In this DeciBio Q&A, Sean McClain, Founder & CEO of Absci, and Zach Jonasson, CFO & CBO, dive into the transformative role of AI in drug discovery, emphasizing Absci’s mission to tackle undruggable targets through generative AI antibody design. Beyond the lab, they discuss Absci’s partnerships and internal pipeline. Sean and Zach also envision a future of predictive models that could redefine traditional drug development, highlighting Absci’s approach to reshaping biotech by bridging deep learning with cutting-edge protein biology.

Thank you for taking the time to talk with us, Sean and Zach. To get started, I’d love to hear more about your background and how that led you to Absci?

Zach: I’m Zach Jonasson, and I’m the CFO and CBO of Absci. I completed my PhD in neuroscience at Harvard, and have been a founder three times over, in addition to building and being a managing partner at 2 venture firms, the last of which was an early investor in Absci— and that’s how I’ve come to know Sean and Absci.

Sean: I’m Sean McClain, and I’m the founder and CEO of Absci. When I founded Absci 13 years ago, it was not originally a drug discovery company. We were looking at building a technology company that would scale. The original technology was meant to allow us to predict antibody binding and efficiency, and we were able to scale this from thousands of antibodies to millions; this progress coincided with transformers taking off at Google around 2018. The idea was then to take protein-protein interaction data and leverage AI to test hypotheses that you couldn’t before. This way you could stop searching for a needle in a haystack and you could make the needle.

Because of the trial-and-error nature of drug discovery, there are specific hypotheses a biologist may have and targets they would like to go after, though the reality is that scientists often can’t go after certain targets due to having the wrong antibody / epitope. With phage display and other display methods, you can’t control how an antibody is designed. With AI, however, you can really focus-in on the design to target the epitope you want. You can target previously “untargetable” structures like ion channels or GPCRs, which may be low-hanging fruit that we haven’t previously been able to unlock. 

We started off thinking partnership was the right direction, mainly working with companies to go after novel targets or drugs. The partnerships were quite successful and were generating a lot of value, so we decided to build out our own pipeline. To help us in constructing our pipeline, we brought in Andreas Busch from Shire (now part of Takeda) as our Chief Innovation Officer. Our pipeline is currently focused around I&I and oncology, and our first asset will hit the clinic very soon. This will be the first de novo AI-discovered antibody to reach the clinic, that we’re aware of. The field itself is super early, but we’ve made so much progress in such a short amount of time by using AI to test hypotheses and solve design problems. The bigger milestone, however, is using AI to predict the biology – what targets to go after, what the efficacy will be, etc. We aren’t quite there yet as we don’t have the data to do so, though we can iterate rapidly, similar to a tech company. We utilize a “lab-in-a-loop” model to turnover wet lab hypothesis testing in ~6 weeks. If we want to solve more complex problems outside of the design realm, we need to scale our wet lab capabilities. We also need to be creative with model design, and leverage 3 key pillars: data, compute, and models; right now, in biology, the bottleneck really is the data.  

Thanks for the background on Absci, Sean. Are there any other antibody modalities that you see an opportunity in now or in the future (e.g., ADCs, bispecifics)? 

Sean: Right now we’re focused on monoclonal antibodies. We’re interested in getting into bispecifics as we think they’re exciting and important to pursue. This is a growing field that we’re planning to integrate into our AI platform, which we’ll discuss more at our R&D day later this year. As far as ADCs, there’s a lot of art that goes into design – this is not a pure science right now, which means there are a lot of ways that AI may help (e.g., with linker, payload, and DAR selection), though this is not something we’re currently focused on. However, we are excited about both bispecifics and ADCs, in addition to T-cell engagers. 

Where do you see the biologics field shifting, given advancements in AI / ML (AlphaFold, for example)? Are there any new areas of the market that Absci will play in given these recent strides, or any areas where Absci might be looking to change its approach?

Sean: I think that the problems we’re trying to solve differ from other protein folding / design problems in the antibody space. What’s difficult to manage on the antibody side is that the CDRs (complementarity-determining regions) you’re designing, which will ultimately bind to your target of interest, are unstructured. This lack of structure means that you must have a specific type of data to effectively design them, so that they may bind to your epitope of interest. We’ve come out with some manuscripts that have shown how we may design a heavy chain CDR to bind a unique epitope on HER2, for example; with AstraZeneca, we’re using that same model to design all 6 CDRs as there’s no pre-existing binder for our epitope of interest. The problem we’re solving is how to design all 6 CDRs to hit that epitope with specificity – that’s really the heart of what we’re focused on. There are a lot of other companies that are focused on similar diseases and biologics, for example, work coming out of David Baker’s lab, and Xaira Therapeutics, but the way we approach the design process is a bit different. 

