Answers are grounded in this paper and cite their sources.
Deep-learning tools for peptide design and structure prediction are increasingly combined into screening workflows, but at the Keap1–Nrf2 interface such work has rarely progressed beyond computational characterization. We report a case study in which an initial-round workflow built from publicly available components─ProteinMPNN, Chai-1, PLIP, and HADDOCK3─was fixed before any experimental data were obtained and its complete prioritized output was carried to synthesis and measurement: 530,299 unique sequences from seven Nrf2-derived backbones were reduced to 5,118 by sequence filtering and to 28 by structure-based filtering, and all 28 were synthesized and assayed alongside 16 analogues and the native Nrf2 12-mer, giving 45 measured peptides. One design, a 12-mer carrying an ENGE variant of the ETGE motif, inhibited the interaction with an apparent IC50 of 120 nM under the assay conditions used, whereas the 14 resampled analogues satisfying the same structural criteria showed low inhibition, indicating that within this workflow the structure-based steps acted as filters for structural plausibility rather than as predictors of activity. Because the set was measured in full and the intermediate sequence sets were retained, we also report how the composition of the surviving population changed across the cascade; the original workflow data are released.
No takes yet. Share an insight, caveat, or question.
Yoon et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: