AI Drug Discovery just cut the first eight years of cancer research down to three or four months, and the oncologist who produced the number spent the rest of his stage time explaining why nobody in a hospital bed will feel it this year. Siddhartha Mukherjee is not a hype merchant — he wrote the Pulitzer-winning history of the disease — and what he described in New Delhi on Saturday is a real, measurable collapse in the slowest and cheapest stretch of the pipeline. Then he pointed at the part of the pipeline that did not move at all, which is the part where patients actually wait.
He was speaking at the 24th Hindustan Times Leadership Summit, as co-founder and chief executive of the drug-discovery company Manas AI, on leave from Columbia, where he taught medicine. Livemint carried the compression as the headline and Hindustan Times carried the mechanics of how the compression is done. Both descriptions agree on the shape of it: the front of the process has been rebuilt around a different question.
The old method was a search. Test billions of existing chemicals, one at a time, and hope something sticks to a target nobody has mapped properly. Manas attacks the mapping first. Its models, constrained by the ordinary rules of physics and chemistry rather than free to hallucinate a molecule that cannot exist, look for the weak spot on a disease-causing protein — then build a compound out of atoms to sit in that weak spot. Mukherjee reaches for a lock-and-keyhole picture because it is the honest one — and the trick that made the lock-finding tractable came from a student of his, who proposed simulating thousands of virtual points across a protein’s surface to work out where a molecule would most likely bind (OU Daily). The discovery step stopped being a search and became a design step, and design steps have timelines you can compress.
Siddhartha Mukherjee — AI Drug Discovery Is Real, and He Named Its Ceiling
The proof he offered was specific. For daraxonrasib, a pancreatic cancer drug, finding the vulnerable site on the protein took human researchers something close to a decade. The keyhole-finding module in his software did the same job in about two hours (Hindustan Times). Manas now ships batches of fifty to a hundred candidate compounds to Chinese chemical partners, which return test data in roughly three weeks. One molecule for a particular cancer, he said, was generated end to end with no human choosing the chemistry, and the company measures it at around two hundred times the potency of the rival medicine already on the market.
That last figure is a company claim from a conference stage, not a trial result, and it should be read that way. The timeline claim is the one with outside corroboration, and it is the one that matters commercially: work that consumed seven to eight years of a medicinal chemist’s career now lands in three to four months, reported by both outlets that covered the talk (Livemint, Hindustan Times).
The Pinch Point — the Molecules Arrive Faster and the Door Never Widened
Here is where the applause should have stopped. Mukherjee described the back end of medicine — putting molecules into human beings and testing them against a living disease — as essentially unchanged, and he was blunt about the consequence. His engines are now producing far more candidates than the system downstream can absorb, and those candidates pile up at the same gate they always did. He calls it the pinch point. The industry calls it phase two, and he described the odds there as a coin flip that no model in his portfolio has learned to weight.
That is the sentence that should travel. More shots on goal does not change the conversion rate if the goalkeeper is the same goalkeeper. It only means more failures get counted, and more money gets burned confirming them. A three-week design-and-test cycle is a genuine triumph of engineering. It is not a cure, and the distance between those two things is measured in funerals that were already scheduled.
Eroom’s Law — Cheaper Discovery Inside a System Built to Grow More Expensive
Pharma’s defining problem for three decades has a name that nobody outside the industry remembers: Eroom’s law, which is Moore’s law spelled backwards, and which observes that the number of new medicines approved each year keeps falling while the cost of each approval keeps climbing. Mukherjee put the whole of medicine’s credibility on this curve. If AI cannot bend it, he argued, the industry’s grandest claim about itself — that this technology produces human good rather than engagement metrics — collapses with it. He is not a miserabilist about the front end: he also cited a study in which radiologists working with AI-assisted screening detected 13.8% more breast cancers than radiologists without it (OU Daily).
The uncomfortable arithmetic is that discovery was never the expensive line item. Compressing the front of the curve by ninety-odd percent moves a small number. The trials, the manufacturing, the regulatory review, the years of watching for a side effect that shows up in year three — that is where the money and the calendar actually live, and none of it has been redesigned. A faster engine feeding an unchanged trial system produces one guaranteed outcome: a backlog with better molecules in it.
Immuneel — $500,000 in America Becomes $35,000 in Bengaluru
The most useful number in the talk was not a timeline. It was a price. Mukherjee co-founded Immuneel Therapeutics in Bengaluru about three years ago with Kiran Mazumdar-Shaw, and its blood-cancer therapy reached the market at roughly $35,000 in India against $500,000 for the equivalent treatment in the United States (Livemint). He says the difference is not charity and not a subsidy. It is an automated manufacturing facility built beside the doctors instead of an ocean away from them, plus binders designed with AI rather than licensed from abroad — the first synthetic binder the company made with its own models, in his telling, is first in its class and made in India.
Patients who had exhausted the standard leukemia protocols are alive because of it, including a boy from Kolkata. Hold that against the three-month discovery figure. One is a laboratory record. The other is a four-hundred-and-sixty-five-thousand-dollar gap between two people with the same disease, decided entirely by which country’s regulatory and manufacturing system they happened to be born inside. The compression is impressive. The price gap is the cure.
Axel’s Take — a Three-Month Molecule Still Awaits a Decade-Scale Verdict
Mukherjee did not stop at diagnosis. He proposed the fix, and the fix is the fight. Trials could be shortened by running against historical or simulated control groups instead of a fresh placebo arm — which he immediately flagged with the only acronym that matters here, bias in, bias out. A model trained on the last twenty years of trials will happily certify a drug that only works on the kind of patient who enrolled in the last twenty years of trials. His answer is conditional approval on the European model: a fast yes, trial conditions still running, and a withdrawal power that can be exercised in two or three years if the confirmatory data disappoints.
That is a reasonable design, and it is also the exact place where this story stops being about AI and becomes about who holds the pen. A three-month molecule arriving at a regulator with no capacity to audit the model that designed it, no legal mechanism to un-approve a drug in under a year, and no published standard for simulated controls is not faster medicine. It is faster paperwork wrapped around the same ten-year gamble. The gate that has to change is not the algorithm. It is the signature.
What India and China are being invited to do is build the fast lane and staff the brake. Mukherjee called it a historic opportunity, and he is right about the size of it — a country with the physicians and the manufacturing density to run trials at a fraction of American cost is the only plausible place the curve gets bent. The opportunity comes with a bill that nobody on that stage was asked to price: the first regulator to approve on model output alone will own every failure that follows, publicly, forever.
So celebrate the number, then read the second half of the sentence. The discovery engine got fast. The verdict engine did not, and the verdict is the only part a dying person is actually waiting on. Manas can design a molecule in three months — and if the trial system stays exactly as it is, the most efficient cancer research operation ever built will spend the next decade queueing behind a door that was never widened, holding better keys.
Sources: Livemint, Hindustan Times, OU Daily.