WorksComputer Science
CryoGaussian: Heterogeneous Cryo-EM Reconstruction Made Explicit with Gaussian Splatting
Abstract
Single-particle cryo-electron microscopy (cryo-EM) determines macromolecular structures from noisy particle images. However, rare, transient compositional states and subtle, continuous conformational variation are difficult to resolve. We present CryoGaussian, a Gaussian splatting method for heterogeneous cryo-EM reconstruction that fits an explicit Gaussian density representation directly to the particle images through a differentiable rasterizer rendering orthographic line integrals modulated by the contrast transfer function. The Gaussians are initialized randomly, with their number derived from the estimated spatial support of the molecular density and their minimum size bounded by the Nyquist frequency of the particle images. To capture the compositional and conformational variation across particle images, we learn particle-image and per-Gaussian latents that jointly modulate the base position, amplitude, and scale of each Gaussian. Each structural state is then reconstructed by fitting the Gaussians conditioned on its inferred latent. On CryoBench, CryoGaussian achieves state-of-the-art reconstruction accuracy for both compositional and conformational heterogeneity. On experimental data, it resolves a rare compositional state of the 50S ribosomal subunit (class C4, 1.3% of particles) with rRNA helix 68 docked in the absence of the central protuberance, and continuous conformational motion of the SF3b and helicase subcomplexes of the pre-catalytic spliceosome. We further optimize the rasterizer for additive cryo-EM projection, reducing training time by 12.8-28.0.
Introduction
Single-particle cryo-electron microscopy (cryo-EM) reconstructs three-dimensional biomolecular density at near-atomic resolution from noisy two-dimensional particle images [1, 2]. Resolving compositional and conformational heterogeneity reveals molecular organization and function, as illustrated by respiratory supercomplexes and proteasome states [3, 4]. Biological interpretation requires both reliable structural detail and an account of the compositional and conformational states of the imaged biomolecules.
Each particle image is a noisy projection of an unknown density modulated by the contrast transfer function (CTF), and the structural state underlying each particle image is also unknown. Signal-to-noise ratios (SNRs) as low as 0.01 obscure fine features [6], while overfitting can introduce misleading density [7]. Discrete classification can miss rare, transient states or average over continuous variation [8, 9]. Density-preserving deformation models including 3DFlex capture motion but are ill-suited to compositional changes [9]. Other continuous representations accommodate compositional and conformational heterogeneity [10–12], yet resolving heterogeneity alone does not ensure faithful per-state reconstruction: even with identical latents, reconstruction procedures recover different structural detail [11]. Accurate heterogeneity estimation and high-quality per-state reconstruction pose distinct challenges.
Gaussian mixtures provide physically motivated representations of molecular density and structural variation [13–15], and 3D Gaussian Splatting (3DGS) makes explicit Gaussians computationally attractive through differentiable rasterization [16]. Cryo-EM, however, records orthographic line integrals modulated by the CTF, so projected Gaussians add without occlusion and require a rasterizer adapted to cryo-EM image formation [17]. CryoSplat adapts Gaussian splatting to this projection model for homogeneous reconstruction, the special case of a single structural state, from randomly initialized Gaussians [5], whereas Gaussian-based heterogeneity methods such as e2gmm, DynaMight, CryoSPHERE, and GaussianEM still require initialization from a reference map or atomic structure [15, 18–20]. These limitations motivate us to employ Gaussian splatting, free of structural priors and grounded in the physics of cryo-EM image formation, to achieve both accurate heterogeneity estimation and high-quality per-state reconstruction.
We introduce CryoGaussian, a Gaussian splatting method for heterogeneous cryo-EM reconstruction that fits an explicit Gaussian density representation directly to particle images. The Gaussians are initialized randomly, without prior structural information, with their number derived from the estimated spatial support of the molecular density and their minimum size bounded by the Nyquist frequency of the particle images. To estimate compositional and conformational heterogeneity, we learn particle-image and per-Gaussian latents that jointly modulate the base position, amplitude, and scale of each Gaussian. To reconstruct each structural state, we fit the latent-conditioned Gaussians at its inferred latent.
Main contributions. (i) We introduce CryoGaussian, which fits an explicit Gaussian density representation directly to particle images from random initialization under data-derived constraints. (ii) We model both compositional and conformational heterogeneity with latent-conditioned Gaussians and reconstruct each structural state from the inferred latents. (iii) We optimize differentiable rasterization for additive cryo-EM projection, whose linearity permits per-Gaussian gradient accumulation and projection scheduling, and distribute training, reducing end-to-end training time by 12.8–28.0× on one GPU and a further 5.1–6.5× on eight GPUs (Table 1). (iv) On CryoBench, we demonstrate accurate heterogeneity estimation and state-of-the-art reconstruction accuracy for both compositional and conformational heterogeneity; on experimental data, we resolve a rare assembly intermediate of the 50S ribosomal subunit (class C4, 1.3% of particles) and continuous motion of the SF3b and helicase subcomplexes of the pre-catalytic spliceosome.
| Dataset | Gradient accumulation | Projection scheduling | 8 GPUs | Time (ms) | Memory saved (%) |
| IgG-1D | 2.39× | 7.69× | 5.41× | 17.31 | 7.98 |
| IgG-RL | 2.27× | 7.96× | 5.79× | 34.33 | 9.56 |
| Ribosembly | 2.46× | 11.37× | 6.48× | 73.89 | 8.97 |
| Tomotwin-100 | 2.62× | 10.23× | 5.14× | 72.90 | 9.74 |
| Spike-MD | 2.38× | 6.68× | 6.16× | 216.49 | 10.17 |
| EMPIAR-10076 | 2.35× | 5.43× | 6.18× | 520.19 | 9.63 |
| EMPIAR-10180 | 2.64× | 8.51× | 5.59× | 376.64 | 9.95 |
Table 1. Training efficiency. On one GPU, CryoGaussian achieves multiplicative speedups over the CryoSplat rasterizer [5] through per-Gaussian gradient accumulation and projection scheduling. On eight NVIDIA RTX PRO 6000 Blackwell GPUs, we measure its further speedup, time per iteration, and memory saved (Table 6).
Methods
CryoGaussian performs heterogeneous reconstruction from particle images with pixel size and dimensions . As is standard in heterogeneous reconstruction [21] [9, 10, 11], the orientations , in-plane translations , and CTFs of the particle images are provided as inputs and held fixed. Figure 1 gives an overview.
Figure 1. CryoGaussian overview. (a) The number of randomly initialized Gaussians is derived from the estimated spatial support of the molecular density (Appendix § 11), and the pixel size sets their minimum size. (b) Separate MLPs decode position, amplitude, and scale modulations of each Gaussian from the concatenated particle-image and per-Gaussian latents (Appendix § 6). (c) Heterogeneity is estimated by jointly learning these latents, and compositional and conformational states are then reconstructed from the inferred latents.
Gaussian density representation and cryo-EM image formation
The biomolecular density for a particle-image latent is
where is the number of Gaussians and is a normalized three-dimensional Gaussian. Each Gaussian has a position , covariance , and amplitude ; comprises the Gaussian parameters, per-Gaussian latents, and decoder weights (§ 2.3).
Under the weak-phase-object and projection approximations, cryo-EM image formation comprises line integration along the electron beam followed by CTF modulation [17]. With selecting the physical image-plane coordinates , the line integral of the rotated density is analytic:
where is a normalized two-dimensional Gaussian. We evaluate these two-dimensional Gaussians on the image grid with spacing and apply CTF modulation:
Here is the discrete Fourier transform and a fixed conversion from projected density to image intensity. Gaussian contributions add independently of depth order, so our rasterizer accumulates them additively instead of the front-to-back alpha compositing of 3DGS [16], as does CryoSplat [5].
Applying the supplied in-plane translations through the Fourier shift theorem gives aligned targets . We minimize the image-domain mean squared error, as in neural and Gaussian-based cryo-EM methods [5, 10, 19]:
where and the inner sum runs over the image grid, subject to the constraints of § 2.2.
Data-derived constraints
The pixel size defines the Nyquist frequency [25]. We use the corresponding sampling scale for the minimum size and number of Gaussians (§ 3.4).
Minimum Gaussian size. Inspired by the sampling-frequency constraint on Gaussian size in Mip-Splatting [26], we bound the full width at half maximum (FWHM) of each Gaussian below by the sampling scale , below which Gaussians add no resolvable detail and fit noise. For every effective principal-axis scale (standard deviation),
Number of Gaussians. Gaussian-mixture cryo-EM methods set the number of Gaussians before optimization [27] [18, 19]. We determine from the pixel size, which bounds the frequency band of the density, and from the spatial support of the density, estimated from a CTF-weighted backprojection of the particle images. For a relative density threshold , the estimated spatial support is the union of the voxel cells of side in which the nonnegative part of is at least times its maximum, including disconnected regions (Appendix § 11.§ 2), and has volume .
The space–bandwidth product characterizes the approximate effective dimension of signals concentrated within a spatial region and a frequency band [28, 29], and in cryo-EM it has been used to set the effective dimension of density representations [30]. We use the rotationally symmetric sampling band , with frequencies in cycles per unit length, and set to the space–bandwidth product of the estimated support (Appendix § 11),
where rounds to an integer. is the effective dimension of densities confined to the estimated support at the sampling scale . We use , which approximates the relative boundary density of the 95-percent mass ellipsoid of a three-dimensional Gaussian (Appendix § 11.§ 3).
The backprojection enters CryoGaussian only through the volume of the estimated support. Gaussian positions are initialized randomly and are not restricted to this support, so no structure of the imaged biomolecules is encoded before optimization, and their structural states, including rare ones (§ 3.2), are inferred directly from the particle images. Inspired by 3DGS-MCMC [31], we fix and periodically relocate low-amplitude Gaussians (Appendix § 6.§ 3).
Heterogeneous reconstruction
Gaussian attribute decoding. Each particle image has a trainable particle-image latent , and each Gaussian has a trainable per-Gaussian latent . Separate multilayer perceptrons (MLPs) decode position, amplitude, and scale modulation from the concatenated latent . The position and amplitude decoders produce
where and are the learned base position and amplitude. The scale decoder modulates the three principal-axis scales , , of the effective covariance , each subject to the FWHM constraint in Equation (5), where is the learned Gaussian rotation (Appendix § 6).
Particle-image latents parameterize the structural state of each particle image, and per-Gaussian latents determine how each Gaussian responds to that state. Each decoder is shared by all Gaussians and all particle images, so that, for a given state, the response of a Gaussian depends only on its per-Gaussian latent. The per-Gaussian latents are optimized jointly with the decoders, which allows Gaussians that vary together to share similar latents without a predefined grouping. Together, the three modulations allow each Gaussian to move, change shape, and appear or disappear across states, so that compositional and conformational variation are represented within one set of Gaussians.
Heterogeneity estimation and per-state reconstruction. Heterogeneous reconstruction comprises two problems under the objective in Equation (4), each evaluated in § 3.1. Heterogeneity estimation, , yields the structural state of each particle image by optimizing the latents jointly with the parameters of randomly initialized Gaussians, which are fitted while the latents change. Per-state reconstruction, , yields the density of each state by fitting randomly initialized latent-conditioned Gaussians, whose parameters include their per-Gaussian latents and decoders, at the inferred latents under the same data-derived constraints. For a selected latent , their decoders define the density , evaluated on a three-dimensional grid without additional optimization. § 3.4 compares this per-state reconstruction with evaluating or continuing to optimize the Gaussians of heterogeneity estimation at the inferred latents.
Efficient cryo-EM Gaussian splatting
Additive cryo-EM projection has no depth ordering or early termination. Our differentiable CUDA rasterizer accumulates gradients per Gaussian instead of pixel-wise atomic scatter, as in Taming 3DGS [32] but without compositing state, and its projection scheduling partitions long tile lists into independent partial projections, the additive counterpart of tile load balancing in 3DGS [33, 34]. Its backward pass recomputes the projection geometry of each Gaussian instead of retaining the forward workspaces, which lowers peak memory. Training is distributed across GPUs, and the image-formation model is unchanged (Table 1; Appendix § 12).
Experiments
Datasets and baselines. We quantitatively compare CryoGaussian with existing methods on CryoBench [21], whose five synthetic datasets provide ground-truth structures and states for per-state reconstruction and heterogeneity estimation. IgG-1D, IgG-RL, and Spike-MD assess conformational heterogeneity through one-dimensional Fab motion, randomized linker conformations, and a molecular-dynamics ensemble, respectively, and Ribosembly and Tomotwin-100 assess compositional heterogeneity through 16 unequally populated assembly states and a mixture of 100 distinct complexes, respectively (Table 4). We evaluate CryoGaussian on experimental cryo-EM data, the assembling 50S ribosomal subunit (EMPIAR-10076) [24] and the pre-catalytic spliceosome (EMPIAR-10180) [35], to assess its ability to resolve compositional and conformational heterogeneity, respectively (Table 5). Data preparation, baseline configurations, the compared methods, and evaluation protocols are detailed in Appendices § 7.§ 1, § 7.§ 4, § 7.§ 5, and § 8, respectively.
| Dataset | Number of particle images | Reference states | Image dimensions (pixels) | Pixel size (Å/pixel) | Signal-to-noise ratio |
| IgG-1D | 100,000 | 100 | 3.0 | 0.01 | |
| IgG-RL | 100,000 | 100 | 3.0 | 0.01 | |
| Ribosembly | 335,240 | 16 | 3.0 | 0.01 | |
| Tomotwin-100 | 100,000 | 100 | 4.5 | 0.01 | |
| Spike-MD | 100,000 | 46,789 | 1.5 | 0.1 |
Table 4. CryoBench datasets used in the main comparison. Dimensions and pixel sizes describe the processed images; the signal-to-noise ratio identifies the noise condition of the released dataset. For Spike-MD, reference states are the 46,789 distinct molecular-dynamics configurations reported by CryoBench.
Table 5. Experimental particle-image datasets. Particle selection and preprocessing are described in Appendix § 7.§ 1. Pixel sizes refer to the deposited particle images.
Implementation. Heterogeneity estimation and per-state reconstruction each use 30,000 iterations of Adam [36], with particle-image and per-Gaussian latent dimensions and selected by ablation (§ 3.4). All datasets use the same dimensions, relative density threshold, and base learning rates. Appendix § 6 details the Gaussian parameterization, Appendix § 7.§ 2 the training settings, and Appendix § 7.§ 3 the training time and peak GPU memory (Table 6).
| Dataset | Box | Estimated Gaussians | Global batch | Time per iteration (ms) | Training time (h) | Forward (particles/s) | Peak allocated (GiB/GPU) | Gradient accumulation | Projection scheduling | 8 GPUs | Memory saved (%) |
| EMPIAR-10180 | 256 | 448,148 | 256 | 376.64 | 6.83 | 956.3 | 12.14 | 9.95 | |||
| EMPIAR-10076 (CryoDRGN-filtered) | 256 | 513,705 | 256 | 520.19 | 8.50 | 841.1 | 14.14 | 9.63 | |||
| EMPIAR-10076 (CryoGaussian-filtered) | 256 | 463,802 | 256 | 464.01 | 7.63 | 950.5 | 12.78 | 9.86 | |||
| IgG-1D (SNR 0.01) | 128 | 10,316 | 256 | 17.31 | 0.30 | 27,250.4 | 0.33 | 7.98 | |||
| IgG-1D (SNR 0.005) | 128 | 10,736 | 256 | 17.94 | 0.31 | 25,501.7 | 0.34 | 8.28 | |||
| IgG-1D (SNR 0.001) | 128 | 31,338 | 256 | 36.51 | 0.64 | 9,738.0 | 0.90 | 8.94 | |||
| IgG-RL | 128 | 11,836 | 256 | 34.33 | 0.46 | 24,235.4 | 0.37 | 9.56 | |||
| Ribosembly | 128 | 33,033 | 512 | 73.89 | 1.21 | 12,287.0 | 1.88 | 8.97 | |||
| Tomotwin-100 | 128 | 57,818 | 256 | 72.90 | 1.22 | 4,942.5 | 1.65 | 9.74 | |||
| Spike-MD | 256 | 115,877 | 256 | 216.49 | 2.79 | 3,680.6 | 3.22 | 10.17 |
Table 6. Training time and memory of CryoGaussian on eight NVIDIA RTX PRO 6000 Blackwell GPUs. We report global batch sizes. Estimated numbers of Gaussians at are rounded to the nearest integer. Peak allocated memory per GPU is reported with the forward workspaces retained, and the memory saved is the reduction of this peak when the workspaces are released before the backward pass (Appendix G.3). The speedups of per-Gaussian gradient accumulation, projection scheduling, and eight GPUs over one GPU and the memory saved are those of Table 1, measured in the same way for the other configurations.
