3DDF-VAE: Dual-frequency variational autoencoder with pose-consistency validation for rare cryo-EM conformation discovery

TY - JOUR

T1 - 3DDF-VAE

T2 - Dual-frequency variational autoencoder with pose-consistency validation for rare cryo-EM conformation discovery

AU - Chen, Yuanbo

AU - Li, Fuwei

AU - Dong, Hao

AU - Wang, Xinsheng

AU - Zhang, Fa

AU - Hu, Bin

AU - Wan, Xiaohua

N1 - Publisher Copyright: © 2026 Elsevier Inc.

PY - 2026/6

Y1 - 2026/6

N2 - Revealing the 3D conformational variability of biomolecules is crucial for understanding their function, while cryo-EM reconstruction of rare states remains difficult due to data imbalance and structural detail loss in existing generative models. We present a dual-stage pipeline consisting of a generative stage and a validation stage. In the generative stage, we employ a 3D dual-frequency variational autoencoder (3DDF-VAE) that separately models low- and high-frequency components of protein density maps to enhance global coherence and local structural detail. In the validation stage, a pose-consistency projection strategy evaluates the generated maps by comparison with the original 2D particles. Experiments on integrin αVβ8, T50S ribosome, and SARS-CoV-2 spike datasets demonstrate that our method produces high-quality density maps, identifies rare conformations, and reconstructs plausible intermediates, while ablation studies confirm the benefits of frequency separation and parameter optimization. This integrated generative–validation framework improves resolution, enhances rare conformation detection, and offers a data-driven approach to explore conformational heterogeneity in complex biomolecular systems.

AB - Revealing the 3D conformational variability of biomolecules is crucial for understanding their function, while cryo-EM reconstruction of rare states remains difficult due to data imbalance and structural detail loss in existing generative models. We present a dual-stage pipeline consisting of a generative stage and a validation stage. In the generative stage, we employ a 3D dual-frequency variational autoencoder (3DDF-VAE) that separately models low- and high-frequency components of protein density maps to enhance global coherence and local structural detail. In the validation stage, a pose-consistency projection strategy evaluates the generated maps by comparison with the original 2D particles. Experiments on integrin αVβ8, T50S ribosome, and SARS-CoV-2 spike datasets demonstrate that our method produces high-quality density maps, identifies rare conformations, and reconstructs plausible intermediates, while ablation studies confirm the benefits of frequency separation and parameter optimization. This integrated generative–validation framework improves resolution, enhances rare conformation detection, and offers a data-driven approach to explore conformational heterogeneity in complex biomolecular systems.

KW - Conformational heterogeneity

KW - Frequency separation

KW - Rare conformation

KW - Single-particle analysis

KW - Variational autoencoder

UR - https://www.scopus.com/pages/publications/105033601610

U2 - 10.1016/j.jsb.2026.108311

DO - 10.1016/j.jsb.2026.108311

M3 - Article

C2 - 41819391

AN - SCOPUS:105033601610

SN - 1047-8477

VL - 218

JO - Journal of Structural Biology

JF - Journal of Structural Biology

IS - 2

M1 - 108311

ER -