EscalNet: Learn isotropic representation space for biomolecular dynamics based on effective energy

Abstract

Deep learning has emerged as a powerful framework for analyzing biomolecular dynamics trajectories, enabling efficient representations that capture essential system dynamics and facilitate mechanistic studies. We propose a neural network architecture incorporating Fourier Transform analysis to process trajectory data, achieving dual objectives: eliminating high-frequency noise while preserving biologically critical slow conformational dynamics, and establishing an isotropic representation space through the last hidden layer for enhanced dynamical quantification. Comparative protein simulations demonstrate our approach generates more uniform feature distributions than linear regression methods, evidenced by smoother state similarity matrices and clearer classification boundaries. Moreover, by using saliency score, we identified key structural determinants linked to effective energy landscapes governing system dynamics. We believe that the fusion of neural network features with physical order parameters creates a robust analytical framework for advancing biomolecular trajectory analysis.