Target-conditioned binder design

Design miniprotein binders across 10,000+ unnatural amino acids.

A design engine that extends RFdiffusion-style geometry generation to a vast, user-defined library of noncanonical chemistry — placing the right exotic residue at the right interface position.

The core bet: a library, not a vocabulary. — we retrieve chemically compatible residues for each binding pocket, then generate their 3-D geometry. No fixed 10k-class output head to overfit.
10k+UAA library scale
8-stageretrieve → diffuse → score
5backbone chemistries
closed-loopactive learning
target · SARS-CoV-2 RBD ◆ LCB1 miniprotein binder ◆ interface → UAA sites drag to rotate · PDB 7JZU

The abstraction problem

Why we don't treat the library as a 10,000-class vocabulary.

A token-per-residue output head learns frequent IDs and collapses on rare or unseen chemistry. We reframe UAA selection as chemical retrieval plus geometry generation — the abstraction that actually transfers.

vs

Fixed 10k-class output head

// residue-ID classification
  • Sparse, biased supervision. Most library members have no protein-context structure or affinity data.
  • Hidden chemistry. A residue ID erases atom counts, topology, stereochemistry — the very relationships that enable transfer.
  • Combinatorial blow-up. A full-library scan at every position is intractable before rotamers or epistasis.
  • Validation gap. Inverse-folding & affinity models are calibrated on canonical residues — high confidence ≠ correct UAA energetics.

Searchable chemical library

// retrieve → pose → refine
  • Score from graphs, not IDs. Each UAA is an atom-and-bond graph + conformer ensemble — unseen chemotypes are scored by structure.
  • Retrieval, not enumeration. ANN search returns a diverse shortlist per pocket — physics is spent only on finalists.
  • Geometry generation. A conditional equivariant / flow model places each candidate's atoms in the real local environment.
  • Tiered, calibrated scoring. A funnel of cheap-to-expensive checks ending in wet-lab confirmation — never one raw energy.

Live design run · concept simulation

The engine, end to end.

Pick a target and press Run design. The pipeline steps from a diffused backbone through chemistry retrieval, pose generation, and a scoring funnel — the same architecture proposed for the production system.

noncanon-engine · design-session
Ready stage 0 / 8
Press Run design to simulate a full target-conditioned UAA minibinder design pass.
stage output appears here — retrieved chemistry, scoring funnel, and the active-learning loop
console● idle