Chemists picked 79 structures off a screen to make in a lab. Number 55 was the one they took forward, according to Insilico Medicine. What had begun as a shape you could turn on a computer acquired a name, rentosertib. A few years later, someone with lung disease actually swallowed it. In September 2026, the first participant received a dose in a Chinese phase 3 trial.
What I like about this story is the rather ordinary question underneath it. A chemist has two hands and a finite number of mornings at the bench. Which compound gets made tomorrow? That’s one place where AI for Science, or AI4S, can change which experiments go ahead before anyone takes a dose. A human could have drawn molecule 55. Whether anyone would have got around to making it before the money ran out is another matter.
What is a year and a half worth?
I put two of Insilico’s accounts side by side. From looking for a target to choosing a candidate for preclinical development, the team spent about 18 months on this project, with a reported budget of roughly $2.6 million. The company gives two and a half to four years as a typical duration for traditional early discovery. In the same units, that’s 30 to 48 months against 18. The difference on paper is a year to two and a half years.
A year is a long time for a lab to wait for an answer. Even if the answer is disappointing, learning it a year earlier leaves time to try something else. That’s the prospect I’d bet on before promising an earlier date at the pharmacy. Nobody paid a second team to run this exact project from scratch without AI. We don’t know how long that team would have taken.
Insilico puts that $2.6 million at about a tenth of a traditional discovery budget. It hasn’t published an itemized comparison. The hospital trials come on a separate bill. There is another expense for which researchers have compared two approaches directly.
Imagine taking each molecule in a library, fitting its shape into a little pocket on a protein and paying for the computing time to score it. The work is like trying a huge keyring against a single lock, and every attempt costs money. A research team tried doing the expensive calculation on a small sample first. A model learned from those examples and helped pick which compounds were worth a closer look. When the researchers checked against molecules already confirmed as hits in experiments, they recovered more than 80% of them for about one fourteenth of the computing cost.
For a calculation that used to cost fourteen dollars, think one dollar, about 93% less for that particular bill. The shortcut also missed some good compounds. If missing one is costly enough, I’d still pay the fourteen. But suppose a library was so expensive to explore that nobody was going to look there at all. A dollar for a first pass is a different proposition. This was a separate screening study. Its computing costs and the $2.6 million rentosertib discovery budget belong on different receipts.
What happens before the first volunteer takes a pill?
The 18 months took the team to the end of step three below. The chemists had one molecule on which they were willing to spend the next part of their budget. The search up to that point is drug discovery. The next team has a different question. How much do we know about the risk to the first volunteer who takes this pill? Gathering the evidence for that decision is preclinical development.
- Find a targetAI flags proteins in patient data. Scientists change their activity in cells to see whether it affects the disease.
- Find a hitAI screens molecules or proposes new ones. Chemists test a shortlist. Reproducible activity gives them a hit.
- Turn a lead into a candidateA promising hit becomes a lead, a key that slips into the keyhole but doesn’t turn yet. A model suggests changes, and chemists file it down round by round, testing each version for activity, solubility and how long the drug stays in the body.
- Prepare for human testingA model can flag risks. Researchers check where the drug goes, what it might harm and whether each batch meets the same quality standard.
Are we making aspirin or insulin?
Aspirin is a familiar example of a small-molecule drug. Its active ingredient has a well-defined chemical structure and can be made through chemical reactions. Rentosertib, the molecule you can turn on screen above, belongs to that family. Now think of the human insulin used to treat diabetes. It’s a protein. Manufacturers typically use bacterial or yeast cells to produce it, then purify it. Insulin is one kind of biologic. Therapeutic antibodies are another. A chemist working on a small molecule might swap out a group of atoms. A team designing a protein drug also has to consider how a chain of amino acids will fold, and one wrong bend can cost the drug its effect. An AI tool such as AlphaFold can predict a structure and help the team decide what deserves a closer look. The shape and the drug’s intended effect still need to be checked experimentally.

