CryoLens
CryoLens is a pre-trained generative model for particle analysis in cryo-electron tomography. Given a single particle, it reconstructs the structure, embeds it, and estimates its pose.
It is a variational autoencoder with a segmented Gaussian splat decoder. Rather than emitting a voxel grid, the decoder describes a particle as a set of 3D Gaussians and renders them into density, which is compared against the real subvolume. Rendering is differentiable, so the model trains end to end — and because density is represented continuously rather than per-voxel, reconstructions stay smooth at the resolutions where tomogram data is noisiest.
It is trained on 5.8 million synthetic particles spanning 103 protein structures, so reconstruction is zero-shot: no per-dataset training is needed to get a structural estimate back, though fine-tuning on your own data helps. It is fast enough to give real-time feedback while picking particles.
Install with pip install cryolens. The model is at czi-ai/cryolens (MIT), with a napari plugin for interactive picking, a REST server for pipelines, and a loader for copick projects. It is also published on the CZI Virtual Cell Platform, with an interactive demo.
Nov 5, 2025