Ultimately, what I think you’re going to see in the field is a lot of companies solving very niche, specific problems, as opposed to one company solving all of them. I think we are starting to observe that now, and I do think it’s important to have a narrow focus as you need to generate specific data sets for these models to ultimately achieve the desired accuracy. The direction we’re headed is antibodies, and more specifically, the de novo design of antibodies towards specific epitopes of interest; we believe that this focus will solve unmet needs in the space and really unlock novel biology in the approach. 

Absci has had a series of partnerships for quite some time, notably the one mentioned with AstraZeneca. However, recently you have been building out your internal pipeline and we wanted to know how has this shift been? Going from pharma collaborations to an internal pipeline is a significant business and operating model change that can be difficult to navigate.

Sean: To build an effective pipeline, you first have to hire people smarter than yourself. I didn’t come from the traditional drug discovery and development background, and I think part of what has made Absci successful is the world-class team that we’ve built out. I like to say that we’re “multilingual” – we’re technologists, but we’re also drug hunters; we’re well-versed in protein biology, but we also know deep learning. I think being “multilingual” in this environment is critical because drug discovery and development are already so complex. Then you add to the complexity by layering in AI, and what you really need then is to invest in talent and infrastructure that’s also very complex. You really do need a team that can be multilingual across the whole spectrum to be successful in developing your own assets. You have to wear the technology “hat”, though you also have to figure out how to apply that technology to create drugs that will ultimately be successful in the clinic. That clinical success does not depend on the technology alone, as you need to nail the correct target with the correct commercial strategy as well. Additionally, the market opportunity has to be there. We all know that if you get the target wrong, the drug will fail, regardless of how good the design is. There’s a lot of complexity that goes into development, and I think we’ve done a really good job of bringing in experts to do this. People like our Chief Innovation Officer Andreas Busch, for example, are figuring out how to solve previously unsolvable problems with our technology, some of which may present a big market opportunity. 

Another component is that partnerships are still really beneficial and key for us. We know that we can’t play in every single therapeutic area; we don’t have the resources to get TA heads in neuro, I&I, oncology, etc., so we’re focused on cytokine biology, but the platform is agnostic and can work across all of these different indications. Therefore, we see partnerships as mechanisms to diversify the assets and the therapeutic areas where we may play. Our partners bring domain expertise, and an example of this is our partnership with Almirall; Almirall brings the dermatology expertise and the targets they’re interested in, which helps us get exposure into dermatology, which is something we wouldn’t be able to do on our own. These partnerships also offer validation of the platform, in addition to upfront capital that you can use to invest in the pipeline. I see partnerships as critically important to our overall strategy, even though we are developing our own assets internally. 

Zach: I would add that, for the assets we’re developing, our mission is to build them to a human-proof point, then transact or out-license them to pharma. This strategy is partnership-oriented but is specific about the stage at which we partner. To Sean’s point, when we partner at the discovery phase, we leverage the partners' target biology expertise and the validation efforts, knowing they’re capable of bringing an asset through the clinical phase into the market. Then, once we launch our own internal programs, we can leverage that talent base that Sean mentioned, which is led by Andreas. That talent base intersects with the AI team, with the mission from then on to bring a molecule forward, and ultimately to transact it at a higher level with more favorable deal terms; that’s another example of this balanced business model. 

Something that we have noticed in these partnerships between AI-first players and traditional pharma is that the AI-first players are developing the asset to a later stage than before and the asset ownership is shifting more towards the AI-first players as well. Do you believe that’s a trend you are seeing in the market and with Absci as well?

Zach: One of the things that’s exciting about our platform is that we’re seeing great efficiency in generating assets. If we can continue to do this with minimal capital investment, then take on the “hard-coded” incremental investments required to bring it through IND or phase 1, the ROI, to Sean’s point (on bringing an asset to a Phase 1 proof of concept and then out-licensing), is drastically more economically favorable. When we consider a risk-return framework, we want to do more of that. We still want to engage in partnerships with pharma, as there are still advantages to doing so other than upfront capital (e.g., expanding our platform into new areas of target classes). The NPV, or the potential risk-adjusted returns, on developing assets, is exciting. That’s assuming, however, that we have the core competency of people like Andreas to help us make the right decisions on target and development. If we can do this efficiently and build out a nice portfolio, that generates a lot of value for our investors. 

One big question that we keep hearing from experts across the market is what is the return on investment (ROI) from using AI in drug R&D. From your position, where do you see the biggest returns? Is it in shortening timelines, generating more assets, etc.?

