There’s a comforting story that’s going around about AI and medicine: feed a model enough data, and out pops the next miracle drug. It’s a tidy narrative. It’s also not quite true, and the people actually building these tools tend to be the first to say so.
One of them is Michael Antonov — best known as the co-founder of Oculus, the virtual reality headset that Facebook bought for roughly $2 billion in 2014. These days he runs Deep Origin, a computational drug discovery company, and in a recent interview with Inside Precision Medicine, he made a case that’s refreshingly unglamorous: AI is genuinely useful in drug discovery, but only when it’s paired with old-fashioned physics, mechanistic biology, and actual lab experiments. Not instead of them.
So what does that combination actually look like?
Drug discovery starts with a target, usually a protein involved in a disease, and a search for a molecule that will bind to it in the right way, without also binding to a hundred other things and causing side effects. Traditionally, that search has meant years of trial and error at the bench.
AI can speed up parts of this. Machine learning models can scan enormous chemical libraries and predict, with reasonable confidence, which candidate molecules are worth testing first. But prediction isn’t proof. A model can suggest a molecule looks promising; it can’t tell you, with certainty, whether that molecule will be toxic in a living system, or how it will behave once metabolised by a real body.
That’s where physics-based simulation and, eventually, the lab come in — modelling how molecules actually move and interact in three dimensions, then testing the results against reality.
Antonov’s point about toxicity is worth thinking about. It’s one of the more stubborn problems in drug development, and, as he put it –
“the predictive tools that exist so far are better than nothing, but far from solved. Roughly speaking, most drug candidates that enter clinical trials still fail, often for safety or efficacy reasons that models didn’t catch in advance. If AI-assisted toxicity screening could meaningfully improve on that, even incrementally, it would save an enormous amount of time, money, and — more importantly — the risk borne by trial participants.”
The other idea he raises is a longer-term one: personalised medicine built around individual patients, where a novel mutation or new virus could, in principle, be fed into a computational pipeline that maps its likely biological effects and points toward candidate treatments.
It’s a compelling vision, and Antonov is careful to frame it as a decade-plus outlook rather than something imminent — worth noting given how often “personalised medicine” gets used as a marketing phrase rather than a grounded prediction.
What I found most useful about this interview wasn’t the AI optimism — it was the caveats. Antonov describes a field crowded with companies of wildly varying quality, plenty of hype, and a real shortage of tools that have been proven against experimental data rather than just benchmarks.
That’s a more honest picture than “AI is revolutionising medicine,” and honestly, more interesting too. The unglamorous work of validating predictions in real biological systems is still where drugs actually get made and AI is a way of making that work faster and better targeted, not a way of skipping it.
Note: this post is based on a single published interview rather than a systematic review of the field, so treat Deep Origin-specific claims (like reported hit-rate improvements) as one company’s account of its own results rather than independently verified figures.
- Grinstein, J.D. “Michael Antonov: From Virtual Worlds to Real-World Drug Discovery.” Inside Precision Medicine, July 1, 2026.
- Deep Origin company website and press coverage — deeporigin.com
- “The Biotech Startups Podcast” (Excedr), four-part interview series with Michael Antonov, 2025 — background on Deep Origin’s platform and mission
