Cryo-EM and cryo-ET heterogeneity analysis that separates subunit rigid-body motion from compositional change into distinct latent spaces.
No providers recorded yet. Browse all providers
A cryo-EM dataset is a pile of noisy 2D projections of many copies of the same complex, and those copies are not identical. Some differ by continuous motion — a domain hinging, a subunit rotating. Others differ by composition — a ligand present or absent, a subunit assembled or missing. Both show up as variance in the particle stack, and methods that learn a single latent space over that variance tend to mix them together, producing a manifold on which motion and occupancy are entangled and hard to interpret.
OPUS-DSD2, from Fudan University with collaborators at Shanghai AI Laboratory, addresses this by giving the two kinds of heterogeneity their own machinery. It augments a 3D convolutional decoder — which reconstructs the density volume from a composition latent code — with a separate multilayer perceptron that models subunit-level rigid-body dynamics. Because rigid motion is expressed through the dynamics module, the composition latent space is freed to capture spatially localized structural variation, and its principal components become interpretable as compositional changes rather than a blend of both effects.
The second contribution is reach into cryo-electron tomography. The predecessor method, OPUS-DSD, was cryo-EM only. OPUS-DSD2 handles cryo-ET data as well, and notably operates directly on noisy template-matching results — the raw output of searching a tomogram for a reference shape — bypassing the subtomogram classification and averaging steps that normally stand between a tomogram and an interpretable structure. That is what makes high-throughput visual proteomics, structural analysis of complexes as they sit inside cells, tractable with this class of method.
The framework is a variational autoencoder over particle images. An encoder maps 2D projections to latent codes; a composition decoder, a convolutional network producing volumes at a default grid of 192³ voxels, reconstructs density; and the dynamics decoder applies learned rigid-body transformations to subunits. A design point carried from the predecessor is the treatment of pose assignment error: rather than assuming consensus refinement produced correct orientations, the method exploits the inevitable errors while limiting their effect on reconstruction quality. Training runs on PyTorch with distributed multi-GPU support, and downstream analysis tools handle latent-space visualization, volume generation along principal components, and particle subset selection for further refinement. Published demonstrations use well-characterized public datasets: the spliceosome (EMPIAR-10180), the SARS-CoV-2 spike (EMPIAR-10492), and the 80S ribosome (EMPIAR-10002), with trained weights and latent encodings distributed for the first two so results can be inspected without rerunning training. The code is GPL-3.0 licensed.
Structural biologists working with conformationally heterogeneous complexes are the users. Typical questions are which discrete compositional states exist in a sample, how a machine moves between them, and which particles belong to which state so that a homogeneous subset can be refined to higher resolution. The cryo-ET path serves in-situ structural biology, where the target is a complex in its native cellular context and the input is a template-matching search over a tomogram rather than a purified particle stack. Like other heterogeneity-reconstruction methods, it is applied per dataset — a property of the problem, since each particle stack defines its own conformational landscape.
OPUS-DSD2 sits alongside methods such as CryoDRGN-AI in the deep-learning approach to cryo-EM heterogeneity, and its specific argument — that dynamics and composition should be modeled by separate mechanisms rather than one shared latent space — targets the interpretability problem that has limited how far these latent spaces can be trusted biologically. Extending the same framework to cryo-ET template-matching output is the more consequential step, because it removes a labour-intensive bottleneck in in-cell structural work. The method remains a preprint awaiting peer review, and the preprint carries no explicit reuse licence.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.