CryoBench evaluation
Per-state reconstruction. Table 2 reports the area under the Fourier shell correlation curve (FSC AUC) per conformation and Table 3 the FSC AUC per image, following the CryoBench protocols (Appendix § 8.§ 2). CryoGaussian attains the highest mean FSC AUC per conformation on all five datasets and per image on IgG-1D, Ribosembly, and Spike-MD. We also average the means over the conformational, compositional, and all five datasets. Under the per-image protocol, the ranking of the compared methods depends on the type of heterogeneity: RECOVAR and SOLVAR [37] rank second on the conformational but fourth and fifth, respectively, on the compositional datasets, whereas CryoKRAQEN [38] ranks second on the compositional but fourth on the conformational datasets. CryoGaussian ranks first in all three averages under both protocols, attaining state-of-the-art reconstruction accuracy on both compositional and conformational heterogeneity. The average of CryoGaussian over all five datasets exceeds that of the next-ranked method by 0.012 per conformation and 0.017 per image. Across 10 training seeds, the standard deviation of the mean FSC AUC of CryoGaussian is at most 0.0012 on each dataset under both protocols (Table 8), an order of magnitude below these margins. Figure 2 shows reconstructed densities and inferred latents. Figure 8–Figure 12 show reconstruction–reference comparisons and FSC curves, and Figure 13 shows densities selected by latent clustering (Appendix § 9.§ 1).
| Method | IgG-1D | IgG-RL | Ribosembly | Tomotwin-100 | Spike-MD | Average of means | |||||||
| Mean (std) | Median | Mean (std) | Median | Mean (std) | Median | Mean (std) | Median | Mean (std) | Median | Conf. | Comp. | All | |
| CryoDRGN | 0.366 (0.003) | 0.366 | 0.349 (0.008) | 0.348 | 0.415 (0.019) | 0.415 | 0.321 (0.034) | 0.322 | 0.340 (0.009) | 0.340 | 0.352 | 0.368 | 0.358 |
| CryoDRGN-AI-fixed | 0.366 (0.001) | 0.366 | 0.355 (0.007) | 0.354 | 0.372 (0.032) | 0.374 | 0.206 (0.041) | 0.212 | 0.301 (0.012) | 0.304 | 0.341 | 0.289 | 0.320 |
| OPUS-DSD | 0.340 (0.006) | 0.341 | 0.346 (0.029) | 0.349 | 0.372 (0.046) | 0.382 | 0.256 (0.038) | 0.266 | 0.228 (0.030) | 0.242 | 0.305 | 0.314 | 0.308 |
| 3DFlex | 0.336 (0.002) | 0.336 | 0.339 (0.007) | 0.339 | – | – | – | – | 0.304 (0.010) | 0.306 | – | – | – |
| 3DVA | 0.351 (0.003) | 0.351 | 0.341 (0.006) | 0.341 | 0.375 (0.038) | 0.372 | 0.095 (0.040) | 0.083 | 0.324 (0.010) | 0.325 | 0.339 | 0.235 | 0.297 |
| RECOVAR | 0.391 (0.001) | 0.391 | 0.372 (0.008) | 0.371 | 0.430 (0.016) | 0.432 | 0.306 (0.093) | 0.330 | 0.363 (0.011) | 0.366 | 0.375 | 0.368 | 0.372 |
| 3D Class | 0.297 (0.019) | 0.291 | 0.309 (0.010) | 0.307 | 0.289 (0.081) | 0.288 | 0.046 (0.026) | 0.037 | 0.307 (0.023) | 0.308 | 0.304 | 0.168 | 0.250 |
| CryoSPIRE* | 0.402 (0.002) | 0.402 | 0.386 (0.014) | 0.389 | 0.427 (0.014) | 0.424 | – | – | – | – | – | – | – |
| SOLVAR | 0.392 (0.001) | 0.392 | 0.377 (0.007) | 0.376 | 0.432 (0.015) | 0.432 | 0.298 (0.095) | 0.322 | 0.355 (0.012) | 0.356 | 0.375 | 0.365 | 0.371 |
| CryoGaussian (Ours) | 0.402 (0.001) | 0.402 | 0.387 (0.010) | 0.387 | 0.432 (0.016) | 0.432 | 0.336 (0.034) | 0.343 | 0.365 (0.012) | 0.367 | 0.385 | 0.384 | 0.384 |
Table 2. Table 2 Per-conformation FSC AUC on CryoBench. Mean (standard deviation) and median across evaluated reference states, with background masking (Appendix § 8.§ 2). Higher is better. Bold and underline mark the best and second best per column, including ties at the displayed precision. Conf., Comp., and All average the means over the conformational (IgG-1D, IgG-RL, Spike-MD), the compositional (Ribosembly, Tomotwin-100), and all five datasets, for methods with values on every dataset, so that averages are comparable across methods. For CryoSPIRE*, values are taken directly from its publication (Appendix § 7.§ 4) because we were unable to find a publicly available implementation. CryoHype and CryoKRAQEN have published per-image values only (Table 3). Dashes mark results not included.
Figure 2. Heterogeneous reconstruction on CryoBench. Representative reconstructed densities (top) and UMAP projections [22] of inferred latents (bottom) are linked by matching numbers. Colors indicate the ground-truth IgG-1D dihedral angle, IgG-RL Fab distance, Ribosembly assembly state, Tomotwin-100 molecular mass, and Spike-MD collective variable CV1. Spike-MD labels indicate receptor-binding domain opening. All 16 Ribosembly states are shown, named as in Qin et al. [23]. Appendix § 9.§ 1 shows all evaluated reconstructions with reference densities and FSC curves (Figure 8–Figure 12), densities of latent clusters (Figure 13), and latent metrics (Figure 5–Figure 7).
| Method | IgG-1D Mean (std) | IgG-1D Median | IgG-RL Mean (std) | IgG-RL Median | Ribosembly Mean (std) | Ribosembly Median | Tomotwin-100 Mean (std) | Tomotwin-100 Median | Spike-MD Mean (std) | Spike-MD Median | Average of means Conf. | Average of means Comp. | Average of means All |
| CryoDRGN | 0.351 (0.028) | 0.356 | 0.331 (0.016) | 0.333 | 0.412 (0.023) | 0.415 | 0.316 (0.046) | 0.321 | 0.340 (0.009) | 0.340 | 0.341 | 0.364 | 0.350 |
| CryoDRGN-AI-fixed | 0.364 (0.002) | 0.364 | 0.348 (0.012) | 0.350 | 0.372 (0.032) | 0.375 | 0.202 (0.044) | 0.207 | 0.301 (0.012) | 0.303 | 0.338 | 0.287 | 0.317 |
| OPUS-DSD | 0.335 (0.026) | 0.339 | 0.343 (0.016) | 0.346 | 0.362 (0.083) | 0.382 | 0.237 (0.049) | 0.251 | 0.229 (0.027) | 0.242 | 0.302 | 0.300 | 0.301 |
| 3DFlex | 0.335 (0.003) | 0.335 | 0.337 (0.007) | 0.337 | – | – | – | – | 0.304 (0.011) | 0.306 | – | – | – |
| 3DVA | 0.349 (0.004) | 0.350 | 0.333 (0.014) | 0.335 | 0.375 (0.038) | 0.375 | 0.088 (0.040) | 0.077 | 0.324 (0.010) | 0.323 | 0.335 | 0.232 | 0.294 |
| RECOVAR | 0.386 (0.005) | 0.388 | 0.363 (0.011) | 0.363 | 0.429 (0.018) | 0.432 | 0.258 (0.109) | 0.254 | 0.362 (0.011) | 0.365 | 0.370 | 0.344 | 0.360 |
| 3D Class | 0.297 (0.019) | 0.291 | 0.309 (0.010) | 0.307 | 0.289 (0.081) | 0.288 | 0.046 (0.026) | 0.037 | 0.307 (0.023) | 0.308 | 0.304 | 0.168 | 0.250 |
| CryoSPIRE* | 0.396 (0.013) | 0.400 | 0.375 (0.020) | 0.391 | 0.422 (0.015) | 0.421 | – | – | – | – | – | – | – |
| CryoHype** | – | – | – | – | – | – | 0.346 (0.033) | 0.353 | – | – | – | – | – |
| SOLVAR | 0.388 (0.003) | 0.389 | 0.369 (0.009) | 0.368 | 0.431 (0.017) | 0.432 | 0.246 (0.104) | 0.238 | 0.354 (0.012) | 0.355 | 0.370 | 0.339 | 0.358 |
| CryoKRAQEN* | 0.375 (0.004) | 0.376 | 0.352 (0.010) | 0.352 | 0.418 (0.019) | 0.422 | 0.335 (0.021) | 0.337 | 0.328 (0.010) | 0.329 | 0.352 | 0.377 | 0.362 |
| CryoGaussian (Ours) | 0.397 (0.004) | 0.398 | 0.366 (0.020) | 0.370 | 0.432 (0.014) | 0.432 | 0.335 (0.035) | 0.341 | 0.364 (0.012) | 0.366 | 0.376 | 0.384 | 0.379 |
Table 3. Per-image FSC AUC on CryoBench. Mean (standard deviation) and median across evaluated particle images, with background masking (Appendix § 8.§ 2); notation as in Table 2. For methods marked * and **, values are taken directly from their publications (Appendix § 7.§ 4), * because we were unable to find a publicly available implementation and ** because we were unable to find implementations of the mechanisms of CryoHype for the other datasets. Table 8 and Table 15 report seed variability and unmasked FSC AUC.
| Dataset | Per-image mean FSC AUC | Per-conformation mean FSC AUC |
| IgG-1D | 0.3976 0.0006 (0.3962–0.3982) | 0.4024 0.0004 (0.4019–0.4032) |
| IgG-RL | 0.3673 0.0011 (0.3656–0.3685) | 0.3879 0.0009 (0.3861–0.3892) |
| Ribosembly | 0.4326 0.0007 (0.4321–0.4339) | 0.4328 0.0008 (0.4313–0.4338) |
| Tomotwin-100 | 0.3366 0.0012 (0.3353–0.3387) | 0.3373 0.0012 (0.3359–0.3392) |
| Spike-MD | 0.3647 0.0008 (0.3634–0.3654) | 0.3651 0.0008 (0.3640–0.3669) |
Table 8. Variability of CryoGaussian across training seeds on CryoBench. Mean standard deviation across 10 training seeds of the mean FSC AUC with background masking, with the range across seeds in parentheses. Seeds vary the random initialization of the Gaussians and latents and the order of particle images, and heterogeneity estimation and per-state reconstruction are both rerun for each seed.
Figure 8. IgG-1D reconstruction at SNR 0.01. (a) CryoGaussian densities reconstructed at the inferred latents for 100 particle images (Ours), each paired with its reference density (G.T., gray). Reconstructed densities are colored by the ground-truth dihedral angle. (b) UMAP projection of inferred latents with the same color scale. (c) Per-image FSC (top) and per-conformation FSC (bottom), evaluated with background masking (left) and without masking (right). Solid curves show the mean, and shaded regions cover the 5th–95th percentiles across 100 evaluated particle images or 100 reference states, respectively.
Figure 12. Spike-MD reconstruction. (a) CryoGaussian densities reconstructed at the inferred latents for 100 selected particle images (Ours), paired with their reference densities (G.T., gray) and colored by collective variable CV1. (b) UMAP projections of inferred latents, colored by CV1 (top) and CV2 (bottom), with collective-variable definitions from Wieczór et al. [69]. (c) Per-image FSC (top) and per-conformation FSC (bottom), with background masking (left) and without masking (right). Solid curves show the mean, and shaded regions cover the 5th–95th percentiles across 100 evaluated particle images or their reference configurations, respectively.
Figure 13. Representative densities selected by latent clustering. For each dataset and noise condition, inferred latents are partitioned by -means into 20 clusters, or 16 for Ribosembly. The observed latent nearest to each cluster center is decoded to obtain a representative density. Each dataset block shows these densities alongside the UMAP projection, with colored markers locating the selected latents. Matching colors link each density to its marker and indicate cluster identity within that dataset.
Heterogeneity estimation. We evaluate the heterogeneity learned by CryoGaussian with the CryoBench metrics (Appendix § 8.§ 1). On all five datasets, CryoGaussian attains the highest neighborhood similarity to the ground-truth heterogeneity embedding at the largest evaluated radius, and the information imbalance from its inferred latents to this embedding lies within 0.04–0.20, where 0 and 1 indicate equivalent and orthogonal information (Figure 5). On Ribosembly and Tomotwin-100, -means clustering of the inferred latents into the ground-truth number of states agrees with the state labels with ARI of 0.974 and 0.823 and AMI of 0.982 and 0.926, ranking first on Ribosembly and second on Tomotwin-100; CryoGaussian is the only method in the top two on both datasets (Table 7). CryoBench further assesses whether latents capture non-structural factors such as pose and CTF [21]. CryoGaussian’s latents are orthogonal to both: their information imbalance in both directions lies within 0.99–1.02 with ground-truth rotations and 0.997–1.000 with CTF parameters on all five datasets (Figure 6 and Figure 7).
| Method | Ribosembly | Tomotwin-100 | ||
| ARI | AMI | ARI | AMI | |
| CryoDRGN | 0.873 | 0.935 | 0.956 | 0.983 |
| CryoDRGN-AI-fixed | 0.624 | 0.771 | 0.791 | 0.906 |
| OPUS-DSD | 0.891 | 0.934 | 0.500 | 0.781 |
| 3DVA | 0.666 | 0.823 | 0.058 | 0.335 |
| RECOVAR | 0.968 | 0.976 | 0.315 | 0.649 |
| SOLVAR | <u>0.971</u> | <u>0.979</u> | 0.280 | 0.625 |
| CryoGaussian (Ours) | 0.974 (0.004) | 0.982 (0.001) | <u>0.823 (0.005)</u> | <u>0.926 (0.001)</u> |
Table 7. Table 7 Clustering accuracy on Ribosembly and Tomotwin-100. Adjusted Rand index (ARI) and adjusted mutual information (AMI) between ground-truth labels and -means cluster assignments of the inferred particle-image latents, with equal to the number of ground-truth states (16 for Ribosembly and 100 for Tomotwin-100). Higher is better. Bold and underline mark the best and second best per column. The compared methods are those of Tables 2 and 3 with publicly available implementations. CryoGaussian values are the mean (standard deviation) over 10 random seeds.
Figure 5. Heterogeneity estimation on CryoBench. Neighborhood similarity as a function of the neighborhood radius (top row) and information imbalance (bottom row) between inferred latents and the ground-truth heterogeneity embedding on (a) IgG-1D, (b) IgG-RL, (c) Ribosembly, (d) Tomotwin-100, and (e) Spike-MD, presented in the format of Figure 2d, Figure 4d, Figure 6d, Figure 7d, and S24 of the CryoBench paper [21]. Stars denote CryoGaussian.
Figure 6. Pose information imbalance on CryoBench. Information imbalance between inferred latents and ground-truth rotations on (a) IgG-1D, (b) IgG-RL, (c) Ribosembly, (d) Tomotwin-100, and (e) Spike-MD, in full view (top row) and enlarged near (1, 1) (bottom row), presented in the format of Figure S22 of the CryoBench paper [21].
Figure 7. CTF information imbalance on CryoBench. Information imbalance between inferred latents and CTF parameters on (a) IgG-1D, (b) IgG-RL, (c) Ribosembly, (d) Tomotwin-100, and (e) Spike-MD, in full view (top row) and enlarged near (1, 1) (bottom row), presented in the format of Figure S23 of the CryoBench paper [21].
Assembling 50S ribosomal subunit
Compositional heterogeneity (EMPIAR-10076). We analyze the assembling 50S ribosomal subunit from an L17-depleted sample [24] using the CryoDRGN-filtered subset of 87,327 particle images [10]. The inferred latents, learned without class labels, distinguish the five major classes and capture finer variation among the 14 minor classes of the published classification [24], including 70S class A (Figure 3). Map-to-map FSC assesses the agreement of the reconstructed assembly states with the deposited reconstructions (Figure 14; Appendices § 8.§ 3 and § 9.§ 2).
Figure 3. Compositional heterogeneity in the assembling 50S ribosomal subunit (EMPIAR-10076). (a, c) UMAP projections of the inferred latents, colored by published major and minor class assignments [24]; black points mark class centers. The inset locates the 1,113 particle images of class C4 [10]. (b) C4 and E4 densities. (d) Major assembly classes B–E and the 70S class A. Dashed lines trace the outline of the mature 50S subunit, and red ellipses mark rRNA helix 68 (H68). CP, central protuberance; LSU, large subunit. Appendix § 9.§ 2 shows the twelve minor classes C1–C3, D1–D4, and E1–E5 with map-to-map FSC against the deposited reconstructions (Figure 14) and the CryoSPARC refinement of C4 (Figure 15). Appendix § 10.7 presents particle selection by CryoGaussian (Figure 20–Figure 22).
