Why 627 BindCraft2 trajectories produced zero binders

Abstract

De novo protein-binder design with deep-learning pipelines such as BindCraft2 has been reported to deliver experimental success rates of 10-100% against well-structured targets. Here we report the opposite tail: two BindCraft2 (v1.0.3) cyclic-peptide campaigns targeting the shallow K81 lactylation epitope of human lactate dehydrogenase A (LDHA) produced zero accepted designs from 627 trajectories. Retrospective analysis of the campaign artifacts shows that 97% of trajectories were rejected at the first ("screen") stage by a single confidence gate (binder pLDDT < 0.60), not because the interface failed to form: the interface metric i_pTM reached 0.69 and the target was confidently folded (pTM = 0.90). Critically, the mid-project decision to disable the "lysine trick" (forced_targeting) was backwards -- that feature moved 17 trajectories past the screen versus 1 without it, and raised interface confidence. We document two preset/metric mismatches that set a macrocycle campaign up to fail regardless of budget: a screen-stage pLDDT gate calibrated for globular binders, and a binder_contacts loss that is degenerate for short peptides. We describe the exact artifacts BindCraft2 writes and distill them into a reusable post-mortem protocol, and we include a complete how-to-use-BindCraft2 guide so the failure analysis also serves as a tutorial.