AI-Designed Molecular Glues Reach the Clinic: Nutshell's NRF2 Degrader

On September 25, 2026, a Shanghai biotech most investors could not name cleared an FDA investigational new drug (IND) application for a molecule that barely existed on paper two years earlier.

Nutshell Therapeutics calls NTS231 a covalent allosteric molecular glue. It does not block a protein. It tags one for destruction. The target is NRF2, a transcription factor that tumours lean on to survive chemotherapy and radiation. The company says NTS231 is the first NRF2 degrader out of China and the second anywhere to reach human testing.

The clearance is small. The signal is not. After a decade in which protein degradation advanced through lucky chemistry, AI-designed degraders are walking into the clinic on purpose.

TIMELINE: molecular glues, from accident to design
─────────────────────────────────────────────────────────────
  2014 ────── 2020 ────── 2024 ─────── 2025 ────── ◉ 2026
  🔬         🧪         💰           🧬          🏭
  lenalidomide  CRBN     Takeda–     first        first PROTAC
  mechanism     glue     Degron      glue         approved;
  decoded       era      $1.2B deal  patients     AI glue
                                    dosed        enters clinic

The accident that started it

In 2014, Benjamin Ebert found that lenalidomide, a multiple myeloma drug then a decade into clinical use, worked by a mechanism nobody had designed. The molecule attaches to cereblon, part of the cell's E3 ubiquitin ligase machinery, and rewires that machinery to destroy a transcription factor the cancer needs. Ebert, now president of Dana-Farber Cancer Institute, had found the first molecular glue.

The discovery reframed a drug class in use since the 1950s. It also exposed a limit. Glues worked only where chemists stumbled onto a productive pairing. Cereblon handed them a handful of targets. The rest of the proteome, including the roughly 80% of human proteins with no deep pocket for a conventional drug to occupy, stayed out of reach.

There is an incredible range of opportunity for discovering new molecular glue degraders.— Benjamin Ebert, president and CEO, Dana-Farber Cancer Institute

Why 2026 is different

Two things changed. The first is regulatory proof. On May 1, 2026, the FDA approved vepdegestrant, developed by Arvinas and Pfizer, for ESR1-mutated breast cancer. It is the first heterobifunctional degrader, or PROTAC (proteolysis-targeting chimera), to clear approval. That moved protein degradation from experiment to endorsed modality, and it gave every glue program in development a regulatory template to copy.

The second is design. Where Ebert found glues by observation, the current generation engineers them. Monte Rosa Therapeutics runs an AI engine that mines protein surfaces for cereblon-compatible targets. Dana-Farber's new platform, published in Nature in August 2026, screens systematically for glues and produced the first metabolically activated one, a degrader that switches on only inside cells with a specific metabolic state.

Those patients are expensive to treat and poorly served. NRF2/KEAP1/CUL3 alterations are mutually exclusive with the EGFR, ALK, ROS1 and BRAF lesions that most lung-cancer drugs target. Immunotherapy performs badly because the tumours run immunologically cold. Chemotherapy buys time, then stops. The pitch: destroying NRF2 removes the tumour's defence rather than attacking it directly.

Nutshell's 24-month shortcut

NTS231 was built on an internal platform Nutshell calls ALLOSTAR, which combines computer-aided drug design with medicinal chemistry. The molecule binds covalently to KEAP1, the protein that normally restrains NRF2, and locks it into a shape that recruits the CUL3 ligase. The complex then degrades NRF2 through the cell's own protein-recycling system.

The company says it reached IND clearance 24 months after target nomination. For a field where discovery-to-clinic timelines usually run five to seven years, that is the company's real claim. The molecule itself looks conventional: a single small molecule, orally available, with non-inferior in vitro activity to VVD-130037, a clinical-stage compound hitting the same mechanism, plus better pharmacokinetics.

Preclinical data covers cell-derived and patient-derived xenograft models across lung squamous, lung adenocarcinoma, oesophageal squamous and head and neck cancers. In chemotherapy-resistant lung squamous models, NTS231 plus paclitaxel inhibited tumour growth more than either agent alone. In an NFE2L2-amplified lung model, the glue paired with a TROP2 antibody-drug conjugate drove tumour regression and suppressed the efflux transporters that normally pump out the payload.

Comparison based on Nature Reviews Drug Discovery (2026) and cited company disclosures

Turning points

Three moments in 2026 will settle whether this is a durable modality or a funding-cycle story.

What to watch

The gap between a cleared IND and a working drug is where most degradation programs have died. As we wrote in September, Enveda's $311M round showed how quickly AI-native drug platforms can raise capital on preclinical promise. Nutshell's next twelve months test whether the promise converts.

Selectivity is the commercial fault line. A glue that removes NRF2 may also remove proteins that look like NRF2, and that risk only shows up in patients. The AI platforms that Monte Rosa, Dana-Farber and Nutshell use are being sold as the answer to that problem. The clinic will audit the claim.

Nutshell has raised tens of millions of dollars and runs a 1,500-square-metre research centre in Shanghai. It is not a large company, and a single IND does not make a franchise. But the direction is unmistakable. Protein degradation spent its first decade proving that a molecule could delete a target it was never designed to touch. Its second decade will run on molecules that were designed for the target from the first drawing.