Figure 14. Assembly-state reconstruction using the CryoDRGN particle selection for EMPIAR-10076. The model is trained on 87,327 particle images. (a) Reconstructed densities for the twelve minor classes C1–C3, D1–D4, and E1–E5, with dashed lines tracing the outline of the mature 50S subunit. (b) Map-to-map FSC between CryoGaussian densities and the corresponding deposited reconstructions [24], for class A and the thirteen assembly classes B–E5. Horizontal dashed lines indicate FSC values of 0.5 and 0.143. (c) UMAP projection of inferred latents colored by published class assignments. (d) The same projection with each class highlighted separately. Colors are shared across panels (b)–(d). Representative-latent selection and map comparison are described in Appendix § 8.§ 3.
Rare assembly intermediate C4. CryoGaussian resolves a rare assembly intermediate with rRNA helix 68 (H68) docked in the absence of the central protuberance (class C4, 1,113 of 87,327 particle images, 1.3%) [10] (Figure 3(b)). Homogeneous CryoSPARC refinement of these 1,113 particle images shows the same two features (Figure 15(a)).
Figure 15. Identification and homogeneous refinement of the 50S assembly intermediate C4. The left and right columns use the CryoDRGN-filtered and CryoGaussian-filtered particles, respectively. (a, b) CryoSPARC refinement uses the 1,113 particles of the released C4 particle selection [10] (a) and their intersection with the CryoGaussian-filtered particle set (1,019 particles; b). (c) Identification of this intermediate in the CryoGaussian latent space, without the released selection. UMAP projections of the inferred latents are colored by published major class assignments; black points mark class centers. Gray boxes mark the regions enlarged on the right, and black polygons outline the selected particles. (d, e) CryoSPARC refinement uses the 1,062 (d) and 954 (e) particles selected in (c). Panels (a), (b), (d), and (e) each show the CryoGaussian density, the CryoSPARC density obtained by homogeneous refinement of the corresponding selected particle images, and FSC curves. Blue curves show gold-standard FSC between independent CryoSPARC half-maps; orange curves show map-to-map FSC between the displayed CryoGaussian and CryoSPARC densities. Horizontal dashed lines indicate FSC values of 0.5 and 0.143. Dashed lines trace the outline of the mature 50S subunit, and red ellipses mark rRNA helix 68 (H68).
Particle selection by CryoGaussian. CryoGaussian also performs the particle selection itself: trained on all 131,899 particle images without class labels, it selects 97,651 particle images (Figure 20; Appendix § 10.7). The C4 intermediate is also resolved from this CryoGaussian-filtered subset, and refinement of the 1,019 of the 1,113 C4 particle images retained in this subset shows the same two features (Figure 15(b)); Figure 21 shows the assembly states reconstructed from this subset and their map-to-map FSC. Figure 22 compares the major assembly classes B–E reconstructed by CryoDRGN with those reconstructed by CryoGaussian from each of the two filtered subsets: regions fragmented in the CryoDRGN densities are continuous in both CryoGaussian reconstructions. CryoGaussian accomplishes both particle selection and the reconstruction of a rare assembly intermediate on experimental data, starting from the unfiltered particle stack.
Figure 20. Particle filtering and reconstruction of EMPIAR-10076 with CryoGaussian. (a, b) Inferred latents from one- and four-dimensional models trained on all 131,899 particle images: a latent histogram with a kernel-density estimate (a) and a UMAP projection (b). (c, d) The same representations colored by published class assignments [24]. The vertical dashed line in (c) marks the lower boundary of the excluded latent interval; the dashed contour in (d) identifies the excluded Gaussian-mixture component. The intersection of the two retained particle sets contains 97,651 images, of which a random 90% (87,885 images) are used for subsequent reconstruction. (e, g) Reconstructed densities for assembly classes B–E and the 70S class A. (f, h) Inferred latents colored by major and minor class assignments; the inset identifies the released C4 particle selection [10]. (i) C4, E4, and E5 densities. (j) Homogeneous reconstruction of the selected C4 particle images with CryoSPARC. Dashed lines trace the outline of the mature 50S subunit, and red ellipses mark H68. Filtering and reconstruction settings are given in Appendices § 7.§ 1 and § 8.§ 3.
Figure 21. Assembly-state reconstruction using the CryoGaussian particle selection for EMPIAR-10076. The model is trained on 87,885 particle images selected as described in Figure 20. (a) Reconstructed densities for C1–C3, D1–D4, and E1–E5, with dashed lines tracing the outline of the mature 50S subunit. (b) Map-to-map FSC against the corresponding deposited reconstructions [24], including classes A and B. Horizontal dashed lines indicate FSC values of 0.5 and 0.143. (c, d) UMAP projections colored by published class assignments, shown jointly (c) and with one class highlighted at a time (d). Colors are shared across panels (b)–(d). Evaluation follows Figure 14.
Figure 22. Assembly-state comparison on EMPIAR-10076. Columns show reconstructed densities for the major assembly classes B–E. Rows correspond to CryoDRGN (top), CryoGaussian using the CryoDRGN-filtered particle subset (middle), and CryoGaussian using the CryoGaussian-filtered subset (bottom). Dashed lines trace the outline of the mature 50S subunit. Red ellipses, drawn at corresponding positions in all three rows of each column, mark regions in which the CryoDRGN density is fragmented and noise-like, whereas both CryoGaussian reconstructions exhibit continuous density. Particle selection is described in Appendix § 7.§ 1.
Pre-catalytic spliceosome
Conformational heterogeneity (EMPIAR-10180). We analyze the Saccharomyces cerevisiae pre-catalytic spliceosome using the CryoDRGN-filtered subset of 139,720 particle images [10]. Reconstructions along the first two principal components of the inferred latents resolve continuous motion of the SF3b and helicase subcomplexes [35] [10] relative to the body and foot of the complex (Figure 4).
Figure 4. Conformational variation in the pre-catalytic spliceosome (EMPIAR-10180). (a) Principal-component projections of inferred latents, with five numbered positions along PC1 (top) and PC2 (bottom). EV denotes the fraction of latent variance explained by each component. (b) Densities reconstructed at positions 1–5, ordered from left to right in each row. Dashed outlines reproduce each row’s first density; arrows indicate the direction of conformational change. Appendix § 9.§ 2 analyzes the full particle stack (Figure 16).
Structural subpopulations. Trained on the full stack of 327,489 particle images, CryoGaussian separates the pre-catalytic spliceosome from structurally distinct populations. Clustering the inferred latents isolates populations lacking resolved SF3b density (14,063 particle images), showing additional density near the U2 core (12,854 particle images), and exhibiting poorly resolved density (10,052 particle images) [10] (Figure 16; Appendix C.3). CryoGaussian resolves continuous conformational dynamics and compositionally distinct subpopulations in one experimental dataset and enables particle selection from the unfiltered stack.
Figure 16. Latent-space analysis of the full EMPIAR-10180 particle stack. A four-dimensional model is trained on 327,489 particle images at 128 × 128 pixels. (a) UMAP projection of inferred latents, with 20 -means clusters numbered 1–20. Five selected groups are colored to match the reconstructed densities in (b); the other groups are gray. (b) Representative densities from two subdivisions of the largest cluster (i, ii), a cluster with poorly resolved density (iii), a cluster lacking density in the SF3b region (iv), and a cluster with additional density near the U2 core (v). (c) Representative densities for all 20 clusters and the two subdivisions 14a and 14b, with particle counts below each density. Representative-latent selection is described in Appendix § 8.§ 3. This full-stack analysis is separate from the CryoDRGN-filtered particle subset used in Figure 4.
Figure labels: UMAP1; UMAP2. 1–20 -means clusters of the model trained on all 327,489 particles; i–v the five sampled for b), each in the color of its volume; the rest in gray; i, ii expected pre-catalytic spliceosome (two sub-centers of cluster 14); iii likely corrupted by imaging artifacts; iv lacking the SF3b subcomplex; v with the U2 core. Cluster particle counts: 1 16,864; 2 17,736; 3 13,864; 4 12,854 (v); 5 10,052 (iii); 6 16,588; 7 22,356; 8 16,389; 9 9,193; 10 18,876; 11 15,729; 12 13,168; 13 22,907; 14 34,127 → 14a, 14b; 15 18,696; 16 14,063 (iv); 17 12,746; 18 14,386; 19 14,296; 20 12,599; 14a 13,559 (i); 14b 20,568 (ii).
Ablations and robustness
Unless noted otherwise, the comparisons below are in mean FSC AUC and hold on all five CryoBench datasets under both protocols; each study is detailed in Appendix § 10.
Number of Gaussians. Mean FSC AUC and, on IgG-1D, neighborhood similarity are highest at the default relative density threshold of 0.02. Mean FSC AUC is 0.002–0.012 lower with more Gaussians at 0.01 and 0.033–0.067 lower with the fewest Gaussians at 0.10 (Table 9; Figure 17).
Figure 17. Sensitivity to the relative density threshold on IgG-1D. Heterogeneity estimation on IgG-1D at SNR 0.01 with for estimating the number of Gaussians. UMAP projections of inferred latents, colored by the ground-truth dihedral angle, are shown on the left; corresponding neighborhood-similarity curves are shown on the right, following Appendix § 8.§ 1. The default threshold yields the highest neighborhood similarity among the tested values. The UMAP projections retain an ordered loop for thresholds from 0.01 to 0.05, whereas produces a fragmented projection and lower neighborhood similarity.
Table 9 Sensitivity to the relative density threshold on CryoBench. selects the estimated spatial support of the density in the CTF-weighted backprojection, and the number of Gaussians follows from the space–bandwidth product of this support (§ 2.2). For each , the number of Gaussians and the per-image and per-conformation FSC AUC with background masking are reported as in Table 2 and Table 3, which correspond to the default . Bold marks the best per row.
| Dataset | 0.01 | 0.02 | 0.03 | 0.04 | 0.05 | 0.10 | |
| IgG-1D | Gaussians | 16,111 | 10,316 | 9,231 | 8,359 | 7,598 | 4,542 |
| Per-image, mean (std) | 0.385 (0.007) | 0.397 (0.004) | 0.390 (0.006) | 0.384 (0.009) | 0.376 (0.013) | 0.361 (0.021) | |
| Per-image, median | 0.386 | 0.398 | 0.391 | 0.385 | 0.379 | 0.366 | |
| Per-conformation, mean (std) | 0.392 (0.002) | 0.402 (0.001) | 0.397 (0.001) | 0.391 (0.002) | 0.384 (0.003) | 0.369 (0.004) | |
| Per-conformation, median | 0.392 | 0.402 | 0.397 | 0.391 | 0.385 | 0.371 | |
| IgG-RL | Gaussians | 20,612 | 11,836 | 8,996 | 7,361 | 6,285 | 3,935 |
| Per-image, mean (std) | 0.359 (0.026) | 0.366 (0.020) | 0.355 (0.025) | 0.352 (0.028) | 0.343 (0.033) | 0.325 (0.044) | |
| Per-image, median | 0.364 | 0.370 | 0.360 | 0.357 | 0.351 | 0.336 | |
| Per-conformation, mean (std) | 0.381 (0.012) | 0.387 (0.010) | 0.379 (0.011) | 0.375 (0.014) | 0.368 (0.016) | 0.351 (0.021) | |
| Per-conformation, median | 0.382 | 0.387 | 0.381 | 0.376 | 0.368 | 0.354 | |
| Ribosembly | Gaussians | 37,387 | 33,033 | 31,647 | 30,420 | 29,285 | 24,076 |
| Per-image, mean (std) | 0.429 (0.018) | 0.432 (0.014) | 0.425 (0.019) | 0.419 (0.023) | 0.409 (0.028) | 0.390 (0.039) | |
| Per-image, median | 0.429 | 0.432 | 0.426 | 0.421 | 0.414 | 0.394 | |
| Per-conformation, mean (std) | 0.430 (0.018) | 0.432 (0.016) | 0.428 (0.018) | 0.423 (0.021) | 0.411 (0.026) | 0.394 (0.038) | |
| Per-conformation, median | 0.431 | 0.432 | 0.428 | 0.424 | 0.415 | 0.396 | |
| Tomotwin-100 | Gaussians | 87,988 | 57,818 | 44,938 | 37,322 | 32,124 | 19,611 |
| Per-image, mean (std) | 0.331 (0.040) | 0.335 (0.035) | 0.321 (0.043) | 0.311 (0.051) | 0.300 (0.061) | 0.268 (0.082) | |
| Per-image, median | 0.337 | 0.341 | 0.329 | 0.320 | 0.312 | 0.287 | |
| Per-conformation, mean (std) | 0.333 (0.038) | 0.336 (0.034) | 0.325 (0.042) | 0.315 (0.048) | 0.305 (0.057) | 0.273 (0.075) | |
| Per-conformation, median | 0.340 | 0.343 | 0.333 | 0.325 | 0.318 | 0.293 | |
| Spike-MD | Gaussians | 142,530 | 115,877 | 109,102 | 103,766 | 99,135 | 81,890 |
| Per-image, mean (std) | 0.355 (0.015) | 0.364 (0.012) | 0.356 (0.015) | 0.348 (0.019) | 0.339 (0.023) | 0.317 (0.031) | |
| Per-image, median | 0.358 | 0.366 | 0.358 | 0.352 | 0.343 | 0.325 | |
| Per-conformation, mean (std) | 0.357 (0.013) | 0.365 (0.012) | 0.358 (0.014) | 0.350 (0.016) | 0.343 (0.020) | 0.323 (0.027) | |
| Per-conformation, median | 0.359 | 0.367 | 0.360 | 0.353 | 0.347 | 0.329 |
Table 9.
FWHM constraint and initialization. Without the FWHM constraint, mean FSC AUC is 0.016–0.030 lower (Table 10) and neighborhood similarity on IgG-1D is lower (Figure 18). Initializing Gaussian positions from a consensus density fitted without latents instead of at random decreases this similarity from 82.7% to 55.8% at SNR 0.01 (Figure 19).
Figure 18. Ablation of the Gaussian FWHM constraint on IgG-1D. Heterogeneity estimation is evaluated at SNR 0.01 and 0.001 with and without the constraint in Equation (5), denoted “physical clamp” in the figure. Left: UMAP projections of inferred latents, colored by the ground-truth dihedral angle. Right: Neighborhood similarity, measured as the percentage of matching neighbors (pMN) between the inferred latents and the ground-truth heterogeneity embedding, following Appendix § 8.§ 1. The constraint yields higher pMN across the displayed neighborhood radii at both noise levels. Colors denote the constraint setting, and line styles denote SNR.
| Dataset | FWHM constraint | Per-image | Per-conformation | ||
| Mean (std) | Median | Mean (std) | Median | ||
| IgG-1D | With | 0.397 (0.004) | 0.398 | 0.402 (0.001) | 0.402 |
| Without | 0.372 (0.013) | 0.374 | 0.386 (0.005) | 0.386 | |
| IgG-RL | With | 0.366 (0.020) | 0.370 | 0.387 (0.010) | 0.387 |
| Without | 0.342 (0.033) | 0.347 | 0.370 (0.018) | 0.371 | |
| Ribosembly | With | 0.432 (0.014) | 0.432 | 0.432 (0.016) | 0.432 |
| Without | 0.402 (0.025) | 0.404 | 0.406 (0.027) | 0.407 | |
| Tomotwin-100 | With | 0.335 (0.035) | 0.341 | 0.336 (0.034) | 0.343 |
| Without | 0.306 (0.047) | 0.314 | 0.313 (0.045) | 0.321 | |
| Spike-MD | With | 0.364 (0.012) | 0.366 | 0.365 (0.012) | 0.367 |
| Without | 0.343 (0.019) | 0.345 | 0.344 (0.019) | 0.347 |
Table 10. Ablation of the Gaussian FWHM constraint on CryoBench. Mean (standard deviation) and median FSC AUC with background masking across evaluated particle images (per-image) and reference states (per-conformation) with and without the constraint in Equation (5). The results with the constraint correspond to Table 2 and Table 3. Higher is better. Bold marks the best per dataset and column.
