Deep learning phase retrieval in X-ray single-particle imaging for biological macromolecules.

Abstract Phase retrieval is an important optimization problem that occurs, for example, in the analysis of coherent diffraction patterns from isolated proteins. All iterative algorithms employed for phase retrieval in this context require some a priori knowledge of the object, usually in the form of a support that describes the extent of the particle. Phase retrieval is a time-consuming task that can often fail, particularly if the support is too loose or of bad quality. This paper presents a neural network that can produce low-resolution estimates of the phased object in a fraction of the time it takes for a full phase retrieval. It can also successfully be used as support for further analysis. Our network is trained on simulated data from biological macromolecules and is thus tailored to the type of data seen in a typical CDI experiment. Other approaches to support finding require very accurate data without missing regions or the full phase-retrieval algorithm to be run for a long time. Our network could speed up offline analysis and provide real-time feedback during data collection.