Particle picking in single particle cryo-EM is an object detection task. Within each micrograph, protein particles of the target molecule must be detected and localized, while ignoring junk, aggregates, carbon edges, ice crystals, contaminants, etc. The target molecule can appear in a multitude of views and may be barely visible or detectable by the human eye.
The purpose of particle pickers based on deep neural networks is to provide a machine-learning approach to solving the particle picking problem. Machine learning algorithms are essentially optimization methods that fit the parameters of a classification model to training data, so that the classification model can be used on new data that has not been seen.
Deep neural network particle picking can be highly specific in picking, more powerfully separating good particles from junk and other noise. Trained models can also be re-used on datasets that are similar to one-another. However, training a deep particle picking can be tricky and requires high-quality initial picks, generally provided manually or from another source that is trusted. Any bias in the training set will appear in the learned model after training, and may therefore affect results. For example, missing a particular view or conformation in the training set will bias the picker against that view or conformation. This is in contrast to the simpler blob and template pickers where as long as a particular view appears similar to other views of the particle, it will be picked up. For this reason, care should be taken in using the deep picking methods.