Automatic particle selection from electron micrographs using machine learning techniques

The 3D reconstruction of biological specimens using Electron Microscopy is currently capable of achieving subnanometer resolution. Unfortunately, this goal requires gathering tens of thousands of projection images that are frequently selected manually from micrographs. In this paper we introduce a new automatic particle selection that learns from the user which particles are of interest. The training phase is semi-supervised so that the user can correct the algorithm during picking and specifically identify incorrectly picked particles. By treating such errors specially, the algorithm attempts to minimize the number of false positives. We show that our algorithm is able to produce datasets with fewer wrongly selected particles than previously reported methods. Another advantage is that we avoid the need for an initial reference volume from which to generate picking projections by instead learning which particles to pick from the user. This package has been made publicly available in the open-source package Xmipp.     

Automated acquisition of cryo-electron micrographs for single particle reconstruction on an FEI Tecnai electron microscope

AutoEM is a software package developed by Zhang et al. [J. Struct. Biol. 1356, 251] for semi-automated acquisition of cryo-electron micrographs from Tecnai series electron microscopes and is used frequently at the lowest level of automation. We report here on the new progress that we have made based on their preliminary work. A fourth low-dose state is created where the system can pre-select all the good holes in a grid square from a single CCD image taken at low magnification, making the system operative at much higher levels of automation. An additional control interface enables the operator to monitor the status of the program and the quality of the data, interact with the program, and direct the execution process according to intermediate results. When data acquisition is in progress, all useful information is automatically saved in certain text files which are easily accessible by a database. More detailed improvements and general advantages are illustrated and discussed. We have started to use the program to perform routine data collection. A number of applications show that the performance of the program is satisfactory and the quality of the micrographs and their power spectra acquired by the program is comparable to those manually collected under the same conditions.     

Fast automatic particle picking from cryo-electron micrographs using a locally normalized cross-correlation function: a case study

Recent progress in single-particle reconstruction methods and cryo-EM techniques has led to the determination of macromolecular structures with unprecedented resolution. The number of particles that goes into the reconstruction is a key determinant in achieving high resolution. Interactive manual picking of particles from an electron micrograph is a very time-consuming, tedious, and inefficient process. We have implemented a fast automatic particle picking procedure in the SPIDER environment. The procedure makes use of template matching schemes and employs a recently developed locally normalized correlation algorithm based on Fourier techniques. As a test, we have used this procedure to pick 70S Escherichia coli ribosomes from a cryo-electron micrograph. Different search strategies including use of a circular mask and asymmetric masks for different orientations of the particle have been explored, and their relative efficiencies are discussed. The results indicate that the procedure can be optimally used to pick ribosomes in a fully automatic way within the limit of selecting less than 10% false positives while missing about 15% of true positives.     

Reference-free particle selection enhanced with semi-supervised machine learning for cryo-electron microscopy

Reference-based methods have dominated the approaches to the particle selection problem, proving fast, and accurate on even the most challenging micrographs. A reference volume, however, is not always available and compiling a set of reference projections from the micrographs themselves requires significant effort to attain the same level of accuracy. We propose a reference-free method to quickly extract particles from the micrograph. The method is augmented with a new semi-supervised machine-learning algorithm to accurately discriminate particles from contaminants and noise.     

SwarmPS: rapid, semi-automated single particle selection software

Single particle analysis (SPA) coupled with high-resolution electron cryo-microscopy is emerging as a powerful technique for the structure determination of membrane protein complexes and soluble macromolecular assemblies. Current estimates suggest that approximately 10(4)-10(5) particle projections are required to attain a 3A resolution 3D reconstruction (symmetry dependent). Selecting this number of molecular projections differing in size, shape and symmetry is a rate-limiting step for the automation of 3D image reconstruction. Here, we present Swarm(PS), a feature rich GUI based software package to manage large scale, semi-automated particle picking projects. The software provides cross-correlation and edge-detection algorithms. Algorithm-specific parameters are transparently and automatically determined through user interaction with the image, rather than by trial and error. Other features include multiple image handling (approximately 10(2)), local and global particle selection options, interactive image freezing, automatic particle centering, and full manual override to correct false positives and negatives. Swarm(PS) is user friendly, flexible, extensible, fast, and capable of exporting boxed out projection images, or particle coordinates, compatible with downstream image processing suites.     

