CryoDRGN-AI

CryoDRGN-AI is a method for ab initio heterogeneous reconstruction in cryo-electron microscopy and cryo-electron tomography, from Ellen Zhong's lab at Princeton University with collaborators at Stanford, SLAC, and Columbia. It was published in Nature Methods in 2025. Every particle image carries an unknown orientation and an unknown conformational state, and conventional pipelines resolve these sequentially, first assigning poses against a consensus reference and only then analyzing heterogeneity. That order fails precisely where the biology is most interesting — samples with strong compositional variation, or datasets dominated by junk particles, where no single reference exists to align against.

CryoDRGN-AI instead infers poses and conformational states jointly, with no starting model. A neural field maps a low-dimensional latent embedding to a 3D density map, and both the per-image poses and the per-image embeddings are optimized directly rather than predicted by an encoder. Pose estimation runs in two stages: a hierarchical grid search over orientations and translations establishes a coarse solution, and stochastic gradient descent then refines it past the resolution limit of the search grid.

This is a reconstruction algorithm rather than a pretrained checkpoint. Each dataset gets its own fitted neural volume, because that fitted volume is the reconstruction, the same relationship a scene-specific neural field has to its scene. That distinguishes it from cryo-EM models that pretrain transferable representations, such as Cryo-IEF and CryoFM: what CryoDRGN-AI ships is a general optimization procedure rather than a transferable checkpoint. The authors did deposit the fitted weights, poses, latent embeddings and configuration for every experiment in the paper on Zenodo under CC BY 4.0, so each published reconstruction can be reproduced or inspected directly.

#Key Features

• Fully ab initio: Poses, conformational states, and the 3D volume are recovered in a single training run from an unfiltered particle stack, with no consensus refinement, initial model, or prior pose assignment.
• Hybrid pose estimation: Hierarchical pose search supplies robustness to noise, then stochastic gradient descent takes over — roughly 20 times faster per image at 128 × 128 resolution and not bounded by the search grid's angular spacing.
• Autodecoder latent variables: Latent embeddings are stored and optimized per image rather than emitted by an encoder, avoiding the memorization failure that amortized inference exhibits on low-signal datasets below −10 dB SNR.
• Robust to junk particles: Outlier and broken particles separate into their own latent-space clusters, which can be removed with an interactive lasso tool and the remainder re-reconstructed.
• Ab initio subtomogram averaging: The same procedure runs on cryo-ET tilt series, relaxing the tilt-geometry constraint during gradient refinement to absorb stage jitter and sample deformation.

#Technical Details

The volume decoder is a coordinate-based neural network with Fourier feature encoding, trained on 4 NVIDIA A100 GPUs with data parallelism. Hierarchical pose search starts from a uniform grid of 4,608 rotations at 15-degree spacing and 49 translations on a 7 × 7 grid, keeps the top 8 rotations, and refines them locally at half the grid resolution; the switch to gradient descent occurs by default after max(2N, 500,000) images. On experimental benchmarks the method performs single-shot reconstruction on deposited particle stacks: the pre-catalytic spliceosome (EMPIAR-10180, 327,490 particles), the assembling bacterial large ribosomal subunit (EMPIAR-10076), and a SARS-CoV-2 spike protein complex under 500 kDa. On the DSL1/SNARE complex (EMPIAR-11846) it recovers a continuous hinging motion from 75,854 particles selected in latent space, where the published workflow needed iterative classification to reach a 6.2 Å consensus map from 49,947 particles. On the V-ATPase complex (EMPIAR-10874) it resolves three rotary states and a previously undescribed state with mEAK-7 bound to the back of rotary state 3; on the ankyrin-1 complex (EMPIAR-11043, over 700,000 particles) it recovers all six known classes plus a rare "supercomplex" state. For cryo-ET, 18,466 subtomograms of the Mycoplasma pneumoniae 70S ribosome (EMPIAR-10499) at 11 tilts each yield three elongation-cycle states in one ab initio run.

#Applications

Structural biologists use CryoDRGN-AI at the front of a processing pipeline, where its tolerance for unfiltered stacks removes several rounds of 2D and 3D classification and the manual, subjective decisions those rounds entail. It is most valuable on flexible or compositionally mixed assemblies — membrane complexes, ribosome assembly intermediates, spliceosomes — and on in situ tomography. It ships inside the cryoDRGN package under GPL-3.0, installs with pip install cryodrgn, and runs as cryodrgn abinit, so no separate model download is required.

#Impact

CryoDRGN-AI extends the cryoDRGN line from heterogeneity analysis given known poses to joint pose and heterogeneity inference, and its concrete result — new states of the V-ATPase and ankyrin-1 complexes found in already-published datasets — argues that existing cryo-EM archives hold structures conventional processing discards. The main interpretive limitation is inherited from the family: the latent space has no physical meaning, so distances between embeddings cannot be read as conformational free energies, and identifying rare states still requires sampling and visual inspection of the learned distribution. Reconstructions also demand substantial compute, with the reported runs using multi-GPU training over many hours per dataset.