Detecting anomalies in X-ray diffraction images using convolutional neural networks.

Our understanding of life is based upon the interpretation of macromolecular structures and their dynamics. Almost 90% of currently known macromolecular models originated from electron density maps constructed using X-ray diffraction images. Even though diffraction images are critical for structure determination, due to their vast amounts and noisy, non-intuitive nature, their quality is rarely inspected. In this paper, we use recent advances in machine learning to automatically detect seven types of anomalies in X-ray diffraction images. For this purpose, we utilize a novel X-ray beam center detection algorithm, propose three different image representations, and compare the predictive performance of general-purpose classifiers and deep convolutional neural networks (CNNs). In benchmark tests on a set of 6,311 X-ray diffraction images, the proposed CNN achieved between 87% and 99% accuracy depending on the type of anomaly. Experimental results show that the proposed anomaly detection system can be considered suitable for early detection of sub-optimal data collection conditions and malfunctions at X-ray experimental stations.

A study on the structure, mechanism, and biochemistry of kanamycin B dioxygenase (KanJ)-an enzyme with a broad range of substrates.

Kanamycin A is an aminoglycoside antibiotic isolated from Streptomyces kanamyceticus and used against a wide spectrum of bacteria, including Mycobacterium tuberculosis. Biosynthesis of kanamycin involves an oxidative deamination step catalyzed by kanamycin B dioxygenase (KanJ), thereby the C2' position of kanamycin B is transformed into a keto group upon release of ammonia. Here, we present for the first time, structural models of KanJ with several ligands, which along with the results of ITC binding assays and HPLC activity tests explain substrate specificity of the enzyme. The large size of the binding pocket suggests that KanJ can accept a broad range of substrates, which was confirmed by activity tests. Specificity of the enzyme with respect to its substrate is determined by the hydrogen bond interactions between the methylamino group of the antibiotic and highly conserved Asp134 and Cys150 as well as between hydroxyl groups of the substrate and Asn120 and Gln80. Upon antibiotic binding, the C terminus loop is significantly rearranged and Gln80 and Asn120, which are directly involved in substrate recognition, change their conformations. Based on reaction energy profiles obtained by density functional theory (DFT) simulations, we propose a mechanism of ketone formation involving the reactive FeIV  = O and proceeding either via OH rebound, which yields a hemiaminal intermediate or by abstraction of two hydrogen atoms, which leads to an imine species. At acidic pH, the latter involves a lower barrier than the OH rebound, whereas at basic pH, the barrier leading to an imine vanishes completely. DATABASES: Structural data are available in PDB database under the accession numbers: 6S0R, 6S0T, 6S0U, 6S0W, 6S0V, 6S0S. Diffraction images are available at the Integrated Resource for Reproducibility in Macromolecular Crystallography at http://proteindiffraction.org under DOIs: 10.18430/m36s0t, 10.18430/m36s0u, 10.18430/m36s0r, 10.18430/m36s0s, 10.18430/m36s0v, 10.18430/m36s0w. A data set collection of computational results is available in the Mendeley Data database under DOI: 10.17632/sbyzssjmp3.1 and in the ioChem-BD database under DOI: 10.19061/iochem-bd-4-18.

Fitmunk: improving protein structures by accurate, automatic modeling of side-chain conformations.

Improvements in crystallographic hardware and software have allowed automated structure-solution pipelines to approach a near-`one-click' experience for the initial determination of macromolecular structures. However, in many cases the resulting initial model requires a laborious, iterative process of refinement and validation. A new method has been developed for the automatic modeling of side-chain conformations that takes advantage of rotamer-prediction methods in a crystallographic context. The algorithm, which is based on deterministic dead-end elimination (DEE) theory, uses new dense conformer libraries and a hybrid energy function derived from experimental data and prior information about rotamer frequencies to find the optimal conformation of each side chain. In contrast to existing methods, which incorporate the electron-density term into protein-modeling frameworks, the proposed algorithm is designed to take advantage of the highly discriminatory nature of electron-density maps. This method has been implemented in the program Fitmunk, which uses extensive conformational sampling. This improves the accuracy of the modeling and makes it a versatile tool for crystallographic model building, refinement and validation. Fitmunk was extensively tested on over 115 new structures, as well as a subset of 1100 structures from the PDB. It is demonstrated that the ability of Fitmunk to model more than 95% of side chains accurately is beneficial for improving the quality of crystallographic protein models, especially at medium and low resolutions. Fitmunk can be used for model validation of existing structures and as a tool to assess whether side chains are modeled optimally or could be better fitted into electron density. Fitmunk is available as a web service at http://kniahini.med.virginia.edu/fitmunk/server/ or at http://fitmunk.bitbucket.org/.