Abstract

Over the past decade, advances in microscope hardware and image data processing algorithms have made cryo-electron microscopy (cryo-EM) a dominant technique for protein structure determination. Near-atomic resolution can now be obtained for many challenging in vitro samples using single-particle analysis (SPA), while sub-tomogram averaging (STA) can obtain sub-nanometer resolution for large protein complexes in a crowded cellular environment. Reaching high resolution requires large amounts of im-age data. Modern transmission electron microscopes (TEMs) automate the acquisition process and can acquire thousands of micrographs or hundreds of tomographic tilt se-ries over several days without intervention. In a first step, the data must be pre-processed: Micrographs acquired as movies are cor-rected for stage and beam-induced motion. For tilt series, additional alignment of all micrographs in 3D is performed using gold- or patch-based fiducials. Parameters of the contrast-transfer function (CTF) are estimated to enable its reversal during SPA refine-ment. Finally, individual protein particles must be located and extracted from the aligned micrographs. Current pre-processing algorithms, especially those for particle picking, are not robust enough to enable fully unsupervised operation. Thus, pre-processing is start-ed after data collection, and takes several days due to the amount of supervision re-quired. Pre-processing the data in parallel to acquisition with more robust algorithms would save time and allow to discover bad samples and microscope settings early on. Warp is a new software for cryo-EM data pre-processing. It implements new algorithms for motion correction, CTF estimation, tomogram reconstruction, as well as deep learn-ing-based approaches to particle picking and image denoising. The algorithms are more accurate and robust, enabling unsupervised operation. Warp integrates all pre-processing steps into a pipeline that is executed on-the-fly during data collection. Inte-grated with SPA tools, the pipeline can produce 2D and 3D classes less than an hour into data collection for favorable samples. Here I describe the implementation of the new algorithms, and evaluate them on various movie and tilt series data sets. I show that un-supervised pre-processing of a tilted influenza hemagglutinin trimer sample with Warp and refinement in cryoSPARC can improve previously published resolution from 3.9 Å to 3.2 Å. Warp’s algorithms operate in a reference-free manner to improve the image resolution at the pre-processing stage when no high-resolution maps are available for the particles yet. Once 3D maps have been refined, they can be used to go back to the raw data and perform reference-based refinement of sample motion and CTF in movies and tilt series. M is a new tool I developed to solve this task in a multi-particle framework. Instead of following the SPA assumption that every particle is single and independent, M models all particles in a field of view as parts of a large, physically connected multi-particle system. This allows M to optimize hyper-parameters of the system, such as sample motion and deformation, or higher-order aberrations in the CTF. Because M models these effects accurately and optimizes all hyper-parameters simultaneously with particle alignments, it can surpass previous reference-based frame and tilt series alignment tools. Here I de-scribe the implementation of M, evaluate it on several data sets, and demonstrate that the new algorithms achieve equally high resolution with movie and tilt series data of the same sample. Most strikingly, the combination of Warp, RELION and M can resolve 70S ribosomes bound to an antibiotic at 3.5 Å inside vitrified Mycoplasma pneumoniae cells, marking a major advance in resolution for in situ imaging.