EscalNet: Learning Isotropic Representation Space for Biomolecular Dynamics Based on Effective Energy
Year: 2026
Type: article
Abstract: Deep learning has substantially advanced the analysis of biomolecular simulation data by providing compact yet informative representations that capture the essential dynamics of a system and expose the underlying mechanisms. Here we combine a neural network with Fourier analysis of trajectory data toward two objectives: preserving the biologically relevant low-frequency components of the motion and constructing an isotropic representation space in which conformational transitions are described by a Markov model. Applied to two Chignolin folding trajectories, the EscalNet (Effective Energy rescale net) representation space consistently gives better clustering and produces a Markov transition matrix that passes a Chapman–Kolmogorov test. Saliency scores further identify the structural determinants that govern the effective-energy landscape underlying the dynamics. Coupling neural-network representations with physical order parameters thus offers a promising framework for biomolecular trajectory analysis.
Author: Guanghong Zuo
Institution: University of Chinese Academy of Sciences
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
Award: G9001394247
OpenAlex ID: W7106705719
