Warp employs BoxNet, a deep convolutional neural network, for particle picking and maintains a central database of training data that any user can contribute to. As new examples are added, the BoxNet model is constantly re-trained using all available data. In addition to this, BoxNet can be re-trained locally on any data you would like to pick more specifically. Thus, while the concept of sample-specific templates does not exist with BoxNet, you can maintain multiple pre-trained models for different samples. More information on how to re-train BoxNet can be found here.

BoxNet selection

Click Select BoxNet model… to open a list of available models. All models contained in [WarpInstallationDirectory]/boxnet2models/ are enumerated here. If you just installed Warp recently, the list of official BoxNet models pre-trained only on the public dataset should be up-to-date. If your installation is a bit older, click Browse Public Repository to download the latest versions. After downloading new models, please re-open the model selection window to see the changes. Once a model is selected, click Use (or follow the re-training guide).

There are 2 types of official BoxNet models: One follows the BoxNet2_* naming scheme and can only distinguish between background and particles. Another is called BoxNet2Mask_* and can distinguish between background, particles, and high-contrast artifacts, e. g. ethane drops. Knowing the exact location of such artifacts allows to dismiss particles located too close to them.

BoxNet settings

The Diameter setting controls how close two particles can be together before the particle with lower probability is dismissed from the picks. The Data Type determines whether the sign of the data is flipped before picking (negative stain), or not (cryo). A Minimum Score threshold is applied to BoxNet’s predictions. When using a model specifically pre-trained on the current protein species, the score can be typically left at its default value of 0.95. Picking new species with the generic version can sometimes require lower threshold values to pick up everything.

If you’re using a BoxNet version that can create masks, you can enforce a Minimum Distance between particles borders (as defined by their diameter) and 💩.

The picked particles can be extracted immediately after the picking, which is especially useful for real-time processing. For that, the Box Size must be set – ca. 150 % of the particle diameter is recommended to avoid interpolation artifacts later. The pixel size is fixed to that used for the entire pre-processing pipeline. If you want to extract particles with a different pixel size, you can do so later using one of the Task Dialogs. Cryo data need to be inverted upon extraction. For SPA packages that require particles to be pre-normalized, e. g. RELION, the diameter used for picking will also be used for background normalization.