Table 2:

Recent notable technical developments for specific cryo-ET pattern-mining tasks.

Software tool	Task	Notes
SuRVoS [44]	Segmentation	Allows users to varythe degree of manual segmentation required.
DSM-Net [46]	Segmentation Classification (supervised) Structure determination	Multi-task deep learning model to train three networks simultaneously on (1) semantic segmentation, (2) classification, and (3) coarse structure determination. Outperformed single-task models and was able to identify structures absent in training data.
VP-Detector [47]	Segmentation Particle picking Classification (supervised)	Uses a CNN to take user-annotated data and automatically segment, localize, and classify particles of interest.
DeepFinder [48]	Particle picking	Efficient multiclass particle picking of structures with various sizes and shapes.
3D-UCaps [49]	Particle picking	Efficient multiclass particle picking was given limited training data.
PySeg [36]	Particle picking Classification (unsupervised)	Particle picking of small membrane-bound complexes.
DSRF3D-v2, RB3D, CB3D [50]	Classification (supervised)	Classifies complexes spanning a range of molecular weights.
DISCA [51]	Classification (unsupervised)	High-throughput, fully unsupervised approach to cluster particles using deep learning.
MPP [52]7	Classification (unsupervised) Structure determination	Iterative alignment to sort a heterogenous set of particles into structure patterns, capable of identifying patterns de novo.
Jim-Net [53]	Classification (unsupervised) Structure determination	Uses pair-matching alignment and hierarchical clustering to achieve comparable accuracy with supervised method to determine structures de novo.
RELION-4.0 Bayesian single-particle tomography structure determination [43]	Structure determination	Structure determination via optimization of a regularized likelihood function. In contrast to previous RELION renditions, features a new weighting system and refinement methods.
FAML [54]	Structure determination	Combines features of FRM and maximum likelihood structure determination methods to achieve higher robustness to noise and artifacts compared to FRM methods and requires fewer subtomograms.
Harmony [126]	Structure determination	Isolates shape information of a particle, allowing for further clustering or alignment.
HEMNMA-3D [55]	Structure determination	Creates a conformational space of a complex that maps how a structure may transition between different configurations in situ.
- [56]	Data augmentation	High density packing by modeling structures with multiple spheres.
- [57]	Data augmentation	Computationally efficient packing by modeling structures as single spheres with gradient descent algorithm.
CryoETGAN [58]	Data augmentation	Deep learning model designed to train on tomograms to generate imitations that can be used to increased training data without the need to simulate artificial tomograms, which may not fully reflect experimental data.
3D-ADA [59]	Domain adaptation	Utilizes deep learning to map particles from images with different parameters to an intermediate latent space.
Cryo-Shift [60]	Domain adaptation	Transforms training data to improve the generalization ability of classifiers to predict on different tomography datasets.