Learning conformational ensembles of proteins based on backbone geometry
Existing protein conformational sampling methods—relying either on evolutionary information or pretrained folding models—suffer from limited applicability, low efficiency, and potential biases. To address these limitations, we propose BBFlow, the first flow-matching generative model that operates exclusively on backbone geometric structure, requiring neither evolutionary sequence information nor pretrained models, and directly learns a conformational ensemble consistent with the Boltzmann distribution from scratch. BBFlow innovatively employs equilibrium backbone geometry both to condition the vector field and to define a learnable SE(3)-equivariant prior distribution, enabling robust modeling of multi-chain proteins and de novo design. Compared to state-of-the-art methods, BBFlow achieves orders-of-magnitude faster training (converging in GPU-days) and significantly accelerated inference, while maintaining competitive performance on both native protein reconstruction and de novo design benchmarks.