ARENA: Adaptive Reseeded Elimination for Novel binder Architectures Team Lunatics- (Tanay Lohia, Sourab Kulkarni)

Method PDF: https://docs.google.com/document/d/1XQL60jLdAoWwg5zZmYAVHaTL-sWGzSUk

== Target Challenge == RBX1's N-terminus (res 1–39) is an intrinsically disordered region (IDR), while the C-terminal RING-H2 is stabilized by three Zn²⁺ ions through 12 load-bearing residues. The IDR makes AI confidence metrics (ipTM, pLDDT) unreliable as sole quality indicators — we supplement with Rosetta energetics and manual structural validation.

== Generation: ARENA == We build on Protein Hunter (Boltz-2 + SolubleMPNN), a compute-light iterative co-design pipeline. Protein Hunter is strongly seed-dependent — poor seeds rarely recover. ARENA solves this via tournament-based progressive elimination during generation:

• R1: 100 seeds × 1 cycle → keep top 40% by ipTM (40 survive)
• R2: 40 seeds × 2 cycles → keep top 20 (trajectory improvement + high ipTM baseline)
• R3: 20 seeds × 2 cycles → full 5-cycle refinement 220 cycle-equivalents per combo, exploring 2.3× more structural basins than equivalent full-trajectory compute.

== Multi-Epitope Hedging == Four parallel ARENA tournaments across mechanistically distinct epitope strategies:

• C1: GLMN-competitive (10 candidates)
• C2: CRL state-selective (5 candidates)
• C3: Extended E2 face (7 candidates)
• C4: Full E2 sweep (4 candidates) All 12 Zn-coordinating residues excluded from every contact list. All hotspot contacts strictly polar/charged. No cysteines in any binder (prevents zinc chelation + disulfide scrambling in reducing E. coli cytoplasm).

== Design Constraints == Size range: 70–150 AA. SolubleMPNN for all sequence design. Deliberately relaxed in-silico thresholds (soft targets: ipTM≥0.85, pLDDT≥0.80) due to IDR suppressing confidence metrics.

== Filtering & Ranking (880 → 50 → 26) == Step 1: Harmonic mean ranking HMean(ipTM, pLDDT) — penalizes imbalance, forcing both metrics high simultaneously. Top 50 selected, interleaved across cohorts (C1>C4>C2>C3). Step 2: Quality filters (ipSAE≥0.20, Rosetta dG<100) + manual curation — Boltz-2 complex inspection, dG_separated, H-bond count, buried surface area, shape complementarity, E2-face coverage. Step 3: Top 13 priority-ranked by qualitative structural review.

== Results == 26 candidates submitted. ipTM: 0.691–0.932. pLDDT: 0.722–0.874. Size: 80–131 AA. 4 epitope strategies. Single A100 GPU.