Isomorphic Labs Files for IPO After $600M Funding: AI to Become Indispensable in Drug Design Within Five Years, Says Chief AI Officer

In 2020, AlphaFold2 emerged and swept CASP14, completely revolutionizing AI + biology.

Based on the milestone research achievements of AlphaFold2, the leader of this projectDemis Hassabis, thereby gaining2024Nobel Prize in Chemistry。

Last year,Isomorphic LabsAnnounced two value deals with multinational pharmaceutical giants Eli Lilly and Novartis30A multi-billion-dollar drug discovery agreement, involving innovative therapies for various diseases and targets, has recently been expanded by Novartis to further broaden the scope of collaboration between the two parties.

Just recently, the company's Chief Artificial Intelligence OfficerMax JaderbergPh.D.Participated in the podcast program of Sequoia Capital,And PartnersStephanie ZhanEngaged in a dialogue, sharing his views onAI+Perspectives on Drug Development.

Before devoting to AI + drug research and development,Max Jaderberg is a leading figure in the field of deep learning, andDemis Hassabis has worked together for over a decade. Previously, he mainly focused on the application of reinforcement learning (RL) in gaming, achieving milestone breakthroughs such as "Capture the Flag" and AlphaStar.

Regarding AI + Drug Discovery, he stated:

1、Isomorphic Labs is attempting to use AI to create a highly versatile drug design engine that can be applied not only to a single target or a single type of drug, but also repeatedly across any different disease areas.

3. If AI is to completely revolutionize biology and the paradigm of drug development, six achievements at the level of AlphaFold2 are required.

4. The GPT-3 Moment in Biology —— It means that AI can design molecules based on physical reality but beyond human intuition, just like the divine move 37 during AlphaGo's match with Lee Sedol.

5. With the gradual success of AI in pharmaceuticals, in five years, the entire industry will be unable to design drugs without AI, just as scientific research cannot be conducted without mathematics.

The following is a compilation of the podcast content (with omissions)

Building a General AI Drug Discovery Engine

Stephanie Zhan：If everything goes smoothly, you have achievedIsomorphic 's vision, then what would the world look like?

Max Jaderberg：We hopeThe AI technology we are building will fundamentally change the way we understand biology and our ability to study chemistry in order to modulate biology.

Therefore, what we are really considering is solving all diseases. In the future, AI will not only help us discover, create, and design new treatments, but also deepen our understanding of the biological world, how our cells work, and the root causes of diseases, thereby opening up new pathways for modulation that we can explore.

So from the very first day of founding the company, what we truly pursued was not developing treatments for specific indications or specific targets, but rather genuinely contemplating how to leverage AI to create a highly versatile drug design engine.It can be applied not only to a single target or even a single therapy, but also to any different disease areas. This is the direction we are currently working on.

Stephanie Zhan：With such a grand goal, how has the way you research and build AI changed from day one?

Max Jaderberg：That's a good question. When I considerAI In some of the current situations in drug design, machine learning models have been widely used in chemistry and biology, but I would call many of these applications from the first generation more localized models.

There may now be some data on the behavior of specific targets or specific classes of molecules, based on which people fit a small multilayer.MLP, to help you generate some predictions, thereby informing the next round of design.

When I consider what we need to achieve the breakthrough drug design engine we've been building, we need about sixAlphaFold-level achievements。

CurrentAlphaFold It is clearly a huge breakthrough in understanding the structure of biomolecules. So what is the structure of proteins? Now we haveAlphaFold 3, which can simulate small molecules andDNA AndRNA The construction process of equivalent substances. This is a fundamental advancement, enabling us to achieve experimental-level accuracy for truly core concepts in biochemistry, thereby unlocking a multitude of thinking and design tasks for chemists.

But what I actually want to say is, we may need more breakthroughs like this. They have somehow achieved experimental-level accuracy in different core concepts of biology and chemistry.

A Devoted Fan of Synthetic Data

Stephanie Zhan：We know the dataVery important in the biopharmaceutical field，Previously Demis HassabisSaid in an interview,AIUnconstrained by data in biology. Could you share your perspective?

