Cyclic peptide ML hits limits: permeability is learnable, affinity and half-life are not
- A broad retrospective benchmark compares many model families and molecular representations on curated public cyclic-peptide data covering permeability, binding and half-life.
- Permeability is the most learnable endpoint; affinity and half-life stay hard. Tree ensembles do well for regression, pretrained chemical language models for classification, with no universal winner.
- Results depend heavily on split strategy, task framing and dataset makeup, and no-skill baselines are essential to read them. An open framework is released. Preprint, not peer reviewed.
💡 Points Based on the abstract
Before trusting any cyclic peptide model, ask how the split was drawn. Random splits across near-identical scaffolds from one lab's series can flatter permeability scores, and that endpoint looks learnable partly because assays are consistent. For your own project, hold out whole chemical series, and compare against a no-skill baseline on your assay before committing synthesis effort.
Assessing the predictive limits of machine learning on cyclic peptides ↗