Where do you start with a scarred lung?
The disease in this story is idiopathic pulmonary fibrosis. Look at it from across the room and you can see hazy markings on a chest X-ray. Look through a microscope and you can see dense scar tissue crowding the fine spaces that used to hold air. A lung has to stretch each time someone draws breath. Scars make that harder.
A pathologist can point to the damage. A chemist needs something to act on. The team gathered fifteen public datasets from blood and lung tissue. Sick cells keep different genes busy from healthy ones. The catch is that a gene can be busy for two quite different reasons. It might be helping cause an injury, or the cell might be trying to repair the damage.
The team used models to compare these changes against known relationships between proteins and findings in the literature. They narrowed the search to a family of proteins that small molecules might be able to reach. A protein called TNIK rose to the top. When they looked at a separate set of data that distinguished individual cells, TNIK was more active in a type involved in scarring. That was a reason to try acting on it. When you see “target” in a drug paper, this is what the researchers hope to affect.
A molecule that binds can still make a poor pill
The computer had a three-dimensional structure for TNIK. With the shape of the keyhole in hand, it could propose small molecules that might fit into it. Chemists picked structures they had a reasonable chance of making and tested them. Some of the early compounds already bound quite tightly. If binding had been the whole assignment, they’d have had a handsome figure for the report.

Then somebody has to swallow the thing. Several early compounds dissolved poorly. Tests with material from human and mouse livers suggested that they would also be broken down too quickly. A pill has to dissolve before its active ingredient can be absorbed, and enough of that ingredient has to last long enough to reach the lung. Those measurements gave the chemists something specific to improve in the next round. With the eventual candidate they could go a step further, to cells derived from a donor with the disease. Under a microscope, a bright green marker grew fainter. You can see the difference in the two images below.

It’s a long way from a dimmer patch of green to easier breathing. I like those images for a more modest reason. A scientist who started with a drawing now had an observation that another lab could check, and a reason to decide whether the next experiment deserved funding.
If it worked in a dish, why look at a mouse lung?
Cells in a dish are in direct contact with the drug. Swallowing a pill asks much more of it. It has to get past the gut and liver, reach the bloodstream and arrive in the lung in sufficient quantity. On the way it can affect other organs. There are two questions here. Does it help with the disease? What else might it harm? The lung sections below can’t tell you whether the drug might injure the liver.

Where do monkeys come in? First you need a species that can answer your question. Some antibodies recognize a human target yet barely interact with its counterpart in a mouse. Observing no harm in that mouse would tell you little about the risk to a person. In some cases researchers consider a nonhuman primate such as a monkey. The drug, the evidence already available and regulatory expectations guide whether animals are needed and which species makes sense. The US FDA is also advancing human cell models, organoids and computational methods where they can reduce unnecessary animal testing.
Here’s a job for AI at that point. Suppose a team has five related compounds and a model flags possible liver injury for two of them. That gives the team a reason to move a liver-cell test up the list. A false alarm is worth recording, too. Before a first-in-human study, researchers have to bring together evidence from testing, information on manufacturing quality, a rationale for the starting dose and a plan for monitoring people. Regulatory and ethics review, and the volunteer’s informed consent, come before dosing.
The AI counted 77. Would you take that list to the lab?
At Revvity Signals, we discussed giving the AI a tool of its own. But the scientist already had an experiment open. For one question about which compounds deserved another test, would we ask them to open another window, move a file and explain the experiment all over again? I argued for putting the conversation beside the existing record. The file was right there.

