Loading…
Tilt interpolation for cryo-electron tomography, synthesizing intermediate projections to improve angular sampling without extra electron dose.
No providers recorded yet. Browse all providers
cryoTIGER — Tilt Interpolation Generator for Enhanced Reconstruction — synthesizes tilt images that were never acquired. Cryo-electron tomography reconstructs a three-dimensional volume from a series of two-dimensional projections taken at successive stage tilts, and its resolution is bounded by how finely those tilts sample the angular range. Sampling more finely means more exposures, and biological specimens tolerate only a limited total electron dose before they are destroyed. cryoTIGER trades that hard physical constraint for a computational one: it interpolates intermediate projections between the tilts that were actually recorded, denser angular sampling at no additional dose.
The insight is that a tilt series is, geometrically, a video — consecutive views of the same scene under smooth camera motion. The authors adapted FILM, a frame-interpolation network from Google Research designed for large motion between video frames, to the specific characteristics of cryo-ET data. The work comes from the Max Planck Institute of Biophysics in Frankfurt with collaborators at Goethe University Frankfurt, posted as a preprint in December 2024 and published in Communications Biology in 2025.
Crucially, the paper is careful about what interpolated data may be used for. Synthesized tilts are used to find particles and segment structures more reliably; the actual particles carried into subtomogram averaging are still extracted from experimentally acquired images, so no invented high-resolution detail enters the final structure.
The network is FILM's UNet-style architecture with five encoder–decoder levels and skip connections, strided convolutions for downsampling and bilinear upsampling in the decoder, totalling about 24.5 million trainable parameters and roughly 1 GB of GPU memory at batch size 4. It uses a scale-agnostic bidirectional motion estimator that shares feature extraction weights across scales, which is what lets a single module handle both large and small inter-tilt motion.
Training used triplets of consecutive tilts — the two outer images as input, the middle as ground truth — cut into 256 × 256 patches. The cryo-ET model was trained on 317,312 triplets drawn from 375 tilt series spanning Dictyostelium discoideum, human T cells, HEK Flp-In T-REx and HEK-293 cells, and human skin fibroblasts, at tilt increments of one, two, and three degrees. The alternative checkpoint is FILM trained on Vimeo-90K, 89,800 sequences of seven frames each.
Against ground-truth tilts, both deep-learning models beat linear interpolation on PSNR, RMSE, and SSIM. On template matching over 80S ribosome and nuclear pore ring datasets, they consistently outperformed linear interpolation and the non-interpolated baseline. Applied to nucleosomes — roughly 250 kDa, in crowded nuclear periphery tomograms — interpolation raised detected particle positions from about 18,000 to about 33,000, an 87.33% increase in found peaks with a 14.97% improvement in precision–recall AUC, from 0.4148 to 0.4769; the resulting map showed more pronounced structural detail on the DNA linker arms at a resolution of 8.3 Å against 8.4 Å without interpolation.
The tool is aimed at structural biologists doing in-situ cryo-ET, particularly on small or low-contrast complexes where particle identification, rather than averaging, is the bottleneck. It offers a route to the benefits of denser tilt sampling for datasets that were already collected, and it slots into existing pipelines through a tutorial Jupyter notebook that takes a dose-filtered, IMOD-aligned tilt series in MRC format and returns an interpolated one.
cryoTIGER is the first application of learned frame interpolation to cryo-ET tilt series, and its broader claim — that incomplete angular sampling, not just the missing wedge, is worth attacking computationally — reframes a problem the field had largely treated as fixed by the microscope. Code is released under GPL-3.0 and the preprint under CC BY-NC-ND, a split worth noting for reuse. The pretrained weights are hosted on an institutional ownCloud share rather than a model registry, and the cryo-ET tilt series used for training are not publicly available, so the released checkpoints cannot be retrained from scratch by outside groups.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.