Wavelet-Enhanced Data-Driven Collective Variables for Efficient Sampling of Protein Folding Landscapes

The development of collective variables (CVs) capable of distinguishing important (meta)stable states and describing slow degrees of freedom (DOFs) is widely recognized as a prerequisite for the success of CV-based enhanced sampling methods, particularly in biomolecular systems where the underlying free-energy landscapes are extremely complex. For complex biochemical processes, selecting suitable geometric CVs for enhanced sampling often relies on chemical intuition, which is challenging and system-dependent. While data-driven CVs have emerged as a promising approach and achieved significant success in recent years, studies indicate that their performance is highly susceptible to fast stochastic fluctuations in molecular simulation trajectories, thereby obscuring the identification of slow DOFs. To overcome this limitation, we turn to a feature extraction strategy that utilizes the discrete wavelet transform to filter out fast-mode motions from MD trajectories, followed by dimensionality reduction of the descriptors. This refined feature subset provides an optimal input for constructing data-driven low-dimensional CVs, allowing for the identification of biologically relevant slow modes with high fidelity. We validate the efficacy of this strategy through a comparative analysis of chignolin and further showcase its applicability to the structurally more complex BBA protein system, which features both helical and antiparallel β-sheet motifs. The results reveal that the CVs derived from this strategy facilitate rapid transitions between folded and unfolded states, thereby significantly accelerating the exploration of protein folding landscapes across systems of varying scales.