TRAILS-MD: Mapping complex conformational landscapes and transition pathways via lightweight lineage-aware adaptive sampling.

Abstract

Molecular dynamics (MD) simulations are fundamentally limited by the timescale gap between rapid atomic fluctuations and slow, activated conformational transitions. While adaptive sampling accelerates the exploration of metastable states, conventional implementations often obscure the dynamical connectivity between discovered basins, confusing broad configurational coverage with the sampling of physically viable transition pathways. Here, we present a lightweight, engine-agnostic adaptive sampling framework (TRAILS-MD) that iteratively launches short trajectory ensembles across either fixed or on-the-fly machine-learned (ML) collective-variable (CV) spaces. A defining feature of our approach is explicit trajectory lineage preservation, which allows for the rigorous reconstruction of connected pathways from highly parallelized, disjointed exploration stages. We demonstrate the efficacy of this framework across a hierarchy of molecular challenges: overcoming low-dimensional barriers in peptide isomerization, navigating entropic bottlenecks in protein folding (Trp-cage and WW domain), mapping transient exit tunnels during ligand unbinding (HSP90), and exploring large conformational changes in adenylate kinase. Finally, we show how lineage-aware exploratory ensembles can seamlessly seed unbiased production trajectories for the construction of Markov state models (MSMs). By bridging the gap between deep-learning-driven coordinate discovery and strict parent-child trajectory tracking, this framework democratizes advanced data-driven sampling for routine computational biophysics workflows.