AutoCryoPicker: An Unsupervised Learning Approach for Fully Automated Single Particle Picking in Cryo-EM Images

Abstract

1 Background 2 An important task of macromolecular structure determination by cryo-electron 3 microscopy (cryo-EM) is the identification of single particles in micrographs (particle 4 picking). Currently, particle picking is laborious, time consuming, and potentially biased 5 due to the need of human intervention to initialize the particle picking. The results 6 typically include many false positives and negatives. Adjusting the parameters to 7 eliminate false positives often excludes true particles in certain orientations. The 8 supervised machine learning (e.g. deep learning) methods for particle picking often 9 need a large training dataset, which requires extensive manual annotation. Other 10 reference-dependent methods rely on low-resolution templates for particle detection, 11 matching and picking, and therefore, are not fully automated. These issues motivate 12 us to develop a fully automated, unbiased framework for particle picking. 13Results 14 We design a fully automated, unsupervised approach for single particle picking in cryo-15 EM micrographs. Our approach consists of three stages: image preprocessing, particle 16 clustering, and particle picking. The image preprocessing is based on image 17 averaging, normalization, cryo-EM image contrast enhancement correction (CEC), 18 histogram equalization, restoration, adaptive histogram equalization, guided image 19 filtering, and morphological operations significantly improves the quality of original 20 cryo-EM images. Our particle clustering method is based on an intensity distribution 21 model which is much faster and more accurate than traditional K-means and Fuzzy C-22 Means (FCM) algorithms for single particle clustering. Our particle picking method, 23 based on image cleaning and shape detection with a modified Circular Hough 24 Transform algorithm, effectively detects the shape and the center of each particle and 25 creates a bounding box encapsulating the particles. 26 Conclusions 27AutoCryoPicker can automatically and effectively recognizes particle-like objects from 28 in noisy cryo-EM micrographs without the need of labeled training data and human 29 intervention and therefore is a useful tool for cryo-EM protein structure determination. 30 Keywords 31Clustering, Intensity Based Clustering (IBC), micrograph, Cryo-EM, singe particle 32 pickling, protein structure determination. -3 - Background 34For decades, X-ray crystallography has been the dominant technique for obtaining 35 high-resolution structures of macromolecules. Single-particle cryo-electron 36 microscopy (cryo-EM) was traditionally used to provide low resolution structural 37 information on large protein complexes that resisted crystallization (e.g., highly 38 symmetric particles of viruses). Though the basic workflow of cryo-EM has not 39 changed considerably over the years, recent technological advances in sample 40 preparation, computation and especially instrumentation have revolutionized the field 41 of structural biology [1] [2] [3], allowing it to solve large protein struct...

DeepCryoPicker: fully automated deep neural network for single protein particle picking in cryo-EM

“…In the first step, 80% of the samples from the collected micrographs are used. Numerous particles are composed and picked from micrograph images using the fully automated framework for particle picking based unsupervised learning approaches that we proposed in our previous models [ 24 , 25 ]. Then, each single particle image is automatically isolated and evaluated as a “good” or “bad” training sample.…”

“…This step is implemented to automatically use different scaling operations (up-sampling or down-sampling) and different scaling factors. We rely on the two unsupervised models (AutoCryoPicker [ 24 ] and SuperCryoEMPicker [ 22 ]) to estimate the dimensions of the original particle patches’ coordinates that are detected in the training particles picking and selection stage. First, for each test micrograph, we calculate the average particle patches’ coordinate form each dataset.…”

“… The second column shows the original test micrographs dimensions, the third column the detected particle patches’ coordinates (dimensions) using AutoCryoPicker [ 24 ] and SuperCryoEMPicker [ 25 ], the fourth column the selected scaling operation and best factor, the fifth column the scaled particle size using the selected scaling factor, the sixth column the scaled micrograph dimensions, and the last column the unified micrograph size …”

“…Therefore, the generation of the training set is fully automated, eliminating the need for manual labeling or labor-intensive particle selection. Second, to accommodate the low-SNR images, a general framework of micrographs preprocessing [ 23 , 24 ] is applied to improve the quality of the low-SNR micrographs. In general, the preprocessing steps increase the particle’s intensity, and pre-grouping the pixels inside each particle makes it easier to be isolated.…”

“…Hence, we propose a fully automated deep neural network for single particle picking based on the fully automated training particle data generation using unsupervised learning algorithms. We use two clustering approaches (regular clustering algorithm using the Intensity-Based Clustering IBC) [ 24 ] and super clustering algorithms using the super k-means [ 25 ]) to automatically generate training particle datasets for training the deep neural networks. To accommodate the low-SNR cryo-EM images, a general framework of micrograph preprocessing that has been used in both our last two models [ 24 , 25 ] is applied to improve the quality of the low-SNR micrographs.…”