Today's Overview
- Diffusion Model Predicts 3D Conformations of NCAA-Containing Cyclic Peptides at Sub-Ångström RMSD Achieves 0.79 Å average RMSD for NCAA-bearing cyclic peptides after stereochemical correction, versus prior ML tools that often invert chiral centers.
Featured
01Diffusion Model Predicts 3D Conformations of NCAA-Containing Cyclic Peptides at Sub-Ångström RMSD
Cyclic peptides stabilized by non-canonical amino acids (NCAAs) are increasingly pursued as orally available therapeutics, but accurate 3D templates are scarce because existing force-fields and generic ML models mishandle exotic side-chains and stereochemistry. AGDIFF, originally built for small-molecule conformer generation, was retrained on the 36,198-member CREMP macrocyclic-peptide ensemble to address this gap.
The all-atom diffusion model ingests only a 2D graph yet recovers chiral centers through a post-generation stereochemical correction step. On benchmark cyclic peptides the pipeline reaches 0.79 Å average heavy-atom RMSD and 6.55° ring-torsion fingerprint deviation, correctly assigning enantiomeric antipodes and yielding Ramachandran-allowed backbone ensembles. All validation is in silico; neither experimental structures nor downstream bioactivity were tested, and performance outside the CREMP chemistry space (e.g., bicyclics or >13-mers) is unknown.
Also Worth Noting
HyperLab’s SBDD pipeline—77 % PoseBuster v2 pose accuracy, 0.70/0.53 affinity correlation—screens 1 M–7 T compounds and, without human post-filtering, delivers five validated hits (IC50 70–600 nM) and three optimized analogs with 200–400 nM IC50 in vitro. link
A MACCS-fingerprint Random Forest QSAR model filtered an external library for in-domain SMYD3 inhibitors, and 250 ns MD singled out CHEMBL4472528 as a stable binder engaging the crystallographic active site. link
Mini-review synthesizes how AI frameworks that integrate heterogeneous high-throughput datasets uncover composite biomarkers, regulatory hubs, and druggable vulnerabilities alongside traditional experimental methods to address metastasis, chemoresistance, and heterogeneity in osteosarcoma. link
Today's Observation
Cyclic peptides that incorporate non-canonical amino acids (NCAA) are a fast-growing modality for disrupting flat protein interfaces, but design cycles collapse when force-fields or earlier ML tools mis-assign chirality or cannot represent exotic linkers. The diffusion conformer generator described today sidesteps these pitfalls by treating the peptide as a full 2-D molecular graph; after a simple stereochemical re-filter it reaches 0.79 Å average RMSD against the CREMP database, a ∼0.4 Å improvement over the best sequence-based neural method and essentially removes inverted chiral centers. Because no fragment library or Ramachandran grid is imposed, thio-ether, D-amino acid, and N-methyl variants are sampled natively, giving medicinal-chemistry teams a quick way to rank macrocycle geometries before docking or free-energy calculations.
Practitioners should note that the benchmark covers only 1,300 curated CREMP conformers, all under 14 residues; larger or head-to-tail cyclized structures may behave differently, and no experimental NMR or X-ray re-docking is shown. For now the model is an in-silico geometries filter: it will accelerate enumeration, but downstream stability, permeability, and target binding will still need standard assays.
The above is personal commentary for reference only. Refer to the original papers for authoritative content.