We propose to interpret Cryogenic Electron Microscopy (CryoEM) data as a supervision for learning parameters of CryoEM microscopes. Following this formulation, we present a differentiable version of Transmission Electron Microscopy (TEM) Simulator that provides differentiability of all continuous inputs in a simulation. We demonstrate the learning capability of our simulator with two examples, detector parameter estimation and denoising. With our differentiable simulator, detector parameters can be learned from real data without time-consuming handcrafting. Besides, our simulator enables new way to denoising micrographs. We develop this simulator with the combination of Taichi and PyTorch, exploiting kernel-based and operator-based parallel differentiable programming, which results in good speed, low memory footprint and expressive code. We call our work as Differentiable TEM Detector as there are still challenges to implement a fully differentiable transmission electron microscope simulator that can further differentiate with respect to particle positions. This work presents first steps towards a fully differentiable TEM simulator. Finally, as a subsequence of our work, we abstract out the fuser that connects Taichi and PyTorch as an open-source library, Stannum, facilitating neural rendering and differentiable rendering in a broader context. We publish our code on GitHub.

abstract: Over the last century, X-ray crystallography has been established as the most successful technique for unravelling the structure-function relationship in molecules. For integral membrane proteins, growing well-ordered large crystals is a challenge and hence, there is room for improving current methods of macromolecular crystallography and for exploring complimentary techniques. Since protein function is deeply associated with its structural dynamics, static position of atoms in a macromolecule are insufficient to unlock the mechanism. The availability of X-ray free electron lasers presents an opportunity to study micron-sized crystals that could be triggered (using light, small molecules or physical conditions) to capture macromolecules in action. This method of ‘Time-resolved serial crystallography’ answers key biological questions by capturing snapshots of conformational changes associated with multi-step reactions. This dissertation describes approaches for studying structures of large membrane protein complexes. Both macro and micro-seeding techniques have been implemented for improving crystal quality and obtaining high-resolution structures. Well-diffracting 15-20 micron crystals of active Photosystem II were used to perform time-resolved studies with fixed-target Roadrunner sample delivery system. By employing continuous diffraction obtained up to 2 A, significant progress can be made towards understanding the process of water oxidation. Structure of Photosystem I was solved to 2.3 A by X-ray crystallography and to medium resolution of 4.8 A using Cryogenic electron microscopy. Using complimentary techniques to study macromolecules provides an insight into differences among methods in structural biology. This helps in overcoming limitations of one specific technique and contributes in greater knowledge of the molecule under study. / Dissertation/Thesis / Doctoral Dissertation Biochemistry 2018

Viruses are the most abundant biological entity in the biosphere and are known to infect hosts from all domains of life. The aim of my work is to identify conserved and non-conserved features among the capsid structures of related and divergent icosahedral viruses via cryogenic electron microscopy, sequence analysis, molecular modeling, and other techniques. Bird polyomaviruses often cause severe disease in their hosts whereas mammalian polyomaviruses generally do not. Avian polyomavirus is a type of bird polyomavirus with an unusually broad host range compared to the restricted tropism of other polyomaviruses. Although most polyomaviruses have a conserved, rigid capsid protein structure, avian polyomavirus has a flexible capsid shell and a non-conserved C-terminus in its major capsid protein. A β-hairpin motif appears to stabilize other polyomaviruses but is missing in avian polyomavirus. The lack of this structure in avian polyomavirus may account for its capsid flexibility and broad host range. A minor capsid protein unique to bird polyomaviruses may be located on the inner capsid surface. This protein may have a role in the acute disease caused by bird polyomaviruses. The solution-state capsid structure of satellite tobacco mosaic virus was unexpectedly different than the previously solved crystalline structure. The conformational differences were accounted for by a shift of the capsid protein about the icosahedral fivefold axis. Conversely, the RNA core was consistent between solution and crystalline structures. The stable RNA core supports previous observations that the viral genome stabilizes the flexible capsid. Halophage CW02 infects Salinivibrio bacteria in the Great Salt Lake. The three-dimensional structure of CW02 revealed a conserved HK97-like fold that is found in all tailed, double-stranded DNA viruses. The capsid sequence of CW02 shares less than 20% identity with HK97-like viruses, demonstrating that structure is more conserved than sequence. A conserved module of genes places CW02 in the viral T7 supergroup, members of which are found in diverse aquatic environments. No tail structure was observed in reconstructions of CW02, but turret-like densities were found on each icosahedral vertex, which may represent unique adaptations similar to those seen in other extremophilic viruses.

<p>Proteins are significant components of living cells. They perform a diverse range of biological functions such as cell shape and metabolism. The functions of proteins are determined by their three-dimensional structures. Cryogenic-electron microscopy (cryo-EM) is a technology known for determining the structure of large macromolecular structures including protein complexes. When individual atomic protein structures are available, a critical task in structure modeling is fitting the individual structures into the cryo-EM density map.</p> <p>In my research, I report a new computational method, MarkovFit, which is a machine learning-based method that performs simultaneous rigid fitting of the atomic structures of individual proteins into cryo-EM maps of medium to low resolution to model the three-dimensional structure of protein complexes. MarkovFit uses Markov random field (MRF), which allows probabilistic evaluation of fitted models. MarkovFit starts by searching the conformational space using FFT for potential poses of protein structures, computes scores which quantify the goodness-of-fit between each individual protein and the cryo-EM map, and the interactions between the proteins. Afterwards, proteins and their interactions are represented using a MRF graph. MRF nodes use a belief propagation algorithm to exchange information, and the best conformations are then extracted and refined using two structural refinement methods. </p> <p>The performance of MarkovFit was tested on three datasets; a dataset of simulated cryo-EM maps at resolution 10 Å, a dataset of high-resolution experimentally-determined cryo-EM maps, and a dataset of experimentally-determined cryo-EM maps of medium to low resolution. In addition to that, the performance of MarkovFit was compared to two state-of-the-art methods on their datasets. Lastly, MarkovFit modeled the protein complexes from the individual protein atomic models generated by AlphaFold, an AI-based model developed by DeepMind for predicting the 3D structure of proteins from their amino acid sequences.</p>