Generating the polymorph landscapes of amyloid fibrils using AI RibbonFold

• 成果类型: Article
• 署名作者: Guo, Liangyue; Yu, Qilin; Wang, Di; Wu, Xiaoyu; Wolynes, Peter G.; Chen, Mingchen
• 署名单位: Changping Laboratory; Rice University; Rice University; Rice University; Rice University
• 刊物名称: PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA
• ISSN/ISSBN: 0027-12926
• DOI: 10.1073/pnas.2501321122
• 发表日期: 2025-04-22
• 关键词: protein-structure prediction; energy landscape; cryo-em; prion; aggregation; kinetics; diseases; LEVEL

摘要

The concept that proteins are selected to fold into a well-defined native state has been effectively addressed within the framework of energy landscapes, underpinning the recent successes of structure prediction tools like AlphaFold. The amyloid fold, however, does not represent a unique minimum for a given single sequence. While the cross-/3 hydrogen-bonding pattern is common to all amyloids, other aspects of amyloid fiber structures are sensitive not only to the sequence of the aggregating peptides but also to the experimental conditions. This polymorphic nature of amyloid structures challenges structure predictions. In this paper, we use AI to explore the landscape of possible amyloid protofilament structures composed of a single stack of peptides aligned in a parallel, in-register manner. This perspective enables a practical method for predicting protofilament structures of arbitrary sequences: RibbonFold. RibbonFold is adapted from AlphaFold2, incorporating parallel in-register constraints within AlphaFold2's template module, along with an appropriate polymorphism loss function to address the structural diversity of folds. RibbonFold outperforms AlphaFold2/3 on independent test sets, achieving a mean TM-score of 0.5. RibbonFold proves well-suited to study the polymorphic landscapes of widely studied sequences with documented polymorphisms. The resulting landscapes capture these observed polymorphisms effectively. We show that while well-known amyloid-forming sequences exhibit limited number of plausible polymorphs on their solubility landscape, randomly shuffled sequences with the same composition appear to be negatively selected in terms of their relative solubility. RibbonFold is a valuable framework for structurally characterizing amyloid polymorphism landscapes.