Sean: I can give you some specific stats based on our asset development process. We used our AI de novo model and lead optimization model to develop a potential best-in-class TL1A asset, where we’ve shown great differentiation compared to the first-generation TL1A assets. We are now hoping to initiate a phase 1 study with it really soon. We were able to do so within 24 months, whereas this process typically takes ~5.5 years (to get a drug into the clinic), including all of the preclinical and IND work. Our investment was roughly $13 - 15 million dollars, whereas traditionally you would normally spend many multiples of that. Doing so at roughly one-tenth of the overall cost is really exciting and is starting to break how the economics of traditional R&D are done. Basically, with that same investment of $50 - 100 million dollars, you can now get 5 or 10 drugs into the clinic, as opposed to 2. Additionally, if you assume that the drugs designed with AI will have a higher probability of success, you add a multiplier, so your R&D investment goes a lot farther than it has ever gone in the past. That’s just one reference as to how the economics are changing, just in observance of our own pipeline. 

What we’re also seeing today is an increase in success in the clinic. We’re really trying to focus-in on the problems that traditional biotech and pharma have not been able to solve, and how we can use AI to unlock those opportunities. If you’re able to unlock novel biology and create really differentiated assets, whether they’re first-in-class or best-in-class, you increase the chance of success in the clinic. I ultimately think this will drive better outcomes for patients. 

Those are some of the ways that I think AI will deliver. The strong ROI comes from unlocking novel biology and creating differentiated assets. On the sheer investment side, you’re getting more shots on goal with the same amount of capital. 

Zach: It’s been so exciting to see where the platform is taking us. We’re using AI to address previously un-addressable targets in a very systematic way, and opening up whole new avenues to design therapies for patients against new disease targets. 

We’re also demonstrating the ability in-house to design novel interfaces against a given epitope. This goes beyond just selecting the epitope; we’re looking at different ways to interface with that epitope, to drive potency, and to potentially capitalize on a unique MOA (mechanism of action). We’re driving towards really differentiated therapeutics which may be more beneficial to patients, which is really exciting. 

Thank you for that overview, that sounds exciting. I would now want to turn to data – it seems like a perennial to have enough high-quality data in this space. As I understand you have moved towards a model where you’re generating most of the data that you need internally. How may this change in the future?

Sean: At the end of the day, I see the data as the moat. These AI models are going to continue being open-source, and so the long-term value will remain with the data / data generation side. I think that data may protect you for a lot longer than a model can, though I do believe that you still need to be pushing the frontier on the model side as well. The way I see models evolving is similar to the semiconductor industry, where every 18 to 24 months a major player comes out with the next chip, and you have to use that newest chip to stay competitive. We do see model innovation as a core competency and something we’ll continue to drive towards. Ultimately, though, where we see the data is where we see the moat. I think the value proposition really lies in the data that we are generating for these models to be trained on; our lab-in-the-loop system may use this data for training in addition to model validation to test how accurate the models are. Using these tools, we can rapidly iterate on which models are working to increase overall accuracy. 

On that topic, what do you think the most valuable types of data are and why? Is it patient-centric data, basic biological data (e.g., genomic / proteomic data), or some other data type? 

Sean: I think it ultimately tracks back to what problem you’re trying to solve. There are drug discovery problems, which are distinct from design problems. Then there’s the whole clinical side, which we aren’t tackling, which requires very different types of data than what is required for drug discovery. On the clinical side, the questions are centered around which patients should be recruited, what ideal endpoints should be, and how trials should be designed. Each of these questions takes different types of data.

I think both types of data (i.e. patient-centric and basic biological) will be important as they’re both key to problems you’ll be trying to solve. You’ll need different types of data within each stage of the development process to get the right outcome. 

Zach: A lot of this is also about how we combine data. At Absci, even in the discovery stage, for example, we’re combining structural data with functional data. We’re asking how we can best combine data to get models trained that can reliably deliver. 

Sean: I think multimodal data, and the multimodal models, will definitely play a big role in the future as well. 

That makes sense. Looking forward now, what are you most excited about in the next 5 years as the space continues to evolve? 

Sean: I’m really excited that, at the end of next year, we’re going to have our own Phase 1 interim readout for what we believe is the first de novo AI-discovered antibody in the clinic. I’m also excited to see how these models will evolve to start predicting outcomes in the clinic (for example, models that can take a target and a drug and predict the overall efficiency in the clinical setting). I think there’s a lot of exciting work to be done around predictive modeling. I’m also curious to see how we can create models that will translate to the clinic effectively. Ultimately, I think we’re going to live in a world where we don’t have to run animal studies anymore, we’ll just use an AI model. I truly believe it’s just a question of when we’ll get there. 

Thank you both for the great conversation. I look forward to staying in touch.