thinking about this product for the beginning and how has it taken shape since. Yeah so we're putting all this different components so we start from the research so as you were saying you know we need to think about not just you know the role models but how do we integrate data experimental data that is generating the lab back into these models. Then you know it comes an infrastructure problem how do we scale these models to run this workflow these agents that can run for hundreds or thousands of hours on GPU now we parallelize it so the results arrive in hours not in weeks. And then there is a product component you know how do we surface these tools both to computational scientists but also to non-computational scientists. Canis and biologists are running kind of this drug discovery programs that are not programmers you need to be able to work with an intuitive interface that really integrates with their existing workflows. And the last one you mentioned you know security I think is also something that we've been taking super seriously and you know sort of from the onset of the company like have been working on our security certifications quickly I think not only to like you know reassure like customers but also like for our own you know for for us to like sleep at night better like I think we we definitely. Understand the sense you know the sensitivity of the data and and the importance to like do right by the people that trust us so yeah definitely a major major concern of ours yeah there's something about your sort of your driving mission that I find incredibly inspiring and I'm going to obviously probably butcher the actual language you guys use but this idea of making frontier biomolecular AI available to all scientists in labs large and small I think is this really remarkable mission to have and when you talk about mission I thought maybe we take a second to explain for you to explain why you incorporated bolts as a public benefit corporation because that's not something you typically see. So why why the PVC in this case as Germany mentioned earlier we started as academics and as academics you know the both project at this goal of you know when seeing kind of this models getting closer and closer we really believe that needed to exist. In organization that would stay at the frontier continue to push the frontier my making accessible to every scientist and so that was part of the founding mission of bolts and we wanted to put it in stone as you know like a mission of the company and so part of you know our public benefit corporation mission is indeed kind of what you mentioned. Now I mean building a company it's it's clear that you you know you're going to have commercial objectives and those are like critical to you know to the prosperity of division and I think that. But you know like this word like putting in stone I think that that Gabri uses I think is is important it's like it's very much like us making that commitment that you know regardless of all the other objectives that we have that that's still going to remain like you know critical to the mission I think it's it's it's it's in the name it's like a constant reminder and yeah I think it's it's powerful and it's on right so let me ask the the evil venture capitalist question. Which is how do you balance this mission that you have of of fostering open science. With starting a company with making a product and selling it so how do you balance the the broader mission with the business model for what you hope you know for the business model for where you hope. So I actually think that the mission and the business models are actually quite a line our mission is to put this. Models and tools in the hands of every scientist and even if you charge every scientist a little and you know just something around the infrastructure of how you run these models if these models are widely adopted and they're used at the very large scale that we believe this model will be used. This could be a great business and so putting the scale at which kind of we believe this models are going to be run and we already seeing kind of this workflows to design your model molecules to design new proteins. Run for hundreds or thousands of hours in GPU and you put that together with you know putting this incredible tools in the hands of every scientist and you can see that you know even just as an infrastructure company of his own this could be a great business. You know on top of that I think we there's all sorts of things that we can do that you know tangential you know to the to the commercial objective like for example like providing grants to academics you know to to use the model to push like adoption and things like that. I think there's a lot of tools I think that you know while you're building a product you can also like you know give back also on the on the other end and try to you know funnel some of the you know some of the financials to to to folks in the in the academic community that you know on the other hand I think provided a ton of ideas and improvements and things like that. So in our world of biotech when someone has a very powerful technology for drug discovery the first instinct is to build your own pipeline of differentiated drugs right and that is in part due to software having been hard as a as a primary business model in our world. You guys have very clearly stated that you will not be a therapeutic company and instead you'll be an infrastructure company what drove the decision and what as convinced you that this time will be different for software and bio what does AI change in that equation. Number one is the first reason is the mission you know I think there's a business model real lines with our mission of putting the stores in the hands of everyone. And I believe you know it's a bet that we've made on the stores really on the one end you know becoming in a very hard to develop and to maintain but also providing a lot of value and so if we are in a world where these models like in the past few years are very hard to develop. So very few therapeutic company can make them but at the same time they do become critical for therapeutic development I see I believe we're going to be in the world where every. We're going to be able to develop these tools internally and so we'll want to turn to companies like bolts to buy these tools as you know inference and you know to be able to compete with you know the very few companies that can develop the tools internally. Yeah and I think to to your point Zach about like also why it's different now I think there's always been like tool you know utility of of computational tools and over several decades I think even in in in farm iron biotech but I think that I think it's one of the first time that you know through molecular design like rational molecular design that you can like you have such a like tight link between like the value of the assets. And what is actually being produced by the tools you know and so I think people are going to invest a lot in that sort of three clinical development and I think if you believe that competition is going to be at the core of that investment. And then it can even become like a differentiator you know commercial advantage