Figure 19. Ablation of Gaussian initialization on IgG-1D. Heterogeneity estimation is evaluated at SNR 0.01 and 0.001 using random initialization or Gaussian positions sampled from a consensus density optimized without latents on all particle images. Consensus initialization uses density-weighted sampling without replacement over thresholded voxels, as detailed in Appendix § 10.§ 3. Left: UMAP projections of inferred latents, colored by the ground-truth dihedral angle. Right: Neighborhood similarity, measured as the percentage of matching neighbors (pMN) between the inferred latents and the ground-truth heterogeneity embedding, plotted against neighborhood radius as a percentage of the 20,000-image evaluation subset (Appendix § 8.§ 1). Higher pMN indicates better preservation of reference conformational neighborhoods. Colors denote Gaussian initialization, and line styles denote SNR. Random initialization yields higher pMN across the evaluated neighborhood radii at both noise levels.
Per-state reconstruction strategy. We examine whether the Gaussians of heterogeneity estimation can replace the randomly initialized Gaussians that CryoGaussian fits at the inferred latents for per-state reconstruction. At the same total number of iterations, mean FSC AUC is 0.008–0.016 lower with heterogeneity estimation alone and 0.005–0.010 lower with per-state reconstruction continued from the Gaussians of heterogeneity estimation (Table 11).
| Dataset | Strategy | Per-image | Per-conformation | ||
| Mean (std) | Median | Mean (std) | Median | ||
| IgG-1D | Heterogeneity estimation alone | 0.381 (0.008) | 0.381 | 0.390 (0.004) | 0.390 |
| Per-state reconstruction continued from | 0.387 (0.007) | 0.389 | 0.395 (0.003) | 0.394 | |
| Per-state reconstruction from randomly initialized | 0.397 (0.004) | 0.398 | 0.402 (0.001) | 0.402 | |
| IgG-RL | Heterogeneity estimation alone | 0.358 (0.027) | 0.362 | 0.379 (0.020) | 0.378 |
| Per-state reconstruction continued from | 0.361 (0.024) | 0.366 | 0.382 (0.018) | 0.381 | |
| Per-state reconstruction from randomly initialized | 0.366 (0.020) | 0.370 | 0.387 (0.010) | 0.387 | |
| Ribosembly | Heterogeneity estimation alone | 0.421 (0.023) | 0.421 | 0.422 (0.026) | 0.422 |
| Per-state reconstruction continued from | 0.426 (0.021) | 0.427 | 0.426 (0.022) | 0.423 | |
| Per-state reconstruction from randomly initialized | 0.432 (0.014) | 0.432 | 0.432 (0.016) | 0.432 | |
| Tomotwin-100 | Heterogeneity estimation alone | 0.320 (0.041) | 0.324 | 0.322 (0.048) | 0.329 |
| Per-state reconstruction continued from | 0.326 (0.037) | 0.332 | 0.328 (0.038) | 0.338 | |
| Per-state reconstruction from randomly initialized | 0.335 (0.035) | 0.341 | 0.336 (0.034) | 0.343 | |
| Spike-MD | Heterogeneity estimation alone | 0.351 (0.026) | 0.354 | 0.353 (0.029) | 0.358 |
| Per-state reconstruction continued from | 0.356 (0.022) | 0.359 | 0.358 (0.027) | 0.361 | |
| Per-state reconstruction from randomly initialized | 0.364 (0.012) | 0.366 | 0.365 (0.012) | 0.367 |
Table 11. Per-state reconstruction strategies on CryoBench. Mean (standard deviation) and median FSC AUC with background masking across evaluated particle images (per-image) and reference states (per-conformation) for three ways of obtaining per-state densities at the same total number of iterations. Heterogeneity estimation alone optimizes and the latents jointly and evaluates at the inferred latents. Per-state reconstruction continued from continues optimizing conditioned on the inferred latents. Per-state reconstruction from randomly initialized follows § 2.3 and corresponds to Table 2 and Table 3. Higher is better. Bold marks the best per dataset and column, including ties at the displayed precision.
Decoders and latent dimensions. Among all 15 combinations of the position, amplitude, and scale decoders and an additional rotation decoder, the combination of the position, amplitude, and scale decoders attains the highest mean FSC AUC. Removing the amplitude, position, or scale decoder from this combination decreases mean FSC AUC by 0.013–0.027, 0.007–0.020, and 0.008–0.015, respectively, and adding the rotation decoder to any combination decreases mean FSC AUC (Table 12 and Table 13). Over particle-image latent dimensions of 2, 4, 8, and 10 and per-Gaussian latent dimensions of 16, 24, and 32, mean FSC AUC is highest at 4 and 24 (Table 14). Although CryoGaussian differs from CryoSPIRE [19] in its heterogeneity model, constraints on the Gaussians, and decoders, these dimensions match the latent and Gaussian feature dimensions of CryoSPIRE.
| Decoders | IgG-1D Mean (std) | IgG-1D Median | IgG-RL Mean (std) | IgG-RL Median | Ribosembly Mean (std) | Ribosembly Median | Tomotwin-100 Mean (std) | Tomotwin-100 Median | Spike-MD Mean (std) | Spike-MD Median |
| Position | 0.343 (0.010) | 0.344 | 0.309 (0.024) | 0.306 | 0.305 (0.047) | 0.308 | 0.208 (0.071) | 0.204 | 0.287 (0.028) | 0.282 |
| Amplitude | 0.358 (0.009) | 0.356 | 0.327 (0.020) | 0.327 | 0.360 (0.042) | 0.362 | 0.259 (0.054) | 0.266 | 0.308 (0.023) | 0.307 |
| Scale | 0.323 (0.017) | 0.320 | 0.286 (0.033) | 0.286 | 0.335 (0.052) | 0.334 | 0.237 (0.068) | 0.232 | 0.269 (0.029) | 0.271 |
| Rotation | 0.314 (0.018) | 0.312 | 0.278 (0.036) | 0.280 | 0.299 (0.048) | 0.300 | 0.201 (0.071) | 0.195 | 0.265 (0.033) | 0.261 |
| Position and amplitude | 0.394 (0.003) | 0.394 | 0.379 (0.015) | 0.378 | 0.421 (0.022) | 0.420 | 0.321 (0.045) | 0.320 | 0.356 (0.018) | 0.356 |
| Amplitude and scale | 0.382 (0.005) | 0.381 | 0.375 (0.020) | 0.377 | 0.425 (0.018) | 0.425 | 0.326 (0.037) | 0.325 | 0.347 (0.022) | 0.347 |
| Position and scale | 0.377 (0.007) | 0.378 | 0.374 (0.021) | 0.372 | 0.418 (0.028) | 0.417 | 0.319 (0.043) | 0.318 | 0.342 (0.019) | 0.342 |
| Amplitude and rotation | 0.356 (0.013) | 0.353 | 0.323 (0.024) | 0.323 | 0.358 (0.046) | 0.360 | 0.257 (0.057) | 0.263 | 0.306 (0.027) | 0.304 |
| Position and rotation | 0.340 (0.014) | 0.340 | 0.307 (0.028) | 0.304 | 0.301 (0.047) | 0.306 | 0.206 (0.074) | 0.201 | 0.285 (0.031) | 0.282 |
| Scale and rotation | 0.318 (0.021) | 0.319 | 0.283 (0.035) | 0.284 | 0.333 (0.053) | 0.331 | 0.235 (0.071) | 0.229 | 0.267 (0.033) | 0.267 |
| Position, amplitude, and scale (CryoGaussian) | 0.402 (0.001) | 0.402 | 0.387 (0.010) | 0.387 | 0.432 (0.016) | 0.432 | 0.336 (0.034) | 0.343 | 0.365 (0.012) | 0.367 |
| Position, amplitude, and rotation | 0.391 (0.004) | 0.391 | 0.377 (0.017) | 0.373 | 0.419 (0.025) | 0.419 | 0.317 (0.046) | 0.319 | 0.352 (0.019) | 0.352 |
| Amplitude, scale, and rotation | 0.377 (0.006) | 0.377 | 0.370 (0.020) | 0.372 | 0.423 (0.022) | 0.423 | 0.324 (0.039) | 0.321 | 0.344 (0.025) | 0.344 |
| Position, scale, and rotation | 0.374 (0.008) | 0.375 | 0.368 (0.022) | 0.366 | 0.417 (0.029) | 0.416 | 0.314 (0.043) | 0.314 | 0.341 (0.022) | 0.340 |
| Position, amplitude, scale, and rotation | 0.394 (0.003) | 0.393 | 0.379 (0.015) | 0.381 | 0.426 (0.019) | 0.428 | 0.327 (0.036) | 0.325 | 0.353 (0.018) | 0.353 |
Table 12. Combinations of Gaussian attribute decoders on CryoBench, per-conformation FSC AUC. Mean (standard deviation) and median FSC AUC with background masking across evaluated reference states for all 15 combinations of the position, amplitude, scale, and rotation decoders, grouped by the number of decoders (Appendix E.5). Each row lists the decoders used, and attributes without a decoder keep their base values (Section 2.3 and Appendix A). The combination of the position, amplitude, and scale decoders is the configuration of CryoGaussian and corresponds to Table 2. Higher is better. Bold marks the best per column.
| Decoders | IgG-1D Mean (std) | IgG-1D Median | IgG-RL Mean (std) | IgG-RL Median | Ribosembly Mean (std) | Ribosembly Median | Tomotwin-100 Mean (std) | Tomotwin-100 Median | Spike-MD Mean (std) | Spike-MD Median |
| Position | 0.334 (0.017) | 0.333 | 0.287 (0.041) | 0.290 | 0.305 (0.052) | 0.306 | 0.207 (0.076) | 0.202 | 0.287 (0.028) | 0.283 |
| Amplitude | 0.348 (0.015) | 0.351 | 0.303 (0.035) | 0.303 | 0.356 (0.041) | 0.353 | 0.254 (0.056) | 0.252 | 0.308 (0.022) | 0.307 |
| Scale | 0.313 (0.024) | 0.320 | 0.263 (0.050) | 0.267 | 0.332 (0.055) | 0.327 | 0.235 (0.063) | 0.234 | 0.269 (0.030) | 0.270 |
| Rotation | 0.305 (0.026) | 0.311 | 0.255 (0.052) | 0.260 | 0.299 (0.051) | 0.297 | 0.198 (0.076) | 0.193 | 0.264 (0.032) | 0.261 |
| Position and amplitude | 0.389 (0.008) | 0.387 | 0.357 (0.026) | 0.355 | 0.420 (0.022) | 0.421 | 0.322 (0.042) | 0.321 | 0.356 (0.017) | 0.355 |
| Amplitude and scale | 0.377 (0.012) | 0.378 | 0.352 (0.027) | 0.352 | 0.424 (0.017) | 0.424 | 0.325 (0.039) | 0.326 | 0.345 (0.022) | 0.346 |
| Position and scale | 0.370 (0.013) | 0.371 | 0.350 (0.034) | 0.350 | 0.416 (0.023) | 0.415 | 0.319 (0.043) | 0.320 | 0.342 (0.019) | 0.342 |
| Amplitude and rotation | 0.344 (0.017) | 0.348 | 0.301 (0.037) | 0.299 | 0.352 (0.041) | 0.348 | 0.250 (0.058) | 0.249 | 0.306 (0.027) | 0.305 |
| Position and rotation | 0.333 (0.021) | 0.331 | 0.283 (0.041) | 0.289 | 0.302 (0.056) | 0.303 | 0.204 (0.079) | 0.198 | 0.285 (0.031) | 0.282 |
| Scale and rotation | 0.310 (0.029) | 0.315 | 0.259 (0.051) | 0.266 | 0.327 (0.059) | 0.323 | 0.232 (0.066) | 0.232 | 0.267 (0.033) | 0.267 |
| Position, amplitude, and scale (CryoGaussian) | 0.397 (0.004) | 0.398 | 0.366 (0.020) | 0.370 | 0.432 (0.014) | 0.432 | 0.335 (0.035) | 0.341 | 0.364 (0.012) | 0.366 |
| Position, amplitude, and rotation | 0.387 (0.010) | 0.384 | 0.356 (0.026) | 0.351 | 0.416 (0.024) | 0.419 | 0.318 (0.045) | 0.316 | 0.352 (0.019) | 0.352 |
| Amplitude, scale, and rotation | 0.373 (0.012) | 0.375 | 0.349 (0.029) | 0.348 | 0.420 (0.019) | 0.420 | 0.321 (0.041) | 0.324 | 0.344 (0.025) | 0.343 |
| Position, scale, and rotation | 0.368 (0.015) | 0.367 | 0.346 (0.035) | 0.346 | 0.413 (0.023) | 0.414 | 0.315 (0.047) | 0.315 | 0.341 (0.023) | 0.340 |
| Position, amplitude, scale, and rotation | 0.390 (0.009) | 0.391 | 0.358 (0.023) | 0.356 | 0.425 (0.019) | 0.428 | 0.324 (0.042) | 0.322 | 0.354 (0.018) | 0.353 |
Table 13. Combinations of Gaussian attribute decoders on CryoBench, per-image FSC AUC. Mean (standard deviation) and median FSC AUC with background masking across evaluated particle images for all 15 combinations of the position, amplitude, scale, and rotation decoders, grouped by the number of decoders (Appendix § 10.§ 5). Each row lists the decoders used, and attributes without a decoder keep their base values (§ 2.3 and Appendix § 6). The combination of the position, amplitude, and scale decoders is the configuration of CryoGaussian and corresponds to Table 3. Higher is better. Bold marks the best per column.
Table 14 Latent dimensions on CryoBench. Mean (standard deviation) and median FSC AUC with background masking across evaluated particle images (per-image) and reference states (per-conformation) for particle-image latent dimension and per-Gaussian latent dimension , with the settings of § 3 otherwise unchanged. and is the configuration of CryoGaussian and corresponds to Table 2 and Table 3. Higher is better. Bold marks the best per column.
| IgG-1D Mean (std) | IgG-1D Median | IgG-RL Mean (std) | IgG-RL Median | Ribosembly Mean (std) | Ribosembly Median | Tomotwin-100 Mean (std) | Tomotwin-100 Median | Spike-MD Mean (std) | Spike-MD Median | ||
| Per-image | |||||||||||
| 2 | 24 | 0.380 (0.008) | 0.381 | 0.341 (0.028) | 0.343 | 0.422 (0.018) | 0.423 | 0.323 (0.042) | 0.328 | 0.342 (0.017) | 0.344 |
| 4 | 24 | 0.397 (0.004) | 0.398 | 0.366 (0.020) | 0.370 | 0.432 (0.014) | 0.432 | 0.335 (0.035) | 0.341 | 0.364 (0.012) | 0.366 |
| 8 | 24 | 0.390 (0.007) | 0.390 | 0.358 (0.025) | 0.360 | 0.418 (0.024) | 0.419 | 0.315 (0.043) | 0.318 | 0.358 (0.016) | 0.361 |
| 10 | 24 | 0.385 (0.010) | 0.385 | 0.349 (0.027) | 0.355 | 0.405 (0.030) | 0.408 | 0.310 (0.049) | 0.313 | 0.354 (0.022) | 0.356 |
| 4 | 16 | 0.388 (0.007) | 0.388 | 0.356 (0.026) | 0.357 | 0.411 (0.024) | 0.413 | 0.312 (0.047) | 0.318 | 0.355 (0.018) | 0.359 |
| 4 | 32 | 0.384 (0.009) | 0.384 | 0.348 (0.028) | 0.354 | 0.423 (0.020) | 0.425 | 0.327 (0.042) | 0.334 | 0.351 (0.019) | 0.352 |
| Per-conformation | |||||||||||
| 2 | 24 | 0.386 (0.004) | 0.386 | 0.363 (0.016) | 0.365 | 0.424 (0.022) | 0.424 | 0.324 (0.043) | 0.330 | 0.342 (0.018) | 0.343 |
| 4 | 24 | 0.402 (0.001) | 0.402 | 0.387 (0.010) | 0.387 | 0.432 (0.016) | 0.432 | 0.336 (0.034) | 0.343 | 0.365 (0.012) | 0.367 |
| 8 | 24 | 0.396 (0.003) | 0.396 | 0.378 (0.013) | 0.382 | 0.419 (0.025) | 0.420 | 0.316 (0.049) | 0.320 | 0.358 (0.016) | 0.361 |
| 10 | 24 | 0.392 (0.005) | 0.390 | 0.373 (0.015) | 0.374 | 0.406 (0.027) | 0.410 | 0.312 (0.050) | 0.314 | 0.355 (0.022) | 0.357 |
| 4 | 16 | 0.393 (0.003) | 0.392 | 0.380 (0.015) | 0.380 | 0.411 (0.025) | 0.414 | 0.313 (0.047) | 0.318 | 0.356 (0.018) | 0.359 |
| 4 | 32 | 0.389 (0.004) | 0.390 | 0.371 (0.017) | 0.374 | 0.425 (0.021) | 0.426 | 0.328 (0.040) | 0.336 | 0.351 (0.020) | 0.352 |
Table 14.
