Understanding the three-dimensional (3D) structure of biomacromolecules is essential for elucidating their biological functions (Cheng, 2018, Zhu et al., 2023). Single-particle cryo-electron microscopy (cryo-EM) enables the high-resolution reconstruction of protein structures in near-native states (Grant et al., 2018, Scheres, 2016, Scheres et al., 2007). In typical cryo-EM datasets, most particles correspond to a few dominant conformations, which can be readily reconstructed due to their high abundance. However, beyond these common conformations, biomacromolecules may assume rare conformations that, despite their low occurrence, are crucial for mediating molecular interactions or acting as functional intermediates. Accurately identifying and characterizing these rare states is therefore crucial for achieving a comprehensive understanding of biomolecular mechanisms. However, their low representation and subtle structural deviations in cryo-EM datasets pose significant challenges for reliable detection and reconstruction (Toader et al., 2023).

Traditional cryo-EM data processing methods, such as 3D classification (Tang et al., 2007, Punjani et al., 2017), are mainly designed to identify dominant structural states by grouping particles into discrete classes based on similarity (Scheres, 2012). While effective for resolving stable conformations, this approach often fails to capture rare conformational states, which are underrepresented in the dataset and poorly separated in feature space (Lederman et al., 2020). As a result, flexible or low-population regions in the final density maps often appear blurred or ambiguous (Kaur et al., 2021), limiting their utility in understanding dynamic molecular behavior.

With the advent of deep learning, a series of data-driven generative approaches have been proposed for conformational modeling (Zhong et al., 2021a, Punjani and Fleet, 2023, Schwab et al., 2024). These methods typically employ a single neural network to directly infer the underlying 3D conformational landscape from 2D cryo-EM images, allowing the generation of rare states from continuous trajectories. In parallel, post-processing strategies (Sanchez-Garcia et al., 2021, He et al., 2023, Liu et al., 2025) have been developed to enhance the visual quality of generated density maps, with the aim of recovering missing high-frequency structural details and improving the resolution of fine features. However, neural networks tend to prioritize learning low-frequency components during training (Wang et al., 2020), resulting in loss of high-frequency details and thus limiting the reconstruction fidelity. Furthermore, the accuracy and reliability of the generated rare conformations remain difficult to verify because of the lack of direct validation mechanisms against experimental data.

To address the challenge of discovering rare conformations, we propose a novel pipeline for protein 3D density map generation and validation, aiming to improve the modeling accuracy and reliability for rare conformations. Specifically, our method features three key innovations:

We introduce a frequency domain separation mechanism and construct a 3D dual-frequency variational autoencoder (3DDF-VAE) model to separately model high-frequency and low-frequency features, thereby enhancing the resolution and detail accuracy of generated structures from a multiscale perspective.

We design a pose-consistency validation strategy based on original pose information, enabling similarity comparison between generated density maps and original 2D cryo-EM images under unified viewing angles to quantify their authenticity and data consistency.

We establish a dual-frequency generation and pose-consistency validation pipeline, covering the full workflow from 3D density map generation and pose-consistency projection to structural scoring and ranking. This design enables systematic exploration of rare conformations and provides a unified structure for generative modeling and validation.

To assess the reliability of our pipeline, we first performed a semi-synthetic validation experiment on the integrin αVβ8 complex (Campbell et al., 2020), where the ground truth conformations are explicitly available. The results show that our approach produces accurate and detailed density maps. Furthermore, the projection-based validation framework consistently pinpoints the most plausible candidates from a large pool of generative samples, underscoring the reliability of our scoring strategy. We then extended our analysis to two real cryo-EM datasets, each exemplifying a different mode of conformational heterogeneity. For the T50S ribosome (Davis et al., 2016), which features discrete assembly intermediates, our model revealed several rare conformations with distinct domain differences. In the case of the SARS-CoV-2 Spike protein (Benton et al., 2021), despite the limited data coverage, the method facilitated the reconstruction of potential intermediate states by leveraging external structural references. Collectively, these findings support the versatility and robustness of our pipeline in discovering rare conformations.