Max Jaderberg：In fact,Regardless of machine learningApplicationIn which field, willThere isData restrictions.

ButI think Demis The key point is, this is not a real bottleneck because we can make progress with existing data, we can generate data, and we can make real progress.

Rather than usMust wait in sitting 50 Year,Only after this world generates enough data can it make an impactfulAI. After all, what people have achieved nowGenuine substantive progress, surpassing anything previously experienced.

Now, does this mean that data biology has no chance?Absolutely not. There is a vast amount in this field.Historical data, but this historical data was not created for machine learning. Therefore, when you go out and think“How to Create Data to Actually Train My Model”When you are thinking in a way that is completely different from how people used to generate data, there is a huge opportunity to explore.

Stephanie Zhan：What kind of data do you think we are currently lacking? Do you think we need synthetic data?

Max Jaderberg：Yes. So I am a big fan of synthetic data. In fact, I have been generating synthetic text data since the beginning of my career.。

We see the same trend in the field of chemistry, where based on quantum chemistry and physics, we canApproximate calculations and create more scalable molecular dynamics simulations. This provides a basis for large-scale synthetic data.

Then we have the model itself, especially generative models, which can actually generate data, and we can use scoring systems to help truly enhance the informational content of that data.

But I think one of the big spaces is going to be what's called“In Vivo Data”。Currently, animal experimental data is very important in drug development, but it is difficult to generate large amounts of data.

Therefore, there is a great opportunity to find new data generation technologies. Some amazing people are working on things like organ-on-a-chip.But you know, things measured entirely on the chip，I think......

In the future, there will be a bunch of new breakthroughs in data generation technologies in the fields of biology and chemistry. You know, this will have a significant impact on how we think about modeling the biochemical world.

Stephanie Zhan：Are you doing any work internally, or are you hoping other participants will fill in some gaps?

Max Jaderberg：We actuallyIsomorphic Labs There are no labs of our own in China, but we collaborate with a great number of companies.We have generated a large amount of data ourselves., withA lot of proprietary data, we have already seen its huge impact.

AlphaFold3 Is Impacting Drug Development

Stephanie Zhan：There is a view that molecular structure modeling and modeling its functions and regulatory functions are very important, but not always the limiting factor in drug development.What are your thoughts on this?

I think the entire industry needs to figure out how to innovate in this space, particularly as we develop predictive models for how these molecules will interact with humans and their toxicity.As these predictive models get better and better, we're going to have to change the way we do clinical trials to take advantage of it.Ultimately delivering treatments to patients who truly and urgently need them.

From my perspective, what's really exciting is, if we create these general models to understand how this molecule interacts with this target, and how any other target interacts, then why can't we use the same model to understand how these molecules interact with other parts of our body?

Stephanie Zhan：AlphaFold 3 What functions are now available for drug designers? How do you use it internally?

Max Jaderberg：Therefore,AlphaFold 3 Enables our drug designers to understand how their molecular designs truly interact with this protein target.

Therefore, our drug designers can make changes to the design and then immediately see how it alters the physical interaction of that molecule with the protein target.

This is really, really powerful. InBefore AlphaFold3, people might not actually know how molecules interact with proteins, and perhaps at some point in a drug design project, specific designs would be used to crystallize structures.

This means that if you're lucky, in six months you go to the real lab and get a resolved 3D structure. But even then, it's just the 3D structure of a single design, not every change you've made.

AlphaFold 3 has completely transformed the way chemists approach this design work. But I want to emphasize that this is far from what we aim to achieve.Because it's not just about what these molecules look like in terms of their interactions, we actually want to know how strongly these molecules interact with this protein.

We want to know other properties of these molecules. We want to understand how these molecules interact with this protein, and how it changes the folding or conformation of the protein, how it alters the function of the protein, and how it actually changes the dynamics of the cell. These are other similar aspects we are working towards.The Breakthrough of AlphaFold.

