Did AI Design Takeda's New Drug?

More than 35% of patients on Takeda’s once-daily psoriasis pill had completely clear skin after sixteen weeks.

That is a big deal.

The drug is zasocitinib. The trial was LATITUDE Atlas, a 606-patient Phase 3 study that put it head-to-head against deucravacitinib, the Bristol Myers Squibb drug sold as Sotyktu. Takeda says zasocitinib beat Sotyktu on the primary endpoint, PASI 100 (completely clear skin with 100% resolution), by more than two and a half times. It also won on the key secondary endpoints.

So yes, the drug works.

Actually, that undersells it. Sotyktu was not some straw man comparator wheeled onto the field to get tackled. It is an approved oral TYK2 drug, used at its approved dose, doing roughly what it did in its own pivotal studies. Zasocitinib beat the real thing on complete skin clearance. In a real head-to-head trial, so that matters!

But the AI framing is also why it annoys me so much.

When the data came out yesterday, BioSpace called it Takeda’s “AI-designed pill.” Benzinga called it an AI drug. That makes for a good headline, and in 2026, I understand why everyone reaches for it. It’s the hot go-to for all of marketing… in anything nowadays, really.

But Takeda did not call it that. Nimbus did not call it that. Schrodinger did not call it that.

The companies closest to this molecule describe something more specific and, frankly, more impressive than the lazy AI label. Nimbus and Schrodinger started the TYK2 program around 2016 and used free energy perturbation, or FEP, to solve a very hard medicinal chemistry problem. TYK2 sits in a family with JAK1, JAK2, and JAK3, which look annoyingly similar from a drug designer’s point of view. The job was to hit TYK2 and spare the others.

FEP is not magic. It is not a robot hallucinating a drug out of the ether. It is physics-heavy computational chemistry used very well.

Schrodinger has called the program the first large-scale use of FEP in drug discovery. Its CEO, Ramy Farid, has also spent the last couple of years warning reporters not to treat AI like pixie dust. His basic view is that the physics is the meal and the machine learning is somewhere closer to garnish.

And in this case, I think he is right.

This is where the story gets a bit uncomfortable for the AI drug discovery crowd. Zasocitinib is a very good drug candidate. It may even become the best oral psoriasis drug in its class. But it did not discover TYK2. Bristol Myers Squibb already dragged that target through the clinic and got Sotyktu approved in 2022. Nimbus came later and built a better molecule against a known, clinically validated target.

That is still valuable. Maybe extremely valuable. Takeda paid $4 billion upfront for Nimbus’s TYK2 subsidiary in 2023, and right now that decision looks a lot less insane than it did that day. The molecule has now gone from Phase 2b, to pivotal Phase 3, to a head-to-head win against the first drug in its class.

The model did something useful. It helped design a molecule good enough to beat the incumbent. That should count for something.

But it should count as a win for computational drug design, not as proof that AI is discovering medicines no human would have found. Those are different claims. One is supported by this data. The other is just headline-grabbing.

And look, I get why people blur the line. “Physics-based free energy perturbation helped optimize an allosteric TYK2 inhibitor against a clinically validated target” is a terrible headline. But an “AI-designed pill beats Sotyktu” is a great one. The problem is that the second version changes the perceived achievement.

It turns a strong medicinal chemistry story into a vague AI victory of sorts.

The full clinical picture is also not public yet. Takeda gave topline results, not the detailed tables. We know zasocitinib won on PASI 100, the hardest and cleanest clearance endpoint. We know it won on key secondaries. We do not yet know the full PASI 75 and PASI 90 breakdowns, the exact Sotyktu number in this trial, the confidence intervals, or whether the practical gap looks as large outside the most stringent endpoint.

That matters because PASI 100 is where Sotyktu has always looked weakest. Its own pivotal trials put the complete clearance at around 10% to 14%. Zasocitinib clearing more than 35% of patients is a real win, but the commercial fight may not be decided on that number alone.

Dermatologists will care about how many patients get meaningfully better, how fast, how durable, how clean the safety looks, and how payers treat another oral option in a market full of very effective biologics.

So my read is pretty simple.

This is a good drug story. It is a good Takeda story. It is a good Nimbus and Schrodinger story. It is probably one of the clearer examples we have of computational chemistry producing a clinically differentiated small molecule.

But it is not the clean AI-discovered drug story the headlines want it to be.

The drug is real. The success so far is real. The AI label is doing more work than it deserves.

Anthropic says its new flagship model accelerated parts of the drug design process by roughly tenfold. The company launched Claude Fable 5, its first publicly available model built on the top-tier Mythos 5 system, and reported that the model produced promising candidates for nine of 14 protein targets while working without human assistance. Anthropic also says one model-generated hypothesis, a potential new antimicrobial target in E. coli, has already been validated in lab testing. [Link]

Frost & Sullivan says AI is moving from experimentation into governed, audit-ready workflows across drug discovery, clinical operations, and manufacturing. The firm's 2026 pharma outlook lists AI-enabled operations as one of five forces reshaping the industry, alongside pricing pressure, patent expiries, supply resilience, and a continued shift toward biologics like GLP-1s, antibody-drug conjugates, and cell and gene therapies. [Link]

OpenFold, the open-source alternative to DeepMind's AlphaFold, added eleven new members including Daiichi Sankyo, Flagship Pioneering, and Benchling. The nonprofit consortium builds freely available AI models for protein structure prediction and drug discovery, and the new members span antibody design, GPCR-based drug design, lab software, and molecular visualization. [Link]