Hannah Wastyk

Non-Canonical Crosslinks Confound Evolutionary Protein Structure Models - Evolution-based protein structure prediction models like AlphaFold and RoseTTAFold have revolutionized structural biology but struggle with rare post-translational modifications (PTMs). This study evaluates their limitations in predicting sactipeptides, a class of ribosomally synthesized and post-translationally modified peptides (RiPPs) with unique sulfur-to-α-carbon thioether crosslinks. - Sactipeptides are particularly challenging because they contain non-canonical crosslinks that are underrepresented in existing structural datasets, making them an ideal benchmark for testing how well deep learning models generalize beyond evolutionary priors. - The study introduces a new zero-shot benchmark for evaluating the ability of structure prediction models to correctly predict the 3D conformations of sactipeptides, specifically the geometry of their thioether crosslinks. - Six leading protein structure predictors—AlphaFold 2, AlphaFold 3, Boltz-1, ESMFold, OmegaFold, and RoseTTAFold 2—were tested on 10 known sactipeptides, five of which have experimentally resolved structures. - Results show that all models exhibit poor performance in correctly predicting sactibonds, with an average GDT-TS of only 11.5% for known sactipeptides and 12.6% for unknown ones. None of the models achieved crosslink distances near the experimentally observed 1.8 Å sulfur-to-α-carbon bond length. - Boltz-1 and RoseTTAFold 2 achieved the highest GDT-TS scores, suggesting that some models partially capture local geometric constraints. However, most models incorrectly predicted disulfide bonds instead of the required thioether crosslinks, highlighting their reliance on evolutionary priors. - Visual analysis reveals that models frequently collapse sulfur atoms into unnatural conformations, demonstrating systematic biases in deep learning-based protein structure predictors when faced with non-standard modifications. - The findings suggest that future structure prediction methods must integrate physics-informed modeling and explicitly account for rare PTMs to achieve accurate 3D structure predictions beyond standard evolutionary data. @rlacombe 📜Paper: biorxiv.org/content/10.1101/… #ProteinFolding #StructuralBiology #DeepLearning #Bioinformatics #PostTranslationalModifications #AI