In this challenge, Theorema proposed and prioritized the use of a shared computational stack (RFdi usion ProteinMPNN Boltz2 MMseqs2) across two complementary strategies: (1) a scale-rst exploration track (larger run volume to maximize search-space coverage before structured triage) and (2) a focused iterative, hypothesis-driven track (fewer targeted runs to improve decision quality). RFdi usion generated candidate backbones, ProteinMPNN generated sequences, Boltz2 evaluated binder/target complex quality, and MMseqs2 was used for novelty gating. We provide 25 best candidate binders. All submitted candidates were designed de novo and screened for novelty/compliance against UniRef50 under competition rules (minimum 25% edit distance / ≤75% identity threshold framework) with invalid motif-sca olded or lead-optimized non-antibody designs excluded. In-silico results showed convergence toward higher-condence designs during iterative renement, including improved epitope productivity for E2-face targeting and top-performing candidates reaching ipTM up to ~0.89 in the strongest batches. Ranking and progression were based primarily on interface metrics: ipTM (campaign target >0.88), with additional artifact-aware controls (including MSA sensitivity checks and repetitive-sequence ltering).

https://drive.google.com/file/d/18gdDrgMQycEh2tWShRj5pU0svDww0vhM/view?usp=sharing