Skip to content

Chai Discovery

Chai Discovery builds two frontier models the factory is designed around — one to predict molecular complexes, one to design new binders. They chain: Chai-2 designs a binder, Chai-1 validates the binder–target complex. Both are Frontier Models, connected on demand via elastic burst.

Chai Discovery in the factory — Chai-1 co-folds complexes, Chai-2 designs binders, and where each connects

Illustrative. Chai-1 is planned (open, Apache-2.0); Chai-2 is gated (partner access). Both burst to a remote GPU.

Chai-1 — complex co-folding   planned

Open (Apache-2.0, commercial-cleared); build in progress — flips to live when it is stood up.

AlphaFold3-class structure prediction of complexes — proteins with other proteins, small-molecule ligands, and DNA/RNA — the co-folding that ESMFold (single-sequence only) cannot do. Where it plugs in:

Capability What Chai-1 adds
Structural Biology Engine Complex co-folding; anchors the design-then-validate chain (RFdiffusion → ProteinMPNN → Chai-1).
Therapeutic Discovery Engine Structure-based design — co-fold a candidate molecule inside its target's pocket to confirm it binds.
Precision Oncology Engine Fusion-protein, drug–target, and neoantigen–MHC complex structures for the tumor board.
CAR-T · Precision Autoimmune agents scFv–antigen and antibody–antigen / HLA–peptide complexes — the structural basis of recognition.
Rare Disease · Neurology agents Whether a missense variant disrupts a protein interface — mechanistic evidence for pathogenicity.
Tuberous Sclerosis (flagship) Co-fold the hamartin–tuberin (TSC1–TSC2) complex and show how a pathogenic variant breaks it — making the gene → mTORC1 → everolimus story structurally concrete.

Chai-2 — de-novo binder & antibody design   gated

Partner access via Chai's platform (not open weights); integration contingent on access.

Zero-shot design of new antibodies and protein binders against a target. Its home is the binder-design family:

Capability What Chai-2 adds
CAR-T Intelligence Agent Design novel scFv / binding domains for CAR-T targets.
Precision Autoimmune Agent Design autoantibody-neutralizing or de-immunized binders.
Precision Oncology · Therapeutic Discovery Biologics against tumor antigens — large-molecule design alongside small-molecule discovery.
Structural Biology Engine The design frontier; hand off to Chai-1 to validate the designed complex.

Working with NVIDIA BioNeMo

Chai and NVIDIA BioNeMo are complementary halves of one design bench: Chai predicts and designs structures; BioNeMo generates and docks the small molecules that populate them. Together they cover a target from both directions.

A ~90-second walkthrough: Chai-1 predicts a complex, BioNeMo fills the pocket with docked small molecules, Chai-1 validates the pose, and Chai-2 adds a biologic lane. Roadmap where noted — decision support, not diagnosis.

  • Chai-1 validates BioNeMo's small molecules. BioNeMo's MolMIM generates candidate compounds and DiffDock docks them; Chai-1 then co-folds the top candidate inside the target pocket — a structure-based cross-check on the docking pose. A different method reaching the same answer raises confidence in the lead.
  • Chai-2 adds the large-molecule lane. Alongside BioNeMo's small-molecule discovery, Chai-2 designs de-novo biologics, and Chai-1 validates either — so a single target can be pursued as a small molecule (BioNeMo) or an antibody / protein binder (Chai) in parallel.

What the pairing enables:

Engine / Agent What Chai + BioNeMo unlock together
Therapeutic Discovery Engine Both design lanes — BioNeMo small-molecule, Chai-2 large-molecule — with Chai-1 as the shared structural validator.
Structural Biology Engine Chai-1 co-folds BioNeMo's docked ligands into a full predicted complex — pose meets structure.
Precision Oncology Engine Offer a tumor board both a small-molecule inhibitor (BioNeMo) and a biologic (Chai-2) against one target, each with structural evidence.

(Honesty: the BioNeMo half — MolMIM, DiffDock — is live but needs the NIM container deployed; Chai-1 is planned and Chai-2 is gated, so the integrated loop is a roadmap capability. All of it is preclinical decision support, bursting to a remote GPU.)

Honest by construction

Chai-1 is planned (open, build in progress) and Chai-2 is gated (partner access) — see the live Capability Maturity Matrix. Both burst to a remote GPU, never pretending to run on the box; and any clinical use is decision support for a qualified clinician, never autonomous diagnosis.