TomoDRGN: Revolutionizing Cryo-ET with Deep Learning for Dynamic Structural Biology
1:15 Introduction to TomoDRGN: TomoDRGN is an extension of cryoDRGN, a deep learning tool initially designed for single-particle cryo-EM, now adapted for cryo-electron tomography (cryo-ET) to analyze structural heterogeneity.
1:45 Motivation: The project aims to understand the assembly and disassembly of macromolecular complexes, processes that are challenging to study due to their inherent heterogeneity and the limitations of traditional structural biology methods.
6:51 Limitations of Discrete Classification: Traditional methods that rely on discrete classification struggle with continuous conformational changes and the combinatorial explosion of states in assembly processes.
10:15 CryoDRGN's Approach: Instead of directly inferring a single 3D structure, cryoDRGN learns a mapping function that generates density maps conditioned on positional and latent (heterogeneity-encoding) variables.
12:27 CryoDRGN Architecture: Utilizes a variational autoencoder framework, with an encoder that maps particle images to a low-dimensional latent space and a decoder that generates density maps.
15:28 Applications of CryoDRGN: Demonstrated on the assembly of the small ribosomal subunit, revealing structural dependencies and the role of the assembly cofactor KsgA.
17:40 MAVEn Tool: Developed for analyzing volume ensembles from cryoDRGN, enabling quantification of structural element occupancies and identification of assembly pathways.
22:27 Voxel-wise Principal Component Analysis: A method to identify dominant modes of motion or conformational changes within specific regions of the density maps.
25:45 Transition to TomoDRGN: Addresses the challenge of analyzing cryo-ET data, where multiple tilted images of the same particle are acquired.
27:33 TomoDRGN Architecture: Employs a dual-encoder network to process tilt series data, mapping multiple tilt images to a single latent representation for each particle.
29:19 Handling Missing Tilts: Uses random tilt sampling during training to accommodate incomplete tilt series data without sacrificing particles or tilt images.
31:55 Model Robustness: Random tilt sampling enhances robustness to overfitting and maintains the ability to discriminate heterogeneity even with a reduced number of tilts.
33:23 Dose and Tilt Angle Weighting: Incorporates dose filtering and weighting of the loss function to account for varying image quality across the tilt series.
36:11 Application to Simulated ATP Synthase: Successfully recovers the rotary motion of ATP synthase from simulated tomographic data.
37:12 Performance on Real Data: Tested on apoferritin and in situ ribosomes, demonstrating accurate reconstruction and the ability to identify distinct structural states, including the presence of holoferritin in an apoferritin sample.
40:46 Analysis of Ribosome Translation States: Identifies different translation states of ribosomes within cells and reveals the presence of membrane-associated ribosomes engaged in co-translational translocation.
43:15 Expedited Cryo-ET Workflow: Demonstrates a rapid workflow combining fast FIB milling, data collection, and TomoDRGN-based particle filtering, achieving sub-6 Å resolution in under two weeks.
46:54 Availability and Resources: TomoDRGN and related tools are available on SBGrid and GitHub, with ongoing workshops and support for users.
52:48 Recommended Starting Point for TomoDRGN: Suggests starting with a sub-30 Å reconstruction for ribosomes to filter out bad particle picks, iteratively improving reconstructions, and analyzing structures in detail when filtering benefits plateau, typically at sub-nanometer resolution.