Image noise and evaluation masks. On IgG-1D, CryoGaussian attains the highest mean per-image FSC AUC among the compared methods at SNR 0.01 and 0.005 (Figure 23), with reconstructions at SNR 0.005 and 0.001 shown in Figure 24 and Figure 25. Removing the evaluation mask changes mean FSC AUC by at most 0.001 on all five datasets (Table 15).
Figure 23. Noise sensitivity of CryoGaussian on IgG-1D. (a–c) UMAP projections of inferred latents at SNR 0.01, 0.005, and 0.001, respectively, colored by the ground-truth dihedral angle. (d) Masked per-image FSC between reconstructed and reference densities. Solid curves show the mean, and shaded regions cover the 5th–95th percentiles across 100 evaluated particle images. (e) Mean per-image FSC AUC as a function of SNR for CryoGaussian and the compared methods. (f, g) Neighborhood similarity, averaged over neighborhood radii of 1–20% of the evaluation subset, and information imbalance of the inferred latents at the three noise levels. (h, i) Corresponding comparisons with the other methods at SNR 0.01, as in Figure 5(a). The legend is shared by (e), (h), and (i), with stars denoting CryoGaussian.
Figure 24. IgG-1D reconstruction at SNR 0.005. (a) Per-image reconstructed densities and paired reference densities. (b) UMAP projection of inferred latents, colored by the ground-truth dihedral angle as in Figure 8. (c) Per-image FSC (top) and per-conformation FSC (bottom), with background masking (left) and without masking (right). Curves and shading show the mean and 5th–95th percentiles across 100 evaluated particle images or 100 reference states, respectively.
Figure 25. IgG-1D reconstruction at SNR 0.001. (a) Per-image reconstructed densities and paired reference densities. (b) UMAP projection of inferred latents, colored by the ground-truth dihedral angle as in Figure 8. (c) Per-image FSC (top) and per-conformation FSC (bottom), with background masking (left) and without masking (right). Curves and shading show the mean and 5th–95th percentiles across 100 evaluated particle images or 100 reference states, respectively.
| Dataset | Evaluation | Masked | Unmasked | ||||
| Mean (std) | Median | Min–max | Mean (std) | Median | Min–max | ||
| IgG-1D (SNR 0.01) | Per-image | 0.397 (0.004) | 0.398 | 0.382–0.404 | 0.397 (0.004) | 0.398 | 0.382–0.403 |
| Per-conformation | 0.402 (0.001) | 0.402 | 0.398–0.404 | 0.401 (0.001) | 0.401 | 0.397–0.404 | |
| IgG-1D (SNR 0.005) | Per-image | 0.391 (0.012) | 0.394 | 0.296–0.400 | 0.391 (0.012) | 0.394 | 0.296–0.399 |
| Per-conformation | 0.398 (0.001) | 0.398 | 0.394–0.402 | 0.397 (0.002) | 0.397 | 0.393–0.401 | |
| IgG-1D (SNR 0.001) | Per-image | 0.302 (0.030) | 0.316 | 0.205–0.336 | 0.292 (0.030) | 0.303 | 0.200–0.331 |
| Per-conformation | 0.328 (0.005) | 0.328 | 0.315–0.339 | 0.318 (0.005) | 0.318 | 0.302–0.329 | |
| IgG-RL | Per-image | 0.366 (0.020) | 0.370 | 0.318–0.400 | 0.366 (0.020) | 0.369 | 0.318–0.400 |
| Per-conformation | 0.387 (0.010) | 0.387 | 0.352–0.410 | 0.387 (0.010) | 0.387 | 0.352–0.410 | |
| Ribosembly | Per-image | 0.432 (0.014) | 0.432 | 0.386–0.452 | 0.432 (0.014) | 0.432 | 0.386–0.452 |
| Per-conformation | 0.432 (0.016) | 0.432 | 0.391–0.454 | 0.432 (0.016) | 0.432 | 0.391–0.454 | |
| Tomotwin-100 | Per-image | 0.335 (0.035) | 0.341 | 0.217–0.387 | 0.335 (0.035) | 0.341 | 0.217–0.387 |
| Per-conformation | 0.336 (0.034) | 0.343 | 0.234–0.390 | 0.336 (0.034) | 0.342 | 0.234–0.389 | |
| Spike-MD | Per-image | 0.364 (0.012) | 0.366 | 0.326–0.386 | 0.364 (0.012) | 0.366 | 0.326–0.386 |
| Per-conformation | 0.365 (0.012) | 0.367 | 0.327–0.386 | 0.365 (0.012) | 0.367 | 0.327–0.386 |
Table 15. Masked and unmasked FSC AUC for CryoGaussian on CryoBench. We report the mean (standard deviation), median, minimum, and maximum across evaluated particle images for per-image FSC and reference states for per-conformation FSC (Appendix § 8.§ 2).
Related Work
Cryo-EM reconstruction. RELION, CryoSPARC, and EMAN2 reconstruct voxelized densities [39, 40, 41], whereas neural fields represent density as a continuous function of Fourier or real-space coordinates [42, 43, 44] [10]. Empirical Bayesian methods regularize reconstruction using signal and noise statistics [45], while Blush and CryoFM incorporate learned density priors [46, 47]. CryoGaussian uses an explicit Gaussian density representation for heterogeneous reconstruction without pretrained density priors.
Heterogeneous reconstruction. Discrete classification reconstructs densities for particle classes [39, 48]. Continuous descriptions use linear subspaces in 3DVA, covariance estimation in RECOVAR and SOLVAR, or latent-conditioned density decoders in CryoDRGN and OPUS-DSD [37, 49] [10, 11, 12]. Hydra and CryoHype model mixtures of distinct complexes with neural fields [50, 51], and CryoKRAQEN learns quantized embeddings [38]. Motion-based approaches use Gaussian displacements, rigid-body transformations, or deformation fields [52] [9, 18, 19]. CryoDRGN-AI directly optimizes particle-image latents, whereas cryoSTAR uses atomic-model constraints to infer latents for per-state reconstruction [53, 54]. CryoGaussian infers compositional and conformational heterogeneity by fitting latent-conditioned Gaussians directly to particle images, and reconstructs each state from the inferred latents.
Gaussian mixtures in cryo-EM. Gaussian mixtures fit density maps and infer heterogeneity from particle images [55, 56, 57] [15]. Their initialization and number of Gaussians commonly incorporate structural information. DynaMight deforms a Gaussian pseudo-atomic model built from a consensus reconstruction [18], e2gmm fits a density map or sets the number of Gaussians from atom counts [27] [15], and CryoSPIRE is a hierarchical, part-based motion model initialized from a consensus reconstruction, whose Gaussians are displaced with their anchors and modulated in amplitude [19]. GaussianEM samples its Gaussians from a homogeneous reconstruction and infers their displacement and density changes with image and Gaussian encoders [20]. CryoGaussian fits a data-derived number of randomly initialized Gaussians directly to particle images and models compositional and conformational heterogeneity through directly optimized particle-image and per-Gaussian latents, without a part-based hierarchy.
Gaussian splatting. Classical volume splatting projects reconstruction kernels into image-plane footprints [58, 59], and 3DGS combines explicit Gaussians with differentiable rasterization and depth-ordered alpha compositing [16]. Adaptations to projection imaging include X-ray image synthesis and tomographic reconstruction [60, 61, 62] and the homogeneous cryo-EM reconstruction of CryoSplat and GEM [63] [5]. Mip-Splatting constrains Gaussian size by the sampling frequency [26], 3DGS-MCMC relocates low-opacity Gaussians [31], and dynamic-scene extensions deform Gaussians over time [64, 65, 66]. CryoGaussian applies CTF-modulated orthographic projection to heterogeneous reconstruction, with a Nyquist bound on Gaussian size, a data-derived number of Gaussians, and Gaussian attributes conditioned on the structural state of each particle image.
Conclusion
We presented CryoGaussian, a Gaussian splatting method for heterogeneous cryo-EM reconstruction that fits an explicit Gaussian density representation directly to particle images from random initialization under data-derived constraints. CryoGaussian attains state-of-the-art reconstruction accuracy across compositional and conformational heterogeneity on CryoBench. On experimental data, it resolves the rare 50S assembly intermediate C4 and continuous motion of the pre-catalytic spliceosome, and enables particle selection from unfiltered particle stacks. These results demonstrate that CryoGaussian achieves both accurate heterogeneity estimation and high-quality per-state reconstruction, with the potential to support mechanistic studies, structure-based drug and vaccine design, and protein engineering (Appendix § 13).
Gaussian parameterization
Base Gaussian attributes
The base amplitude is parameterized as . Without latent conditioning, and , and the covariance is
where each effective principal-axis scale satisfies the Gaussian FWHM constraint in Equation (5). We express scales in physical units throughout, using for a learned base scale before applying the constraint.
Gaussian positions are sampled from an isotropic Gaussian distribution centered in the reconstruction box. We initialize all three base scales to and each rotation matrix to the identity matrix .
Gaussian attribute decoders
The position and amplitude decoders follow Equation (7). The scale decoder produces log-scale modulation from and modulates each base scale separately:
Applying the Gaussian FWHM constraint to each principal-axis scale gives the effective scales and the effective covariance of § 2.3,
where is the learned Gaussian rotation.
Decoder configuration. Each Gaussian attribute decoder takes as input and uses one hidden layer of width 32 with a ReLU activation. Latent dimensions are specified in § 3. Separate decoders produce position modulation, amplitude modulation, and scale modulation; all three are trained during heterogeneity estimation and per-state reconstruction. The position decoder predicts a three-dimensional position modulation that is added directly to the base Gaussian position. Amplitude modulation follows the sigmoid parameterization in Equation (7).
Gaussian relocation at a fixed number of Gaussians
Inspired by the relocation strategy of 3DGS-MCMC [31], we redistribute existing Gaussians while retaining the predetermined number of Gaussians . Gaussians are selected for relocation when they have low base amplitudes. For each selected Gaussian, we sample a Gaussian from those not selected for relocation, with replacement and with probabilities proportional to their base amplitudes. The relocated Gaussian copies the base position, scales, and rotation of the sampled Gaussian. If Gaussian is sampled times, its base amplitude before relocation is divided equally among Gaussian and the relocated Gaussians, so that each receives .
Experimental settings
Datasets and preprocessing
Table 4 summarizes the CryoBench particle images used in our experiments. Table 5 summarizes the source particle-image datasets; particle selection and preprocessing are described below.
CryoBench data. We use the released CryoBench datasets [21] and their supplied pose and CTF metadata, without additional particle selection. IgG-1D, IgG-RL, and Tomotwin-100 each contain 1,000 particle images per state; Ribosembly has unequal state populations. IgG-1D, IgG-RL, and Ribosembly are Fourier-downsampled from to , and Tomotwin-100 from to . In each case, the physical field of view is preserved. Spike-MD is used at its released image dimensions. Its reference-state count follows the sampled molecular-dynamics configurations rather than a partition into discrete structural classes.
Experimental data. For the main EMPIAR-10076 and EMPIAR-10180 reconstructions, we follow the particle selection and preprocessing procedures provided with CryoDRGN [10], using the corresponding pose and CTF metadata. We Fourier-downsample the images from EMPIAR-10076 and EMPIAR-10180 to , following CryoDRGN [10]. This preserves the physical field of view and yields pixel sizes of 1.6375 and 2.12375 Å/pixel, respectively. EMPIAR-10076 contains 131,899 source images; its CryoDRGN-provided selection contains 87,327 images. For EMPIAR-10180, the main reconstruction uses 139,720 particle images from the CryoDRGN-filtered subset. We exclude corrupted particle 308902 from the 327,490-image source stack, leaving 327,489 images for the separate full-stack analysis in Figure 16, Fourier-downsampled to at 4.2475 Å/pixel following Zhong et al. [10].
CryoGaussian particle selection for EMPIAR-10076. For the separate filtering analysis in Figure 20, we follow the filtering strategy of Zhong et al. [10] with one- and four-dimensional models trained on all 131,899 particle images. The four-dimensional model uses the same latent dimension as our heterogeneous reconstruction. For the one-dimensional latents, we form a Gaussian kernel density estimate and exclude the interval between the two local minima adjacent to its highest peak, retaining 100,528 images. For the four-dimensional latents, we fit a five-component Gaussian mixture and exclude the component with the largest mean latent norm, retaining 105,211 images. The intersection of the two retained sets contains 97,651 images. We train the subsequent reconstruction on a random 90% of this intersection, comprising 87,885 images. Published class labels are used only for post-hoc evaluation of the selection.
Training settings
Heterogeneity estimation and per-state reconstruction use the same particle images, pose and CTF metadata, and particle selections, and each use 30,000 iterations (§ 3). We set the base learning rates to for Gaussian positions; for Gaussian scales, rotations, base-amplitude parameters, and trainable particle-image latents; and for per-Gaussian latents and the position, amplitude, and scale decoders. These rates are defined for a global batch size of one and scaled by the square root of the global batch size, consistent with the learning-rate scaling in Grendel and gsplat [67, 68]. The Gaussian position learning rate decays to 1% of its initial value, following 3DGS and X-Gaussian [60] [16]. We use Adam [36] with momentum coefficients (0.9, 0.999), , and zero global weight decay. Inspired by 3DGS-MCMC [31], we relocate Gaussians every 100 iterations starting at iteration 500 and stopping at iteration 25,000 (Appendix § 6.§ 3). Latent dimensions are specified in § 3 and the relative density threshold in § 2.2. For IgG-1D, IgG-RL, Tomotwin-100, Spike-MD, EMPIAR-10076, and EMPIAR-10180, global batch sizes are 256 for heterogeneity estimation and 128 for per-state reconstruction. For the larger Ribosembly particle stack, we use 512 and 256, respectively. CryoGaussian results are from one training run. To assess variability across training seeds, heterogeneity estimation and per-state reconstruction are additionally rerun with 10 training seeds for the clustering accuracy in Table 7 and the seed-variability analysis in Table 8.
Training time and memory
We run CryoGaussian on eight NVIDIA RTX PRO 6000 Blackwell GPUs. Table 6 reports global batch sizes, times per iteration, training times, forward throughput, and peak allocated memory per GPU for heterogeneous reconstruction, together with the speedups and memory reductions of Table 1. Table 1 isolates the two rasterizer optimizations of § 2.4 and distributed training with three implementations of the rasterizer that evaluate the same image-formation model and differ only in how projections and gradients are scheduled. The CryoSplat rasterizer [5] scatters gradients pixel-wise with atomic updates and processes each tile list in one pass, the second accumulates gradients per Gaussian (Appendix § 12.§ 2), and the third additionally partitions long tile lists into independent partial projections (Appendix § 12.§ 3) and is the rasterizer of CryoGaussian. Speedups are ratios of mean time per iteration, forward and backward passes and the optimizer update included, over training runs with the global batch sizes of Table 6 and otherwise identical settings. The two rasterizer speedups are measured on one GPU, so that they exclude communication between GPUs, and are applied cumulatively, the first from the CryoSplat rasterizer to the second and the second from the second to the third, so that their product, 12.8–28.0 across the seven datasets, is the speedup of the CryoGaussian rasterizer over the CryoSplat rasterizer. The distributed-training speedup is the ratio of the mean time per iteration of the CryoGaussian rasterizer on one GPU to that on eight GPUs at the same global batch, with each GPU processing one eighth of the batch, and corresponds to a strong-scaling efficiency, the speedup divided by the number of GPUs, of 64–81%.
Releasing the forward workspaces before the backward pass (Appendix § 12.§ 3) lowers peak allocated memory per GPU by 8.0–10.2% at these global batch sizes, and Table 6 reports the peak with the workspaces retained and this reduction. These optimizations make the numbers of Gaussians given by Equation (6) and the full box sizes of the experimental datasets practical to optimize, with training times of 6.8–8.5 h on eight GPUs for EMPIAR-10076 and EMPIAR-10180 (Table 6).
Baseline configurations
The FSC AUC comparisons include CryoDRGN, CryoDRGN-AI-fixed, OPUS-DSD, 3DFlex, 3DVA, RECOVAR, 3D Classification, CryoSPIRE, CryoHype, SOLVAR [37], and CryoKRAQEN. Dataset coverage follows Table 2 and Table 3 under the CryoBench evaluation framework [21]. Methods with publicly available implementations are run following CryoBench [21]. For CryoSPIRE [19] and CryoKRAQEN [38], the included results are taken directly from their publications, as we were unable to find a publicly available implementation at the time of writing. For CryoHype [51], the included result is taken directly from its publication. Its public release provides the configuration for the Tomotwin-100 experiment only, and its IgG-1D result is obtained with an additional smoothness loss for which we were unable to find a specification in the publication or a configuration in the release. Obtaining results for CryoHype on IgG-1D and the other datasets would require configurations and an implementation of this term that the publication does not specify, so we do not include them. CryoHype and CryoKRAQEN are compared under the per-image protocol only (Table 3). Dashes denote results not included for a given dataset and protocol.
