Podcast Transcript

David Brühlmann [00:00:23]:
We’re back with David Younger, who’s the founding CEO of A-Alpha Bio. In part one, we have been tracing the data problem behind AI-enabled protein engineering from more than one angle. The science of measuring interactions at scale and the harder question of how an industry that guards its data gets anyone to share it. There is much more ground to cover, so let’s dive back in.

Let’s talk about business. I’m curious, David, why do you believe that technology is not enough on its own to solve the problem? And what’s finally made a new business model the necessary piece?

David Younger [00:02:17]:
Even with a technology like AlphaSeq, where we’re able to measure a million affinities at a time, we believe that the quantity of data that’s needed to solve these problems is still very difficult to reach. I’m not going to say out of reach. There’s always some amount of scale that would get a single company to that level, but it’s pretty far off. Again, from a structural perspective, we think that it’s going to take on the order of hundreds of thousands or millions of structures. That is a daunting task. In a zero-shot optimization, local affinity landscape perspective, we think it’s going to take probably tens to hundreds of thousands of affinity landscapes. Again, a hugely ambitious task.

And then as we start to think about even more complex problems, right now just kind of simple binders, but we want to engineer things that are potent and pH sensitive or dual-specific to multiple targets. As we start to layer on these additional complexities, our models will have to be more performant and we’ll need even more training data behind them. Again, even at the scale of AlphaSeq and being able to measure affinities in the way we do, we still think that these are really challenging numbers to get to if a single company is kind of fully footing the bill. And as a result, we have thought very hard about business model innovation and how we can incentivize essentially the pooling of data, pre-competitive data, such that we can really advance this industry in a way that is currently being bottlenecked. And our answer to this challenge is a data consortium. So the general mechanics of this consortium are that model builders, pharmas, AI labs can join this consortium.

And they essentially pay quarterly for the opportunity to design an AlphaSeq experiment. So a million affinity measurements every quarter. And instead of them just getting their data back, we pool all of the data from the consortium and we send it to every consortium member. A-Alpha also contributes our own experimental designs that we perform. We add that to the mix and we send that out to the consortium members as well. So the ability for this type of a model to scale is really, really exciting. So we recently launched this consortium with five founding members. With five founding members, each one of those members will receive about 28 million affinity measurements every year, which is a pretty vast dataset when we think the total number of affinity data for antibodies in the public domain is about half a million. So we’re talking about large, large quantities of data here. I should also say the total quantity of affinity data in the public domain is about half a million, and I think upwards of 75% of that is AlphaSeq data that we have made available in our publications. So for the consortium, at 5 members, each one of those members will receive about 28 million affinity data points per year. By the time the consortium grows to 20 members, that will grow to 100 million data points per year.

As the consortium grows, we’re able to provide this really dramatic increase in scale of data to the members of this consortium. And we hope that the amount of value that these members are getting is far, far, far, far larger than what they’re having to pay in. Hopefully it’ll just be a no-brainer in order to get the type of training data that companies need to solve these problems. We hope that this is just a no-brainer solution for companies moving forward.

David Brühlmann [00:05:35]:
If you can multiply, shall I say, the effect, it’s definitely the way to go, especially if you have this large number of data you need to generate in order to have meaningful models. But this begs another question, because when it comes to sharing data in academia, usually it’s not such a problem. But I do know from my experience, when it’s especially these sensitive molecule projects, you know, a new portfolio project, then companies are very cautious. How do you go about this challenge to make sure that only what’s necessary is shared and that companies can keep, I’d say, the most critical aspects of their molecules?

David Younger [00:06:13]:
So one of the really exciting and unique things about this consortium model is that it’s all prospective data, right? We are not asking companies to go back historically into their treasure trove of proprietary data to find what they can share. This is all forward-looking. So this is all data where members of the consortium are going into the design of these experiments with eyes fully wide open. These are data that are going to be shared with other consortium members. For solving generalizable problems in computational protein design, whether that’s de novo design, zero-shot optimization, other challenging tasks, there is an enormous amount of pre-competitive data that is needed.

Companies don’t need to be looking at their proprietary targets. They don’t need to be looking at their proprietary molecules. They need an enormous diversity of binders and an enormous diversity of targets to train generalizable models that they can then apply those models to their proprietary targets later on. So we are talking here exclusively about pre-competitive training data to achieve model generalizability. We will also work with companies, and members of the consortium are actively working with us. In addition to their work in the consortium, we are also generating private data that relates to their proprietary programs.

So again, I think if we come back to kind of the 2 sides of the data bottleneck, there’s the side of the data bottleneck that’s upstream, right? We don’t have nearly enough training data for ML models. And then the other side is we don’t have the ability to generate data in high enough throughput to validate large numbers of designs that are coming out of models. That output largely is proprietary. A company is going to be asking their models to design binders against their proprietary targets, and they want to test tens of thousands of those potential candidates. That is not a type of experiment that a company would do within the consortium. That’s something that they would work with us on separately. It would be entirely proprietary. Those data would never be shared with anyone. They would never go into A-Alpha models. But for this much larger bolus of data that’s needed to train generalizable models, what we have found is that there’s very little pushback, that companies can produce pre-competitive data that they are comfortable sharing with others.

