Computational Reconstruction in Cryo-Electron Microscopy


In just a few years, Cryo-EM has gone from being the ugly duckling of structural biology to one of the hottest techniques in science. New generations of microscopes and detectors have enabled us to break through previously unimaginable resolution barriers and achieve close to atomic resolution even of complex biological molecules. However, the resulting two-dimensional images are exceptionally noisy since the electron dose must be kept low to avoid damage, and averaging data over many images poses other problems, in particular due to motions in molecules. I will present how we are solving these challenges with new computational approaches based on Bayesian statistics, in particular Regularized Likelihood Optimization, where the key idea is that we no longer directly derive a single model from experimental data, but ask which model (out of many) is most likely to have generated the observation. This turns out to be particularly important for cryo-EM, where the method produces an entire ensemble of sub-structures instead of the single one commonly obtained in X-ray crystallography. By implementing these algorithms on graphics processors, we are now able to run 3D reconstructions on desktops instead of supercomputers. This in turn is making it possible to repeat the reconstructions iteratively, which enables us to handle large biomolecular complexes (for instance ribosomes) as multiple semi-independent bodies, and extract the inherent molecular flexibility and motions from the cryo-electron micrographs. Finally, I will describe some very recent work where we try to combine experimental and theoretical methods to study ensembles of molecular conformations, and how we would like to use these types of reconstruction methods for wider classes of experimental data, including e.g. neutron scattering and light microscopy.