Reconstruction methods in our comparisons
Our CryoBench comparisons use the same supplied ground-truth poses as fixed inputs [21]. The compared methods represent structural variability through discrete states, continuous subspaces, deformation fields, or neural representations. 3D classification estimates a finite set of volumes [21, 40], whereas 3DVA learns a linear subspace [49]. RECOVAR combines regularized covariance estimation with adaptive kernel regression for per-state reconstruction [11]; SOLVAR directly optimizes a low-rank covariance representation and uses RECOVAR’s reconstruction procedure with its estimated principal components [37]. 3DFlex learns latent-conditioned deformation fields that transport a canonical density [9]. CryoDRGN learns image embeddings with an encoder and a latent-conditioned neural field [10], while CryoDRGN-AI-fixed directly optimizes particle-image latents [21, 53]. OPUS-DSD uses a convolutional encoder–decoder with a structural disentanglement prior [12]. CryoHype uses a transformer hypernetwork to condition implicit neural representation weights on each particle image [51], and CryoKRAQEN combines quantized embeddings, kernel-guided assignment, and a triplane representation [38]. Among Gaussian approaches, CryoSPIRE jointly optimizes particle-image latents and Gaussian features in a model initialized from a reconstructed density. Feature clustering identifies structural parts and anchors for a hierarchical representation of part-level motion, local deformation, and occupancy variation [19]. CryoGaussian models compositional and conformational heterogeneity through latent-conditioned Gaussian attributes, starting from randomly initialized Gaussians. Particle-image and per-Gaussian latents jointly condition this representation, and each structural state is then reconstructed conditioned on the inferred particle-image latents with a representation optimized from randomly initialized Gaussians.
Evaluation protocols
The synthetic-data evaluation distinguishes heterogeneity estimation from per-state reconstruction, following the evaluation framework of Jeon et al. [21]. The evaluation of reconstructions from the experimental datasets (EMPIAR-10076 and EMPIAR-10180), which provide no ground truth, is described in Appendix § 8.§ 3.
Heterogeneity estimation
Neighborhood similarity (percentage of matching neighbors, pMN) and information imbalance compare the inferred latents with the ground-truth heterogeneity embedding of each particle image, as defined in CryoBench [21]. These comparisons are made on all five datasets with the embedding that CryoBench defines for each of them, the sine and cosine of the dihedral angle for IgG-1D, the center of mass and the sine and cosine of the dihedral angle for IgG-RL, the size rank of the atomic models for Tomotwin-100 and for neighborhood similarity on Ribosembly, the voxel intensities of the reference densities for information imbalance on Ribosembly, and the two collective variables of the receptor-binding domain for Spike-MD. Neighborhood similarity is reported as the mean and standard deviation over the evaluation subsets of CryoBench as a function of the neighborhood radius, over the range of radii shown for each dataset in the CryoBench paper. Following the CryoBench protocol, information imbalance is reported as the mean and standard deviation over , 3, 10, and 30 nearest neighbors (, 3, and 10 for Tomotwin-100). As in CryoBench, it is also computed between the inferred latents and the ground-truth rotation of each particle image, represented by its flattened rotation matrix, and its CTF parameters, represented by the two defocus values and the sine and cosine of the astigmatism angle after standardization. Clustering accuracy on Ribosembly and Tomotwin-100 follows CryoBench, and for CryoGaussian it is reported as the mean and standard deviation over 10 random seeds.
Per-image and per-conformation FSC evaluation
For density maps and , Fourier shell correlation is defined as [21]
Here, is the Fourier shell at radial spatial frequency ; hats denote Fourier transforms, and denotes complex conjugation. For masked comparisons, the evaluation mask is applied to both maps before the Fourier transform. Following CryoBench’s per-image evaluation implementation [21], we compute FSC AUC by trapezoidal integration:
where are the sampled frequencies in cycles/voxel and . We retain this frequency scale without rescaling AUC to a unit interval; its theoretical maximum over the full frequency range is 0.5 [21].
Per-image FSC compares a density reconstructed at an evaluated particle’s inferred latent with that particle’s reference density. It therefore retains the correspondence between an image and its underlying state. Per-conformation FSC instead selects one representative latent for each evaluated reference state: the latent nearest to the mean of the observed latents assigned to that state is decoded and compared with the corresponding reference density. These two evaluations differ in their sampling and reference pairing [21].
For CryoGaussian, densities are evaluated from the final Gaussian density representation . Per-image and per-conformation FSC and their AUC summaries follow the CryoBench evaluation protocol and implementation [21], with background masking unless otherwise stated. The density panels in Figure 8–Figure 12 use inferred latents for particle images; the FSC panels report the two evaluations separately. Solid curves show the arithmetic mean, and shaded regions cover the 5th–95th percentiles across the evaluated images or reference states. Figure 13 instead selects an observed latent nearest to each -means cluster center, using 20 clusters except for Ribosembly, which uses 16. Colors identify cluster membership within each dataset. For Spike-MD, we generate reference densities from the corresponding molecular-dynamics structures using ChimeraX 1.6.1 molmap, following the dataset-generation protocol of Jeon et al. [21]. For this dataset, the reference states are the distinct molecular-dynamics configurations, each shared by approximately two particle images [21], and the per-conformation evaluation selects one representative latent per configuration as for the other datasets.
Evaluation on experimental data
Assembly-state selection and comparison. We train CryoGaussian on the full EMPIAR-10076 particle stack for particle selection (Figure 20; Appendix § 10.7). Assembly-state comparisons use both the resulting CryoGaussian-filtered subset and the CryoDRGN-filtered subset. Published class assignments [24] identify the particle images associated with each reference state. We decode the observed latent nearest to the class-mean latent to obtain a representative density, following the selection procedure of Zhong et al. [10]. Map-to-map FSC compares these densities with the corresponding deposited reconstructions. The published assignments are used to select and evaluate representative densities after training. For the assembly intermediate with H68 density but no resolved central protuberance (class C4) [10], the CryoDRGN-filtered reconstruction uses a single representative latent: the observed particle latent nearest to the mean inferred latent of the 1,113 particles identified by the released CryoDRGN state indices. We decode this latent using the same representative-latent convention and density rendering pipeline as the other assembly classes. We additionally perform homogeneous refinement in CryoSPARC. Panel (a) of Figure 15 uses all 1,113 particles identified by the released CryoDRGN state indices; panel (b) uses their intersection with the 97,651-particle set retained by CryoGaussian filtering, yielding 1,019 particles. This intersection uses the full filtered set before the random 90% training subsampling. The identification of this intermediate does not require the released state indices. On the UMAP projection of the inferred latents, colored by published major class assignments, we manually outline, at the boundary of the class C cluster, a group of particles that includes particles assigned to class E by the published classification (panel (c)). The outlined regions contain 1,062 particles for the CryoDRGN-filtered subset and 954 for the full CryoGaussian-filtered set. Panels (d) and (e) show the CryoGaussian densities decoded at the observed latent nearest to the mean inferred latent of the selected particles, together with the CryoSPARC refinement of these particles. Gold-standard FSC assesses consistency between the independent CryoSPARC half-maps, while map-to-map FSC assesses agreement between the CryoGaussian and CryoSPARC densities.
Spliceosome latent selection. For the main EMPIAR-10180 reconstruction using the CryoDRGN-filtered subset, we apply principal component analysis to the inferred latents and decode the five positions marked along each of the first two principal components in Figure 4. Separately, we train CryoGaussian on the full particle stack (Figure 16), fit 20 -means clusters to the inferred latents, and subdivide the largest cluster into two groups. Representative densities are decoded at the observed latents nearest to the cluster centers.
Additional reconstruction results
CryoBench reconstruction
Table 2 complements the per-image comparison by evaluating one representative latent for each reference state. For CryoGaussian, mean per-conformation FSC AUC is 0.402 on IgG-1D, 0.387 on IgG-RL, 0.432 on Ribosembly, 0.336 on Tomotwin-100, and 0.365 on Spike-MD. Table 7 reports clustering accuracy of the inferred latents on Ribosembly and Tomotwin-100, and Table 8 reports the variability of the mean FSC AUC across training seeds. Figure 5 reports neighborhood similarity and information imbalance between the inferred latents and the ground-truth heterogeneity embedding on all five datasets for CryoGaussian and the compared methods, computed and presented following CryoBench (Appendix § 8.§ 1). In CryoBench, neighborhood similarity at the smallest radius reflects how well images of the same structure are grouped and at larger radii how the relations between structures are preserved, and information imbalance near (0, 0) indicates equivalent information, near (1, 1) orthogonal information, and below the diagonal that the inferred latents predict the ground-truth neighborhoods more closely than the reverse [21]. The neighborhood similarity of CryoGaussian is comparable to or higher than that of the compared methods over most of the evaluated radii and the highest at the largest radius on all five datasets, where it is 95.1% on IgG-1D, 45.0% on IgG-RL, 75.5% on Ribosembly, 23.4% on Tomotwin-100, and 37.2% on Spike-MD. Its information imbalance is (0.059, 0.041) on IgG-1D, (0.238, 0.201) on IgG-RL, (0.101, 0.098) on Ribosembly, (0.102, 0.076 on Tomotwin-100, and (0.435, 0.123) on Spike-MD, in the order , close to (0, 0) on IgG-1D, Ribosembly, and Tomotwin-100 and below the diagonal on IgG-RL and Spike-MD. Figure 6 and Figure 7 report the information imbalance between the inferred latents and the ground-truth rotation and CTF parameters of each particle image for CryoGaussian and the compared methods, computed with the pose and CTF embeddings, neighborhood sizes, and evaluation subset of CryoBench (Appendix § 8.§ 1) and presented in the format of Figures S22 and S23 of the CryoBench paper [21]. measures how well neighborhoods in the ground-truth rotation or CTF parameters predict neighborhoods in the latent space, the reverse, and values near 1 in both directions correspond to orthogonal information as defined in CryoBench, where neither predicts the other. For CryoGaussian, the information imbalance with the ground-truth rotations is (0.9995, 1.0097) on IgG-1D, (0.9919, 1.0185) on IgG-RL, (1.0004, 1.0039) on Ribosembly, (0.9993, 1.0061) on Tomotwin-100, and (1.0000, 1.0029) on Spike-MD, and with the CTF parameters (0.9988, 0.9970), (0.9989, 0.9982), (1.0001, 0.9993), (1.0001, 0.9973), and (0.9990, 1.0001) on the same datasets, in the order . As in CryoBench, each value is the mean of the information imbalance over the neighborhood sizes , and its standard deviation over these sizes, shown as error bars in the figures, is at most 0.004 for CryoGaussian. The values for the compared methods lie between 0.92 and 1.09 for rotations and between 0.96 and 1.00 for CTF parameters, so that the latents of all methods are orthogonal to rotation and CTF parameters in this sense, as observed in CryoBench.
Figures 8–12 pair per-image reconstructed densities with their corresponding reference densities and show masked and unmasked FSC curves. The paired displays relate per-state reconstruction to the reference heterogeneity represented by the inferred latents. Figure 13 provides a complementary selection based on latent clustering, with each representative density decoded at an observed latent nearest to a cluster center. Lower-SNR IgG-1D results are discussed separately in Appendix § 10.8.
Heterogeneous reconstruction on experimental data
Assembling 50S ribosomal subunit. Figure 14 shows the reconstructed minor assembly states and their map-to-map FSC against deposited densities using the CryoDRGN-filtered particle subset.
Figure 9. IgG-RL reconstruction. (a) CryoGaussian densities reconstructed at the inferred latents for 100 particle images (Ours), paired with their reference densities (G.T., gray). Colors indicate the distance from the Fab domain to the center of mass of the rest of the structure. (b) UMAP projection of inferred latents with the same color scale. (c) Per-image FSC (top) and per-conformation FSC (bottom), with background masking (left) and without masking (right). Solid curves show the mean, and shaded regions cover the 5th–95th percentiles across 100 evaluated particle images or 100 reference states, respectively.
Figure 10. Ribosembly reconstruction. (a) CryoGaussian densities reconstructed at inferred latents for particle images from all 16 reference assembly states (Ours), paired with their reference densities (G.T., gray). (b) UMAP projection of inferred latents. Colors in (a) and (b) identify reference assembly states. (c) Per-image FSC (top) and per-conformation FSC (bottom), with background masking (left) and without masking (right). Solid curves show the mean, and shaded regions cover the 5th–95th percentiles across 101 evaluated particle images in the top row and 16 reference states in the bottom row.
Figure 11. Tomotwin-100 reconstruction. (a) Per-image CryoGaussian densities (Ours) and paired reference densities (G.T., gray) for 100 distinct complexes. (b) UMAP projection of inferred latents; colors in (a) and (b) indicate molecular mass in kDa. (c) Per-image FSC (top) and per-conformation FSC (bottom), with background masking (left) and without masking (right). Solid curves show the mean, and shaded regions cover the 5th–95th percentiles across 100 evaluated particle images or 100 reference complexes, respectively.
Representative latents are selected as described in Appendix § 8.§ 3. Figure 15 compares the intermediate with H68 density but no resolved central protuberance with homogeneous CryoSPARC reconstructions using the released CryoDRGN state selection (a), its intersection with the CryoGaussian-filtered particle set (b), and particles selected in the CryoGaussian latent space (c–e). The intermediate is identified in this latent space without the released state indices: refinement of the selected particles also shows H68 density and no resolved central protuberance. The separate particle-selection analysis is described in Appendix § 10.7.
Pre-catalytic spliceosome. The principal-component traversals in Figure 4 describe conformational variation within the CryoDRGN-filtered subset. Figure 16 separately examines the full particle stack using the preprocessing in Appendix § 7.§ 1. Its cluster-selected densities include differences in SF3b density and additional density near the U2 core.
Ablations and robustness
Number of Gaussians and relative density threshold
The number of Gaussians is determined by the estimated spatial support through Equation (6), and the support by the backprojection and the relative density threshold through Equation (21). We vary over and let the number of Gaussians follow. Raising removes the voxels of lowest relative density from the estimated support (Equation (25)), and across the five CryoBench datasets this range of changes the number of Gaussians from 0.33 to 1.74 times its value at the default threshold. Figure 17 evaluates heterogeneity estimation on IgG-1D at SNR 0.01, and Table 9 evaluates per-image and per-conformation FSC AUC on all five datasets as in Table 2 and Table 3.
On IgG-1D, the default gives the highest neighborhood similarity, and thresholds of 0.01–0.05 retain conformational ordering in the UMAP projections. Increasing to 0.10 reduces the estimated support and the number of Gaussians, yielding a fragmented projection and lower neighborhood similarity. Conversely, lowering to 0.01 increases the number of Gaussians without improving neighborhood similarity over the default.
On all five datasets and under both protocols, FSC AUC is highest at and decreases with every further increase in , by 0.033–0.067 in mean FSC AUC between and , with the largest decrease on Tomotwin-100, where the number of Gaussians falls from 57,818 to 19,611. Lowering to 0.01 enlarges the estimated support and the number of Gaussians without raising FSC AUC above the default on any dataset. Reconstruction accuracy is thus highest at the default threshold, decreases as the estimated support is restricted to denser regions of the backprojection, and does not increase when weaker density is included.
Gaussian FWHM constraint
We compare heterogeneity estimation on IgG-1D with and without the Gaussian FWHM constraint at SNR 0.01 and 0.001 (Figure 18). The constraint in Equation (5) sets a lower bound of on Gaussian FWHM, preventing Gaussian widths from decreasing below this sampling scale. At SNR 0.01, the inferred latents retain an ordered loop in the UMAP projection with the constraint, whereas removing it produces folded, overlapping branches. At SNR 0.001, both projections become diffuse. Neighborhood similarity, evaluated in the inferred latent space against the ground-truth heterogeneity embedding, is higher with the constraint across the displayed neighborhood radii at both noise levels. These comparisons support constraining Gaussian FWHM when learning conformational heterogeneity from noisy particle images.
Table 10 evaluates per-image and per-conformation FSC AUC on all five CryoBench datasets with and without the constraint, as in Table 2 and Table 3. On all five datasets and under both protocols, removing the constraint decreases FSC AUC, by 0.016–0.030 in mean and 0.016–0.028 in median FSC AUC, and increases its standard deviation across evaluated particle images and across reference states. The constraint raises reconstruction accuracy on every dataset in addition to neighborhood similarity on IgG-1D.
