The Research And Application Of New Algorithm Based On Fourier-Mellin Transformation For Single Particle Cryo-EM

In 2017,Jacques Dubochet,Joachim Frank and Richard Henderson,the three scientists,were awarded the Nobel Prize in chemistry for their achievements in the technique of cryo-electron microscopy(cryo-EM)and especially the pioneering work in methodology research.From that we can seen cryo-EM had been the field at the forefront of fundamental research.To avoid the damage-effect to biological sample by electron-beam radiation,the sample is embedd in ice with amorphous state and the image is recorded with low dose,which makes high resolution microscopic image with high noise and low signal.The process,that extracts the structural information from the cryo-EM image,likes the process that monitors the voice of a creeping ant in a factory with background noise of machine.Thus,the processing of cryo-EM image always fulls of challenges.In general,the processing of cryo-EM image begins with an initial model.The quality of the initial model decides the rate of convergence,the resolution of the final result and even the faithful of the result in most case.In this paper,the scientific question,ad initio model reconstruction in cryo-EM,is studied.Combined the techniques in pattern recognition and artificial intelligence,a new method based on Fourier-Mellin transformation and high performance particle swarm optimal search algorithm is implemented for fast ad initio model reconstruction.With those algorithms,the large-scale parallel software is programmed.The new method and software is applied to different symmetry EM data sets for ad initio model reconstruction and the near-atom resolution structure reconstruction of icosahedral Cypovirus.The experimental result shows that the new method and software can fast build initial model with a few data and provide high resolution structure on the same time.The main work in this paper is listed as follow.1)Research of new algorithm and software development to fast reconstruct faithful initial model.(a)A method based on Fourier-Mellin transformation to compute the rotational angle in plane.Classification 2D in single particles reconstruction is the same problem as image registration in pattern recognition,and it need resolve the rotational and translational parameters between image and template.For those particles in the same plane with rotational relationship,the main task of classification 2D is to find the rotational angle and then align those particles to improve the signal noise ratio.In this paper,a new method based on Fourier-Mellin transformation extract the amplitude information of particle and template and then resample the information with log-polar coordinate.After that,it converts the rotational relationship into translation relationship and the rotational angle is resolved by Fourier shift theory.(b)A EM image matching method based on symmetry phase-only correlation matching filter.In the method of classification based on template matching,the result effected by weaken weight of low-frequency signal due to contrast transform function of EM.According symmetry phase-only correlation matching filter matching method,our new method only use the phase information of low-frequency signal and provide identifiable symbol among those class with small difference,which can calculate the relative displacement between image and template.(c)A sub-pixel shift parameter calculation approach based on Gauss regression.The shift parameters,only resolved with integer in symmetry phase-only correlation matching filter,must be resolved by interpolation while the result from the interpolation based on centrobaric method is bias to integer.In this paper,a new method fit the nine point around the maximum correlation value point with elliptical Gaussian intensity function and resolve the sub-pixel shift parameter by linear regression.(d)Three dimensional refinement method based on Particle Swarm Optimal high performance search algorithm.The determination of orientation-center parameters(?,?,?,?,y)of biomolecule in EM image is a Multi-objective Optimum problem.The five dimension of the search can be reduced into two dimension by Fourier-Mellin transformation and symmetry phase-only correlation matching filter.The two remainder dimension is searched by heuristically search and high performance PSO algorithm in artificial intelligence is used.The search begins from several points in defined area and the parameters are accepted followed Monte Carlo criterion to avoid local optimum and to find the global best.(e)A software AlignFM based on MPICH and openMP to search orientation-center with high performance.Those algorithms mentioned above are programed by Fortran language and the computation is paralleled by MPICH and openMP.The process from ad initio model reconstruction to high resolution structure reconstruction is integrated in the package.2)The application of new algorithm and software package.(a)ad initio model reconstruction from different symmetry EM data sets and high resolution structure reconstruction.The correctness of the software is test on projection image data from high resolution structure of Escherichia coli 70S Ribosome.And then ad initio mode reconstruction is test on EM data sets of icosahedral cypovirus,?-galactosidase,thermoplasma acidophilum 20S proteasome,plasmodium falciparum 80S ribosome with 5000 particles in each data set.(b)Research of the structure of icosahedral virus with near-atom-resolution.A transcribing cypovirus structure about 3.8A resolution is determined by the new method and software.From the 3.8A ranscribing cypovirus structure,the near-atom structures of two transmethylase on the capsid that participate in RNA capping are determined and the space location of their active sites are identified.Three channels in which RNA arrive the active site of two transmethylase in RNA capping are identified.