David Brühlmann [00:08:32]:
And what is your ideal customer for this consortium?

David Younger [00:08:36]:
It’s a mix. I mean, our ideal customer is anyone who is training generalizable AI protein design models. And in the launch of our consortium, we announced some of our founding members. Our goal with the launch with those founding members was to really represent the different types of companies that we think will be able to benefit. So for example, GSK joined as a big pharma actively working on training models internally, a perfect customer profile for us. And we’re very excited to be working with other big pharmas moving forward. Cradle and Boltz joined the consortium. So these are model builders who are training models that they’re then deploying in various ways to support pharmas and biotechs across the industry.

We are very excited about working with model builders who are deploying their models in that type of way. And then Dyno Therapeutics also joined the consortium, and they are training models internally and then leveraging those models through service offerings to various companies. So these are essentially 3 different profiles of companies, all of which we think can benefit enormously from participation in the consortium.

David Brühlmann [00:09:49]:
Let’s look forward. If the consortium works, if you can attract more customers, how will antibody discovery or development look like in 2030?

David Younger [00:10:00]:
With the current consortium, we are focused on a particular type of antibody. We’re focused on VHH antibodies, and we’re focused on data that’s generated through AlphaSeq. So that is local affinity landscapes and structural data. We recognize that this is still a fairly narrow focus. And so over time, we plan to both increase the number of consortia, right? So right now it’s VHH-focused. We’re going to, in the future, launch an scFv, so a dual-chain consortium, and continue to extend out to different scaffolds and different problem statements.

We also recognize that binding data is not everything here. We need to also, over time, incorporate other types of data into this consortium model. So when we look towards 2030, for example, I hope that we’re in a world where companies no longer feel bottlenecked by their data, where people really feel like there is a path towards generating data in abundance such that we can see models improve and we can start to really identify where are the kind of subtle gaps in those model performances to generate data that’s specific to solving, to filling those cracks.

I think right now we’re really at a point where there’s just an enormous need to raise that tide and lift all boats, just with an abundance and diversity of data. In 2030, I think that we will be in a place where companies have access to sufficient data to solve a lot of these problems. I still don’t think that we will have solved the data bottleneck. I think that there’s still going to be a need for more data. I also think that we’re going to get more clever about how we generate data at scale and generate the right datasets to solve these very important problems related to specific therapeutic goals.

David Brühlmann [00:11:43]:
As we’re wrapping up, David, what other questions should I have asked?

David Younger [00:11:48]:
I think one thing that we’ve talked a bit about to others is contrasting with federated learning models, how something like a consortium model contrasts with federated learning. That’s a topic we could chat about if that’s interesting.

David Brühlmann [00:12:01]:
Sure, go ahead.

David Younger [00:12:01]:
We’ve been asked a number of times how the consortium differs from federated learning models. So federated learning essentially being where participating companies will contribute data to a third party. That third party will train a model on all of those data and then distribute just that model back to all of the participants. I think that there are a few things here. I think one is federated learning is sort of definitionally looking at historical data. And so it’s not data that was generated with the intent of training generalizable machine learning models.

It’s not data that was trained kind of knowing that it was going to be pooled and distributed to others. So there are often concerns about confidentiality of that data and different members of that federated model getting different incremental value from other people’s data. So I think that that’s one challenge. I think another is that a federated learning model doesn’t naturally incentivize collaboration. I think one of the other things that I’m really, really excited about with the consortium is how it incentivizes collaboration.

So you can imagine that if you have this group of companies who are all individually designing datasets on a regular basis, and then getting data from every one of those consortium members, if they learn something new about how to design an experiment, and they say, okay, I’ve found something really exciting here. I see that if I design an experiment this way, it’s incredibly powerful for the models that I train. The incentive of that company is to share that insight with the community because they want all of the other consortium members to generate data in that way too, because then they’ll get far more of it. And so again, I think that in addition to generating far more data, generating data at scale, one of the ways that I think a model like this consortium is really going to progress the field is by pushing people, incentivizing people who might not otherwise collaborate at a much deeper level than has historically been possible.

David Brühlmann [00:14:04]:
In a nutshell, what is the most important takeaway from our conversation?

David Younger [00:14:09]:
I think the most important takeaway is coming back to what are the misconceptions of the field. We need far more data than exists today, both because the data that exists today is too limited in quantity, but also because those publicly available datasets are limited in quality for training machine learning models. So I think that’s one thing. And then second is that, you know, I think that with a combination of innovative technology and innovative business model, solving these problems is possible.

David Brühlmann [00:14:40]:
Thank you so much, David. Where can people connect with you?

David Younger [00:14:45]:
People can connect with me on LinkedIn or through the A-Alpha Bio website. Very happy to follow up with anyone who’s interested.

David Brühlmann [00:14:54]:
Excellent. There you have it, smart biotech scientists. I’ll leave the links in the show notes. And David, once again, thank you so much for being on the show today.

David Younger [00:15:03]:
Thank you, David. It was a great conversation.

David Brühlmann [00:15:05]:
If there is one thing this conversation makes clear, it’s that breaking the data bottleneck in protein engineering won’t come from a better model alone. It takes new incentives for companies to actually share what they know. David Younger’s approach at A-Alpha Bio offers one path forward. Thank you so much for listening today.