Identifying structural states that reveal druggable pockets is central to structure-based drug design. However, the intrinsic flexibility of biomolecules makes these conformations difficult to pinpoint. Molecular dynamics (MD) simulations offer a powerful way to capture this flexibility, but the vast amount of high-dimensional data they generate is challenging to interpret directly. Clustering provides an effective strategy to simplify these data sets, enabling the construction of kinetic models, the mapping of conformational pathways, and the identification of representative ensembles. Existing clustering approaches, however, face a critical trade-off; k -means offers speed and scalability but restricts states to rigid centroid-based regions, while hierarchical agglomerative clustering (HAC) can capture more complex geometries at the cost of quadratic time and memory requirements. To address this, we introduce the hierarchical extended linkage method (HELM), a hybrid framework that combines the efficiency of k -means with the flexibility of hierarchical clustering. HELM utilizes n -ary functions to robustly define and stabilize local micro-clusters before efficiently assembling them into a complete hierarchy. We applied HELM to protein-DNA binding and protein-folding systems. Remarkably, it enabled full clustering of a ∼1.5-million-frame HP35 folding trajectory: the largest trajectory of its kind processed in a single run using standard hardware. HELM yields robust cluster counts supported by Davies-Bouldin and Calinski-Harabasz indices, while dramatically reducing runtime and memory demands. By efficiently recovering metastable states and representative conformations, HELM bridges computational efficiency with structural insight, offering a powerful tool to accelerate drug discovery workflows. HELM is freely available as part of the MDANCE package (https://github.com/mqcomplab/MDANCE).
Santos et al. (Sun,) studied this question.