Besides the developments described above, RELION-3 comprises two major algorithmic advances that have already been described elsewhere.
Firstly, Bayesian particle polishing implements a Gaussian Process regression algorithm for estimating beam-induced motion tracks for individual particles, and an improved B-factor estimation algorithm for resolution-dependent weighting of individual movie frames. Although not strictly necessary, Bayesian polishing is typically performed after motion-correction at the micrograph (patch) level by MotionCor2 (Zheng et al., 2017 (link)). To expose the metadata of this algorithm better to the Bayesian polishing approach, we also implemented our own version of the MotionCor2 algorithm. Unlike the UCSF implementation, our version does not use GPU-acceleration but runs on multi-core CPUs using OpenMP multi-threading. For movies, and using 12 cores, our program runs at comparable speeds to the UCSF implementation. More details about the Bayesian polishing and our implementation of the MotionCor2 algorithm have been described by Zivanov et al, (2018) (link).
Secondly, multi-body refinement implements an automated and iterative approach for focused refinement with partial signal subtraction on multiple independently moving parts of a complex (Nakane et al., 2018 (link)). This approach is useful for improving densities of flexible multi-domain complexes, and yields insights into the type of inter-domain motions that are present in those.
In addition, many smaller programs and functionalities have been added throughout the single-particle processing workflow. We highlight the following. The motion correction, CTF estimation and auto-picking jobs now output PDF files with metadata plots for all micrographs. We have improved our 3D initial model generation by adopting a stochastic gradient descent (SGD) procedure that follows the implementation in cryoSPARC much more closely than our previous version. We achieved considerably speedups in 2D and 3D classification by using subsets of particles in the first few iterations, which was inspired by a similar implementation in cistem (Grant et al., 2018 (link)). We have added a program calledrelion_star_handler that implements a range of useful operations for metadata files in the star format (Hall, 1991 (link)). And we have implemented a program called relion_align_symmetry that aligns the symmetry axes of a 3D map in arbitrary orientation according to the conventions in RELION.
Firstly, Bayesian particle polishing implements a Gaussian Process regression algorithm for estimating beam-induced motion tracks for individual particles, and an improved B-factor estimation algorithm for resolution-dependent weighting of individual movie frames. Although not strictly necessary, Bayesian polishing is typically performed after motion-correction at the micrograph (patch) level by MotionCor2 (Zheng et al., 2017 (link)). To expose the metadata of this algorithm better to the Bayesian polishing approach, we also implemented our own version of the MotionCor2 algorithm. Unlike the UCSF implementation, our version does not use GPU-acceleration but runs on multi-core CPUs using OpenMP multi-threading. For movies, and using 12 cores, our program runs at comparable speeds to the UCSF implementation. More details about the Bayesian polishing and our implementation of the MotionCor2 algorithm have been described by Zivanov et al, (2018) (link).
Secondly, multi-body refinement implements an automated and iterative approach for focused refinement with partial signal subtraction on multiple independently moving parts of a complex (Nakane et al., 2018 (link)). This approach is useful for improving densities of flexible multi-domain complexes, and yields insights into the type of inter-domain motions that are present in those.
In addition, many smaller programs and functionalities have been added throughout the single-particle processing workflow. We highlight the following. The motion correction, CTF estimation and auto-picking jobs now output PDF files with metadata plots for all micrographs. We have improved our 3D initial model generation by adopting a stochastic gradient descent (SGD) procedure that follows the implementation in cryoSPARC much more closely than our previous version. We achieved considerably speedups in 2D and 3D classification by using subsets of particles in the first few iterations, which was inspired by a similar implementation in cistem (Grant et al., 2018 (link)). We have added a program called
Free full text: Click here
