Structural Biology from Cells to Atoms Optical microscopy D Image Alignment and Classification Yao Cong cryotomography cryomicroscopy 8nm 00 nm 50nm 1nm 5nm Crystallography 0.1nm 0.35nm 0.6nm 0.3nm Shanghai Institute of Biochemistry and Cell Biology, CAS National Center for Protein Sciences, Shanghai 013/7/11 IBP X-ray microscopy 3DEM Structure Deposition Statistics cryotomography cryomicroscopy & X-ray Crystallography NMR Pipeline in Biological Cryo-EM Single Particle Biochemical preparation Cryo-em sample preparation Imaging Data collection http://www.ebi.ac.uk/pdbe/emdb/index.html!"#$%&'())"*)!+%,()-+.'"/0)130)4#%5&$')6/78&$'7%() What is Observed in Single Particle Imaging Image processing Reconstruction Structural analysis Modeling Wah Chiu cryo-em Micrograph of Single Particles D views of mass density of individual proteins Same molecule (Single particle); particles assume random orientations in vitrified ice Joachim Frank electron beam is damaging to biological samples! minimize the dose on the specimen! low signal to noise ratio in individual image
cryo -EM Micrograph of Single Particles Model-Based refinement strategy Classification D alignment 3D reconstruction Class averaging D alignment Low SRN in individual image --> collect many data to accumulate the signal Images Steve Ludtke D Image Alignment Introduction D alignment: determines the 3 in-plane relative transformation parameters of two images (1 rotat. & trans.) D Image Alignment Introduction Application: in model-based single particle analysis Classification Class averaging 3D reconstruction Accuracy: a critical factor to obtain a faithful 3D reconstruction structure Efficiency: a limiting factor for the improvement of resolution Classification of Images Classification strategies are divided into supervised and unsupervised Supervised (Model-based classification): divide according to similarity with reference Unsupervised (Reference-free classification): divide according to intrinsic properties. D image alignment methodology Model-based classification Self-Correlation Function method () Resampling to Polar Coordinate method (RPC) D Fast Rotational Matching method (FRMD) Cross-Common Line method Polar Fourier Transform (PFT) Reference-free classification Multivariate statistical analysis (MSA) Maximium Likelihood method Frank et al., 1978; van Heel and Frank, 1981
Self-correlation function method () : inverse FT of the amplitude of the image of an image is invariant with respect to the translation of this image Resampling to polar coordinate method (RPC) 1D rotation search: FFT accelerated D translational search: exhaustive search T.S. Baker, R.H. Cheng, JSB, 116, (1996) 10 L. Joyeux, P. A. Penczek, Ultramicroscopy 9(00) 33 Joachim Frank FRMD Description FRMD Description Idea: Rotate both objects around their own center of mass, while translate one object along the positive x axis, until find the best matching position. Cross-correlation is a measure of similarity of two images The correlation function is a function of rotations and 1 distance: Here: FRMD: rota.+1 trans. D FFT accelerate rotational parameters search Avoid expensive zero padding associated with linear space FFTs FRMD Description How can we implement the FRMD search? Y Cong, et al. JSB, 15 (005) 104
Efficiency test images (RNA polymerase) What effects the efficiency of D alignment? Angular sampling K=4 SNR=0.1477 (a) reference image (b) raw image 104 x 104 pixels (100º; 6, -4) SNR = K=10 4000 particle images FRMD RPC 11 79.98 447.0 193.39 6 174.89 736.5 01.19 3 54.93 155.7 14.87 1.4 044.64 875.3 38.34 11 104.16 1958.55 193.39 6 91.69 353.71 01.19 3 1063.13 6708.58 14.87 1.4 4183.73 13694.5 38.34 * timings in sec. l = (k + 1) linear scan parameter was set to 4 or 10 pixels Class averaging accuracy test images Class averaging accuracy comparison Pixel Error = d sin!" +!x +!y 1.4º sampling (g) (SNR=1.3716) (h) (SNR=0.3350) (i) (SNR=0.1477) (j) (SNR=0.0847) (k) (SNR=0.0544) More noisy (a-c) Scatter plot of pixel error as a function of SNR for FRMD, RPC and (d) Average pixel error as a function of SNR for the three methods. (l) (SNR=0.0373) (m) (SNR=0.08) (n) (SNR=0.016) (o) (SNR=0.0173) Total 60,000 test particle images. SNR (0.008~153.1). (p) (SNR=0.0138) SNR = exhibits more intrinsic sensitivity to noise than FRMD and RPC. Hemocyanin 3D reconstruction Hemocyanin: O carrier, In scorpion Pandinus imperator (arthropods): 4-mer, 1.75MDa It can be activated to have the tyrosinase phenoxidase (PO) activity, which is involved in hair, skin and eye coloring Emperor Scorpion The accuracy of FRMD and RPC is comparable especially in low SNR region. Experimental data 3D reconstruction test SNR = JEOL 300FSC electron cryo-microscopy 0000 particles boxed out 1.4º sampling, 1.8 Å /pixel FRMD
How to call FRMD in EMAN1? refine 6 mask=18 frmd=90,160 maxshift=10 hard=5 sym=c1 pad1=360 ang=.5 classkeep=0.7 classiter=3 refine proc=# of CPUs The first parameter, 90 specifies the approximate maximum radius of the particle; The second value is the number of angular sampling points at each radius when resampling the D particle image into polar coordinates; maxshift=10 specifies the 1D translation search up to 10 pixels, which actually closely related to the center of mass determination uncertainty during boxing process. maxshift is usually suggested to use when FRMD method is specified, so as to limit the translational search range in FRMD. Yao Cong, Steven J. Ludtke, Methods in Enzymology. (010) Acknowledgements Wriggers @ Scripps, UT Julio Kovacs Chiu @ NCMI Steve Ludtke Wen Jiang Qinfen Zhang David Woolford Henz Decker @ U. Mainz Thorsten Schweikardt Cong @ SIBCB Liangliang Kong Yanxing Wang Yunxiang Zang Zhanyu Ding Xiangyang Liu Jinhuan Chen Zhicheng Cui D image alignment methodology Model-based classification Self-Correlation Function method () Resampling to Polar Coordinate method (RPC) D Fast Rotational Matching method (FRMD) Cross-Common Line method Polar Fourier Transform (PFT) Reference-free classification Multivariate statistical analysis (MSA) Maximium Likelihood method