Unsupervised random forests for interpretable representation analysis of protein conformational landscapes

Understanding biomolecular function from molecular simulations requires extracting low-dimensional, interpretable representations from high-dimensional trajectories. While recent machine learning techniques have improved representation learning, they often rely on supervision, extensive hyperparameter tuning, or yield opaque latent embeddings. Here, we introduce unsupervised random forest (URF), a self-supervised adaptation of random forests for interpretable and scalable feature selection from unlabeled simulation data. URF assigns feature importances based on internal structure without requiring labels or predefined tasks. An internal metric learning coefficient enables autonomous quality assessment and parameter selection. Through a memory-efficient implementation, URF scales to large-scale datasets. Across 10 systems, ranging from folded to disordered proteins, URF matches or exceeds the performance of about 15 baseline approaches. The features selected by URF not only provide physical insight but also enhance downstream analyses such as Markov model. URF, thus, offers a general-purpose, interpretable, and unsupervised framework for representation learning in biomolecular simulations.