Baited Classification with 2D Template Matching
Edward Petrossian, Sarah Loerch · Microscopy and Microanalysis · 2024
Cryo-electron microscopy (cryo-EM) has revolutionized our ability to visualize cellular structures at near-atomic resolution. However, the dense packing of biomacromolecules within cells poses challenges for accurate molecule identification. Cryogenic electron tomography (cryo-ET) is the method of choice to study cellular architecture at high resolution [1]. However, the resolution of tomograms reconstructed from tilt series is often limited by radiation damage, restricting the distinguishability of small or structurally similar particles. Subtomogram averaging has emerged as a valuable tool for improving resolution by averaging signals from thousands of particle images, which has achieved resolutions of ∼3 Å [2]. Nevertheless, this method is throughput limited and distinguishing compositionally and conformationally distinct states of a molecule, especially rare ones, remains a challenge. Here, we present a novel approach utilizing 2D template matching (2DTM) [3] to improve the detection and classification of molecules in cryo-EM images. 2DTM is an emerging method for the detection of molecules in cryogenic electron microscopy (cryo-EM) images of cells that leverages the preservation of high-resolution molecular features in single-tilt images. We introduce the concept of "baited classification" for the classification of conformationally different states. Biological molecules often exist in compositionally and conformationally distinct states that report on the functional status of the molecule. Distinguishing such similar structures or features often requires the preservation of high-resolution features. To simplify, one can consider compositionally distinct structures as discrete states – a binding partner may either be present or not. In contrast, conformational changes are on a continuous scale. 2DTM has already been used to classify ribosomal 60S maturation intermediates, which are, in simplified terms, compositionally discrete states [4]. As ribosomes mature, different proteins join or leave the maturing 60S ribosome and it undergoes structural remodeling. In this simplified model, the ensemble can be represented by a series of template structures. In contrast, conformational changes are on a continuous scale. It is conceivably difficult to represent the entire ensemble of conformational states present in a cell with template structures. Here, we propose a "baited classification" approach. We initially identify molecules based on a common core structure by calculating pixel-wise cross correlations between 2D projections of the template and the high-resolution target image, which yields signal-to-noise ratios (SNR) that are subjected to a significance test. We then compare each target to a set of templates that defines the conformational space of the molecule. To dissect these conformations and identify interactions with other factors, we use an in situ classification approach that we initially developed to distinguish different ribosome maturation states [4]. We hypothesize that targets of similar conformations will cluster together based on their similarity or dissimilarity to each of the templates (Fig. 1). Thus, for identification, a cluster need not be represented by a template. We are merely using different templates to define the multidimensional conformational space of ribosomes. Each cluster will locate to a unique coordinate that is defined by its relative (dis)similarity to each of the templates. Through multidimensional Gaussian analysis, we then determine the probability of each target for belonging to a specific cluster. Finally, we extract targets for each cluster individually and reconstruct their 3D structures. We term this approach “baited classification”. Novel classes can be identified with reduced template bias, as the target is not represented in the set of templates first place. Here we use this approach to identify biologically sensible classes of ribosomes in semi-purified specimen. Ribosomal subunits rotate with respect to each other during translation, and the head region of the mammalian small ribosomal subunit (SSU) swivels with respect to the body region of the SSU. Additionally, auxiliary factors, such as elongation factors and tRNA, bind at specific stages throughout the elongation cycle. We show that we can use baited classification to identify classes that differ conformationally and compositionally from the series of templates we used. We cross-validate the presence of this class using single-particle analysis. We then apply this approach to images from cells. In situ classification strategy. We compare each identified location with multiple different templates. Through pairwise comparison of the SNRs obtained for each template, we obtain a multidimensional distribution (only 2 dimensions shown for clarity). Similar targets have similar distances to each of the templates, reflecting their relative similarities. We then extract each cluster for 3D reconstruction. Notably, a cluster does not need to be represented by a template (e.g. cluster 4). It is separated by its unique similarities and differences to each template. (Created with Biorender)