Early numbers say the mechanism underlying the search for viable molecules is real
IT IS the 19th century. You visit a young physician, Arnold Cahn, about a stubborn stomach complaint. The diagnosis: intestinal worms.
His assistant returns from the pharmacy with a bottle of white powder. You swallow it. The worms are unmoved, but your fever breaks. You have just helped discover acetanilide, the fever drug that would lead chemists to paracetamol.
And for most of history, breakthroughs like this arrived by luck.
But with the rapid development of artificial intelligence in the last few years, scientific discovery in life science has been meaningfully changed for the better.
For example, in 2020, DeepMind’s AlphaFold learnt to predict the three-dimensional shapes of proteins, a problem that had stumped scientists for decades. Within two years, it mapped more than 200 million structures, against roughly 170,000 solved by hand over sixty years. That is the promise of AI, and we’ve seen similar breakthroughs in biology, genetics, biochemistry and other branches of life science. AI has proven itself as a useful tool in scientists’ arsenal, and shown that challenges long thought intractable can be solved.
AI systems are expected to reshape how new drugs are discovered and developed. Nvidia’s Jensen Huang has said that life sciences is turning from a biological problem into an engineering one, and Anthropic’s Dario Amodei has suggested that AI could compress a century of discovery into a decade.
As optimism runs hot on the potential of AI, investors must bypass the hype to form a logical investment thesis. The challenge with backing any frontier technology ultimately boils down to determining whether you’re making a calculated bet or a fool’s wager.
Making a case for the tech #
Critics of AI in drug development have a favourite statistic, and it is a good one: zero.
Since 2019, some US$60 billion has been committed to AI-driven drug discovery; roughly 175 AI-originated programmes have entered human trials. Not a single one has yet won approval from America’s Food and Drug Administration.
Taken alone, the figure invites an obvious conclusion – that the industry has mistaken a compelling narrative for a viable one.
Taken in context, it points to a subtler explanation: The technology may work, but the approval process it must pass through was never built to move any faster for it.
The more revealing statistics lie earlier in the pipeline. Molecules identified by AI are clearing Phase I trials at a rate of 80 to 90 per cent, against a historical baseline near 52 per cent for those discovered by conventional means.
In Phase II, they succeed roughly 40 per cent of the time, matching or slightly exceeding the traditional range of 29 to 40 per cent.
Phase III, the stage that actually yields approvals, is only now being tested at scale, with an estimated 15 to 20 AI-derived programmes expected to enter pivotal trials in 2026. Insilico Medicine’s Rentosertib leads the field, having produced a clinically meaningful result in Phase IIa.
The absence of an approval, then, is best read as an artefact of timing rather than a verdict on the technology.
The pattern of capital allocation lends the thesis further weight, without in itself proving it.
In May, Isomorphic Labs, Alphabet’s drug-design venture, closed a US$2.1 billion funding round, among the largest ever raised by a private life-sciences company, participated by Thrive Capital, Alphabet, GV, Temasek, MGX and Britain’s sovereign AI fund.
Xaira Therapeutics launched in 2024 with more than US$1 billion behind it, courtesy of Arch Venture Partners and Foresite Labs.
Such sums are not the product of retail enthusiasm, but of institutions whose business is the disciplined pricing of risk. Their wager is not on a slogan but on a mechanism: that computation can narrow, meaningfully and repeatably, the search space for viable molecules.
Three developments converged to make this moment.
Protein-structure prediction moved from unsolved to solved, chiefly through AlphaFold and its successors. The computational cost of large-scale molecular simulation fell by an order of magnitude. And the pharmaceutical industry itself, confronted by looming patent expiries and a trial-failure rate that has hovered near 90 per cent for decades, turned from sceptical gatekeeper to willing customer.
Novartis, Sanofi and Eli Lilly, among others, have signed platform partnerships well ahead of any proof of approval – a signal, from firms with every incentive towards caution, that the underlying economics are already judged sound.
None of this eliminates risk. Recursion’s decision in 2025 to discontinue several pipeline candidates is a useful corrective: Platform promise does not exempt any single molecule from the base rates of human biology.
What AI narrows is the search space, not the ultimate uncertainty of Phase III.
But an investment thesis resting on rising early-stage success rates, disciplined institutional capital, an identifiable structural catalyst and a near-term test of validity is not speculation dressed up as strategy. It is, by the standard definition, a calculated bet.
Where the value concentrates #
The path from idea to approved drug runs through three stages.
Discovery, where molecules are found and designed, is now mostly done on a computer. Preclinical validation tests candidates in the lab. Then come clinical trials, where a drug must prove itself first in animals, then in people.
Most of the investor attention still lands on the first stage, but the picture there is busier than it looks. Model quality was the moat for perhaps a few years; it no longer is.
Isomorphic Labs, DeepMind’s drug-discovery spin-off, still guards its most capable engine, but open-source rivals such as Boltz and Chai are narrowing the gap quickly enough that no proprietary model stays ahead for long.
What does not commoditise is data – the hard-won, hard-to-copy kind: patient cohorts, rare-disease pathways, experiments that a rival cannot simply repeat.
Proprietary data is the one asset that a well-funded competitor cannot simply buy or replicate on a shorter timeline, which is what makes it a genuine moat rather than a temporary lead.
Noah’s Farm, which has quietly assembled a dataset linking food composition to taste perception that few competitors could replicate, illustrates the principle better than any pharmaceutical example: The dataset is the business, more than the model built on top of it.
More interestingly, large pharmaceutical companies are sitting on a huge repository of proprietary data that has accumulated in decades without a useful way to harness it. This is the exact problem companies such as Drylab AI aims to solve, to help pharmaceutical companies train and own their biology model rather than renting intelligence elsewhere.
The second stage, validation, is where a quieter but more decisive contest is playing out.
Fully automated “self-driving” labs are neither new nor as close as they sound. Today’s systems handle mostly simple, repetitive tasks, with scientists firmly in the loop. The real advantage is subtler.
A model’s suggestions are only as good as the results fed back into it. So whoever closes that loop tightest, running experiments and folding each lesson into the next design, improves fastest, and the edge compounds.
Robotics is the parallel – real-world scale matters because the awkward edge cases never show up in simulation. Singapore-based ChemLex runs an autonomous chemistry lab and has begun taking on pharmaceutical clients abroad.
The third stage, clinical trials, holds the hardest truths. Innovation and external funding have clustered at the front end because the further along you go, the harder it gets.
Historically, about nine in 10 drugs that reach human trials still fail, and no amount of computation makes that stage disappear – living bodies are more variable than any model can anticipate.
A newer crop of companies is using “virtual” trials to model outcomes and flag safety problems before costly testing. It is early and unproven, but it aims at the most expensive part of the journey.
Although AI has not made drug discovery certain, it has made the search deliberate, replacing intuition and chance with a process that can be tested, repeated and improved. That distinction is what separates a calculated bet from a fool’s wager, and the evidence here supports the former.
The early numbers say the underlying mechanism is real, not promised: Phase I and Phase II success rates already match or exceed historical norms. The capital behind the thesis was committed by investors paid to be sceptical, not swept up in a narrative.
And the timing is not arbitrary – a genuine scientific bottleneck broke, computational costs genuinely fell, and an industry with every reason to resist instead signed on early.
Isaac Kwa is senior associate, investment, and Tuan Anh Nguyen is associate, investment, at Vertex Ventures South-east Asia and India
Decoding Asia newsletter: your guide to navigating Asia in a new global order. Sign up here to get Decoding Asia newsletter. Delivered to your inbox. Free.
Share with us your feedback on BT's products and services