Protein folding remains a formidable challenge despite significant advances, particularly in sequence-to-structure prediction. Accurately capturing thermodynamics and intermediates via simulations demands overcoming time scale limitations, making effective collective variable (CV) design for enhanced sampling crucial. Here, we introduce a strategy to automatically construct complementary, bioinspired CVs. These uniquely capture local hydrogen bonding─explicitly distinguishing protein-protein from protein-water interactions─and side-chain packing, taking into account both native and non-native contacts to enhance state resolution. Using these CVs in combination with advanced enhanced sampling methods, we simulate the folding of Chignolin and TRP-cage, validating our approach against extensive unbiased simulations. Our results accurately resolve complex free-energy landscapes, reveal critical intermediates such as the dry molten globule, and demonstrate agreement with reference data. This interpretable and portable strategy underscores the critical role of microscopic details in protein folding, opening up a promising avenue for studying larger and more-complex biomolecular systems.
Rizzi et al. (Mon,) studied this question.