Rapid Synthesis of Cryo-ET Data for Training Deep Learning Models

Abstract

Deep learning excels at cryo-tomographic image restoration and segmentation tasks but is hindered by a lack of training data. Here we introduce cryo-TomoSim (CTS), a MATLAB-based software package that builds coarse-grained models of macromolecular complexes embedded in vitreous ice and then simulates transmitted electron tilt series for tomographic reconstruction. We then demonstrate the effectiveness of these simulated datasets in training different deep learning models for use on real cryotomographic reconstructions. Computer-generated ground truth datasets provide the means for training models with voxel-level precision, allowing for unprecedented denoising and precise molecular segmentation of datasets. By modeling phenomena such as a three-dimensional contrast transfer function, probabilistic detection events and radiation-induced damage, the simulated cryo-electron tomograms can cover a large range of imaging content and conditions to optimize training sets. When paired with small amounts of training data from real tomograms, networks become incredibly accurate at segmenting in situ macromolecular assemblies across a wide range of biological contexts.

Simulating the cellular context in synthetic datasets for cryo-electron tomography

“…Although higher order structures such as organelles (mitochondrion, centriole, ...) or the nucleolus would require specific and very sophisticated models, the simulation of low-order structures proposed here can already be used to study many interesting cellular processes such as membrane morphology [23], [24], filament interaction under pathological conditions [22], macromolecule clusters [17], macromolecules in the liquid-like state [19] and membrane-bound protein nanodomains [18], [55]. In addition, local feature simulation has been shown a path to overcome the current limitations of DL approaches for processing (segmentation, macromolecular localization, ...) cryo-tomograms [32], [34]. However, validating tools analyzing experimental cryo-ET data is particularly challenging.…”

“…In [33] some membrane-bound proteins are modelled but just by simple blobs in 2D. A concurrent pre-print [34] generates a more diverse scenario than previous approaches, which includes vesicles and membrane-proteins but not curved filaments. More importantly, this pre-print shows that coarse synthetic data already allows to train usable CNN models.…”

Template Learning: Deep Learning with Domain Randomization for Particle Picking in Cryo-Electron Tomography

“…Recent works 27,40 that involve simulating cryo-ET data for practical applications (e.g., deep learning) utilize an iterative brute-force random placement algorithm, involving the rotation of a duplicate of a molecule in each iteration to find a suitable non-overlapping position within the sample. Nonetheless, random placement leads to unstructured empty spaces between the molecules that prevent achieving highly dense samples.…”

“…In 48 , no distractors were employed in the data simulation to train deep learning models on subtomogram segmentation. Two recent studies 27,40 utilized a small set of objects, serving as analogs to distractors in the simulation of data for training deep neural networks in regression tasks, including signal restoration and segmentation. These objects included gold fiducials, actin bundles, vesicles, and randomly placed small spheres that were relatively sparsely distributed in the volume.…”

“…In a subsequent experiment, we explored if there are benefits of using many distractors compared to a limited number, similar to previous approaches 27,40 . Subsequently, we reintroduced three distractors—small (PDB 1S3X), medium (PDB 5A20), and large (PDB 6UP6) proteins—while maintaining the previously established volumetric density ratio balance between distractors and templates in the simulated tomograms.…”