De Novo Binder Workflow

The canonical 6-step framework the platform orchestrates. Use cases, CAR-T format filters, and limitations all on one page.

Where this fits in the docs
  • /docs/knowledge-map — 5-layer mental model — the macro view.
  • /docs/de-novo-workflow (this page) — the 6-step operational workflow — the narrative of how to actually run a de novo design.
  • /docs/recipes — 'I have X, I want Y' — concrete compositions of platform tools.
  • /docs/readout-guide + /docs/scoring-audit — per-score reading manual + honest critique of the metric set.

The six steps

W1Input antigen structure
What: Provide the target antigen as a deposited PDB or an AlphaFold-predicted structure. Defined epitope is strongly preferred.
Why: Everything downstream rides on the input. Garbage-in, garbage-out — designing against a misfolded or unannotated antigen produces structurally-plausible binders that miss the real surface.
Platform path: /import — drop a PDB, paste a UniProt id, or pick from the curated target catalog.
W2Define binding hotspot
What: Specify the antigen residues or surface patch you want the binder to engage. A residue list like `A:42,A:45,A:50` or a lasso-selected contiguous patch.
Why: The single most consequential choice in the entire workflow. Two binders engineered against different epitopes of the same antigen behave like different drugs in the clinic.
Platform path: /epitope-picker — interactive 3D viewer with curated presets (BCMA membrane-distal, CD20 ECL2, GPRC5D tumor-selective region, …).
W3Backbone generation
What: Sample diverse antibody backbone conformations conditioned on the antigen + hotspot. Output is geometry only — no amino-acid identities yet.
Why: Backbone diversity is what RFdiffusion buys you. Sampling a single backbone is no better than picking from a static library; the value is in the cloud of conformations.
Platform path: /pipeline Stage 2 — RFantibody / BoltzGen / partial-diffusion modes for warm-start regenerate from a parent.
W4Sequence design
What: Inverse-folding: given a backbone, produce N candidate amino-acid sequences that should reliably fold into it.
Why: Backbones without sequences aren't testable. ProteinMPNN-class models invert the structure → sequence mapping cheaply and with high success rate; AntiFold tunes that for antibody CDRs.
Platform path: /pipeline Stage 3 — ProteinMPNN by default, AntiFold for antibody-aware variants, IgDesign for CDR-focused redesign.
W5In-silico filtering
What: Score every (binder, antigen) complex and rank. Dimensions: interface quality (ipTM / iPAE), buried surface area, CDR geometry, developability (TAP), stability (ThermoMPNN), aggregation (AggreScan3D), humanness (BioPhi), cross-reactivity (counter-screen).
Why: Designs that fold are not the same as designs that bind, are stable, are developable, and are safe. Each filter answers one of these orthogonal questions. The funnel narrows; survival rate is the platform's main wet-lab proxy.
Platform path: /pipeline Stage 4 + 5 — Boltz-2 + Protenix ensemble for confidence, TAP / ThermoMPNN / AggreScan3D / BioPhi as per-score columns surfaced on /candidates. Audit lives at /docs/scoring-audit; per-score detail at /docs/readout-guide.
W6Experimental validation handoff
What: Top-N candidates (typically 20-100) leave the platform: synthesize, display-screen (yeast or phage), measure binding (SPR / BLI), run specificity panel, then functional assay (CAR-T cytotoxicity, TCE cell-based).
Why: The platform's whole purpose is to make this step cheaper and faster — not to replace it. Until wet-lab feedback closes the loop, every score above is an unvalidated prediction.
Platform path: Export FASTA + PDB + scores from /candidates. Display-construct exporter (with linkers, tags, signal peptide) is the planned next addition. Wet-lab results flow back via /api/v1/wet-lab/measurements — see /docs/wet-lab-loop.

Use cases

#Use caseDescriptionPlatform path
U1Epitope-specific designForce the binder onto a chosen surface patch. Example targets: CD20 extracellular loop, BCMA membrane-distal epitope, GPRC5D tumor-selective region./epitope-picker → /pipeline
U2CDR grafting / redesignHold the antibody framework fixed; redesign HCDR3 or all CDR loops on top of a wet-lab-validated parent. The warm-start regenerate mode./pipeline with --mode partial_cdrs --cdr-regions hcdr3 --parent-vhh PARENT.pdb
U3De novo VHH / scFv / mini-binderGenerate novel binders from scratch when no validated starting point exists. Smaller binders fit in vivo CAR mRNA cassettes more easily than scFv./vhh-design (VHH) · /mini-binder (60-240 aa) · /pipeline (full)
U4Avoid cross-reactivityDesign against a unique surface patch and counter-screen against homologs in silico. For BCMA: counter-screen vs TACI + BAFF-R (TNFRSF family)./pipeline --counterscreen --counterscreen-target BCMA
U5Mutation-resistance predictionSubstitute clinically-observed antigen mutations (e.g. BCMA position-27) and re-score; retain candidates whose binding is preserved./mutation-resistance — submits a Boltz-2 prediction per mutation + WT baseline

CAR-T format filters

Additional gates that apply on top of binder-level scoring when the Layer-0 format intent is CAR-T (any CAR-T — scFv-CAR, VHH-CAR, in-vivo or ex-vivo).

#FilterStatus / note
C1High tumor bindingAffinity within the therapeutic window. Too low = no killing; too high = no off-rate, T-cell exhaustion. ipTM + iPAE proxy this in silico.
C2Low tonic signaling (binder-level)Binder should not cluster CAR on the T-cell surface without antigen present. Currently proxied by a heuristic v0 — surface hydrophobicity + net charge + Fv-Fv self-association — flagged as unvalidated.
C3Proper CAR surface expressionFolds correctly + traffics to the membrane. TAP-derived approximation today; a dedicated CAR-context expression model is the future Wave 3 deliverable.
C4Low antigen-independent clusteringTied to C2 — folded into the tonic-clustering composite for now.
C5Activity under low antigen densityFunctional CAR-T activity when the tumor expresses BCMA at low copies. Not modeled in silico — Phase 2+ research item; needs wet-lab data first.
Honest limitations
Three things this workflow does NOT do — be explicit about them in every conversation about candidates.

RFdiffusion output ≠ final drug candidate

What the workflow produces is a set of structurally-plausible binders ranked by interface scores. It is NOT a Kd measurement, NOT an expression validation, NOT a safety profile. Display screening, affinity maturation, specificity testing, expression / developability characterization, and functional assays still happen.

Affinity is not always better

Tighter binding ≠ better drug. CAR-T affinity above ~nM regime can amplify on-target / off-tumor toxicity (the Carvykti boxed-warning story). Composite scoring should treat affinity as bounded, not maximized.

In silico scoring is uncalibrated until wet-lab data lands

Single-model ipTM disagrees between Boltz-2 and AF-Multimer on these BCMA candidates (Wave 1 sanity check: Spearman ρ = 0.48). Cross-model agreement is the best we have until wet-lab Kd / cytotoxicity comes back and seeds the calibration runner (/docs/wet-lab-loop).

Mapping to the W/U/C/V gap matrix
Internal engineering uses W (workflow) / U (use cases) / C (CAR-T filters) / V (validation) prefixes to track which boxes are checked. This page is the human-readable side of that matrix.
PrefixMaps to
W1–W6The 6 workflow steps above
U1–U5The 5 use cases above (4 from the original framework + 1 from internal review)
C1–C5The 5 CAR-T-specific filters above
V1Display-construct exporter (yeast / phage / mammalian) — currently deferred