Biophysical, Kinetic, and Algorithmic Audit of De Novo Miniprotein Binder Campaigns: A 500-Agent Jesse Expansion of Citrini Therapeutics

Biophysical, Kinetic, and Algorithmic Audit of De Novo Miniprotein Binder Campaigns: A 500-Agent Jesse Expansion of Citrini Therapeutics We present a kernel-certified Lean 4 formalization (30 theorems in module JesseMath.Biophysics.CitriniDeNovoBinderAudit compiled under leanprover/lean4:v4.34.1 with zero sorry declarations and zero custom axioms beyond ZFC, enforced at compile time by #assert_zfc_axioms), a standalone 500-assertion Python verification oracle (harness/citrini_de_novo_binder_oracle.py), and a 500-agent self-sandboxed Jesse swarm audit (POD-I through POD-X) of the Citrini Research and Anthropic de novo miniprotein binder campaign against five Buthidae voltage-gated sodium channel (NaV1.4/NaV1.6) peptide ligands (Cn2, Ts1, AaH II, AaH I, and LqhIII; 318,484 generated designs, 553,052 1-seed screens, 28,451 5-seed ranked pairs, and 211 wet-lab synthesized designs). Primary Empirical Discoveries and Strategic Extensions: Cross-Family (α+β) Multi-Target Binding via Oligomeric Rmax Deconvolution (Pillar I): Construct AaHII-B4 (a199ee01, Set 3, Genie 3 + SolubleMPNN) bound both the North African Site-3 α-ligand AaH II (KD = 21.00 nM at Adaptyv Bio across 4 replicates; KD = 6.31 nM at Twist Bioscience) and the Brazilian Site-4 β-ligand Ts1 across all three Adaptyv Bio replicates (KD = 12.20, 8.04, 7.17 nM; mean 9.14 nM) and Twist SPR (KD = 28.10 nM), with zero off-target binding to CLEC12A or ovalbumin (OVA). Normalizing for the tetrameric stoichiometry (Nolig = 4) of the aggregated commercial Ts1 reagent lot (Cusabio CSB-EP323347TON, which produced 11.00× Rmax on Twist's positive control antibody) yields Rmax,obs / (4 Rmax,mono) = 1.05×. Unimolecular koff Invariance and 4 Salvaged Nanomolar Binders (Pillar II): Adaptyv Bio's C-terminal twin-Strep capture on negatively charged carboxymethyl dextran (CMDP) chips induced a 46.07× to 50.63× kon suppression on official AaH II binders AaHII-B1..B3 relative to Twist Fc-capture. Auditing unimolecular dissociation invariance (koff) salvages four additional nanomolar binders (bbe37cf8 at 145 nM, 47dd0a8a at 65 nM, cf8652bc at 134 nM Twist / 260 nM Adaptyv, and 9a8cdc7a at 68.1 nM Reverse BLI with 81.2% LR blockade), expanding validated hits from 17 to 22 across 21 distinct sequences. Resolved Picomolar Kinetics (96.35 pM) and Isoelectric Resurfacing of Cn2-B1 (Pillar III): Individual 1:1 sensorgram fits for Cn2-B1 (888c8f04) resolve KD = 56.93 pM (Rep 1) and 135.77 pM (Rep 2), with mean 96.35 pM outperforming the affinity-matured human scFv benchmark LR (110.00 pM) zero-shot. Its non-saturating ovalbumin signal stems from an outlier basic isoelectric point (pI = 7.8 vs. 4.0–4.8 across the other 16 binders), corrected by peripheral Lys→Glu resurfacing (Δq = −8, pI: 7.8 → 5.1), while Cn2-B2 (66c2cc29, 3.4 nM, pI = 4.0) blocks both LR (90%) and RU1 (66%). Goodhart Score Inversion and 6.64× Atomistic Generator Enrichment (Pillar IV): Within single-target sets that produced binders, official hits had a lower mean composite final_score (0.9240 vs. 0.9288) but +447.0 Å2 higher buried surface area (2,113.2 vs. 1,666.2 Å2). The three atomistic SE(3) diffusion and flow engines (RFdiffusion3, Genie 3, Proteina-Complexa) achieved a 21.43% official hit rate (12/56) and 28.57% salvaged hit rate (16/56) versus 3.23% (5/155, 6.64× enrichment) and 3.87% (6/155, 7.38× enrichment) for the rest of the campaign. Apo-Monomer Pre-Organization, Ion-Channel Receptor Decoys, Two-Compartment PK/PD, and 2-Wave Active Learning Economics (Pillars V–VII): We formalize the thermodynamic conformational penalty ΔGconf = R T ln(1 + Kapo) explainings Set 2's 0/30 outcome on rigid 4-disulfide scaffolds, soluble NaV1.4/NaV1.6 S3–S4 extracellular loop receptor decoys, heterobivalent VHH/Fc-miniprotein fusions resolving the glomerular filtration versus subcutaneous depot washout mismatch, and a 2-wave active learning portfolio allocating 85% of synthesis capacity to RFdiffusion3, Genie 3, and Proteina-Complexa under an apo-monomer scRMSD ≤ 1.5 Å and ΔSASA ≥ 1,950 Å2 gate.