New algorithm improves 3D imaging of complex biomolecules

MIT Technology Review : Determining the three-dimensional structure of large proteins and other biomolecules has proven difficult. X-ray crystallography works well for molecules that form crystals, but many proteins do not. Now, Marcus Brubaker of the University of Toronto and his colleagues have developed an algorithm that improves a different, previously less effective, imaging technique called electron cryomicroscopy. The technique involves freezing molecules in a thin film and imaging them with transmission electron microscopy. A 3D composite of multiple 2D images of the same molecule can then be constructed. However, the process is time-consuming due to the amount of noise in the images and uncertainty concerning the molecules’ orientations when they were imaged. Thanks to two algorithmic innovations, Brubaker’s team was able to reduce the imaging time from 2 weeks to just 24 hours. The key improvements were using a machine learning process to sift through the noise in the images to glean the useful information and incorporating importance sampling. The latter relies on the fact that molecules in thin films are usually positioned on their sides, so the algorithm can skip evaluating potential head-on structural orientations.