Loading…
Cryo-EM heterogeneous reconstruction in real space, decoding density maps from a conformational latent space disentangled from pose and CTF.
No providers recorded yet. Browse all providers
Two cryo-EM particle images of the same protein conformation, recorded from different directions and at different defocus, do not resemble each other. An encoder trained to place particles in a conformational latent space encodes that difference too, so the landscape mixes real structural change with viewing geometry and contrast transfer function (CTF) effects. A cluster in it may mark an orientation rather than a state of the molecule.
HetSIREN attacks that entanglement during training. After the first forward pass, the poses within a batch are shuffled and sent back through the decoder, producing a second set of projections of the same conformations at different orientations; a loss term pulls the two sets of latent vectors together, so the encoder learns to return the same point regardless of orientation. A third pass shuffles the CTFs for the same purpose. What survives is structure. The other half is a decoder working entirely in real space, built from meta-sinusoidal layers in which a dense ReLU hypernetwork computes the weights a sine-activated layer then uses — an adaptation of SIREN implicit representations that holds up across a dataset of many structures rather than one signal.
The method comes from the Xmipp/Scipion image-processing group at CNB-CSIC in Madrid, with datasets from collaborators at UCSF, Academia Sinica, the Chan Zuckerberg Imaging Institute and Altos Labs, and ships as Scipion Flexibility Hub protocols rather than a standalone program. Unlike CryoDRGN-AI it is not ab initio: per-particle CTF parameters and angular assignments are required inputs, which HetSIREN then refines. With CryoNeRF and OPUS-DSD2, it is one of a family of neural methods fitted to a single particle stack.
The encoder is a residual convolutional network by default, or a multilayer perceptron with three hidden layers of 1,024 neurons; latent dimensionality is user-set and defaults to 10. The decoder stacks three hidden meta-sinusoidal residual layers, with neuron and hyperneuron counts tied to the latent dimension to keep memory small, followed by a physics-based decoder that applies the image formation model. Multiresolution training minimizes error across filtered and downsampled image pairs, reaching a high-resolution solution in one training step. On EMPIAR-10028, the Plasmodium falciparum 80S ribosome, a 10D landscape learned from roughly 50,000 curated particles isolated a state lacking the 40S subunit that accounts for about 600 particles, some 1.3% of the data; DeepRes local resolution of the decoded maps improved on the cryoSPARC refinement of the full stack by 2.7 Å on average. For the SARS-CoV-2 Spike beta variant, separate 8D landscapes at 4 °C and 37 °C showed 1-Up and 2-Up receptor-binding-domain states in the cold sample against a predominantly 3-Down landscape at 37 °C, where the minority 1-Up state opened less far (21.6° ± 0.9° versus 24.9° ± 0.7°).
The users are structural biologists holding a curated particle stack with angular and CTF assignments, who want to know what states it hides. Two Scipion protocols cover the workflow: one trains the network and infers its latent space, the other applies a trained network to new particles. Because the decoder produces a high-resolution volume from any single point in the landscape, sparsely populated states that discrete classification would average away stay interpretable — and an equilibrium that shifts with storage temperature, as the Spike analysis found, is exactly what matters when formulating a recombinant vaccine antigen.
HetSIREN contributes three separable ideas: SIREN-based hypernetworks for map decoding, latent-space disentanglement of nuisance parameters, and explicit real-space regularizers. The work was peer reviewed and published in Nature Communications, and the code is GPL-3.0 across the Flexutils-Toolkit package and the scipion-em-flexutils plugin, both archived on Zenodo. Its constraints are inherent to the approach: a fresh network is trained on each particle stack, so there are no pretrained weights to download and latent dimensionality is a per-dataset choice, and the need for prior pose and CTF estimates places it after a conventional refinement rather than in place of one.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.