The Infinite Search Space Problem
I spent my morning thinking about the number 10 to the power of 60. That is the estimated number of drug-like molecules that could potentially exist. For context, there are only about 10 to the power of 22 stars in the observable universe. For a century, we have been trying to find life-saving needles in that cosmic haystack by hand, one pipette at a time, failing 90% of the time once we actually hit human trials. It is a miracle we have medicine at all, honestly.
But now, the venture capital world is behaving as if that haystack is finally being digitized. We are seeing a massive migration of dollars away from traditional 'bespoke' biotech—where you hire a hundred PhDs to stare at protein folding for a decade—toward automated IP generation platforms. These companies aren't just looking for one drug; they are building 'discovery engines.' I find myself wondering if we are fundamentally changing what a 'drug company' even is, or if we are just creating a very expensive way to generate more sophisticated failures.
If the value of a company shifts from its physical lab results to its underlying algorithm, does the 'science' become secondary to the 'architecture'? It is a strange tension. In the traditional model, the value was the molecule. In the new model, the value is the map of the territory. I’m curious to see if a map can ever be accurate enough to predict how a complex, messy human body will react to a synthetic intruder.
Shifting the Burden of Proof
The economics of this are starting to look more like software than science. In 2023, we saw AI-driven biotech firms like Recursion and Exscientia command valuations that would make a traditional mid-stage clinical firm weep with envy. Investors are betting on the 'platform'—the idea that once you solve the chemistry for one disease, you can simply 'copy-paste' the logic onto the next one. It’s a seductive thought. If you can shorten the discovery phase from five years to five months, the internal rate of return (IRR) on a drug doesn't just improve; it teleports to another dimension.
However, I can't help but notice the 'Valley of Death' hasn't moved. You can use all the GPUs in Northern California to design a perfect molecule, but you still have to put it into a human being. Biology is notoriously indifferent to how beautiful your code is. Phase II clinical trials remain the great equalizer. I’m fascinated by this disconnect: the front-end of the pipeline is moving at light speed, while the back-end—the actual testing—is still stuck in the slow, bureaucratic, biological reality of the 20th century.
Photo by Jahra Tasfia Reza on Pexels
What happens when we have a thousand 'perfect' candidates sitting at the door of the FDA? The bottleneck is shifting. We used to be limited by our imaginations and our hands. Soon, we might be limited only by our patience and the number of volunteers willing to let us test an algorithm's best guess on their liver enzymes. It’s a weirdly optimistic problem to have.
The Death of the Long-Tail Gamble
Venture capital is losing its appetite for the 'long-tail'—those high-risk, high-reward bets on a single orphan drug. Instead, they want the 'factory.' If you own the platform that generates the candidates, you own the toll booth on the road to the pharmacy. This commercial restructuring is turning Big Pharma into a collection of logistics and marketing hubs that simply buy 'de-risked' assets from the AI startups.
I wonder if this kills the 'garage scientist' vibe of early biotech. When discovery requires a $100 million compute cluster, the barrier to entry isn't just brilliance; it's sheer hardware. Does this consolidate the future of our health into the hands of whoever has the best relationship with Nvidia? It feels like we are trading one kind of gatekeeper (the academic establishment) for another (the compute elite).
There is also the question of 'Black Swan' discoveries. Some of our best drugs—Penicillin, Viagra, Minoxidil—were accidents. Serendipity is a huge part of medical history. Can an AI be 'accidental'? Can a system designed for optimization ever find the thing it wasn't looking for? I'm worried that by making the process more 'efficient,' we might be narrowing our field of vision to only what the training data says is possible.
What This Actually Means
We are witnessing the 'SaaS-ification' of human health. The transition from wet-labs to dry-labs isn't just a technical upgrade; it's a fundamental change in the incentive structure of medicine. We are moving toward a world where the most valuable asset isn't a patent on a chemical, but the proprietary data loop that predicted that chemical would work.
This shift might finally break 'Eroom’s Law'—the observation that drug discovery is becoming slower and more expensive over time despite technological gains. If these platforms work, the cost of entering a trial could plummet. But if they don't—if they just produce more convincing hallucinations of cures—we are going to see a valuation crash that makes the dot-com bubble look like a minor accounting error.
Ultimately, I’m rooting for the machines. Not because I love algorithms, but because the current system of 'guess and check' is failing too many people with rare diseases that aren't 'profitable' enough to study by hand. If we can automate the curiosity, maybe we can finally treat the things we’ve ignored for a century. We are standing at the edge of a world where medicine is designed, not found. That is a terrifying and beautiful thought.
Quick Answers
Is AI actually discovering new drugs yet?
Yes, several AI-designed candidates are currently in Phase I and Phase II human trials, particularly for conditions like idiopathic pulmonary fibrosis and various cancers.
Does this mean drugs will get cheaper?
In theory, lower R&D costs should lower prices, but in reality, pricing is determined by market value and insurance negotiations, not the cost of the 'discovery' compute time.
What is a 'Dry-Lab' valuation?
It’s a market trend where companies are valued based on their computational platforms and data processing capabilities rather than their physical inventory of chemicals or successful lab results.
