1/ 🧪 a de novo enzyme that oxidizes one chosen C-H bond on a steroid: @UWproteindesign, @OfficialUoM and collaborators designed heme hydroxylases that hit estradiol at either the C6 (93% d.e.) or C16 (99% d.e.) position. 👇
1
4
5/ ⚠️ a bioRxiv preprint, not peer reviewed. yields and turnover are still low for real synthesis, and this is in vitro with purified enzyme and H2O2. crystal structures, to 1.9 Å, did match the designed active sites closely.
1
1
6/ ✅ preprint: doi.org/10.64898/2026.11.12.… for synthesis chemists: would a designed enzyme at ~37% yield be worth it to reach a site classical catalysts can't?
1/ 🧬 can a protein folding model tell a real antibody binder from a lookalike decoy? a new @Roche Prescient Design benchmark says better than chance, far from reliable. it ships as an open dataset built from patents. 👇
1
1
10
5/ ⚠️ a bioRxiv preprint, not peer reviewed. labels come from patent claims, not fresh assays; decoys are constructed, not measured non-binders. epitope localization stayed weak: the best model touched annotated epitope residues in only 62.8% of cases.
1
2
6/ ✅ paper + dataset: doi.org/10.64898/2026.10.24.… if you screen antibodies on folding confidence: what hit rate would make a zero-shot digital screen worth it for you?
2
1/ 🧬 a new structure model reports 79.07% on FoldBench antibody-antigen interfaces vs 47.90% for AlphaFold 3: NovaAtom-Lite from @Geodesiclab, benchmarked by the company. 👇
2
21
5/ ⚠️ the company's own release and run, not peer reviewed. baselines come from other groups' reports, scored on differing target counts (160 to 172 on antibody-antigen). access is through Geodesic's API and NovaDDE workspace; the release doesn't mention open weights.
1
4
6/ ✅ the release: geodesiclab.com/news/introdu… for antibody work: would you re-run your own targets on it before switching models?
7
1/ 🧪 9 of 19 AI-designed peptides bound their target in the lab, and 2 beat the natural ligand's apparent affinity. @try_litefold's LiteMol-1 designed them from the BHRF1 protein sequence alone. 👇
1
1
8
5/ ⚠️ a company technical report, not peer reviewed: one target, one capture assay, apparent affinities only, no cell or function data. two positives carry poor-fit flags, and BLI signal size did not consistently track apparent KD.
1
6
6/ ✅ report: litefold.ai/research/bhrf1 sequences and measurements are planned for @proteinbase. for peptide campaigns: do you trust co-folding scores enough to rank designs before a first assay?
