After Oculus was sold to Facebook for two billion dollars a decade ago, Michael Antonov, one of the founders, could have become a major tech investor, a popular podcaster with his own mega-theory of everything, a hedonist, or all of the above. Instead, he gravitated towards the emerging field of longevity biotech, where the uncertainties are as big as the potential.
Michael runs a sprawling investment fund, Formic Ventures, that focuses primarily on longevity biotech. He also co-founded his own company, Deep Origin, which has a soberingly realistic approach to foundation AI models in biology.
How did you end up in longevity?
I got into this space about 10 years ago. This was after I exited Oculus to Facebook, and I was still working at Oculus for a handful of years while looking for other interesting things to do.
Once, I was presenting at the same event as Aubrey de Grey. I was talking about virtual reality, and he was talking about longevity. To me, that sounded very interesting, so we ended up talking a bit afterwards. I realized there’s been a lot of progress in genomics and in biology in general.
So, I was looking for the most meaningful thing I could do in my life, and, at that point, it was clear where VR was going. Yes, it would take a lot of engineering, and many smart people were working on it, but I didn’t feel like it was critical for me to keep contributing there.
Aging, on the other hand, felt like the biggest challenge facing mankind. I had some resources for investing from the Oculus exit, and I felt I could learn things.
I got very curious about the science of aging, where it was at the moment, and what might be possible in the future. This led me to participate in the community. I was going to a lot of mixer events, conferences, and also started learning science. I ended up taking several years’ worth of biology, biochemistry, and related courses at the Berkeley Extension because I realized I wanted to know how we work internally.
Eventually, I decided that I wanted to invest in this space. I was looking at tools that speed up research and longevity-related therapeutic. One approach to improving healthspan is by developing drugs, and in the aging field, even though we want to target aging, we usually have to make a drug for a specific disease.
At some point, I met Alex Zhavoronkov, who was instrumental in helping me find my way in the ecosystem. We became good friends. I ended up investing in his InSilico Medicine, and we still connect a lot.
A few years later, I ended up starting Formic Ventures and looking at what companies I could invest in. Finally, I realized I wanted to get more into scientific tools, and that is what made me start Deep Origin.
So, the first thing you built in this space was an investment company. When I looked into it, I saw that you had invested in about 40 companies, which is an unusually wide net. Can you tell me more about the company and its philosophy?
Generally speaking, I looked at companies which could make an impact and took an approach that is different, let’s say, from the typical way of doing things. Fundamentally, that is because I want to make a difference rather than just money.
It’s not focused purely on profit, but, of course, the question of whether the company can be successful and profitable is an important one. Also, there were companies where I was willing to take larger risks because I really liked the team and the direction.
I’ll give you two examples. One of the companies I invested in is Turn Bio. They work on cellular reprogramming, and they were in this space significantly before Altos, and before this field became so hot. I still believe in their approach. At the point when I first invested, it was high risk, but it felt like there was meaningful data and direction.
If you believe that epigenetic reprogramming is possible, it seemed like a worthwhile goal to move forward. That’s one example of a company which is aligned with my vision – a unique approach (at that time) plus a meaningful impact. And when it succeeds, it certainly has a big potential.
Another good example is Nanotics. They use nanoparticles to take cytokines out of the circulation, which is very different from putting the drug in. So, it’s a different modality, high-risk novel approach, but it’s a kind of thinking that we need in this space.
Looking broadly at the Formic profolio, not all of my investments are in the longevity space, although about 70% of them are. I have also invested in some companies created by “the Oculus mafia,” meaning people who we started the company together with.
Do you think there’s any place for VR in the longevity field?
Not a lot, it’s mostly unrelated. However, VR is very good at training, and there have been some companies that utilize this in the medical space. As an example, Osso VR trains surgeons, and they have shown that the results of the training for the same time period were measurably better. VR can also be good for scientific visualization. But it’s not directly advancing science, it’s just a different way to interact with the world. For instance, you can look at the drug in 3D space and how it docks. Chemists may be seeing things a bit clearer by looking at them from different angles.
From your experience and the breadth of your investments, where do we stand now with longevity biotech? How optimistic are you? How do you judge the trends in the last couple of years?