Gaussian initialization
IgG-1D provides a ground-truth one-dimensional conformational motion for evaluating whether inferred latents preserve reference conformational neighborhoods. We compare random and consensus initialization for heterogeneity estimation at SNR 0.01 and 0.001, giving four runs in total (Figure 19).
Random initialization follows Appendix § 6. For consensus initialization, we first obtain a consensus density by optimizing the Gaussian representation without latents on all particle images for 30,000 iterations, with the learning rates, data-derived constraints, and relocation schedule of heterogeneity estimation in Appendix § 7.§ 2 otherwise unchanged, which yields a single density fitted to all particle images irrespective of their structural states. We initialize Gaussian positions by density-weighted sampling without replacement over thresholded voxels. Only the positions are initialized from this density; the other Gaussian attributes use the same initialization as in the random-initialization runs. The positions are optimized during training.
Each run jointly learns latents for the particle images and the Gaussian density representation. We compute neighborhood similarity in the inferred latent space following Appendix § 8.§ 1; UMAP projections visualize the latents colored by the ground-truth dihedral angle.
Averaged over 20 neighborhood radii (1–20% of the evaluation subset), pMN is 82.73% for random initialization and 55.84% for consensus initialization at SNR 0.01. At SNR 0.001, the corresponding values are 30.89% and 27.10%. Random initialization yields higher pMN across all evaluated neighborhood radii at both noise levels.
Per-state reconstruction strategy
Table 11 compares three ways of obtaining per-state densities on the five CryoBench datasets at the same total number of iterations and otherwise identical settings. Heterogeneity estimation alone optimizes and jointly for the combined 60,000 iterations of heterogeneity estimation and per-state reconstruction and evaluates the densities of at the inferred latents. Per-state reconstruction continued from continues the optimization of for the per-state reconstruction iterations conditioned on the inferred latents . Per-state reconstruction from randomly initialized conditioned on follows § 2.3 and corresponds to the results in Table 2 and Table 3. The two per-state reconstruction strategies use the same heterogeneity estimation and its inferred latents , so that their difference reflects per-state reconstruction alone. Per-state reconstruction from randomly initialized attains the highest mean FSC AUC on all five datasets under both protocols, by 0.008–0.016 over heterogeneity estimation alone and 0.005–0.010 over per-state reconstruction continued from . On every dataset and under both protocols, the increase in mean FSC AUC from continued optimization to reconstruction from randomly initialized exceeds the increase from heterogeneity estimation alone to continued optimization, so that the larger part of the improvement at the same total number of iterations arises from reconstruction from random initialization rather than from further optimization of .
Combinations of Gaussian attribute decoders
Tables 12 and 13 evaluate all 15 combinations of the position, amplitude, scale, and rotation decoders, from one decoder to all four, on the five CryoBench datasets with the settings of § 3 otherwise unchanged. The combination of the position, amplitude, and scale decoders is the configuration of CryoGaussian (§ 2.3 and Appendix § 6.§ 2), in which the Gaussian rotation has no decoder. The rotation decoder has the same architecture as the other decoders and produces an axis–angle vector from , so that replaces in Equation (10), where denotes the skew-symmetric matrix of a vector. Its output layer is initialized to zero, so that the decoded rotation equals at initialization, and is still learned per Gaussian. In every combination, the attributes without a decoder keep their base values, so that a Gaussian without the position decoder stays at , one without the amplitude decoder keeps , one without the scale decoder keeps its base scales under the FWHM constraint, and one without the rotation decoder keeps (Appendix § 6).
The combination of the position, amplitude, and scale decoders attains the highest mean and median FSC AUC and the lowest standard deviation on all five datasets under both protocols, by 0.007–0.010 in mean FSC AUC over the best combination of two decoders and 0.044–0.081 over the best single decoder. Removing one decoder from this combination decreases mean FSC AUC on every dataset under both protocols, by 0.013–0.027 for the amplitude decoder, which is the largest decrease on every dataset, by 0.007–0.020 for the position decoder, and by 0.008–0.015 for the scale decoder. Removing the position decoder decreases mean FSC AUC by 0.012–0.020 on the conformational datasets IgG-1D, IgG-RL, and Spike-MD and by 0.007–0.010 on the compositional
datasets Ribosembly and Tomotwin-100, and removing the scale decoder by 0.008–0.009 on the conformational and 0.011–0.015 on the compositional datasets. The same ordering holds with one and two decoders, where the amplitude decoder alone attains the highest mean FSC AUC on every dataset, the combination of the position and amplitude decoders is the highest combination of two decoders on the conformational datasets, and the combination of the amplitude and scale decoders is the highest on the compositional datasets, in agreement with the roles of the three attributes in Section 2.3, where position modulation moves Gaussians between states and amplitude and scale modulation change their presence and shape. Adding the rotation decoder decreases mean FSC AUC in every combination on every dataset under both protocols, by 0.001–0.006 when added to one or two decoders and by 0.006–0.012 when added to the combination of the position, amplitude, and scale decoders, and the rotation decoder alone attains the lowest mean FSC AUC on every dataset. The configuration of CryoGaussian is accordingly the combination of the position, amplitude, and scale decoders, with no rotation decoder.
Latent dimensions
Table 14 varies the particle-image latent dimension at and the per-Gaussian latent dimension at on the five CryoBench datasets with the settings of Section 3 otherwise unchanged. The configuration of CryoGaussian, and , attains the highest mean and median FSC AUC and the lowest standard deviation on all five datasets under both protocols. Decreasing to 2 decreases mean FSC AUC by 0.016–0.025 on the conformational datasets IgG-1D, IgG-RL, and Spike-MD and by 0.008–0.012 on the compositional datasets Ribosembly and Tomotwin-100, whereas increasing to 8 or 10 decreases it by 0.006–0.017 on the conformational and by 0.013–0.027 on the compositional datasets. Decreasing to 16 decreases mean FSC AUC by 0.007–0.010 on the conformational and by 0.021–0.023 on the compositional datasets, whereas increasing to 32 decreases it by 0.013–0.018 on the conformational and by 0.007–0.009 on the compositional datasets. The conformational datasets are more sensitive to a smaller particle-image latent and a larger per-Gaussian latent, and the compositional datasets to a larger particle-image latent and a smaller per-Gaussian latent.
Particle selection
We examine EMPIAR-10076 using both the CryoDRGN-filtered and CryoGaussian-filtered particle subsets. Figure 20 documents the CryoGaussian particle selection (Appendix § 7.§ 1). Of the 97,651 retained particles, a random 90% (87,885 particles) are used for subsequent heterogeneous reconstruction. Figure 21 shows the reconstructed minor assembly states and their map-to-map FSC against deposited densities for the CryoGaussian-filtered subset. Representative latents are selected as described in Appendix § 8.§ 3. Figure 22 compares reconstructed densities for major assembly classes B–E with CryoDRGN and across the two particle selections used for CryoGaussian reconstruction, and marks regions in which the CryoDRGN density is fragmented and noise-like whereas both CryoGaussian reconstructions are continuous. Both particle selections recover the intermediate with H68 density but no resolved central perturbance. Homogeneous CryoSPARC refinement of the intersection between the released state selection and the CryoGaussian-filtered particle set shows the same features (Figure 15(b)), as does refinement of the 954 particles selected in the CryoGaussian latent space (Figure 15(c, Figure 15e)). Starting from the unfiltered particle stack, and without the released state indices, CryoGaussian accomplishes the particle selection, the identification of this intermediate in its latent space, and its reconstruction.
Sensitivity to image noise
We use the three released IgG-1D noise conditions at SNR 0.01, 0.005, and 0.001 [21]. Figure 23 combines latent-space evaluation and per-image FSC for these conditions. Mean masked per-image FSC AUC is 0.397, 0.391, and 0.302, respectively. CryoGaussian achieves the highest mean per-image
FSC AUC among the compared methods at SNR 0.01 and 0.005. The decrease is more pronounced at SNR 0.001. The corresponding mean per-conformation FSC AUC values are 0.402, 0.398, and 0.328 (Table 15). Neighborhood similarity averaged over neighborhood radii of 1–20% is 82.7%, 76.4%, and 30.9%, and information imbalance from inferred latents to the ground-truth heterogeneity embedding is 0.041, 0.060, and 0.383 at the three noise levels. At SNR 0.01, the information imbalance between the ground-truth heterogeneity embedding and the inferred latents is 0.059 and 0.041 in the two directions, and the comparison with the other methods in Figure 23(h, Figure 23i) corresponds to Figure 5(a).
Estimated numbers of Gaussians for these noise conditions are reported in Table 6. Figure 24 and Figure 25 show the lower-SNR reconstructed densities and FSC curves in detail. Baseline sources and latent-space evaluation procedures for the composite comparison are specified in Appendices § 7.§ 4 and § 8.§ 1.
Sensitivity to masking
We assess the sensitivity of per-image and per-conformation FSC AUC to background masking by evaluating the same reconstructed densities with and without the evaluation mask (Table 15), following the masked and unmasked comparison in CryoBench [21]. Removing the mask changes the mean FSC AUC by at most 0.001 at the displayed precision, except for IgG-1D at SNR 0.001, where the mean decreases by approximately 0.010 for both per-image and per-conformation evaluation. These results indicate limited sensitivity of the reported FSC AUC to background masking, with a larger effect at the lowest SNR.
Estimating the number of Gaussians
This appendix gives the basis of the space–bandwidth rule for the number of Gaussians, the backprojection used to estimate spatial support, and the interpretation of the relative density threshold.
Space–bandwidth basis
We derive the space–bandwidth product in Equation (6) using the multidimensional space–frequency concentration problem [28, 30]. Let be a measurable spatial region of finite volume and a frequency region of finite volume. We use the Fourier convention
The concentration operator on is
To justify the trace, define by
Then . Zero extension and Plancherel’s theorem give , while . Thus is positive and trace class, has eigenvalues , and its trace is
The trace is the space–bandwidth product. It provides a nominal effective dimension for spatially and spectrally concentrated signals; its interpretation as a count of well-concentrated modes is approximate, whereas the trace identity is exact [29].
For the ball ,
Treating the above-threshold voxels as a union of physical voxel cells gives . At ,
The factor is the volume factor of the frequency-domain ball; the spatial support need not be spherical or connected. We obtain the number of Gaussians by rounding the space–bandwidth product to an integer.
The concentration identity concerns a linear space of signals, and we use its value as the number of Gaussians in the nonlinear Gaussian representation. The relative density threshold specifies the estimated spatial support, and the sampling bandwidth specifies the spatial scale at which the product is evaluated. The Gaussian FWHM constraint uses the same spatial scale.
Backprojection for spatial support estimation
We compute a CTF-weighted backprojection, the closed-form linear reconstruction of the particle images at their supplied orientations, using regularized gridding in CryoDRGN (version 3.4.4) [10]. The backprojection uses all particle images in the corresponding dataset. Particle images are corrected for the supplied in-plane translations. Fourier downsampling, when used, preserves the physical field of view. After any downsampling, the pixel size is also the voxel spacing used for spatial support estimation. The supplied orientations are used to backproject the particle images. Let denote the voxel centers of the resulting density on a grid with spacing . For a relative density threshold , the estimated spatial support is defined by
assuming , and is the union of the voxel cells of , with . Only the number of Gaussians derived from this support is used in subsequent reconstruction, and Gaussian attributes are not initialized from the backprojected density.
Interpretation of the relative density threshold
The relative density threshold defines the estimated spatial support used in the space–bandwidth product, and its numerical value is reported in § 2.2. Thresholding relative to the positive maximum makes this estimate invariant to a positive multiplicative rescaling of the backprojected density. A lower threshold includes more weak-density regions in the estimated spatial support.
Support-volume interpretation. Let . The voxel set admits an equivalent characterization as a maximizer of a discrete excess-mass criterion:
Each voxel contributes ; selecting all nonnegative contributions gives . This criterion balances retained peak-normalized density against support size. Since and , the support-size penalty also controls the space–bandwidth product used to set .
Gaussian reference. A single three-dimensional Gaussian provides an analytical reference for . Its density superlevel sets are highest-density regions: each encloses a given probability mass within the minimum volume [70]. Let and let denote the cumulative distribution function of a chi-square random variable with three degrees of freedom. Then
The 95-percent mass ellipsoid has relative boundary density ; the relevant chi-square quantile is approximately 7.815 [71]. Thus, approximates this relative boundary density and encloses approximately 0.9502 of the Gaussian mass. An invertible affine coordinate transformation gives the same relation for any nondegenerate three-dimensional Gaussian. This calculation supplies an analytical reference for the fixed threshold; the fraction of the integrated positive backprojected density enclosed by is data dependent.
Dependence on the threshold. For a fixed backprojection and voxel grid, two thresholds satisfy
The change in estimated spatial support is therefore determined exactly by the voxels whose normalized densities lie between the two thresholds. Its effect on the number of Gaussians follows through . The space–bandwidth identity gives conditional on ; it does not determine . Figure 17 and Table 9 evaluate the empirical effect of varying on heterogeneity estimation and per-state reconstruction; the corresponding results are discussed in Appendix § 10.§ 1.
Efficient cryo-EM Gaussian splatting
This appendix describes the gradient reductions and the partitioning of Gaussian lists for image tiles underlying § 2.4. We use the same Gaussian density, CTF modulation, and reconstruction objective as in § 2.1, where the projection operator in Equation (2) is the line integral along the beam direction . Let denote the contribution of Gaussian to the predicted image before CTF modulation: it is the th summand in Equation (2), multiplied by in Equation (3). Rasterization groups the projected Gaussian footprints into image tiles. A Gaussian is assigned to a tile only when its footprint overlaps that tile, as in Speedy-Splat [72], rather than to every tile overlapping the axis-aligned bounding box of the footprint. The following reductions refer to the Gaussian–pixel interactions selected by this assignment.
Gradients through CTF modulation
The CTF-modulated prediction is related to the projection by a linear operator. Applying its adjoint to the image residual gives the gradient supplied to the rasterizer:
where the overline denotes complex conjugation; the CTFs used here are real-valued. The factor follows from Equation (4). CTF modulation couples spatial locations in the image domain, while the rasterizer receives after this operation has been differentiated. It can therefore accumulate gradient contributions over each projected Gaussian footprint using .
Per-Gaussian gradient accumulation
For differentiation, regard the effective Gaussian attributes for each particle image as separate intermediate variables, . These attributes are functions of and, in heterogeneous reconstruction, , as defined in § 2.1. Let be the image-grid locations assigned to Gaussian by the rasterizer. With these interactions held fixed, the gradient reduction is
A pixel-wise scatter computes contributions at each pixel and uses atomic updates to accumulate them into Gaussian gradients. Our gather implementation assigns one thread block to each Gaussian for each particle image. Threads in that block accumulate contributions to the projected Gaussian attributes and reduce them before writing the result, avoiding concurrent writes from different blocks to that Gaussian’s gradient in this reduction. The projection derivatives then propagate these gradients to the three-dimensional Gaussian attributes in Equation (27). The per-splat backward pass of Taming 3DGS [32] performs this reduction under depth-ordered compositing, where it also requires the compositing state of the pixels in front of each splat. In additive projection, Equation (27) involves only and the footprint of the Gaussian.
For attributes shared across particle images, the resulting gradients are summed across those images. Gradients for attributes decoded separately for each particle image are kept separate until they are propagated through the Gaussian attribute decoders to the trainable representation parameters.
Partitioned additive projection
Let be the Gaussian list assigned to image tile for particle image . A long list can be divided into disjoint consecutive sublists whose union is . For each pixel location in that tile, linearity gives
Each inner sum forms a partial projection that can be evaluated independently, and the partial projections are then accumulated. This exposes additional parallelism when a small number of tiles contain long Gaussian lists. Under depth-ordered compositing, 3DGS implementations balance such lists by redistributing work across thread blocks [33, 34]. Linearity makes the partition direct here. Batching particle images also supplies parallel work and amortizes preprocessing and scheduling overhead. We retain batched projection when the batch supplies sufficient parallelism and partition long tile lists when parallelism across particle images is limited.
The partitioning of Gaussian lists for image tiles and the gradient reductions above retain the same Gaussian–pixel interactions. The forward pass allocates workspaces for the projected geometry of the Gaussians, the Gaussian–tile keys and their sorted order, and the per-tile ranges, which 3DGS rasterizers retain for the backward pass because back-to-front compositing requires the sorted lists and the per-pixel transmittance [68] [16]. In additive projection, Equation (27) requires neither, so the rasterizer releases these workspaces once the projections are formed, retaining the Gaussian parameters, the viewing matrices, the projections, and the computation graph to the Gaussian attribute decoders, and the backward pass recomputes the projected geometry of each Gaussian from the retained quantities and gathers gradients over the pixel footprint determined by its projected center and radius. This trades recomputation for memory, as in activation recomputation [73], and the memory held between the two passes scales with the number of Gaussians rather than with the total length of the tile lists. Releasing the workspaces lowers peak allocated memory per GPU by 8.0–10.2% on the datasets of Table 1 (Appendix § 7.§ 3).
