AI Protein Design Draws Pharma Deals and New Enzymes, October 2026

A month when design overtook prediction

In mid-October 2026, AI drug discovery news shifted away from predicting how proteins fold and toward designing new proteins and selling those designs to drugmakers. Within about a week, a Danish obesity specialist licensed an AI protein-design platform, GSK widened its use of a two-year-old design startup, and David Baker's institute published AI-designed enzymes that cut disease-linked proteins. AstraZeneca also signed a large target-discovery deal. Reporting on all of these points the same way: the field's money and attention are moving toward generative design. Coverage is also consistent that none of this work has yet shown a clinical result.

Pharmaphorum's October roundup puts AstraZeneca's alliance with insitro at the top. The same roundup flags news from Iambic, DaltonTx, NVIDIA and Chai Discovery.2 For readers following protein structure prediction and AI protein design, though, the more revealing stories sit just below that headline.

Generate and Zealand: protein design meets the obesity race

On October 13, Zealand Pharma and Generate Biomedicines announced a collaboration and license agreement. The two companies plan to design multispecific protein therapeutics for obesity and other metabolic diseases.32 Generate gets $30 million upfront and up to $35 million in early preclinical milestones. Total milestones reach nearly $1.2 billion, plus tiered royalties. The deal covers two initial programs, with an option to add a third.32 Generate will run research through lead-candidate selection. Zealand then takes over clinical development, regulatory filings and worldwide commercialization.34

Coverage agrees on the terms but stresses different things. Reuters, as syndicated, framed the pact as part of the race to find obesity drugs beyond the GLP-1 blockbusters from Novo Nordisk and Eli Lilly. It also noted that Generate's shares rose about 5% after hours.40 BioSpace gave it a more skeptical frame. It placed the deal right after a Phase 2 readout for Zealand's Roche-partnered amylin analog petrelintide that analysts considered underwhelming. Patients lost up to 9.2% of body weight against 2% on placebo, and analysts found the blood-sugar results lacking.35 Zealand's own announcement had called those results positive.32

The reasonable reading is that this is a diversification bet, not a sign of distress. Zealand's science chief described proteins as an addition to the company's small-molecule and peptide toolbox, opening routes to polypharmacology.32 Generate keeps adding partners. It already has collaborations with Novartis and Amgen, and its most advanced AI-designed asset, the anti-TSLP antibody GB-0895, is in Phase 3 for severe asthma.3539 Neither company has disclosed targets or mechanisms, so the scientific content of the programs is still unknown.37

Chai Discovery: a wet-lab test before the contract

Chai Discovery's deal with GSK, announced October 9, may be the clearest sign of how large drugmakers now buy AI design tools. GSK designed the test itself. Chai's models produced designs against GSK targets zero-shot, meaning without training on those specific targets. GSK's own wet-lab testing then found binders to every target it tested, and the evaluation led to a broader collaboration.44 GSK's Christopher Austin said Chai's technology complements internal models that GSK keeps retraining on its own lab data.44

Endpoints counted this as Chai's sixth pharma deal in nine months, with GSK joining Eli Lilly, Pfizer, Novartis, argenx and Bristol Myers Squibb. Endpoints put Chai's valuation at $3.8 billion.42 The Pharma Letter traces that valuation to a $400 million Series C in July. It also reports that Chai-3, first disclosed with the Pfizer license in June, doubles the antibody-design success rate of Chai-2.41

The more careful coverage points to what was left out. Neither company disclosed how many targets were tested, how many designs were screened, what share bound, or any financial terms.48 One analysis makes a sharp point about the missing denominator. A binder for every target means little if each target needed hundreds of designs, and it means a lot if each needed only a few.46 That critique is fair. Still, the strongest evidence here is GSK's behavior: a drugmaker with its own models ran its own experiments and then expanded the relationship. The phrase "complements our own internal models" also suggests big pharma is building a mix of in-house and outside tools rather than handing design over to one vendor.

