Target fixed at A19-127 (the IgV domain, not the whole ectodomain — see PD-L1); binder length sampled per-design over 55-65 aa rather than fixed at one length, since a hard length cap biases the backbone search toward whatever topology happens to fit that exact span. Hotspots (A54, A56, A115, A121, A123) bias the fold toward contacting that patch without pinning a single binding mode — RFdiffusion still explores multiple approach angles across the 3,000 backbones. 3,000 is oversampling on purpose: hit rate after the full filter-and-wet-lab funnel is typically low single-digit percent, and backbone generation is the cheapest stage to buy volume at. Output is Cα/backbone frames only — no side chains, no sequence — so PD-L1/Campaign/PDL1-R1/Run/MPNN-001 threads the sequence next. Reference/RFdiffusion 2023; the underlying strategy traces back to Reference/Cao 2022.
Stage E of PD-L1/Campaign/PDL1-R1.