FlowDyn: Diffusion-Based Generative Modeling of Protein Conformational Ensembles for Structural Biology

Proteins rarely exist as static entities; instead, they explore dynamic conformational ensembles that underpin function, regulation, and interactions. Current machine learning approaches focus primarily on predicting a single static structure, overlooking the conformational variability critical to biological activity. We introduce FlowDyn , a diffusion-based generative framework for modeling protein conformational ensembles directly from sequence and structural priors. FlowDyn leverages denoising diffusion probabilistic models in internal coordinate space, enhanced with geometric constraints and energy-informed regularization, to generate diverse yet stereochemically valid conformations. Benchmark studies show that FlowDyn captures functionally relevant transitions such as loop rearrangements, domain shifts, and allosteric motions, while maintaining bond and angle fidelity. Comparative analysis against AlphaFold and molecular dynamics simulations demonstrates that FlowDyn offers a unique balance between diversity and physical plausibility. Furthermore, the learned latent space trajectories provide interpretable pathways connecting structural states. These capabilities enable downstream applications in flexible docking, allosteric mechanism discovery, and variant effect prediction. By explicitly modeling protein dynamics, FlowDyn advances structural machine learning beyond static prediction toward a richer, ensemble-centered understanding of biomolecular behavior.