Method Description PDF: https://drive.google.com/file/d/1sfpSaxuQp8FLdHHDlcwHnIQPBU_cknhD/view?usp=sharing

We used a two-stage de novo binder design workflow combining generative backbone design, sequence optimization, and structure-based reranking. First, we generated 1,000 RBX1 binder candidates using RFDiffusion followed by ProteinMPNN sequence design. These designs were prescreened with AlphaFold2 and ranked using a weighted composite score inspired by findings from Overath et al. (2025), particularly the value of interface-focused confidence metrics for binder prioritization. The AF2 prescreen used the following weights: PAE interaction (0.55), ipTM (0.25), pLDDT (0.10), ipSAE (0.07), LIS (0.02), and ESM-PLL (0.01).

In a second stage, we evaluated the top 30 AF2 candidates with AlphaFold3 and reranked them using an AF3 metric order likewise motivated by trends reported in Overath et al. (2025), especially the strong performance of AF3-derived interface metrics such as ipSAE_min and LIS. Our AF3 reranking order was: af3_ipsae_min, af3_lis, af3_chain_pair_iptm, af3_pae_interaction, and af3_plddt_binder. In practice, this AF3 subset also contained all designs with AF2 ipTM ≥ 0.52.

Our hypothesis was that a broad generative search followed by progressively stricter structural reranking would enrich for binders with confident interfaces, favorable local packing, and robust complex formation. The final submission therefore combines the AF3-reranked top 30 with the next-best AF2-ranked candidates to complete the top 100, balancing higher-fidelity AF3 rescoring with broader coverage of the AF2-screened design space.