Stephanie Zhan：Interesting. So, what goals and plans are you focusing on in the projects developed internally in China?

Max Jaderberg：We have some exciting drug design projects focused on immunology and oncology. We've made some incredible progress there, and it's really exciting to see. Especially how these models have changed the way we actually conduct drug design in these projects.

Stephanie Zhan：The company also collaborates with Eli Lilly and Novartis.Collaboration, could you share what these partnerships are like?

Max Jaderberg：Yes, we signed collaborations with two companies in January last year. They brought us some really, really challenging problems. These are targets that have been worked on for more than a decade by the field and by companies like Novartis.

So these are not"Oh, we'll give it a try" are truly difficult things.Last year was an incredible year, both for our internal projects and for these partner projects, as we were truly able to see how these models work. It enabled us to genuinely discover new chemicals and find new ways to modulate targets that people have been studying for a long time. Recently, we expanded our collaboration with Novartis, which I believe is real proof of some early successes in these partnerships.

Stephanie Zhan：You and Demis HassabisEstablished a decade-long cooperative relationship with himWhat does cooperation feel like?

Max Jaderberg：Demis He is an extraordinary person, a true dreamer, and also very approachable. After just five minutes of conversation, he can truly reveal the depth of his ambitious thoughts.And set an extremely ambitious goal from the very beginning.So I think he has this powerful ability to inject great energy into the team.

Stephanie Zhan：You have built a truly outstanding team, composed of the best talents from many different fields such as artificial intelligence, chemistry, and biology. Can you share how you view this?

Max Jaderberg：The application of AI in the field of drug design has not been around for long.Therefore, the chance of finding someone who is both a world-class expert in drug discovery and also a world expert in machine learning or deep learning is basically...Zero.

AIIn BiologyGPT-3Moment

Max Jaderberg：GPT-3 is essentiallyIs a generative model. For me,GPT-3 The moment crosses the boundary between the two, generating something that looks like text, but I don't believe it was generated by humans.

Therefore, when I consider applying it to biology,InGPT-3 The moment of re-creating the actual appearance of things in reality,This means that the AI-generated molecules are stable,Can work and exist in our physical reality. But in fact, AI can generateThings beyond human comprehension, yet they do exist in the real world.

This is just so exciting. In fact, you know, we're starting to see internally that through our generative models, we're creating what human drug designers would say,"Mm. I'm not quite sure. I prefer this one. Then you test it in physical reality, the generative model is correct, and humans are wrong."

Stephanie Zhan：This is very interesting. I like the37-Step Analogy. When Models Possess Creativity and Surpass Humans.

Max Jaderberg：TheStep 37 was the astonishing move made by AlphaGo in its match against Lee Sedol. On move 37 of the game, it shocked the world and the Go community because it could not be explained by humans.

It looks like a mistake. You know, in the thousands of years of human Go history, no one has ever made this move. It turns out that when you expand the game, this is what allowsThe key move in that match where AlphaGo defeated Lee Sedol.

Stephanie Zhan：So, when will we see our first one in clinical practice?AI What about the generated drugs, and the Phase I, II, and III trials?

Max Jaderberg：We have made astonishing progress in our drug design project. When we started, we obtained a large number of AI The assets, these molecules, when they enter the clinical stage, how can we really start thinking about participating in clinical development to deliver these molecules to patients as quickly and safely as possible.

What I'm thinking about here is new approaches to working with regulators, what new approaches will be taken to integrate our predictive models, not just in understanding how this molecule works against the disease, but as we've discussed, how it interacts with the rest of the body. I think there will be opportunities to consider streamlining and speeding up that process.

Max Jaderberg：In a sense, pharmaceutical companies will useAI。I think that in five years you're not going to be able toAI Designing a drug in this scenario is like trying to conduct scientific research without using mathematics.

AI Will become a fundamental tool in biology and chemistry——At least inIsomorphic In the world, it is already a program that everyone uses. So it won't be,“Oh, is it pharmaceuticals or artificial intelligence? In a sense, it will be one and the same, and the entire industry will adapt to this.