Future directions
CryoGaussian could serve as a reconstruction component in agentic workflows that repeatedly select particles, reconstruct structural states, and evaluate their agreement with experimental observations. Scientific agents such as Coscientist demonstrate how language models can coordinate specialized tools and use feedback to plan experiments [74]. In cryo-EM, such a workflow could invoke CryoGaussian to compare alternative particle selections and reconstruct candidate structural subpopulations. Accelerating projection and gradient computation with our CUDA rasterizer and distributed training could reduce the cost of these repeated analyses, allowing more structural hypotheses to be examined at a given computational cost. Where local map quality supports atomic interpretation, reconstructed states could be passed to automated model-building tools such as ModelAngelo [75], connecting heterogeneity analysis to molecular models for downstream study.
A further direction is to connect experimentally resolved structural states with molecular design. ProteinZero improves inverse-folding models through online reinforcement learning with computational feedback [76], while PocketXMol uses atomic prompts to unify structure prediction and the design of pocket-interacting small molecules and peptides [77]. These capabilities motivate workflows in which reconstructed conformations inform structural targets or binding environments, and subsequent experiments test the resulting designs. CryoGaussian could provide the reconstruction component as new particle images become available, enabling comparisons across experimental conditions and structural states. Evaluating this integration would establish whether accelerated reconstruction can make iterative structure-guided design more practical.
References
- [1]Yifan Cheng, Nikolaus Grigorieff, Pawel A. Penczek, and Thomas Walz. A Primer to Single-Particle Cryo-Electron Microscopy. Cell, 161(3):438–449, 2015.DOI
- [2]Ka Man Yip, Niels Fischer, Elham Paknia, Ashwin Chari, and Holger Stark. Atomic-resolution protein structure determination by cryo-EM. Nature, 587(7832):157–161, 2020.DOI
- [3]Wan Zheng, Pengxin Chai, Jiapeng Zhu, and Kai Zhang. High-resolution in situ structures of mammalian respiratory supercomplexes. Nature, 631(8019):232–239, 2024.DOI
- [4]Yuanchen Dong, Shuwen Zhang, Zhaolong Wu, Xuemei Li, Wei Li Wang, Yanan Zhu, Svetla Stoilova-McPhie, Ying Lu, Daniel Finley, and Youdong Mao. Cryo-EM structures and dynamics of substrate-engaged human 26S proteasome. Nature, 565(7737):49–55, 2019.DOI
- [21]Minkyu Jeon, Rishwanth Raghu, Miro Astore, Geoffrey Woollard, Ryan Feathers, Alkin Kaz, Sonya M. Hanson, Pilar Cossio, and Ellen D. Zhong. CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM. In Advances in Neural Information Processing Systems, volume 37, 2024.arxiv.org/abs/2408.05526
- [22]Leland McInnes, John Healy, and James Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv preprint arXiv:1802.03426, 2018.arxiv.org/abs/1802.03426
- [23]Bo Qin, Simon M. Lauer, Annika Balke, Carlos H. Vieira-Vieira, Jörg Bürger, Thorsten Mielke, Matthias Selbach, Patrick Scheerer, Christian M. T. Spahn, and Rainer Nikolay. Cryo-EM captures early ribosome assembly in action. Nature Communications, 14(1):898, 2023.DOI
- [24]Joseph H. Davis, Yong Zi Tan, Bridget Carragher, Clinton S. Potter, Dmitry Lyumkis, and James R. Williamson. Modular Assembly of the Bacterial Large Ribosomal Subunit. Cell, 167(6):1610–1622.e15, 2016.DOI
- [25]Claude E. Shannon. Communication in the Presence of Noise. Proceedings of the IRE, 37(1):10–21, 1949.DOI
- [26]Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447–19456, June 2024.arxiv.org/abs/2311.16493
- [27]Muyuan Chen, Michael F. Schmid, and Wah Chiu. Improving resolution and resolvability of single-particle cryoEM structures using Gaussian mixture models. Nature Methods, 21(1):37–40, 2024.DOI
- [28]David Slepian. Prolate Spheroidal Wave Functions, Fourier Analysis and Uncertainty—IV: Extensions to Many Dimensions; Generalized Prolate Spheroidal Functions. Bell System Technical Journal, 43(6):3009–3057, 1964.DOI
- [29]Frederik J. Simons and Dong V. Wang. Spatiospectral concentration in the Cartesian plane. GEM - International Journal on Geomathematics, 2(1):1–36, 2011.DOI
- [30]Eugene Katsevich, Alexander Katsevich, and Amit Singer. Covariance Matrix Estimation for the Cryo-EM Heterogeneity Problem. SIAM Journal on Imaging Sciences, 8(1):126–185, 2015.pubmed.ncbi.nlm.nih.gov/25699132
- [31]Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D Gaussian Splatting as Markov Chain Monte Carlo. In Advances in Neural Information Processing Systems, volume 37, 2024.arxiv.org/abs/2404.09591
- [32]Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Francisco Vicente Carrasco, Markus Steinberger, and Fernando De La Torre. Taming 3DGS: High-Quality Radiance Fields with Limited Resources. In SIGGRAPH Asia 2024 Conference Papers, 2024.arxiv.org/abs/2406.15643
- [33]Hao Gui, Lin Hu, Rui Chen, Mingxiao Huang, Yuxin Yin, Jin Yang, Yong Wu, Chen Liu, Zhongxu Sun, Xueyang Zhang, and Kun Zhan. Balanced 3DGS: Gaussian-wise Parallelism Rendering with Fine-Grained Tiling. arXiv preprint arXiv:2412.17378, 2024.arxiv.org/abs/2412.17378
- [34]Xinzhe Wang, Ran Yi, and Lizhuang Ma. AdR-Gaussian: Accelerating Gaussian Splatting with Adaptive Radius. In SIGGRAPH Asia 2024 Conference Papers, 2024.DOI
- [35]Clemens Plaschka, Pei-Chun Lin, and Kiyoshi Nagai. Structure of a pre-catalytic spliceosome. Nature, 546(7660):617–621, 2017.DOI
- [36]Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations, 2015.arxiv.org/abs/1412.6980
- [37]Roey Yadgar, Roy R. Lederman, and Yoel Shkolnisky. SOLVAR: Fast covariance-based heterogeneity analysis with pose refinement for cryo-EM. arXiv preprint arXiv:2602.17603, 2026.arXiv
- [38]Gao, W., Wu, Y., & He, X. (2026). CryoKRAQEN: Kernel-Regularized Annealing for Quantized Embedding Networks in Cryo-EM Heterogeneous Reconstruction. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 28298–28307.
- [39]Sjors HW Scheres. RELION: implementation of a Bayesian approach to cryo-EM structure determination. Journal of structural biology, 180(3):519–530, 2012.DOI
- [40]Ali Punjani, John L Rubinstein, David J Fleet, and Marcus A Brubaker. cryoSPARC: algorithms for rapid unsupervised cryo-EM structure determination. Nature methods, 14(3):290–296, 2017.DOI
- [41]Guang Tang, Liwei Peng, Philip R Baldwin, Deepinder S Mann, Wen Jiang, Ian Rees, and Steven J Ludtke. EMAN2: an extensible image processing suite for electron microscopy. Journal of structural biology, 157(1):38–46, 2007.pubmed.ncbi.nlm.nih.gov/16859925
- [42]Huaizhi Qu, Xiao Wang, Yuanyuan Zhang, Sheng Wang, William Stafford Noble, and Tianlong Chen. CryoNeRF: reconstruction of homogeneous and heterogeneous cryo-EM structures using neural radiance field. bioRxiv preprint, 2025.DOI
- [43]Yue Huang, Chengguang Zhu, Xiaokang Yang, and Manhua Liu. High-resolution real-space reconstruction of cryo-EM structures using a neural field network. Nature Machine Intelligence, 6(8):892–903, 2024.DOI
- [44]David Herreros, Carlos Perez Mata, Chari Noddings, Deli Irene, James Krieger, David A. Agard, Ming-Daw Tsai, Carlos Oscar Sanchez Sorzano, and Jose Maria Carazo. Real-space heterogeneous reconstruction, refinement, and disentanglement of CryoEM conformational states with HetSIREN. Nature Communications, 16(1):3751, 2025.DOI
- [45]Sjors H. W. Scheres. A Bayesian View on Cryo-EM Structure Determination. Journal of Molecular Biology, 415(2):406–418, 2012.DOI
- [46]Dari Kimanius, Kiarash Jamali, Max E. Wilkinson, Sofia Lövestam, Vaithish Velazhahan, Takanori Nakane, and Sjors H. W. Scheres. Data-driven regularization lowers the size barrier of cryo-EM structure determination. Nature Methods, 21(7):1216–1221, 2024.DOI
- [47]Yi Zhou, Yilai Li, Jing Yuan, and Quanquan Gu. CryoFM: A Flow-based Foundation Model for Cryo-EM Densities. In The Thirteenth International Conference on Learning Representations, 2025.arxiv.org/abs/2410.08631
- [48]Sjors H. W. Scheres, Haixiao Gao, Mikel Valle, Gabor T. Herman, Paul P. B. Eggermont, Joachim Frank, and Jose-Maria Carazo. Disentangling conformational states of macromolecules in 3D-EM through likelihood optimization. Nature Methods, 4(1):27–29, 2007.pubmed.ncbi.nlm.nih.gov/17179934
- [49]Ali Punjani and David J. Fleet. 3D variability analysis: Resolving continuous flexibility and discrete heterogeneity from single particle cryo-EM. Journal of Structural Biology, 213(2):107702, 2021.DOI
- [50]Axel Levy, Rishwanth Raghu, David Shustin, Adele Peng, Huan Li, Oliver Clarke, Gordon Wetzstein, and Ellen Zhong. Mixture of neural fields for heterogeneous reconstruction in cryo-EM. Advances in Neural Information Processing Systems, 37:569988–57017, 2024.arxiv.org/abs/2412.09420
- [51]Jeffrey Gu, Minkyu Jeon, Ambri Ma, Serena Yeung-Levy, and Ellen D. Zhong. CryoHype: Reconstructing a thousand cryo-EM structures with transformer-based hypernetworks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 35280–35290, 2026.arxiv.org/abs/2512.06332
- [52]Takanori Nakane, Dari Kimanius, Erik Lindahl, and Sjors H. W. Scheres. Characterisation of molecular motions in cryo-EM single-particle data by multi-body refinement in RELION. eLife, 7:e36861, 2018.DOI
- [53]Axel Levy, Rishwanth Raghu, J. Ryan Feathers, Michal Grzadkowski, Frédéric Poitevin, Jake D. Johnston, Francesca Vallese, Oliver Biggs Clarke, Gordon Wetzstein, and Ellen D. Zhong. CryoDRGN-AI: neural ab initio reconstruction of challenging cryo-EM and cryo-ET datasets. Nature Methods, 22(7):1486–1494, 2025.DOI
- [54]Yilai Li, Yi Zhou, Jing Yuan, Fei Ye, and Quanquan Gu. CryoSTAR: leveraging structural priors and constraints for cryo-EM heterogeneous reconstruction. Nature Methods, 21(12):2318–2326, 2024.DOI
- [55]Kawabata, T. (2008). Multiple Subunit Fitting into a Low-Resolution Density Map of a Macromolecular Complex Using a Gaussian Mixture Model. Biophysical Journal, 95(10), 4643–4658.
- [56]Takeshi Kawabata. Gaussian-input Gaussian mixture model for representing density maps and atomic models. Journal of structural biology, 203(1):1–16, 2018.DOI
- [57]Muyuan Chen, Bogdan Toader, and Roy Lederman. Integrating molecular models into CryoEM heterogeneity analysis using scalable high-resolution deep Gaussian mixture models. Journal of Molecular Biology, 435(9):168014, 2023.DOI
- [58]Westover, L. (1990). Footprint Evaluation for Volume Rendering. Proceedings of the 17th Annual Conference on Computer Graphics and Interactive Techniques, 367–376.
- [59]Zwicker, M., Pfister, H., van Baar, J., & Gross, M. (2001). EWA Volume Splatting. Proceedings of IEEE Visualization 2001, 29–36.
- [60]Cai, Y., Liang, Y., Wang, J., Wang, A., Zhang, Y., Yang, X., Zhou, Z., & Yuille, A. (2025). Radiative Gaussian Splatting for Efficient X-Ray Novel View Synthesis. Computer Vision – ECCV 2024, 15059, 283–299.
- [61]Ruyi Zha, Tao Jun Lin, Yuanhao Cai, Jiwen Cao, Yanhao Zhang, and Hongdong Li. R²-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction. In Advances in Neural Information Processing Systems, volume 37, 2024.arxiv.org/abs/2405.20693
- [62]Chinmay Talegaonkar, Yash Belhe, Ravi Ramamoorthi, and Nicholas Antipa. Volumetrically Consistent 3D Gaussian Rasterization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10953–10963, 2025.arxiv.org/abs/2412.03378
- [63]Huaizhi Qu, Xiao Wang, Gengwei Zhang, Jie Peng, and Tianlong Chen. GEM: 3D Gaussian Splatting for Efficient and Accurate Cryo-EM Reconstruction. arXiv preprint arXiv:2509.25075, 2025.arXiv
- [64]Yang, Z., Gao, X., Zhou, W., Jia, S., Zhang, Y., & Jin, X. (2024). Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
- [65]Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian Splatting for Real-Time Dynamic Scene Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.arxiv.org/abs/2310.08528
- [66]Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-Gaussian Embedding-Based Deformation for Deformable 3D Gaussian Splatting. In Computer Vision – ECCV 2024, volume 15073 of Lecture Notes in Computer Science, pages 321–335. Springer, 2025.arxiv.org/abs/2404.03613
- [67]Hexu Zhao, Haoyang Weng, Daohan Lu, Ang Li, Jinyang Li, Aurojit Panda, and Saining Xie. On Scaling Up 3D Gaussian Splatting Training. In The Thirteenth International Conference on Learning Representations, 2025.arxiv.org/abs/2406.18533
- [68]Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An Open-Source Library for Gaussian Splatting. Journal of Machine Learning Research, 26(34):1–17, 2025.arxiv.org/abs/2409.06765
- [69]Miłosz Wieczór, Phu K. Tang, Modesto Orozco, and Pilar Cossio. Omicron mutations increase interdomain interactions and reduce epitope exposure in the SARS-CoV-2 spike. iScience, 26(2):105981, 2023.DOI
- [70]Rob J. Hyndman. Computing and Graphing Highest Density Regions. The American Statistician, 50(2):120–126, 1996.DOI
- [71]NIST/SEMATECH. Critical Values of the Chi-Square Distribution. e-Handbook of Statistical Methods, Section 1.3.6.7.4, n.d. Accessed September 2026.DOI
- [72]Alex Hanson, Allen Tu, Geng Lin, Vasu Singla, Matthias Zwicker, and Tom Goldstein. Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025.arxiv.org/abs/2412.00578
- [73]Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training Deep Nets with Sublinear Memory Cost. arXiv preprint arXiv:1604.06174, 2016.arxiv.org/abs/1604.06174
- [74]Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. Autonomous chemical research with large language models. Nature, 624(7992):570–578, 2023.DOI
- [75]Kiarash Jamali, Lukas Käll, Rui Zhang, Alan Brown, Dari Kimanius, and Sjors H. W. Scheres. Automated model building and protein identification in cryo-EM maps. Nature, 628(8007):450–457, 2024.DOI
- [76]Ziwen Wang, Jiajun Fan, Ruihan Guo, Thao Nguyen, Heng Ji, and Ge Liu. ProteinZero: Self-Improving Protein Generation via Online Reinforcement Learning. Transactions on Machine Learning Research, 2026.arxiv.org/abs/2506.07459
- [77]Xingang Peng, Ruihan Guo, Fenglin Guo, Ziyi Wang, Jiayu Sun, Jiaqi Guan, Yinjun Jia, Yan Xu, Yanwen Huang, Muhan Zhang, Jian Peng, Xinquan Wang, Chuanhui Han, Zihua Wang, and Jianzhu Ma. Unified modeling of 3D molecular generation via atomic interactions with PocketXMol. Cell, 189(7):1904–1922.e28, 2026.DOI
























