CalphaEBM: Protein Cα Energy-Based Model

CalphaEBM is a physics-based, machine-learned modular energy function for Cα protein coordinates that enables stable Langevin dynamics simulation. The model decomposes the effective free energy (potential of mean force) into four interpretable terms.

Key Features

• Modular energy: Local geometry, excluded volume, secondary structure, and packing
• Langevin dynamics: Stable conformational sampling simulation
• Denoising score matching: Train without partition function
• Safety features: Smooth switching, bounded radii, force caps
• Multiple output formats: DCD, NPY, PT, PDB for compatibility with MD tools