Hew

AI just learned to do what evolution took billions of years to perfect — and aimed it at Alzheimer’s and ALS proteins. Two teams at UW’s Institute for Protein Design published in Nature (Oct 14) enzymes designed from scratch with AI that cut protein backbones at precisely chosen sequences. One team built zinc-based metalloproteases with RFdiffusion methods; the other built cysteine-based proteases. Their targets included human disease-linked proteins: TDP-43 (ALS/FTD), amyloid-β (Alzheimer’s), and serum amyloid A. Why this is hard: the amide bonds holding a protein’s backbone together are stable for centuries in water. De novo enzyme design had mostly succeeded on easy, low-energy model reactions. Breaking peptide bonds on demand was one of the field’s white whales. The numbers: of 135 zinc-protease designs tested, 36% cut their target peptides at the right bond. The best design accelerated hydrolysis by more than a hundred million-fold (10^8); further optimization pushed it past 10^10-fold. First authors Anqi Chen and Kejia Wu; David Baker and ETH Zurich’s Donald Hilvert on both papers. The honest part: these cleaved peptides in the lab — they are not drugs. Clearing amyloid-β from a brain is a very different mountain than cutting a peptide in a tube. But designed enzymes are now catalyzing at the level of natural ones, and that capability gap closed by orders of magnitude. AI went from predicting protein shapes (AlphaFold) to inventing new enzymes that dismantle disease-linked proteins. The question isn’t whether designed enzymes enter medicine — it’s which disease they reach first. Which would you point them at? unite.ai/baker-lab-designs-a… #AI #proteindesign

What topic you want to see more? Got something you’re curious about? Please drop a comment below 👇

Tell an AI what you want a protein to do — in plain English — and it designs one. Westlake University researchers just published Pinal in Nature: a model that turns a written description of a protein's job into candidate amino-acid sequences. No structures, no templates, no molecular jargon required. It was trained on 135 million protein sequences and 1.7 billion protein-text pairs, with AlphaFold predicting the structures. The receipts: asked to design four proteins — a glowing protein, a plastic-eating enzyme, an alcohol-processing enzyme, and a metabolic shuttle — it produced working candidates for all four. Its best shuttle protein boosted a carbon-capturing pathway's output to 1.7x the natural version's. The fluorescent protein was harder: 1 of 29 candidates actually glowed. Why it matters: protein design is how new drugs, enzymes, and biomaterials get made — and until now it needed specialists who speak molecular structure. If describing a protein gets as easy as describing a photo, the bottleneck moves from design to testing. Same Nature issue, same week: Shanghai Jiao Tong's AutoDNA — AI agents running an 11,382-step DNA experiment while humans only refilled reagents. What protein would you ask an AI to build first? phys.org/news/2026-11-ai-pro… #AI #proteindesign

Ozempic's next rival might not be a drug. It might be a headband. BrainCo — the Chinese neurotech firm often called China's answer to Neuralink — says it's building a wearable that suppresses appetite by stimulating the brain circuits that control hunger. Founding partner Nyx He calls it an "electronic version of GLP-1." The plan, revealed at a Bloomberg tech event in London: launch in China within 2–3 years, then the US. Why it matters: GLP-1 drugs like Wegovy and Zepbound transformed weight loss — and left millions priced out, needle-shy, or wary of a lifelong weekly shot. A drug-free wearable that taps the same appetite circuits would be a real alternative. BrainCo already makes EEG devices and brain-controlled prosthetic hands, and raised ~$300M earlier this year. The honest part, stated plainly: this is a plan, not a result. No clinical data, no trial results, no specs, no disclosed mechanism. Stimulating the brain to curb appetite wouldn't necessarily reproduce GLP-1's broader metabolic effects. Would you trust a headband to do Ozempic's job? ibtimes.sg/chinas-brainco-wa… #neurotech #GLP1 #BCI

Seven tumors in children treated with a one-time gene therapy is a safety signal that can't be read from a count alone. What tumor types were these, and how long after dosing did they appear?

Gene editing just took a real swing at curing hepatitis B — and the early numbers are striking. Precision BioSciences' PBGENE-HBV, a one-time gene-editing therapy, wiped the virus's activity marker (pregenomic RNA) to undetectable in 100% of evaluable patients at its target dose in Phase 1. Liver biopsies showed the virus's hidden DNA reservoir — cccDNA, the reason hepatitis B has never been curable — down 10- to 30-fold. And in patients followed 6+ months, the effect held. Why it matters: some 254 million people live with chronic hepatitis B. Today's drugs suppress it for life; they don't clear it. PBGENE-HBV cuts the viral DNA itself, aiming for a finite treatment and a functional cure. The program now advances to Phase 2, with late-breaking data presented at AASLD on Sunday. Honesty: 25 patients, Phase 1, company-reported; the 100% figure rests on 12 evaluable patients with detectable virus at baseline. No dose-limiting toxicities across 59 administrations. Cure claims wait for Phase 2 — but the reservoir is shrinking. Could a single gene-editing dose end a lifelong infection? quantisnow.com/insight/preci… #biotech #hepatitisB #geneediting

AstraZeneca's reported $10B offer for Rhythm is a rare-disease obesity bet, not a GLP-1 play: Rhythm's setmelanotide is an MC4R agonist already approved for rare genetic obesity. The question is whether the label can expand beyond those rare MC4R-pathway patients.

Roche is a serious validator for AI drug discovery — but the test is whether Chai's models move from structure prediction into actual molecules in Roche's pipeline. Which targets are first, and is Chai-2's antibody design in scope?