De-novo mini-protein binder candidates against human DLL3 (EGF1 / EGF4 epitopes), generated end-to-end by the Catalyst AI engine (target → epitope → design → validation → FTO pre-screen). ⚠️ VALIDATION CORRECTION (2026-06-18). An honest re-audit found these candidates are NOT cross-model validated. The high AlphaFold2 i_pTM reported below (0.72–0.73) is the trajectory (design-time co-fold) score — AF2 scoring its own design objective, which is optimistic by construction. When AF2 is asked to independently re-predict binding (the proper validation, with a positive control: barnase:barstar = 0.73 ✓), l84 scores ~0.06 — i.e. AF2 independently predicts NO binding. Only Boltz-2 (independent) scores it high (0.83). The two models therefore disagree under a true apples-to-apples independent comparison. Consistent with this, zero designs across all five BindCraft runs (egf1/egf1_v2/egf1_v3/egf4/egf4_pep) passed BindCraft's own MPNN independent re-validation. These are Boltz-2-only, AF2-independent-rejected, trajectory-stage candidates — not validated binders. They are kept here as honest negative/learning results, not as a ready asset. See §5.
| ID | Length | Epitope | AF2 traj (co-fold) | AF2 independent | Boltz-2 i_pTM | FTO (epitope) | Sequence |
|---|---|---|---|---|---|---|---|
| DLL3-EGF1-l84 | 84 | EGF1 (217–241) | 0.73 | ~0.06 ✗ (no binding) | 0.83 | clean — no overlap w/ claimed EGF3-4/5-6 | ETGSVVVTRENMTEEEKVKEAKMESSNRERIYHKKMSDSSWKRTVYVVRKMGNKVVIKKYKSEPVWIGGVWQPVQYYVGRKYYD |
| DLL3-EGF4-l55 | 55 | EGF4 (339–374) | 0.72 | not re-run (same trajectory-only status) | 0.84 | overlaps claimed EGF3-4 | SMNINEMKHMMRRAVEGFFRSGRARYIATFPGTRIIIHRIPGAHPDEYEIHWREF |
| DLL3-EGF4-l87 | 87 | EGF4 | 0.72 | not re-run (same status) | 0.82 | overlaps claimed EGF3-4 | RDEPIQYFKWMERWLWEKVYMFRSHPVMHNRFYRMVKMARWMRSNARHMREDDIEELSLRFNDTFVVVMKQWKQVMNGQPLVPDDGL |
The two-model consensus was NOT actually met. The earlier "both ≥0.6" claim compared the AF2
trajectory number against Boltz-2's independent number — apples to oranges. Under a true
independent comparison, AF2 rejects l84 (~0.06; positive control barnase:barstar = 0.73 ✓ confirms
the method) while Boltz-2 accepts (0.83). The models disagree, and BindCraft's own MPNN
re-validation passed zero designs across all five runs. l55/l87 are in the same trajectory-only
state (not independently re-run, but no design from their runs passed validation either).
There is no validated lead. The charge-engineering work in §6 (a real, reusable engine capability)
produced an l84 variant with lower kidney-risk charge that *also* held up on Boltz-2 — but it inherits
l84's unresolved AF2 disagreement, so it is not a validated binder either. The honest takeaway: on this
small, acidic, cysteine-dense EGF surface, de-novo design did not yield an independently cross-validated
binder. See §5 for the path forward (selection / wet, per the expert panel).
membrane-proximal, high-confidence (AlphaFold pLDDT 80) surface, chosen over the DSL Notch-binding
face (low confidence, patent-dense).
(GADPDSAYICHCPPGFQGSNCEKRVDRCSLQPCRNGGLCLDLGHALRCRCR).
Patent landscape (which family claims which epitope):
antibody-construct patents that characterise binders by epitope and claim EGF3-4 and EGF5-6
regions, often with "competes for binding / binds the same epitope" breadth.
(G203/R205/P206).
round.
Structural epitope-overlap pre-check (our "computational competition assay"): from the Boltz-2
predicted complex we extract the DLL3 residues each binder actually contacts and map them to the
claimed regions. For l55 the epitope = DLL3 residues 339–374 (the EGF4 domain) + a few EGF4-5
linker residues → it binds within the EGF3-4 region (l87 similar, by design).
What that does and does NOT mean (claim-read of US11591396, Amgen tarlatamab): the blocking
patents do not (and legally cannot) claim the natural EGF4 region — they claim **specific
invented molecules + uses. US11591396 claim 1 requires ALL of: (i) enumerated CDR
sequences (SEQ ID 111–116), (ii) a bispecific DLL3×CD3 antibody format, and** (iii) binding
the EGF3-4 epitope. l55/l87 are de-novo non-antibody mini-proteins — no CDRs, not bispecific,
structurally/sequence-distinct — so they satisfy only the epitope limitation and **do not read on
the composition claim.** "EGF4 is claimed" therefore means "we can't make *their antibody*," not "we
can't make an EGF4 binder." Post Amgen v. Sanofi (US Supreme Court, 2023), broad
functional/epitope antibody genus claims are additionally hard to enforce — further favouring a
structurally-distinct de-novo binder.
Residual FTO to resolve with counsel (not disqualifying): (a) method/use claims ("treat
SCLC with an anti-DLL3 agent") can be broader — must be checked; (b) any separate **pure-epitope or
pure-competition** independent claim in other DLL3 patents; (c) a formal FTO opinion on the full
claim text. The EGF1 round runs in parallel as an extra-clean-epitope hedge. The engine flags
*which epitope a candidate hits, which patent family it overlaps, and the claim type*; the legal
opinion is counsel's.
proteins).
ideally sub-100 nM as a starting point for maturation.
domain.
(the RLT vector).
prediction.** Boltz-2 likes them (0.82–0.84); AF2, re-predicting independently, rejects l84 (~0.06,
with a passing positive control). BindCraft's own MPNN re-validation passed zero designs across all
five runs. De-novo design on this small, acidic, cysteine-dense EGF surface **did not produce an
independently cross-validated binder.** That is the honest result.
build — independent-reprediction validation discipline, the FTO/patent module, and the
developability/charge-engineering modules. The engine catching its *own* over-stated "0.73 AF2"
before a CRO/pharma DD team did is a credibility *gain*, not a loss.
(a) Selection over design — the CRO VHH display-screening campaign
([CRO_RFP_DLL3_VHH.md](CRO_RFP_DLL3_VHH.md)) is the primary route to a real DLL3 binder; (b) prove
the engine/process (repeatability across targets + one wet-validated binder) rather than a hero
molecule on a crowded, hard target; (c) if any in-silico candidate is ever advanced, the bar is
two INDEPENDENT models ≥0.6 (never a trajectory co-fold) plus a cheap pooled display wet test
as the real-or-dead gate, before any biodistribution spend.
*Engine-generated; honestly re-audited (NOT cross-model-validated — AF2-independent rejects, see
header + §5); FTO-flagged. Selection (CRO display) is the route to a real binder; counsel does the FTO
opinion; a dosimetrist + isotope partner do the RLT build.*