Model-based particle picking for cryo-electron microscopy

We describe an algorithm for finding particle images in cryo-EM micrographs. The algorithm starts from a crude 3D map of the target particle, computed from a relatively small number of manually picked images, and then projects the map in many different directions to give synthetic 2D templates. The templates are clustered and averaged and then cross-correlated with the micrographs. A probabilistic model of the imaging process then scores cross-correlation peaks to produce the final picks. We give quantitative results on two quite different target particles: keyhole limpet hemocyanin and p97 AAA ATPase. On these particles our automatic particle picker shows human performance level, as measured by the Fourier shell correlations of 3D reconstructions.     

A clarification of the terms used in comparing semi-automated particle selection algorithms in cryo-EM

Many cyro-EM datasets are heterogeneous stemming from molecules undergoing conformational changes. The need to characterize each of the substrates with sufficient resolution entails a large increase in the data flow and motivates the development of more effective automated particle selection algorithms. Concepts and procedures from the machine-learning field are increasingly employed toward this end. However, a review of recent literature has revealed a discrepancy in terminology of the performance scores used to compare particle selection algorithms, and this has subsequently led to ambiguities in the meaning of claimed performance. In an attempt to curtail the perpetuation of this confusion and to disentangle past mistakes, we review the performance of published particle selection efforts with a set of explicitly defined performance scores using the terminology established and accepted within the field of machine learning.     

An approach to automated particle picking from electron micrographs based on reduced representation templates

Reduced representation templates are used in a real-space pattern matching framework to facilitate automatic particle picking from electron micrographs. The procedure consists of five parts. First, reduced templates are constructed either from models or directly from the data. Second, a real-space pattern matching algorithm is applied using the reduced representations as templates. Third, peaks are selected from the resulting score map using peak-shape characteristics. Fourth, the surviving peaks are tested for distance constraints. Fifth, a correlation-based outlier screening is applied. Test applications to a data set of keyhole limpet hemocyanin particles indicate that the method is robust and reliable.     

A novel method for improvement of visualization of power spectra for sorting cryo-electron micrographs and their local areas

In a context of automation of cryo-electron microscopy, we developed a novel method for improving visibility of diffraction rings in the power spectra of cryo-electron micrographs of vitreous ice (without carbon film or high concentration of diffracting material). We used these enhanced spectra to semi-automatically detect and remove micrographs and/or local areas introducing errors in the global 3D map (drifted and charged areas) or those unable to increase global signal-to-noise ratio (non-diffracting areas). Our strategy also allows a detection of micrographs/areas with a strong astigmatism. These images should be removed when using algorithms that do not correct astigmatism. Our sorting method is simple and fast since it uses the normalized cross-correlation between enhanced spectra and their copies rotated by 90 degrees. It owes its success mainly to the novel pre-processing of power spectra. The improved visibility also allows an easier visual check of accuracy of sorting. We show that our algorithm can even improve the visibility of diffraction rings of cryo-electron micrographs of pure water. Moreover, we show that this visibility depends strongly on ice thickness. This algorithm is implemented in the Xmipp (open-source image processing package) and is freely available for implementation in any other software package.     

A strategy for determining the orientations of refractory particles for reconstruction from cryo-electron micrographs with particular reference to round, smooth-surfaced, icosahedral viruses

Cryo-electron microscopy and three-dimensional image reconstruction are powerful tools for analyzing icosahedral virus capsids at resolutions that now extend below 1 nm. However, the validity of such density maps depends critically on correct identification of the viewing geometry of each particle in the data set. In some cases-for example, round capsids with low surface relief-it is difficult to identify orientations by conventional application of the two most widely used approaches-"common lines" and model-based iterative refinement. We describe here a strategy for determining the orientations of such refractory specimens. The key step is to determine reliable orientations for a base set of particles. For each particle, a list of candidate orientations is generated by common lines: correct orientations are then identified by computing a single-particle reconstruction for each candidate and then systematically matching their reprojections with the original images by visual criteria and cross-correlation analysis. This base set yields a first-generation reconstruction that is fed into the model-based procedure. This strategy has led to the structural determination of two viruses that, in our hands, resisted solution by other means.     

FindEM--a fast, efficient program for automatic selection of particles from electron micrographs

The FindEM particle picking program was tested on the publicly available keyhole limpet hemocyanin (KLH) dataset, and the results were submitted for the "bakeoff" contest at the recent particle picking workshop (Zhu et al., 2003b). Two alternative ways of using the program are demonstrated and the results are compared. The first of these approximates exhaustive projection matching with a full set of expected views, which need to be known. This could correspond to the task of extending a known structure to higher resolution, for which many 1000's of additional images are required. The second procedure illustrates use of multivariate statistical analysis (MSA) to filter a preliminary set of candidate particles containing a high proportion of false particles. This set was generated using the FindEM program to search with one template that crudely represents the expected views. Classification of the resultant set of candidate particles then allows the desired classes to be selected while the rest can be ignored. This approach requires no prior information of the structure and is suitable for the initial investigation of an unknown structure--the class averages indicate the symmetry and oligomeric state of the particles. Potential improvements in speed and accuracy are discussed.     

