Particle-Picking in Cryo-Electron Microscopy Images with Online Resources

Aims: Use deep learning online resources to identify and pick single particle views in micrographs to enable high quality three-dimensional reconstruction for macromolecular structure determination.

Study Design: Using the keyhole limpet hemocyanin dataset, a public cryo electron microscopy (cryo-EM) dataset containing two dimensional projections of the particles in two views (top and side) in several orientations, and a recent deep learning algorithm made available in a GitHub repository, we design the procedure to pick both views with high degree of confidence, using only online resources and running in a standard laptop.

Methodology: The defocus images are subject to a pre-processing stage to increase its contrast and ameliorate its radiometric range. This is followed by a training stage, that needs a few images annotated with examples of both views of interest – top and side view – identified using user-friendly tools available online. The annotated subset of images is divided for train and validation purposes, and the algorithm runs to produce a set of weights, that can be used for inference in any other similar image, locating all the instances of the particles in both views in seconds.

Results: From the 57 images used to evaluate the performance of the algorithm, 63% had both precision and recall better than 90%. The global precision on the test dataset was 91.6% and recall 98.1%. Considering each view separately, top views detections attain a precision of 91.7%, with 100% for recall; side views recall remains on 96.7%, while precision attain 91.5%.

Conclusion: Deep learning methods are a promising tool for extracting large amounts of single particle views as needed for quality 3D structure reconstruction, as they can be implemented with minimal computer skills and trained to achieve state-of-the-art automatic discrimination, as described in this study.