01By the numbers — where AI drug discovery is now
"Phase" here means the stage at which a drug is tested in people. Phase I uses a small number of people and mainly checks safety. Phase IIa adds more people and starts to check whether the drug works. Phase III uses a large number of people to confirm whether it truly works. The higher the number, the later the stage, and the more patients are involved.
The number to watch is the 80-90% pass rate through that first Phase I. In older drug-making, only about half (around 52%) made it through Phase I. AI has sharpened how candidates are picked, so from the very start only "the ones with a good chance of working in people" are chosen.
02Insilico Medicine — the leading edge of generative-AI drug design
The most symbolic advance of the past year is Insilico Medicine's rentosertib.
- Disease it targets: idiopathic pulmonary fibrosis (IPF) — a hard-to-treat disease of unknown cause in which lung tissue gradually stiffens, making it harder to breathe
- What it aims at (the target): a protein called TNIK. The body part a drug is designed to home in on is called its "target"; TNIK is a kind of enzyme thought to be involved in how the disease progresses
- How it was made: designed with the company's "Chemistry42" — a system in which AI dreams up drug candidates
- Progress: finished Phase IIa (a mid-sized human test) in early 2026, showing signals that it looks safe and looks effective
- Why it matters: the first drug designed from scratch by AI to pass Phase IIa
- Next stage: getting ready for the large Phase III test that confirms the result
This is not just a story of "one drug succeeding." It is the first public success of a new way of making drugs, in which AI finds what the drug should aim at (the target), designs the starting material that fits that target (the lead compound), and selects the candidate to test in people — while human researchers focus on the later confirmation work.
03Big pharma activity — partnerships and consolidation
2025-2026 was also the year big pharmaceutical companies switched from "watching from the sidelines" to "using AI for real." The main partnerships and acquisitions are listed below.
Pfizer × Boltz
A partnership built on a large AI model. Boltz's AI for predicting the three-dimensional shape of proteins was built into Pfizer's drug-making workflow. Once you know a protein's shape, it becomes easier to design a drug that fits it.
Bayer × Cradle
Adopting Cradle's AI for designing proteins. It becomes easier to build proteins that do a chosen job, which widens the search for candidate drugs that use antibodies (proteins the body makes to grab foreign matter).
Eli Lilly × NVIDIA
A tie-up to build a large, drug-discovery-dedicated computing setup (an "AI factory") together. They lined up many of NVIDIA's high-performance computing chips (H100/H200 GPUs) to create the foundation for running AI at speed.
Recursion × Exscientia (acquisition)
In 2025, Recursion acquired Exscientia. Two companies that search for drugs with AI became one, combining the power to design drug candidates with AI and the power of "phenotypic screening" — trying candidates on cells one after another and spotting promising ones from changes in their appearance or behavior. It is a sign of the industry reorganizing.
04Technical breakthrough — AlphaFold3 / OpenFold3
AlphaFold3 (built by Google DeepMind / Isomorphic Labs) and its open version, OpenFold3 (open-source, meaning anyone can use it), were announced in late 2024 and came into real use in research from 2025. These are tools that use AI to predict the three-dimensional shape a molecule, such as a protein, will fold into.
- Predicting how well a drug sticks to its target: it can calculate how tightly a drug candidate (a ligand — a small molecule that latches onto the target) binds to the protein it aims at, about 1,000× faster than older methods
- More targets become reachable: proteins whose shape used to be hard to determine can now have that shape predicted by AI and be used as drug targets
- The range of treatable targets widens: targets once written off as "impossible to make a drug for" can now be attempted
- RNA and DNA too: AlphaFold3 can now handle biological molecules beyond proteins, broadening where it can be used
As a result, AI has reached a stage where it can speed up the whole chain — finding the target → refining the candidate → designing the human test — as one connected flow.
05Open issues — Phase III data and regulatory approval
| Issue | Current state | Outlook |
|---|---|---|
| Phase III data scarcity | Many Phase IIa successes; few Phase III completions | Multiple Phase III readouts expected 2026-2027 |
| First "fully AI" approval | No clear FDA-approved drug explicitly tagged as AI-derived yet | First case projected 2026-2027 |
| Regulatory guidelines | FDA, EMA, PMDA developing AI-derived-drug evaluation frameworks | Country guidelines finalized 2026-2027 |
| Data bias | Training-data bias may produce effect differences for some patient groups | Diversity-dataset construction in progress |
| IP / patents | Legal debate over patentability of AI-designed molecules | Case law accumulating at USPTO and elsewhere |
Industry forecasts expect more than 200 AI-related approvals between 2025 and 2030, and see this as something that will reshape pharma research and development from the ground up.
06Structural shift — pharma R&D paradigm redefined
The progress of AI drug discovery is changing the very structure of how pharmaceutical companies do research and development.
- The researcher's job changes: AI takes over compound design, and researchers concentrate on "interpreting and checking the candidates AI puts out"
- More targets within reach: the range of what can be made into a drug widens, making hard-to-treat and rare diseases realistic to tackle
- The partnership model changes: tie-ups with AI startups move to the center of R&D strategy
- The competitive picture changes: the gap widens between firms that have an AI drug-discovery platform and those that do not
- How human trials are run changes: AI starts to help with designing the trial, choosing which patients take part, and analyzing the yardsticks used to measure good or bad results (endpoints)
07Connections to other trends
AI drug discovery connects to several other trends.
- AI clinical practice in China and Middle East — the importance of these regions as trial venues grows; AI-discovered candidates clinically accelerate there
- AI reshaping the social structure — redefining the researcher role affects pharma employment structure
- Novartis CEO joining Anthropic's Board (separate piece) — a symbol of strategic fusion between safety-focused AI and pharma
The past year was a turning point for the industry: AI drug discovery moved from an "experimental stage" to a "clinical validation stage." The success of Insilico's rentosertib in Phase IIa is the first publicly disclosed case showing that a molecule designed by generative AI actually works in the human body.
For the pharmaceutical industry, AI is no longer just a "tool that speeds up the work." It is a "change that rebuilds how drugs are made in the first place." The old way of making drugs — ten years, billions of dollars, and frequent failure — has its time and cost cut sharply by AI. The past year says that era has truly arrived.
The next five years decide each company's structural position in the industry. Investment in AI drug discovery, partnerships, building an in-house setup — whether these decisions get made now is what sets the view five years from here.
Sources
- HUSPI, "AI Drug Discovery State of the Art 2026"
- Drug Target Review, "AI in Drug Discovery: 2025 Annual Report"
- Insilico Medicine press release (rentosertib Phase IIa)
- Company press releases (Pfizer-Boltz, Bayer-Cradle, Eli Lilly-NVIDIA)
- Google DeepMind / Isomorphic Labs (AlphaFold3 launch 2024)