Image reconstruction from cryo-electron micrographs reveals the morphopoietic mechanism in the P2-P4 bacteriophage system.

The satellite bacteriophage P4 does not have genes coding for any major structural proteins, but assembles a capsid from the gene products of bacteriophage P2. The capsid assembled under control of P4 is smaller (45 nm) than the normal P2 capsid (60 nm). The low resolution (4.5 nm) structures of P2 and P4 capsids were determined by cryo-electron microscopy and image processing. The capsid of P2 shows T = 7 symmetry with most of the mass clustered as 12 pentamers and 60 hexamers. The P4 capsid has T = 4 symmetry with a similar distribution of mass to P2, but the hexamer geometry has changed. The major capsid protein has a two-domain structure. The major domains form the capsomers proper, while connecting domains form trivalent contacts between the capsomers. The size determination by P4 appears to function by altering hexamer geometry rather than by affecting the interdomain angle alone.     

A binary segmentation approach for boxing ribosome particles in cryo EM micrographs

Three-dimensional reconstruction of ribosome particles from electron micrographs requires selection of many single-particle images. Roughly 100,000 particles are required to achieve approximately 10 A resolution. Manual selection of particles, by visual observation of the micrographs on a computer screen, is recognized as a bottleneck in automated single-particle reconstruction. This paper describes an efficient approach for automated boxing of ribosome particles in micrographs. Use of a fast, anisotropic non-linear reaction-diffusion method to pre-process micrographs and rank-leveling to enhance the contrast between particles and the background, followed by binary and morphological segmentation constitute the core of this technique. Modifying the shape of the particles to facilitate segmentation of individual particles within clusters and boxing the isolated particles is successfully attempted. Tests on a limited number of micrographs have shown that over 80% success is achieved in automatic particle picking.     

Zernike phase contrast cryo-electron tomography of sodium-driven flagellar hook-basal bodies from Vibrio alginolyticus

Vibrio alginolyticus use flagella to swim. A flagellum consists of a filament, hook and basal body. The basal body is made up of a rod and several ring structures. This study investigates the structure of the T ring which is a unique component of the V. alginolyticus sodium ion-driven flagellar basal body. Using Zernike phase contrast (ZPC) cryo-electron tomography, we compared the 3D structures of purified hook-basal bodies (HBB) from a wild-type strain (KK148) and a deletion mutant lacking MotX and MotY (TH3), which are thought to form the T ring. ZPC images of HBBs had highly improved signal-to-noise ratio compared to conventional phase contrast images. We observed the outline of the HBBs from strains KK148 and TH3, and the TH3 mutant was missing its T ring. In the wild-type strain, the T ring was beneath the LP ring and seemed to form a ring shape with diameter of 32 nm.     

Polyhedral 3D structure of human plasma very low density lipoproteins by individual particle cryo-electron tomography1

Human VLDLs assembled in the liver and secreted into the circulation supply energy to peripheral tissues. VLDL lipolysis yields atherogenic LDLs and VLDL remnants that strongly correlate with CVD. Although the composition of VLDL particles has been well-characterized, their 3D structure is elusive because of their variations in size, heterogeneity in composition, structural flexibility, and mobility in solution. Here, we employed cryo-electron microscopy and individual-particle electron tomography to study the 3D structure of individual VLDL particles (without averaging) at both below and above their lipid phase transition temperatures. The 3D reconstructions of VLDL and VLDL bound to antibodies revealed an unexpected polyhedral shape, in contrast to the generally accepted model of a spherical emulsion-like particle. The smaller curvature of surface lipids compared with HDL may also reduce surface hydrophobicity, resulting in lower binding affinity to the hydrophobic distal end of the N-terminal β-barrel domain of cholesteryl ester transfer protein (CETP) compared with HDL. The directional binding of CETP to HDL and VLDL may explain the function of CETP in transferring TGs and cholesteryl esters between these particles. This first visualization of the 3D structure of VLDL could improve our understanding of the role of VLDL in atherogenesis.     

