Good micrographs have a homogeneous background level (without any smooth gradient along the micrograph), are not astigmatic, have no drift, and have high-resolution structural information (visible Thon rings in high frequencies) (see Fig. 2). Astigmatic micrographs could, in principle, be processed. However, in practice they are avoided since most programs cannot track correctly the astigmatism angle through all the sequence of iterative alignments.

Micrographs can be semi-automatically screened by estimating their Power Spectrum Density (PSD) and their Contrast Transfer Function (CTF) (8). The PSD is an estimate of the energy distribution of the micrograph over frequency. Astigmatic micrographs have elliptical rings, while non-astigmatic micrographs have circular rings. Drifted micrographs have masked Thon rings (they do not appear to be complete). The CTF is normally described by a number of parameters, among which the most important are the microscope operating voltage and the defocus (9). However, Xmipp provides a full 2D characterization of the CTF, as well as, the background noise (8), which also plays a role in the accurate determination of the defoci and cannot simply be “filtered out.”

In order to correctly estimate the PSD it is important that the digital micrograph does not have empty borders or labels as in Fig. 3.

The information content of the micrograph is an important issue to consider. We need to estimate at which resolution the information content of the micrograph fades into the noise (let us call to this frequency). This frequency is characterized by a strong decay of the envelope of the CTF (Xmipp reports the frequency at which the CTF envelope drops below 1% of its maximum value). 3D Reconstruction algorithms may recover a few Angstroms more of resolution beyond this frequency, but not many, and the final volume resolution is strongly determined by the maximum resolution visible in the micrographs.

Related to the maximum attainable resolution are the sampling rate and the downsampling factor. The sampling rate relates spatial frequencies measured in 1/Angstroms with digital frequencies measured in cycles/sample (the maximum digital frequency is 0.5 cycles/sample (10)). The relationship is w  =  f  T, where wis the digital frequency, fis the spatial frequency in 1/Angstroms, and Tis the sampling rate in Angstroms/pixel. When we downsample the images, we increase the pixel size. Downsampling brings in two important benefits: first, images are obviously smaller with the subsequent gain in disk space and computing time; second, downsampling reduces the noise in the images by cutting out in Fourier space a region where no signal is present.

In principle, we can increase the pixel size by a factor Kas long as (downsampling by a factor larger than this introduces an effect called aliasing in which high-frequency components may be strongly degraded). In practice, it is normally preferred to have a safety region such that (this is so to avoid the aliasing of the signal with the noise). For instance, if , and , then we can downsample by a factor . Note that the downsampling factor may not be an integer number. In the Xmipp protocols, downsampling is performed in Fourier space (and that is why it can accept non-integer factors). Downsampling in Fourier space provides the best accuracy even for integer downsampling factors since the micrograph PSD is multiplied by a rectangular window (11). On the other side, downsampling by simply averaging neighboring pixels (an operation often called binning) introduces strong distortions in the high-frequency components (11).

Considering (being the effective sampling rate after downsampling), ideal micrographs have . Micrographs with (also called oversampled) can be safely downsampled since no special gain in resolution will be obtained by so finely sampled images and the noise present may hinder the quality of the final reconstruction. Micrographs with may suffer from aliasing resulting in reconstruction artifacts (they are said to be undersampled).

Manual picking is sometimes criticized for biasing the dataset towards those shapes that the user better recognize or has in mind as the possible projections of her structure. In order to avoid this bias, automatic or semiautomatic particle picking algorithms are sometimes preferred as an objective way of choosing particles. At the same time, the particle picking process, which in general is a tedious and time-consuming task, is computationally assisted and accelerated. Algorithmically chosen datasets contain a nonnegligible amount of wrongly picked particles (contaminants, true particles on carbon, particle conglomerates, etc.). These datasets have to be carefully scrutinized to remove wrongly picked particles from the onset. This can be done by a manual revision of the automatic picking results, and/or by classifying in 2D the selected particles and eliminating those classes corresponding to wrongly picked particles.

Xmipp allows semiautomatic particle picking (16). The algorithm has been designed to keep a low false-positive rate (FPR), i.e., picking as few wrongly picked particles as possible. In order to keep the FPR as low as possible, the algorithm must be trained by the user in the kind of particles he is interested in. The first step of the training requires the user to manually pick about 100–200 particles. Then, the algorithm learns a set of features describing the particles being picked. In the next step, the algorithm tries to automatically select particles from a micrograph that has not been manually processed. The first attempt will pick a number of true particles, but many other true particles may be left. The user is required to pick those “unseen” particles. He is also required to correct the algorithm by removing those particles that have been wrongly picked automatically. This correction information helps the algorithm to distinguish true particles from other objects looking alike. This process of the algorithm trying to correctly pick particles and the user correcting for errors is repeated on more micrographs until the user is satisfied with the algorithm results. At each step, the algorithm learns from its errors and next time it will try to better distinguish between particles and nonparticles. No automatic particle picking algorithm is absolutely infallible. As a rule of thumb, they can be applied when the imaging conditions are not specially challenging.

When extracting the particle projections, there are a number of actions we can take:

Automatic picking algorithms have a nonnegligible FPR (i.e., they pick locations in the micrograph that do not actually correspond to true particles). Particle screening is a successful way of identifying them.

For each image, we calculate the gray values histogram, square the gray values and compute the radial average of these squared values. The gray histogram and the radial average are stacked into a multivariate vector associated to each projection. Now, we perform a PCA analysis of the whole set of projections and project the multivariate vector onto the PCA space spanned by the first two eigenvectors. Then, we analyze the multivariate normality of these projections. The normality is measured as the Mahalanobis distance of the PCA projection to the space origin. Typical projections have a small distance, while nontypical projections have a larger distance. The dataset is sorted according to their distance (called z-score). Normally, wrongly picked particles, particles on edges, contaminants, etc., have a large distance and are sorted to the end of the list. At this point, the user may discard the particles he considers that do not correspond to the structure under study, or that for some reason do not follow the general trend.