AI drug discovery moves from prediction to a learning loop
TechBio 3.0
A proposed system that links AI design with repeated laboratory testing.
multimodal AI
AI that combines several kinds of biological and chemical information.
generative chemistry
Methods that use AI to design new drug molecules.
What happened
A new review in Nature Portfolio's npj Drug Discovery, a scientific journal, describes a proposed framework called “TechBio 3.0.” It is not one new product. It connects multimodal AI, generative chemistry, and automated laboratory experiments. The intended loop is simple: predict, design, test, and learn.
Why drug discovery needs a new loop
Developing a medicine is slow and expensive. The review says the average cost now exceeds $2 billion. The journey from finding a target to approval can take more than 12 years. Almost 90% of candidates entering clinical trials never win approval.
The problem is biological complexity. One molecule can hit its intended target. It can also affect other proteins, cells, or organs. It may change how the body absorbs or removes medicines. A model that sees only molecular shape can miss these effects.
From prediction to connected experiments
The review describes three stages. TechBio 1.0, especially from 2015 to 2019, used machine learning to predict properties such as binding, solubility, and early toxicity. TechBio 2.0, from about 2020 to 2023, added generative models that could design new molecules and speed candidate selection.
TechBio 3.0 connects those tools to the laboratory. Multimodal AI combines chemical structures, protein interactions, cell responses, and other biological data. Generative models can include safety limits while designing molecules. Automated labs then make and test selected compounds. Their results return to the models. The system can choose the next experiment and flag uncertainty.
What evidence exists
The review cites Insilico Medicine, an AI drug-discovery company. It advanced a lung-disease program from target identification to a preclinical candidate in about 18 months. The review compares that with a traditional estimate of three to five years.
It also discusses Rentosertib, an AI-designed drug candidate. In a phase IIa trial for idiopathic pulmonary fibrosis, 71 patients received treatment. The highest-dose group showed a mean forced-vital-capacity improvement of 98.4 milliliters. That exceeded an approximately 80-milliliter threshold used for meaningful improvement. No treatment-related serious adverse events were reported in that group.
These results are encouraging, but they are not approval. The review also describes Isomorphic Labs, a DeepMind spin-out, using AlphaFold-based structure prediction with generative chemistry.
What remains unknown
The biggest question is translation. Can better predictions create safer, effective medicines for many diseases? AI can generate molecules far outside its training data. It may misjudge uncertainty. A molecule must also be chemically makeable. Lab capacity, robot errors, and test reproducibility can limit the loop.
The review states that, as of mid-2026, no AI-designed small-molecule drug had FDA marketing approval. The FDA is the U.S. medicine regulator. A phase III program for Rentosertib was expected in the second half of 2026, but the source does not provide its result.
What to watch next
The important test is not whether AI can draw a molecule. It is whether the full loop repeatedly produces medicines that help patients. Future evidence should include clinical outcomes, independent validation, reproducible experiments, and clear explanations for regulators.
AI drug research works best when labs teach it
📰 Full story: AI drug discovery moves from prediction to a learning loop
A new scientific review explains why AI needs repeated lab tests.
TechBio 3.0
A plan for connecting AI ideas with repeated lab tests.
multimodal AI
AI that reads several kinds of information together.
Rentosertib
A medicine candidate designed with AI and tested in people.
💡 The gist
- AI can suggest new medicine ideas.
- Lab tests can teach the AI what happened.
- Faster research does not guarantee safe medicine.
A new review describes TechBio 3.0 (a plan for connected drug research). It links computer predictions, new molecule designs, and laboratory tests.
Older AI systems often did one job. They predicted a molecule's properties. Later systems could design new molecules. But the computer and the laboratory were not always connected.
In TechBio 3.0, the steps form a loop. AI suggests a medicine candidate. Scientists or machines make it. Tests check how it works and whether it may be harmful. The results go back to the AI. Then the AI chooses a better next idea.
Multimodal AI (AI that reads several kinds of information) can see more clues. It can study molecule shapes, protein connections, and cell changes. This matters because one molecule may help one body part. It may also affect other parts. Looking at more clues may reveal risks sooner.
The review describes Insilico Medicine (a company that uses AI to study medicines). One lung-disease program reached a candidate for testing before human trials in about 18 months. A traditional estimate was three to five years.
The review also discusses Rentosertib (a medicine candidate designed with AI). It reached a study in people with a lung-scarring disease. In the study, 71 people received treatment. The highest dose improved a breathing measure by an average of 98.4 milliliters. The review says this passed an approximate 80-milliliter meaningful-change mark. No treatment-related serious side effects were reported in that group.
These facts show progress. They do not prove that AI can finish medicine research alone. By mid-2026, no small-molecule medicine designed this way had approval from the U.S. FDA, America’s medicine regulator. Researchers still need more tests, larger studies, and longer follow-up.
A molecule must also be possible to make. Lab machines can make mistakes. Tests may not repeat the same result. AI can be unsure when a molecule looks unlike its training examples.
The next question is simple. Can this loop repeatedly create medicines that help patients? Watch clinical results, independent checks, and clear safety evidence.
A computer and a lab learn about medicines
📰 Full story: AI drug discovery moves from prediction to a learning loop
A computer suggests ideas, and tests check them.
TechBio 3.0
A way for computers and labs to learn together.
AI
A computer helper that suggests ideas.
Rentosertib
A possible medicine designed with AI.
TechBio 3.0 (a way for computers and labs to learn together) is a new idea.
AI (a computer helper) suggests a medicine candidate. A candidate is something that might become medicine. People and machines test it. They check if it helps. They also check if it can hurt.
The test answers go back to the AI. The AI uses them for its next idea. This keeps happening. It is like a learning circle.
Rentosertib (a medicine candidate designed with AI) reached testing in people. That is a hopeful step. It is not a medicine people may buy yet. Scientists still check safety and help.
The lab must make the medicine too. The tests must give honest answers. So, fast computer ideas are not enough. People still need to check the work.