Baker lab's molecular scissors

The biggest scientific news came from UW Medicine's Institute for Protein Design. Two Nature papers described proteases, enzymes that cut proteins at chosen sites, designed largely from scratch with AI.1114 The workflow combined several tools. A variant of RFdiffusion2 generated backbones, ProteinMPNN and a related tool assigned amino-acid sequences, and AlphaFold 3 checked whether the predicted structures matched the intended designs.15 Ars Technica reports that the designs resembled nothing in existing structure databases, which shows the tools were not simply copying known proteins.11

The headline numbers differ somewhat from outlet to outlet, mostly because they measure different steps:

• Cysteine proteases: 13 of 69 designs were active, with rate gains of up to about 30 million-fold.1520
• Zinc proteases: more than a third of 135 designs cut their targets. GeekWire cites a speed-up of 100 million-fold.1415
• After optimization: Ars Technica gives the best optimized enzyme as 10^9.11 Another account says four amino-acid substitutions pushed the rate gain past 10 billion-fold.20 Phys.org describes a 30-fold boost to the top zinc enzyme after refinement.15

All accounts agree on what matters for drug discovery. The enzymes cut sequences in amyloid-beta and TDP-43, proteins linked to Alzheimer's disease and ALS, and they worked inside mammalian cells.20 Because an enzyme can cut many target molecules one after another instead of holding onto just one, a researcher on the project said such drugs might work at very low concentrations.14

The researchers were also open about limits. Retargeting an existing design to a new sequence did not work well. In that case they had to start the design process over.11 None of these enzymes has been tested in people.15

Prompteina: protein design in plain English

Also from the University of Washington, Prompteina has been described as the first model that generates all-atom protein structures from plain-English design goals. The work was posted to arXiv on October 9 and accepted to NeurIPS 2026.13 It builds on Proteina-Complexa, a binder-design framework from NVIDIA and Baker Lab researchers that reported a 91% wet-lab hit rate across a 133-target screen.13 In lab tests, nine Prompteina-designed binders attached to GM2A, a target linked to Tay-Sachs disease for which no designed binders had been reported. The best had an affinity of 107 nanomolar.13 Code, data and weights were released publicly.13 The method is still unproven beyond that one target. But it points to a future in which the main design step becomes writing a precise specification.

Structure prediction becomes plumbing

Structure prediction did not disappear this month. It became background infrastructure. The AlphaFold Protein Structure Database listed more than 261 million predicted models as of September 24.23 A September release added predicted complexes for more than 2,800 viruses to support pandemic preparedness.29 Benchling launched Model Hub, which runs AlphaFold 2, Chai-1, Boltz-2, OpenFold 3 and Protenix side by side and plans to add generative design models.30 When competing folding engines are offered as a menu inside lab software, the competition has moved somewhere else.

Two recent studies show why caution still matters. A preprint benchmarking Boltz-2 on 317 protein-ligand pairs found no measurable link between the model's self-reported confidence and its actual error in predicting binding affinity.24 A systematic review of nine major structure models found that eight relied mainly on benchmark testing. None reported prospective wet-lab validation.25 These findings help explain why GSK tested Chai's designs at the bench and why the Baker papers lean on experimental data.

Where the money is going, and what it doesn't prove

The AstraZeneca–insitro alliance shows the target-discovery side of the same trend. It includes $40 million upfront and about $1.46 billion in milestones for work in MASH and cardiomyopathies.3 Most outlets put the total near $1.5 billion. One cites Bloomberg at up to $1.86 billion, a gap the published terms do not explain.48 Danaher, meanwhile, plans an autonomous AI lab at Abcam, expected to run at scale in early 2027. The lab will design antibodies and other reagents, test them, and feed the results back into its models.1

The main lesson of October is that the design-and-test loop has become the product. Generate, Chai, Danaher and the Baker lab all describe some version of a model generating designs, a lab testing them, and the data flowing back into the model.32441

The main warning comes from Insilico Medicine's Alex Aliper. He compares AI to driving fast on a highway, where you still have to slow down for downtown traffic once you reach clinical trials.7 Every deal described here pays out mostly on milestones that have not yet been met. AI protein design now clearly works in the lab. Whether it works in patients is still unknown.