Sixty years of computers in drug discovery, and what has reached the clinic so far

In 1964, Corwin Hansch, a chemistry professor at Pomona College in California, and his colleague Toshio Fujita published an equation in the Journal of the American Chemical Society. It linked measurable properties of a molecule, such as lipophilicity (its preference for oil over water), bulk and charge distribution, to its biological activity. The method became known as QSAR, for quantitative structure-activity relationships. The premise: estimate a compound's effect before making it, once the molecule is described in numbers.

A French-language solo episode of Tech Anatomy, published in September 2026, follows that premise across sixty years, treating every later method as a more elaborate way of turning molecules into data a computer can process. Along the way, the question changes from whether computers can reason about chemistry to whether what they produce helps patients more often than conventionally discovered drugs do.

From equations to shapes

The first decades owed little to artificial intelligence in today's sense. In the 1970s at Squibb, David Cushman and Miguel Ondetti designed captopril, an antihypertensive, by reasoning about the shape of the enzyme pocket the drug had to occupy instead of screening thousands of compounds blindly. The image behind it, a key fitting a lock, dates back to Emil Fischer in 1894; Daniel Koshland later refined it into induced fit, where both partners deform slightly as they bind.

In 1982, Irwin Kuntz's group at the University of California, San Francisco released DOCK, the first molecular docking program. It tries a candidate molecule in many orientations inside a protein's 3D structure and scores the fit. The clearest validation of structure-based design came in the 1990s with HIV protease inhibitors, designed against the crystal structure of an enzyme the virus needs, which became part of the combination therapies that turned HIV into a chronic condition. The episode notes that there was "not an ounce of artificial intelligence in it yet" (translated from French). It was crystallography, physics and medicinal chemistry.

From sorting molecules to generating them

Two unrelated events in 2012 set up the next phase. Jack Scannell and colleagues described what is now called Eroom's law: since 1950, the number of new drugs approved per billion dollars of R&D spending had halved roughly every nine years in inflation-adjusted terms. The same year, Merck ran an open QSAR competition on the Kaggle platform. A team from Geoffrey Hinton's deep learning group in Toronto, trained in neural networks rather than chemistry, won it.

Up to that point, in the episode's words, "the computer was a sorter". It ranked molecules that already existed in a catalogue. Around 2017 and 2018 the verb changed. Rafael Gómez-Bombarelli, Alán Aspuru-Guzik and colleagues described in ACS Central Science a model that learns a continuous map of chemical space and can propose structures that are on no list. Models such as REINVENT and ReLeaSE then rewarded the generator for molecules closer to a target profile, a technique called reinforcement learning.

Deals followed, such as Sanofi with Exscientia in 2017 for up to €250 million in potential payments, a headline value that includes milestones. In September 2019, Insilico Medicine reported in Nature Biotechnology that its GENTRL system had designed inhibitors of a kinase called DDR1 in 21 days, then tested them in cells and mice. A few months later, W. Patrick Walters and Mark Murcko pointed out in the same journal how closely the lead compound resembled an existing kinase inhibitor.

The first candidates in humans

In January 2020, Sumitomo Dainippon Pharma and Exscientia announced that DSP-1181, a candidate for obsessive-compulsive disorder "created using Artificial Intelligence", had entered a Phase 1 study in Japan. The companies said the exploratory research phase took less than 12 months, against a typical 4.5 years. The compound went no further: trade press reported in 2022 that it had fallen short in Phase 1, and a CAS analysis found that the molecules in the related patent shared their shape with haloperidol, approved in 1967. Exscientia merged with Recursion in a deal announced in August 2024.

In parallel, protein structure prediction and design earned the 2024 Nobel Prize in Chemistry, covered in a separate Tech Anatomy article on AlphaFold.

The episode closes on rentosertib, an Insilico molecule that inhibits a protein called TNIK, in idiopathic pulmonary fibrosis, a disease that progressively stiffens the lungs. In a Phase 2a trial of 71 patients published in Nature Medicine in June 2025, forced vital capacity, the volume of air a patient can exhale, rose by 98.4 mL over 12 weeks in the 60 mg once-daily group and fell by 20.3 mL on placebo. Insilico states that both the target and the molecule came from its AI tools. In July 2026 it announced a 52-week Phase 3 trial of 320 patients in China.

What the success rates show so far

The most cited attempt to measure the effect is a 2024 analysis in Drug Discovery Today by Madura Jayatunga and colleagues at Boston Consulting Group, covering the pipelines of AI-native biotech companies. In Phase 1, which mainly tests safety, they found success rates of 80 to 90 percent, well above historical averages. In Phase 2, the first real test of efficacy, the rate was about 40 percent on a limited sample, in line with historical averages.

Competing readings

The first reading takes the Phase 1 figure at face value, as the authors do: generative chemistry is good at optimising the properties that decide early trials, such as absorption and toxicity. Efficacy depends mostly on whether the chosen target actually drives the disease, and most AI programmes applied design tools to targets chosen conventionally. On this view, rentosertib matters because the company says AI chose the target as well.

The second reading is more cautious about the Phase 1 advantage itself. Early AI companies often worked on validated targets and on chemistry close to known drugs, as the GENTRL and DSP-1181 critiques suggest, and both choices lower early risk regardless of the method. The samples are small, and the first candidates entered trials only in 2020, while a drug typically takes a decade or more to reach approval, so the absence of an approved AI-discovered drug says little either way.

What the record does not establish

"AI-discovered" has no standard definition. Takeda's zasocitinib, a psoriasis drug whose application the FDA accepted under priority review in September 2026, came from a collaboration between Nimbus Therapeutics and Schrödinger built largely on physics-based computational methods. Some coverage calls it AI-designed, some does not. Company timelines, such as 12 months instead of 4.5 years, have not been measured independently. And the rentosertib result is a secondary endpoint in groups of 17 or 18 patients; the trial's primary endpoint was safety.

What to watch

The FDA's decision on zasocitinib is expected in early 2027. The rentosertib Phase 3 trial started recruiting in September 2026, with primary completion estimated for October 2029. The indicator that would separate the two readings above is the Phase 2 success rate as the sample of AI-derived molecules grows: if it stays at the historical average, the gain lies mainly in chemistry and speed; if it rises, target selection is improving too. Neither outcome is visible yet.

Listen to the episode

L'histoire de l'IA dans la découverte de médicaments

Listen · 17 min