I think the good part is that there has been progressively more capital in the space over the last five to six years. There are more funds, more people involved in it and believe in it, it’s a more active space. That’s a holistically good thing.
Specifically, biotech has been in a little bit of a trough. There was this market crash last year, and it hasn’t fully recovered. So, it’s still harder now to raise money. It certainly affects current longevity companies. Otherwise, I think it’s a positively developing industry.
Personally, I probably have grown less optimistic than I was eight years ago. That’s because we all know drug development takes a decade, but we don’t feel it in our bones. As we come into the industry, and look at the exciting research developments at conferences, the progress can feel quicker than it really is. We are making progress, just not as fast as I’d like.
I really want to think about how we as an industry find ways to speed it up. Is it scientific tooling, such as modeling of biology that will truly make a difference – which is why we created Deep Origin? Is it robotic automated labs? I don’t know.
Can we grow organs on chips and convince the FDA that those results are at least as compelling as long-term trials? We need some solution that would fundamentally speed up the progress.
I think the final wildcard is AI. There may not be quite enough data and proof for how much difference AI will make, but what is possible is also unknown, and there’s a lot of positive excitement. It may well be that with the help of AI we will make big leaps, but it’s unclear at the moment.
What do you think about AI’s role in drug discovery? How big of a gamechanger can it be?
AI is very helpful in drug discovery. It’s actually an area where we work. In Deep Origin, we have a very strong physics and AI team. We specifically work on physics-based technologies, on things like molecular dynamics, docking, energy fields, and AI training of custom models to do it even better and faster, or in broader sets of molecules.
AI helps a lot in structural biology. This is where you see that AlphaFold can predict some parts of protein interactions, there’s more and more structure. In general, AI is able to predict phenotypes and chemical outcomes, like drug properties, toxicity aspects, and what not.
To what level is that good is really a function of the data we have and, in some cases, of previous models we’re picking from. As a result, it’s a very powerful tool to speed everything up, but there are enough gaps in data and other technologies for it not to be completely lifechanging. That’s where we are today.
Everyone hopes that the next generation of AI will be just magical. At every step, something gets better, but it’s not magical. We still have lots of biological problems.
Zooming in on this data problem: specifically for big foundation models of biology, which is something Deep Origin is also working on, how serious is it, and what can we do to make things better?
It’s a significant problem. The amount of data you need is, in some sense, problem-specific. If you’re looking at a high-level patient phenotype, you need one kind of data, but when you’re looking at molecular structures, you might need biochemistry, crystallography, or other data. Those are different classes of data, and the volume of it that you will need will also be different.
We need a lot of lab automation at scale. That would be instrumental for insights into deep biology.
We also need better data integration. Often data is in different institutes, it’s siloed in, not accessible to researchers. You need to apply, maybe to form partnerships. By the nature of it, that means data is not as widely shared as we’d like it to be, which is one of the reasons the field is not moving as fast as we want it to.
We do need orders of magnitude more data to truly model biology. All kinds of data – molecular structure, tissue microscopy, multi-omics of different cell and tissue types taken from people of different ages. All those things, we don’t have enough of. We really need to build up the dataset, as a world community, and be able to share it. That would enable simulations and better predictions.
I understand that one of the things Deep Origin is trying to do is to automate the process of drug discovery.
Deep Origin’s vision is to enable finding cures faster through deeper understanding of biology. To put it in one line, our mission is to organize, model and simulate biology. That’s what we do.
The name Deep Origin speaks about going to the origins of life, which are atoms and molecules. You have to go up from atoms and molecules – how do they come together? Understanding the structure and so on.
We are building a platform that has two branches. On one hand, it supports data collection, management, and processing from the wet lab through analysis. That is, how do you record your experiment, how do you analyze it, how do you get data out of it?
The second one, beyond that, is simulation. Simulation is, basically, if we have this data, or if we have some knowledge, what would happen under certain conditions? How would proteins interact? How would physics work? How can you actually simulate a cell?
Ultimately, the way these two things are connected is that even if you have a good simulation stack, you need lots of data to validate it. Both are meaningful. So far, what we have built is a suite of tools which we provide to biotechs and license to pharma.