and I think there's also you know I think maybe for the first time like certain categories of problems that are going to become easier maybe even more than easier like maybe only solvable via compute. In a way that you know in the past I think you could always go back to the lab to do something and I think we're right around this time now we're like you know we're reaching sort of experimental accuracy on some things and at some point like if you can do that across like many assets then you can sort of run you know measure many things in parallel. At the same you know through one big computational design and that's not something that's really tractable experimentally so I think we're going to go from like you know this sort of like trying to match experimental to like you know sort of compute superiority and it's not going to be on everything by any means you know I'm not I don't mean to say that like we're not going to have to go to the lab anymore I think we definitely are going to continue to go live especially for like more like functional measurements and things like that. But I think that there are some categories of problems where it's going to computation is going to become the obvious route now don't think that that's quite in the case before and it's largely a function of like just a sheer performance I think of the tools. Yeah one of the striking things you have in sort of your your vision that you articulate is this idea that you're hoping for a future where scientists can go from a therapeutic hypothesis to designing a human ready molecule without leaving their computer screen. That's a pretty audacious vision just to piggyback on what you were just saying Jeremy where do you think we are along that journey and where do you think you know we have a long way to go. Yeah first to clarify when we say that without living the screen we actually don't mean without testing things experimentally. But you know for example we have seen in the past few years you know the development of autonomous robotics labs a lot of automation and experimental validation as well and so if you think about combining kind of those autonomous lab with platforms and tools. Like pulses that can you know decide what experiments to run integrate back the results of those experiments into kind of new pipelines to optimize. Then you can allow basically every scientist you know without having to hire a number of computational scientists hire you know a rent a wet love and so on. And with a much smaller budget be able to go from an idea to potentially in molecules to put into animals or humans very much faster than before. Yeah I love this idea of a business that provides you know the core infrastructure to bring AI into the the drug discovery development process. In a way to help create the the infrastructure for what will be the future virtual lab which is one of the things that I think the industry has has long been hoping for. And to do it in a way via your mission and your model to do in a way where you are a rising tide lifts all boats business which is again also relatively rare thing in our field but it's an inspiring mission to have. One of the things that I really admire in terms of your your vision for what this company will do is to be broadly accessible to all scientists. And Jeremy I think you mentioned earlier this idea of you're sort of expanding your audience. Can you talk a little bit about you know if we sort of you know forward the movie a bit and bolts is widely available. To scientists and companies you know labs large and small what will the day you know sort of the average day of a scientist look like in a world where bolts is part of their of their common infrastructure. I think in many ways it's going to look very similar in terms of like what the hypothesis that there's testing are maybe the way that they're going about it is different. You know you are still going to like try to figure out okay what's an interesting target I can go after what's and you know what's the type of modality that I think is relevant here. So I want to design a small molecule that want to design an antibody and you're going to you know I think generate hypotheses generate you know potential candidates with you know with these tools. And along the way you're going to try to also cover as many you know properties of interest that you have. So instead of like you know sort of treating things in a more like step by step where you say okay I'm going to find something that binds and then I'll optimize it a little bit so it's more soluble or it's more specific and less toxic. Maybe you're going to sort of have those things defines more at the onset of you know of the design process and then inevitably you're going to end up in the lab you're going to you know have to go test these things. But the hope is that you know the day to day looks like hypothesis generation hypothesis testing in the lab and then iterating on that you know as rapidly as possible. And I think hopefully like being able to do that maybe on many projects you know at the same time if you can like you know when you when you allow things to go quicker. I think generally you know we've seen that even with like GPUs and whatnot like it's not that people do less things it's you know because it's quicker they just end up doing more things in parallel. Because their time hasn't changed but the amount of things that they can do within within that time has changed and so yeah if I have like one hope is that. You know we start to see you know biologists and you know drug makers I think experiment with many more hypothesis at the same time and. Can I explore many different disease areas at the same time if you talk about drug discovery like the general view is drug discovery it takes a long time. It takes a lot of money and it is a fundamentally risky prospect right the probability that a drug will make it all the way through into become an approved medicine the odds are stunningly low that you will be successful. And let's just for a second assume that there are a series of bottlenecks that make drug discovery so challenging so time consuming so risky and so expensive. One is you know that the you know the molecule that you design in a computer will will perform as predicted in an experiment. Another bottleneck is that it will once you have this predictability problem between in vivo in vitro models and in vivo models i.e. what happens in a petri dish or what happens in a mouse or what happens in a monkey doesn't ultimately translate to what actually happens in a human when you do the human clinical trial. Where in your view do you see AI having an impact in helping you know remove for at least improve those bottlenecks. Is it going to help with the predict the predictability that a molecule behaves in a monkey will it help with the predictability that a molecule behaves in a human. Help us to help us develop a little bit of intuition around how I will ultimately help in the in the ultimate endeavor of making a medicine available to patients. I think it will happen across but I think the way that at least we're saying it a bolt is that we're going from the bottom up so I think we are already starting to be able to model