Industries / Health / Pharmaceuticals
Drug discovery
Finding a molecule that acts on a cause of disease and could become a medicine.
6
Milestones achieved, from 1910 to 2020.
2
To watch for: announced, not yet confirmed outside.
1
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1910
Screening compounds by number
Ehrlich's chemists made arsenic compounds one after another and tested each against the germ that causes syphilis. The 606th preparation, arsphenamine, went on sale as Salvarsan in 1910. Two assistants had first judged it useless; Sahachiro Hata repeated their tests and found it better than every other drug tried.
LessonA poor test method can bury a working compound, and a careful repeat can find it.
Credit
Paul Ehrlich, Alfred Bertheim (who first made the compound, in 1907) and Sahachiro Hata
1928
A mould that kills bacteria
Alexander Fleming found in 1928 that a Penicillium mould had antibiotic properties, and Howard Florey and Ernst Chain carried the work further. The three shared the 1945 Nobel Prize "for the discovery of penicillin and its curative effect in various infectious diseases".
Credit
Alexander Fleming, Howard Florey, Ernst Chain
1964
Drugs designed from biology
James Black set out to block the beta receptor, a protein on heart cells that a stress hormone docks onto, so as to cut the heart's demand for oxygen. The result was propranolol in 1964, and the same method gave cimetidine for stomach ulcers in 1975. Gertrude Elion and George Hitchings designed drugs from differences in how normal cells, cancer cells and germs build nucleic acids, the molecules of DNA and RNA, including 6-mercaptopurine for leukaemia (1951) and acyclovir for herpes (1977). The Nobel committee wrote that drug development had earlier rested mainly on chemically modifying natural products, and that these three introduced "a more rational approach".
Credit
James Black, Gertrude Elion, George Hitchings (Nobel Prize 1988)
1997
The rule of five
C. A. Lipinski and colleagues found that a molecule is more likely to be poorly absorbed when it has more than 5 hydrogen-bond donors, more than 10 hydrogen-bond acceptors, a molecular weight over 500, or a calculated fat-solubility score (cLogP) over 5. They also found that leads from high-throughput screening, the robotic testing of large compound libraries, tended to be larger, more fat-soluble and less water-soluble than leads found before it.
LessonTesting more compounds faster pulled chemists toward molecules the body absorbs poorly, and a simple filter helped pull them back.
Credit
C. A. Lipinski, F. Lombardo, B. W. Dominy, P. J. Feeney
2012
Fewer drugs per dollar
J. W. Scannell and colleagues showed that the number of new drugs approved per billion US dollars spent on research and development had halved roughly every 9 years since 1950, a fall of about 80-fold after inflation. This happened while the science, technology and management behind each step improved.
LessonBetter tools at each step did not make the whole process more productive.
Credit
J. W. Scannell, A. Blanckley, H. Boldon, B. Warrington
2020
Protein structures predicted
Demis Hassabis and John Jumper presented AlphaFold2, an AI model that predicts the 3D shape of a protein. By the Nobel committee's count it has predicted the structure of virtually all 200 million known proteins and has been used by more than 2 million people in 190 countries. A protein's shape guides the design of molecules that bind to it.
Credit
Demis Hassabis and John Jumper (Nobel Prize in Chemistry 2024, shared with David Baker for protein design)
Watch for
An AI-found drug in patients
Insilico Medicine says generative AI found both the target, a protein called TNIK, and the molecule, rentosertib, for idiopathic pulmonary fibrosis, a scarring lung disease. In a Phase 2a trial of 71 patients over 12 weeks, side-effect rates were similar to placebo, and the highest-dose group's lung capacity (forced vital capacity) rose by a mean of 98.4 ml against a fall of 20.3 ml on placebo.
Achieved whena larger, longer trial meets its efficacy goal, or a regulator approves the drug.
Claimed by
the trial's sponsor, Insilico Medicine, and co-authors, in Nature Medicine, August 2025 issue
Antibody designs in a month
Anthropic reports that its model Claude, working with the biotech company Divergence Labs, turned the company's goals into a design plan and generated about 35,000 antibody designs in about a month. In early lab tests, 3 designs bound IgE, the molecule behind allergic reactions, and 20 could pull it off immune cells; 3 of those appeared to do so better than the standard medicine.
Achieved whenan outside lab repeats the binding results, or a drug built from the designs enters a registered trial.
Claimed by
Anthropic and Divergence Labs, October 2026
Open
01
A ledger of AI-found drugs
A public table of every drug candidate that a company has called AI-discovered or AI-designed. Each row links to the candidate's trial registry entries, its phase and status, any posted results, and, where the structure is public, the most similar molecule known before the programme began. The table refreshes itself from the registries every week.
The briefCredit · unclaimed
Open 01 · the brief
A ledger of AI-found drugs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of every drug candidate that a company has called AI-discovered or AI-designed. Each row links to the candidate's trial registry entries, its phase and status, any posted results, and, where the structure is public, the most similar molecule known before the programme began. The table refreshes itself from the registries every week.
Why it is still open
The most-quoted success rates for AI-found drugs come from a 2024 analysis of the pipelines of AI-native biotech companies. It reported 80–90% success in Phase I and about 40% in Phase II "on a limited sample size". It was a one-time count, and its result depends on which companies count as AI-native.
Built with what exists
the ClinicalTrials.gov API (version 2); ChEMBL and PubChem for structures; RDKit for similarity scores; a language model such as Claude or Gemini to read company pipeline pages and press releases, with a person checking every row it adds. The advance that makes this a small-team job is language models that can read hundreds of web pages and fill a fixed table, citing the page for each value.
First step
Write the rule for what counts as AI-discovered. Then enter 20 candidates, starting with rentosertib (NCT05938920), and publish the table.
Done looks like
At least 100 candidates, each with a registry ID or a stated reason for none. A second person checks 30 random rows and agrees with at least 29. Phase I and Phase II success rates are recomputed with the counting rule published beside them, and the table updates weekly for three months.