A Bayesian method for classification of images from electron micrographs

Particle classification is an important component of multivariate statistical analysis methods that has been used extensively to extract information from electron micrographs of single particles. Here we describe a new Bayesian Gibbs sampling algorithm for the classification of such images. This algorithm, which is applied after dimension reduction by correspondence analysis or by principal components analysis, dynamically learns the parameters of the multivariate Gaussian distributions that characterize each class. These distributions describe tilted ellipsoidal clusters that adaptively adjust shape to capture differences in the variances of factors and the correlations of factors within classes. A novel Bayesian procedure to objectively select factors for inclusion in the classification models is a component of this procedure. A comparison of this algorithm with hierarchical ascendant classification of simulated data sets shows improved classification over a broad range of signal-to-noise ratios.     

A composite-likelihood approach for detecting directional selection from DNA sequence data

We present a novel composite-likelihood-ratio test (CLRT) for detecting genes and genomic regions that are subject to recurrent natural selection (either positive or negative). The method uses the likelihood functions of Hartl et al. (1994) for inference in a Wright-Fisher genic selection model and corrects for nonindependence among sites by application of coalescent simulations with recombination. Here, we (1) characterize the distribution of the CLRT statistic (Lambda) as a function of the population recombination rate (R=4Ner); (2) explore the effects of bias in estimation of R on the size (type I error) of the CLRT; (3) explore the robustness of the model to population growth, bottlenecks, and migration; (4) explore the power of the CLRT under varying levels of mutation, selection, and recombination; (5) explore the discriminatory power of the test in distinguishing negative selection from population growth; and (6) evaluate the performance of maximum composite-likelihood estimation (MCLE) of the selection coefficient. We find that the test has excellent power to detect weak negative selection and moderate power to detect positive selection. Moreover, the test is quite robust to bias in the estimate of local recombination rate, but not to certain demographic scenarios such as population growth or a recent bottleneck. Last, we demonstrate that the MCLE of the selection parameter has little bias for weak negative selection and has downward bias for positively selected mutations.     

The 17A structure of the 420 kDa lobster clottable protein by single particle reconstruction from cryoelectron micrographs

Crustaceans form clots by the rapid crosslinking of a hemolymph clottable protein (CP) to form long, branched polymers. Clotting limits hemolymph loss from wounds as well as playing a part in the innate immune response. CP is a 420 kDa homodimer with a large quantity of associated lipid, primarily the carotenoid pigment astaxanthin. The three-dimensional structure of CP from the lobster Panulirus interruptus has been determined to 17 A resolution by single particle reconstruction from electron micrographs of the protein embedded in vitreous ice. The most prominent feature of this structure is a large cavity spanning the length of the molecule, which is the likely lipid binding pocket. The EM structure has been used in a low resolution molecular replacement search with data from orthorhombic CP crystals, and a solution is presented which describes the crystal packing.     

Initial evaluation of a direct detection device detector for single particle cryo-electron microscopy

We report on initial results of using a new direct detection device (DDD) for single particle reconstruction of vitreous ice embedded specimens. Images were acquired on a Tecnai F20 at 200 keV and a nominal magnification of 29,000×. This camera has a significantly improved signal to noise ratio and modulation transfer function (MTF) at 200 keV compared to a standard CCD camera installed on the same microscope. Control of the DDD has been integrated into Leginon, an automated data collection system. Using GroEL as a test specimen, we obtained images of ∼30 K particles with the CCD and the DDD from the same specimen sample using essentially identical imaging conditions. Comparison of the maps reconstructed from the CCD images and the DDD images demonstrates the improved performance of the DDD. We also obtained a 3D reconstruction from ∼70 K GroEL particles acquired using the DDD; the quality of the density map demonstrates the potential of this new recording device for cryoEM data acquisition.     

Smartnotebook: A semi-automated approach to protein sequential NMR resonance assignments

Complete and accurate NMR spectral assignment is a prerequisite for high-throughput automated structure determination of biological macromolecules. However, completely automated assignment procedures generally encounter difficulties for all but the most ideal data sets. Sources of these problems include difficulty in resolving correlations in crowded spectral regions, as well as complications arising from dynamics, such as weak or missing peaks, or atoms exhibiting more than one peak due to exchange phenomena. Smartnotebook is a semi-automated assignment software package designed to combine the best features of the automated and manual approaches. The software finds and displays potential connections between residues, while the spectroscopist makes decisions on which connection is correct, allowing rapid and robust assignment. In addition, smartnotebook helps the user fit chains of connected residues to the primary sequence of the protein by comparing the experimentally determined chemical shifts with expected shifts derived from a chemical shift database, while providing bookkeeping throughout the assignment procedure.