well. Things on the molecular scale able to understand how different like our interactions is now we're moving. From molecule from just pairs of molecules to pathways trying to understand how molecules can have a certain effect on pathways and maybe of target effects and other things like that. We're going to move to cells and you know beyond that and so I think you know we're going to go bottom up and try to kind of reduce a lot of those bottlenecks by basically bringing and developing kind of better molecules from the get go. I think one of the big problems of you know why clinical trials. I show expensive obviously is that you know many molecules failed clinical trials or even if they succeed clinical trials and not very effective. And so they are potentially not worth a lot of money. I think if we're able to you know bring you know much better candidates at the onset of clinical trials we're going to be able to on the one end be able to tackle a lot more diseases that right now. Unfortunately are not feasible whether those are rare diseases or those are for example tropical diseases. And on the other hand we're going to be able probably to do things like personalized medicines by being able to really personalize the type of therapeutics to particular. And so I think the amount of things that we'll be able to do is it's very past as long as we have this understanding of biology at the right level of the problem. And I think you know we've been talking a lot about molecular design I think through a lot of this discussion. But of course like you know a lot of drug discovery is also just like sheer understanding of of disease and understanding of like what targets to go after and like that's that's always going to be you know probably the biggest reason that things fail. And so you know I think the hope is that you know while we're very much focused you know on this problem of like designing molecules like really at the fundamental problems that we're solving are like more. Our more principle is like you know does this interact with this and with with those kind of tools like you can also try to just understand biology better right like every was talking about modeling beyond just individual molecules and modeling like maybe more complex systems and. If you can do that I think you can also like help people like understand biology better makes better decisions as to like you know what to pursue. And and so I think it's going to come from both angles both our ability to design better but also our ability to help understand biology better derive more insights. I think that's going to be equally equally important as you move beyond the stage of molecular design as you get into drug optimization as you get into predicting toxicity in humans. Those tasks require more and more in vivo data data sets that are hard to generate at scale. What's your strategy for for getting to those types of data sets whether it's through you know in house data generation or data deals that are getting more and more invoked these days or simply tapping into some open source project. What's your tickling on evolving on that front I think it's going to be a combination of them and in many cases we're also going to leverage the data that our customers have and you know for example. Today we also announce a partnership with with Pfizer where we're going to use. There are data to fine tune our models to serve you know Pfizer Pfizer scientists best and so I think that's also going to be part of our strategy to give every company kind of the best tools possible. But for sure you know we're going to already investing and we're going to invest a lot on on data going from every scale. I think one of the things that's fascinating about this company is this idea that you started as a project as an open source project and you are now very purposefully building a product. And one of the sort of the key tradeoffs at least in our view that when you move from a from a project into something that is a formal product is it has to have sort of three very very critical components one is it needs to be bullet proof. It needs to be weather proof and it needs to be user proof and what I mean by that is bullet proof like it needs to work right because obviously people aren't going to use a product that doesn't work at the models need to do with the or purported to be able to do. Weather proof in that it needs to be secure like I need to know I need to believe if I'm a if I'm a scientist that a Pfizer or in my own lab that mine my data is secure that you know it's going to benefit me to use your models. It's not going to benefit other users to use my my proprietary data and it ultimately needs to be user proof in the sense that if you want this to be widely accessible. I think you need to be able to have interfaces that are intuitive and that are useful to a broad range of users. And so this is one of the things we've been fascinated to see the one of the things we've been fascinated to see is the evolution of your thinking from taking that. Project that has been so incredibly successful and turning it into a product that is bullet proof weather proof and user proof and so we're really excited to see where this goes. Gabriela why don't you walk us through the various things that you are announcing as part of of today's news. So first of all we're announcing bolts PBC our company with the the seat round led by and regional with Zeta and amplify. Then we're also launching bolts lab our platform and in it announcing our first new agents for both small molecule and protein design and so any scientist can now go and sign up to to both lab and start using this to design. Newsmo molecules new protein finally we're also announcing a multi year partnerships with Pfizer to develop new set of the art models and giving their scientists access to this models for bolts lab for them to develop new medicines. Amazing well congratulations on on the launch and and all that you all have accomplished as a company in the short time that the that bolts has been in existence. If we want access to product if we want to go into our lab tomorrow and and and tell our our PI or tell our boss hey we got to get bolts how do we do that just visit our website bolts up bio and you can send up to bolts lab. Well pleasure to have you both on the A6 and Z podcast and congratulations again on your launch. Thank you and thanks for your support. Thanks for listening to this episode of raising health if you like this episode be sure to like comment subscribe leave us a rating or review and share it with your friends and family. Follow us on X at a 16 Z and subscribe to our sub stack at a 16 Z dot sub stack dot com. Thanks again for listening and I'll see you in the next episode. As a reminder the content here is for informational purposes only should not be taken as legal business tax or investment advice or be used to evaluate any investment or security and is not directed at any investors or potential investors in any a 16 Z fund. Please note that a 16 Z and its affiliates may also maintain investments in the company's discussed in this podcast for more details including a link to our investments. Visit a 16 Z dot com forward slash disclosures.