BPS2026 – Fourier-Net: A “one-stop shop” for CV generation and featurization

Advances in computational biology and biophysics have been driven by molecular simulation methods that probe the conformational landscapes of biomolecules. Among these, molecular dynamics (MD) integrates equations of motion to generate dynamical trajectories that capture time-dependent behavior. However, conventional MD struggles to sample states separated by large energy barriers. Enhanced sampling methods address this limitation by biasing the simulation along carefully chosen collective variables (CVs) that represent key features of the system. The quality of a CV critically determines sampling efficiency: a poorly defined CV may fail to distinguish metastable basins or resolve transition states. Machine learning (ML) approaches have recently emerged to automate CV discovery, surpassing many intuitive choices. Yet, they often suffer from limited interpretability or rely on user-defined input features, requiring prior knowledge of the system. To overcome these challenges, we introduce Fourier-Net, a neural network framework that generates CVs from short classical MD trajectories within metastable basins. Fourier-net combines rigorous data preprocessing with a hybrid data-driven and physics-informed architecture, enabling the automated extraction of CVs even for large, complex systems without predefined features. Its design also affords direct interpretability, revealing key nonlinearities within a singular function, effectively operating as a featurizer. Moreover, Fourier-Net’s tailored loss function produces a multi-task CV capable of simultaneously distinguishing metastable basins and resolving transition states—a first for CV generation methods. Using alanine dipeptide in both vacuum and solvent as initial benchmarks, we demonstrate that Fourier-Net generates one-dimensional CVs that are more informative than those produced by state-of-the-art alternatives, highlighting its potential as a comprehensive solution for enhanced sampling and biomolecular feature extraction.