Here’s the thing nobody says plainly enough: AI has genuinely transformed the front end of drug discovery. It has not touched the back end. And the back end is where drugs go to die.
Let me explain what I mean.
The front end: where AI actually works
Drug discovery has four jobs that happen before you test anything in humans:
- Figure out which protein in the body is causing the disease (target identification)
- Find a molecule that sticks to that protein (hit finding)
- Make that molecule better — stronger, safer, more stable (lead optimization)
- Predict whether it’ll kill you before you take it (ADMET prediction)
AI is legitimately excellent at all four. This isn’t hype.
Insilico Medicine proved it. They used their AI platform to identify a target for pulmonary fibrosis, design a molecule from scratch, and get it into human trials in 30 months. The industry norm is 4-6 years. That’s a real, published, peer-reviewed result.
AlphaFold — DeepMind’s protein structure prediction AI — mapped 200 million protein structures. Scientists had manually determined about 100,000 over the previous 50 years. That’s a 2000x expansion in structural knowledge.
Generative AI can now invent molecules that have never existed on Earth. Not by searching through nature, but by designing them from scratch, the way an architect designs a building. Diffusion models — the same technology behind DALL-E and Stable Diffusion — have been adapted to generate 3D molecular structures.
These are real capabilities. They work. They’re being used right now by companies like Recursion, Isomorphic Labs, Generate:Biomedicines, Atomwise, and dozens of others.
The back end: where AI can’t help (yet)
Here’s the uncomfortable part.
Once you have a promising molecule, you still have to:
- Test it in cells (does it actually work?)
- Test it in animals (does it kill them?)
- Test it in humans, Phase I (is it safe?)
- Test it in humans, Phase II (does it work?)
- Test it in humans, Phase III (does it work better than what we already have?)
- Get it approved by the FDA (is the data good enough?)
- Manufacture it at scale (can you make a billion pills?)
AI barely touches any of this.
Phase I success rates for AI-designed molecules are actually impressive — 80-90% versus 52% historically. That’s real. But Phase II and Phase III? No improvement. The biology of whether a drug actually works in a human body hasn’t changed because a computer designed the molecule.
As of July 2026, zero AI-discovered drugs have received FDA approval. The most advanced one — Insilico’s rentosertib — just completed Phase IIa. It still needs Phase IIb, Phase III, manufacturing validation, and full regulatory review. That’s years away.
The $7 billion question
Since January 2026, pharmaceutical companies have committed more than $7 billion to AI drug discovery partnerships. Insilico alone signed deals with Servier, Eli Lilly, SK Biopharmaceuticals, and Takeda.
Seven billion dollars. Zero approved drugs.
That’s not a scandal — it’s an investment thesis. These companies are betting that AI-designed molecules will have higher success rates in late-stage trials because they were better designed from the start. That’s a reasonable bet. But it hasn’t been proven yet.
The total AI drug discovery market is about $2.6 billion in 2025. McKinsey estimates generative AI could save pharma $60-110 billion annually across the value chain. Those are projections, not results.
The “black swan” problem
Here’s a critique that nobody in the AI drug discovery world likes to talk about.
Milad Alucozai pointed out something sharp: Lipitor, Gleevec, and Cyclosporine — three of the most important drugs in modern medicine — all violate Lipinski’s Rule of 5, the classic molecular filter that has guided medicinal chemistry for decades.
AI drug discovery is essentially automating those filters at billion-dollar scale.
“Filters eliminate garbage,” Alucozai wrote. “They don’t create gold.”
The argument: breakthrough drugs work because they do something unexpected in a living human system. Train a model on historical data and you get a system that biases toward consensus. It’s constitutionally incapable of producing biological surprises.
That’s a structural limitation, not a bug you can fix with more data.
What’s actually new in 2025-2026
A few things have shifted:
The Recursion-Exscientia merger (August 2024) was the biggest structural event. Recursion brought industrial-scale phenotypic screening — millions of cell images generated weekly. Exscientia brought lead-optimization and a Bristol Myers Squibb partnership. The merger signaled that pure-play AI biotechs need scale to survive.
AlphaFold3 (May 2024) expanded from predicting single protein structures to predicting protein-DNA, protein-RNA, and protein-ligand interactions. That’s the actual drug-binding stuff that matters. But its accuracy drops for novel interactions where training data is sparse.
Diffusion models became the dominant framework for molecular design. RFdiffusion for proteins, DrugGEN for small molecules. Same principle as image generation — start with noise, iteratively refine into a valid structure.
FDA published draft guidance (January 2025) on AI in drug development. They reviewed 300+ AI-containing submissions and introduced a risk-based credibility assessment framework. This is a start, not a framework. Regulatory acceptance of AI-designed drugs is still evolving.
173 AI-originated drug programs are now in clinical development, up from about 24 in late 2023. 15-20 are expected to reach pivotal trials in 2026.
The honest picture
Here’s where we actually are:
The front end of drug discovery — finding targets, designing molecules, predicting properties — has been genuinely transformed by AI. Months instead of years. Millions instead of billions. This is real.
The back end — clinical trials, efficacy in humans, regulatory approval, manufacturing — hasn’t changed. AI doesn’t make biology more predictable. It doesn’t make clinical trials faster. It doesn’t shortcut the FDA.
The industry is funding a massive acceleration of the cheapest, fastest part of drug development while the expensive, slow parts remain unreformed.
That’s not a failure. It’s a sequencing problem. The discovery wins need to translate into clinical wins, and those haven’t arrived yet.
The bottom line
AI is a super-powered magnifying glass and a lightning-fast calculator. It narrows millions of possibilities down to hundreds worth testing. It designs better starting molecules. It predicts problems earlier.
But biology still has the final say. Always will.
The first AI-designed drug to get FDA approval will be a genuine milestone. Projections say 2026-2027. When it happens, it’ll validate the entire approach. Until then, the honest answer is: AI has made the easy part much faster. The hard part is still hard.
Sources: Clinical Trial Vanguard, IntuitionLabs, OncoDaily, PDP Spectra, BioMed Nexus, Nature, Drug Target Review, arXiv, FDA. Full research report in the vault.