The engine designing a brand-new, functional, advanceable therapeutic candidate end-to-end — everything that does not require a wet lab. Paired with ENGINE_PROOF_PDL1.md (a non-blocking targeting module) and ENGINE_STATUS.md, this is the stronger case: a de-novo binder that blocks its target (a real mechanism, not just engagement). In-silico only — i_pTM ≠ Kd; not a drug; a bounded, de-risked hypothesis for a cheap wet test. (Validation tools AlphaFold-class are commercial-OK Boltz-2 + AF2; nothing non-commercial is shipped.)
The engine first failed honestly on hard targets (DLL3 EGF1, B7-H3 IgC, PD-1) — good design-time
co-fold scores, but no sequence passed independent AF2 validation; the engine refused to ship them.
A 4-expert panel diagnosed the cause as epitope SELECTION, not method: a de-novo mini-protein needs a
concave surface with a buriable hydrophobic ANCHOR (a "socket"); flat polar antibody/loop epitopes (a
"lid") fail. We built an epitope-designability pre-screen (now an engine tool), validated it on real
structures (PD-L1 face designable 75.6 → it produced l74; B7-H3 IgC epitope NOT designable 63.8 → it gave
0 designs), and used it to pick a designable, advanceable target — IL-4Rα (IL-4-binding face,
designability 69.4). That target produced a passing design at trajectory 1, where the un-screened hard
targets gave 0 across ~20. **The screen turns target/epitope selection into an auditable, GPU-saving
decision — itself a moat.**
IL4RAb_l101_mpnn7A 101-aa de-novo mini-protein (non-antibody) binding the IL-4-binding face of human **IL-4 receptor α
(IL4R / CD124)**.
GEEVVKLVEQWAQGAIERIELKIRKMEYTYGPDNKEAQAKLLEELKAFRDAVIEKMREFVRKKYEEEGMPKEEYEKAMKKVEELEAKWPEFVAKVDPEKGK
| Gate | Result |
|---|---|
| Binding — two INDEPENDENT models | AF2 independent 0.74 (pLDDT 0.90) and Boltz-2 (MSA-backed) 0.901 — both clear 0.6 |
| Interface energetics | ΔG −47.7, shape complementarity 0.76, buried area 1657 Ų, 20 interface residues, 4 H-bonds, Hotspot-RMSD 1.24 |
| Function (MoA) | binder footprint overlaps 82% (18/22) of the IL-4-binding face → strongly predicts IL-4 blockade. Functionally positioned as an IL-4Rα antagonist (not just an engager). Caveat (do NOT overclaim): dupilumab's value is dual IL-4 *and* IL-13 blockade because IL-4Rα is the shared subunit; this binder's in-silico footprint supports IL-4 blockade — whether it *also* blocks IL-13-driven IL-4Rα signalling must be tested in an IL-13→STAT6 assay. Not yet a dupilumab-equivalent dual blocker. |
| Specificity | vs the closest paralog IL-13Rα1, Boltz-2 0.294 — far below the on-target 0.901 (margin 0.607) → selective for IL-4Rα |
| Developability | net charge −3 (near-neutral), solubility 0.96, no aggregation; minor cationic patch + borderline instability (index 43.7) — a clean-enough backbone, the engine de-selected a dirtier sequence (mpnn1, net −8) for this one (mpnn7) |
| FTO (pre-screen, not legal advice) | low risk — a de-novo non-antibody does not read on the CDR-sequence claims of the approved anti-IL-4Rα antibody (dupilumab class), though it hits the same functional region; method-of-use claims are a separate counsel check |
Grounding: target structure + the real IL-4-binding face from the IL-4:IL-4Rα complex (PDB 1IAR);
hotspots taken from that interface, not guessed. Design: BindCraft (AF2-hallucination → ProteinMPNN →
independent AF2 validation), self-hosted GPU; second model Boltz-2 NIM (MSA-backed).
IL-4Rα blockade is a validated, huge mechanism: dupilumab (anti-IL-4Rα mAb) is an IL-4Rα antagonist
that blocks both IL-4 and IL-13 (the shared subunit), with Dupixent ~$17.8B 2025 net sales. So the
biology and market are de-risked. But position honestly:
antibody is very hard to displace, and a 101-aa mini-protein has a likely short systemic half-life
(systemic use would need half-life extension — albumin-binder / Fc / PEG / XTEN — which re-complicates
CMC, immunogenicity and PK).
fast-to-engineer, potentially low-cost, IP-whitespace binder whose short half-life is an *advantage* for
local indications — topical atopic dermatitis, inhaled asthma, intranasal CRSwNP, local
EoE — where you avoid competing with a systemic antibody and sidestep the half-life problem.
genuine therapeutic antagonist candidate — *if* the wet gates (§3), especially functional STAT6
blockade, pass.
Done (in-silico): two-independent-model binding, interface energetics, function (epitope occupancy of
the IL-4 site), paralog selectivity, developability, FTO pre-screen. This de-risks "does the engine's
output have a plausible, selective, functional, developable binding mode." It does.
Wet GO/KILL (pre-registered, cheapest-first) — no AI can answer these:
dual IL-4/IL-13 pathway blocker" (the latter is a far stronger story; IL-4-only still has value, weaker).
severe aggregation.
favour local delivery first (§2).
higher-stakes than a short oncology course (ADA can neutralise efficacy over time). Add: **MHC-II /
T-helper-epitope prediction** (a dedicated MHC-II predictor — the driver of anti-drug-antibody risk;
the engine's mhcflurry_predict is MHC-I peptide-level, a partial signal only), a human PBMC
re-stimulation assay, and an ADA risk read before any clinical positioning.
epitopes — and it refuses to ship designs that fail independent validation (DLL3/B7-H3/PD-1).
The honest one-liner: **"the engine designed a selective, developable, FTO-clean de-novo IL-4Rα antagonist
that *strongly predicts IL-4 blockade* in silico (IL-13 dual-blockade still to be tested) — and told us
exactly which ~$15–30k wet test makes it real." Positioning: a functional antagonist proof case —
in silico, wet-ready** (not yet a dupilumab-equivalent dual-pathway blocker).
*Engine-generated; two-INDEPENDENT-model validated; functionally characterised; FTO-flagged. In-silico
only — wet validation is the truth. A bounded, advanceable hypothesis, not a drug.*