2 sequences threaded per backbone (3,000 × 2 = 6,000) at sampling_temperature 0.1 — low on purpose. RFdiffusion already supplied structural diversity across 3,000 backbones, so stage F doesn't need much extra sequence diversity per backbone; a low temperature keeps ProteinMPNN close to its per-position argmax, which packs better and folds more confidently than the noisier sequences a higher temperature (0.2-0.3) would give. SolubleMPNN weights bias against hydrophobic surface positions specifically because E. coli cytoplasmic expression (stage I) has no chaperone help for a sticky exterior — see PD-L1/Note/Expression results for what happened to the designs that still ran hot on surface hydrophobicity anyway. Reference/ProteinMPNN 2022.
Stage F of PD-L1/Campaign/PDL1-R1.