Efficient and Accelerated Protein Free Energy Landscape Reconstruction via Conditional Variational Autoencoders

Accurate free energy landscape (FEL) construction is vital for deciphering protein function but is hindered by computational bottlenecks and sampling inefficiencies, often leading to unreliable state characterization. To address this bottleneck, we introduce a conditional variational autoencoder (CVAE)-based framework integrating dimensionality reduction, clustering, and data balancing. By learning conformational distributions in low-dimensional latent space and adjusting weights for undersampled data, our approach enables more accurate FEL estimation and rapid optimization. Our method significantly reduces computational time compared to traditional enhanced sampling, while maintaining comparable accuracy and offering greater generalizability across different protein systems. Validation on four systems, chignolin, adenylate kinase, ribose-binding protein, and c-Abl tyrosine kinase, representing a diverse range of molecular complexity, demonstrates robust FEL resolution across scales. This adaptable, component-based approach makes FEL analysis more accessible for complex molecular systems, facilitating rapid investigation of protein dynamics and creating new opportunities for therapeutic development.