Later, I checked the design using a public file from a different project. It contained 94 activity records for a protein called CDK2. I wanted compounds that could halve the measured activity at a low concentration, the value labelled IC₅₀ in the file. At a cutoff of 10 nM, the assistant returned 43 records. I checked every identifier and value. All 43 matched.
I raised the cutoff to 100 nM. The assistant’s summary now said 77 compounds. On the very same screen, the three bars below that cutoff held these counts.
1 + 42 + 35 = ?
See why I asked for a recount
78 records. I asked the assistant to check the original file and it revised its answer to 78. After getting all 43 records right on the previous question, it had given me every reason to copy down 77.
That is why I gave the chart and the source table space on the screen. On a good day, the assistant goes through dozens of records and hands over a shortlist. When the numbers disagree, the scientist can open the cited file and settle the matter at the row in question. That saves a file transfer and keeps a miscount out of the notes that go to the bench tomorrow.
Some things take a full year
- Phase 1A small group helps establish tolerability and how long the drug stays in the body. One rentosertib study enrolled 78 healthy volunteers. Some cancer trials enroll patients from the start.
- Phase 2In people with the disease, is there an early sign of benefit? What dose makes sense? Safety monitoring continues
- Phase 3A larger trial compares benefits and risks against a control group. For rentosertib, that means following patients for 52 weeks.
There is time to save in the hospital, too. During recruitment, clinicians check medical records against trial criteria. In a small NIH study with two clinicians, highlighting the relevant passages with AI cut time on that task by about 40%. A patient’s twelve weeks of follow-up still takes twelve weeks.
In the phase 2a trial, 71 people with pulmonary fibrosis were randomly assigned to four groups. They received different dosing schedules or placebo for twelve weeks. At a lung function test, you fill your lungs and blow into the mouthpiece until you have nothing left. The volume of air that comes out is called forced vital capacity. It takes much longer to explain the name than to understand the movement.

At twelve weeks, the group taking 60 mg once daily could blow out an average of 98.4 mL more than at the start. The placebo group averaged 20.3 mL less. What I wanted to know was whether someone could walk to the shops without stopping to rest. The researchers did measure how far people could walk in six minutes. Changes in walking distance were similar across groups. It was a small, short trial. That question is still worth asking.

The main question here was safety. Sixteen of the 71 participants stopped treatment early. Reasons for stopping included liver problems and diarrhoea. The phase 3 trial is looking at how quickly lung capacity falls over fifty-two weeks. You might gain a year choosing a molecule at the start and still have a good reason to wait every week of that year at the other end.
After phase 3, the team may be in a position to apply for marketing approval. Regulators weigh the benefits and risks for the intended patients and review manufacturing quality. If the drug is approved, rare harms may only become apparent as more people take it for longer. Reports from patients and clinicians keep coming in, sometimes alongside dedicated phase 4 studies.
I keep thinking about the experiments that never get funded. Molecule 55 made it to a hospital. Someone else’s number 55 may still be a drawing on a chemist’s computer. If that chemist opens next year’s budget and finds they can afford to make it after all, I’d count that as a pretty good day.
Research sources and image credits
Research and data
- Discovery and phase 1 · The 71-person phase 2 study · Phase 3 registration · First phase 3 dose
- The 55th molecule · Discovery time and budget · Comparison timeline · Sponsor’s cost estimate · Independent computing-cost study
- NIH on lead optimization · Biologics · AlphaFold
- First-in-human studies · ICH on animal studies · FDA alternatives · Clinical phases and postmarket monitoring
- ChEMBL · CDK2 · NIH · TrialGPT
Insilico Medicine funded the clinical trial. Its project account lists 79 molecules selected for synthesis. A later table counts 78 synthesized and tested.
Image credits
- Laboratory photograph, NIH / NCATS; spirometry demonstration, Veronica Burkel / CDC
- Molecular drawing and coordinates for the animation from PubChem 164938183 (hydrogens omitted)
- Insulin illustration, David S. Goodsell / RCSB PDB · CC BY 4.0
- IPFeditor · Chest X-ray · Lung histology · CC BY-SA 3.0
- Crops from Ren et al., target ranking (1b) · cell map (Extended Data 2d) · binding pose (2a) · patient cells (Extended Data 4a) · mouse lung control and highest dose (3c) · CC BY 4.0
- Lung function chart redrawn from Xu et al., Fig. 2a; product screenshots from the author’s Signals design case.
Phase 2 changes in lung function
| Group | Mean (mL) | 95% CI (mL) |
|---|---|---|
| Placebo | −20.3 | −116.1 ~ 75.6 |
| 30 mg once daily | −27.0 | −88.8 ~ 34.8 |
| 30 mg twice daily | +19.7 | −60.5 ~ 99.9 |
| 60 mg once daily | +98.4 | 10.9 ~ 185.9 |