Right now, some of the tools we built solve very specific problems. For example, we have docking and virtual screening solutions. If you have a protein you want to drug, we can provide a state-of-the-art solution for that narrow problem.
But we’re actually looking at how multiple dimensions of research come together. If you’re going to design a novel drug molecule, you may also need to run biological screens. In which database will they be stored? How are they linked? How are aspects of biology represented? We want to support this whole process.
Yes, we also want to automate workflows for bioinformatics processes, and we will connect to labs, but before you can do advanced analysis or AI, you need to collect the data and the data needs to be in a certain understandable format.
We actually want to make parts of our platform open going further, so that users can have more standards and easier ways to access the data. The main question is how we get consistent quality data so that researchers can answer questions, and how do you run simulations that help them dig even deeper.
Tell me about your work on foundation models for biology.
We actually have foundation models for some of our docking and chemical properties predictions, for some structure work. That said, “foundation model” is a very overloaded term. Probably every time that someone tells me they’re starting a company or raising money to build a foundation model for biology, my question is, what does it do?
You can have an LLM model which is trained on text and works on text. Alternatively, you can train a model on structures, which is what AlphaFold 3 did.
I’m not an AI expert, we have really strong AI people on our team, like Garik Petrosyan, but you’re basically combining multiple training sets together for different purposes. You have a neural net to predict protein and maybe DNA structure, you have sequence, and you train them together in a way that some of the knowledge is transferable.
So, if you now want to predict something new, to adapt your model to a new domain, you do some extra training. You may not have had enough data to get a good output on this new dataset, but because you’re combining it with a bigger model, you’re now able to get high quality results, because some internal learning from these other learning modalities is being applied in your new case.
In my understanding, foundation models are a generalization of that. You’re picking a set of subdomains, and they are able to predict a set of other things. But I do not yet believe in a universal biological foundation model that will solve everything for you.
You can probably try to train it, and it will make a useful suggestion, or it will hallucinate. Imagine if you merge LLM and structural biology. Now, you can ask it questions about pathways, which is text, and you can also ask it, what would a protein look like in a given context? And it might be able to give you both a text and an image as an answer.
But the question is, will it be a good protein for the given task? Most likely, the answer is going to be “no” because the model didn’t have enough similar proteins or enough precision in training. It may be a good, smart guess but you will need to keep working on it. So, foundation models by themselves are not the complete answer.
Our approach is different. We have some foundation parts in AI, but we’re combining them with physics-based tools. If you look at molecular simulations that people have been doing for the last 20, 30 years, they use energy fields, and we’ve gotten better and better at it. But we are not perfect, it just takes too long to compute. That’s a big challenge.
But now, you can start combining it with AI. You can apply those coarse-grain approaches to it. You can train specific neural nets which will do a given task, like binding a molecule to a pocket, very well.
Your typical off-the-shelf foundation model that you find on the Internet will probably not do a good job at that because it hasn’t been trained specifically on that problem, but if you’re making a drug, you need a specific, high-quality answer.
We are combining those physics-based tools with AI, with generalized LLMs and other things, into a solution to a given problem set. It can’t answer every question in the world, but for a class of structural biology or drug discovery problems, it will use state-of-the-art tools to give you the world’s best answer (we believe) for a given set of narrow problems.
What do you think is the main bottleneck for this approach specifically and foundation models in general?
There are several. If we look at a general molecular interaction problem, such as folding, how things combine, how they bind, how reactions are happening, one problem is that we don’t always have data on energy fields.
For instance, we don’t have very good datasets of energy fields for RNA (we have much better ones for proteins). When we don’t have this, we can’t run simulations as well.
The second challenge is that some things are either too hard to compute or we simply lack knowledge. This applies, for instance, to some quantum effects. For example, if you want to break bonds, you need to do quantum chemistry, which is hard.
And the third class of problems is not having enough data – basically, how many outcomes of a given experiment have you measured? The way we approach it is, first, we try to collect the best data we can out of open or collaborative sources for our physics models. Like I said, we are trying to augment our physics with AI, so you can sometimes run simulations faster and get better data from those simulations.
