CryoDECO

Single-particle cryo-EM has a chicken-and-egg problem at its core. To sort a mixed particle stack into its constituent complexes you need 3D references to compare each image against; to build those references you first need the particles sorted. Ab initio methods break the loop by optimizing both at once from a random start, but that blind search collapses into a local minimum when a sample holds many species — or one flexible machine sampling many states — and the run simply fails.

CryoDECO (DEconstructing Extreme COmpositional and COnformational heterogeneity) removes the blindness rather than the coupling. It is an encoder–decoder framework in which the encoder is a Vision Transformer initialized from Cryo-IEF, a cryo-EM foundation model pretrained by contrastive learning on a corpus of unlabeled particle images. Because that encoder already separates particle images by the structure and state they depict, the latent space is semantically organized before a single volume has been computed: classification is disentangled from reconstruction instead of racing it. The decoder, a coordinate-based neural field, then synthesizes density and searches for particle orientations against that organized manifold.

The framework was developed at Westlake University with Zhejiang University, and its codebase adapts CryoDRGN-AI. It supersedes the group's earlier CryoSolver prototype. Importantly, CryoDECO is not a zero-shot tool: the pretrained weights are a warm start and a regularizer, and the full network is fine-tuned on each particle stack it is pointed at.

#Key Features

• Foundation-model warm start: The encoder inherits general cryo-EM image priors, so the reconstruction loop begins inside a valid structural manifold rather than at a random point in it.
• Tunable latent topology: The latent width is a topological knob matched to the sample — wide for discrete mixtures, tight to force continuous motion onto a smooth, interpolatable manifold.
• Joint pose recovery: Orientations come from a coarse-to-fine hierarchical search over SO(3), refined continuously by gradient descent; poses from a prior CryoSPARC refinement can be supplied instead.
• In silico purification: Latent clusters are exported as per-class .star files that re-import into CryoSPARC for independent high-resolution refinement.

#Technical Details

The encoder is a ViT-Small — 12 layers, 384-dimensional embeddings, 12 heads, 14 × 14 patches over 224 × 224 images — trained from scratch under the MoCo v3 contrastive recipe on approximately 134 million particle images, Cryo-IEF's original 65-million-image corpus plus 308 further public EMPIAR datasets. The decoder is a 3-layer residual MLP with 256 hidden units acting as a neural field in the Hartley domain, taking Gaussian random Fourier features of rotated frequency coordinates concatenated with the per-particle latent code; LayerNorm inside the MLP blocks stabilized training. Optimization runs in three stages: a warm-up pass over 10,000 images at fixed random poses, a HEALPix-discretized hierarchical pose search, then 100 epochs of SGD refinement that jointly updates network weights, latent codes and a learnable pose lookup table. On the Tomotwin-100 benchmark of 100 simulated species, CryoDECO reached an Adjusted Rand Index of 0.622 and Adjusted Mutual Information of 0.853, against 0.086/0.275 for CryoDRGN-AI and 0.116/0.374 for CryoDRGN2; top-1 k-NN classification was 94.0% versus 44.0%, and 93.1% versus 86.5% on Ribosembly. Classification accuracy on Tomotwin-100 rose from 41.3% at a latent width of 4 to 94.0% at 128, the evidence behind the recommendation of 128 for compositional mixtures and 4 to 64 for conformational ones.

#Applications

On the EMPIAR-11693 "EM ladder" — apoferritin, β-galactosidase, tobacco mosaic virus and PP7 virus-like particles imaged together — CryoDECO separated all four populations cleanly enough that independent CryoSPARC refinements reached 2.48, 3.00, 2.83 and 3.38 Å, each slightly better than the original study reported. Applied to an unpurified Chaetomium thermophilum cell extract, it pulled out the pyruvate dehydrogenase complex core, fatty acid synthase, the oxoglutarate dehydrogenase E2 core and the pre-60S ribosomal subunit at 3.65–4.02 Å, with no templates and no prior knowledge of what the sample contained. Traversing the latent space of experimental data visualized a rotation of the tri-snRNP head region and the swinging arm of αVβ8 integrin.

#Impact

The methodological claim is the durable one: initialization, not the optimizer, is what limits ab initio heterogeneity analysis, and a general-purpose cryo-EM encoder transfers well enough to fix it. Shifting purification from the bench to the pipeline is what the authors call "panoramic structural biology". The caveats are real. Every dataset requires a full training run — roughly 9 to 39 hours on two A40 GPUs in the reported experiments — and the user must still choose the latent width from expectations about the sample. The work is a preprint on the LangTaoSha server and has not been peer reviewed; the code is GPL-3.0, while the released encoder checkpoint sits in the Cryo-IEF HuggingFace repository under academic-use-only terms. Coverage is limited to single-particle data, with cryo-electron tomography left as future work.