So, not all the data needs to come from experiments, some of it can come from simulations, but we also try to collect data when we can. We have a small wet lab, which we do experiments in, but, in general, we are looking at a more hybrid approach. We’re not a brute force data company, we’re more “physics combined with AI” company.
Can you give me a specific example of how this hybrid physics/AI approach works?
Let’s take a look at a hard problem. There’s a technology called molecular dynamics. We have our own version of it which we think is particularly good but it’s a well-studied field. It models molecules and simulates them over time – for example, how a full protein folds and unfolds, what shapes it takes.
It’s a very useful technique. The problem is it takes too long, because you have to do computations of many atoms and all their relationships for each femtosecond, and then you need millions and millions of these time steps to get a picture of that millisecond where useful biological activity takes place.
The problem is that there’s just not enough computing power, but if you can get enough of it, you get very useful answers for some cases.
Imagine running this simulation a number of times, but instead of always trying to do big tasks, you train your neural net. Then, this neural net can give you suggested answers for certain problems faster than if you actually ran an expensive simulation each time.
So, now, for a set of problems, you’ve made it faster. We’re looking for these kinds of patterns. I’m simplifying a bit, but you can imagine doing it for every bit of biology.
You also have a foundation that funds academic research.
Yes, The Antonov Foundation makes grants in areas I care about. A meaningful part is supporting longevity research by good scientists. So, I have donated to Buck, to SENS, I’ve supported some of Vera Gorbunova’s projects. We’re talking about people whom I know who are doing interesting things.
We’ve looked at projects with promising outcomes and provided them with some capital. It’s really impact-focused.
When you are looking at the whole longevity field and your place in it, what are the main hurdles we will have to overcome? What needs to change ASAP?
More flexibility in regulation on the FDA side would help. That’s definitely a very expensive process, and it needs to be more streamlined. There have been some good steps already, for instance, with drugs for very specialized groups, which can be fast-tracked. But in my view, it’s not enough. As Milton Friedman suggested, we need to put in conscious effort to always fight the tendency to overregulate. Here, some political push is required.
In addition to that, I’d love to see more standardization and automation in biological data and processes. There are too many standards and vendors, which makes it harder to do reproducible experiments. We need a bigger push on the industry level to open data structures, representations, and protocols for experiments.
This brings me to the question I honestly forgot to ask: how does the reproducibility problem reflect on AI in biology in terms of data acquisition?
This is a big issue because, typically, models are trained on some external data sets, and those were possibly collected under heterogeneous conditions. This makes the whole thing less reliable; it needs to be somehow adjusted for. If we really want high-quality predictability, we need highly reproducible robotic labs with well-controlled conditions at scale.
Not a lot of companies today have it. Some pharma does. InSilico has an interesting new lab, but that’s not a universally adopted practice yet. We need much more coherent standardized protocols for really high-quality AI predictions.
You come from the tech industry. It seems that tech people’s interest in longevity is booming and that their outlook on longevity is more positive and optimistic than that of the general public. What do you think of this budding synergy?
I agree that the interest is growing and that tech investors, and maybe especially crypto investors, tend to be more optimistic. On the other hand, sometimes, they don’t realize what they’re getting into, but this is a good thing because often you have to believe that something is possible to keep pursuing it. In the end, because of their support, we are making more progress.
Look at Altos Labs as an example. A number of wealthy people came together to support this initiative. So, overall, more resources and optimism are going into longevity. It’s good, it’s happening, but it’s not a linear process, it feels very stochastic.
About them not realizing what they’re getting into – do you mean that people who made their money in tech industry, where timelines are different, might not have enough stamina to invest in longevity biotech and wait for a decade or more to see the return?
Some will have what it takes, and others won’t. Some people will get discouraged, maybe after a couple of bad bets, but many are on a mission, because this is something that matters. They see that having money doesn’t do that much for their lives, and they want to help the world. So, this is going to continue. As for me, I’m in for the long run.
I assume you don’t regret your decision to go into longevity head-first?
No, although I wish it was moving faster. I wish it was easier, but it is a fruitful and necessary endeavor, and, in general, the field is on an upswing. We’re growing.
