This presentation introduces recent advancements in the CryoDragon software package for cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET) data analysis, specifically addressing conformational and compositional heterogeneity.
The first part focuses on CryoDragon ET, a method for heterogeneous subtomogram averaging designed for in-situ data. It extends the deep generative model approach to multiview inputs (subtilts), enabling robust analysis of lower signal-to-noise ratio data from cellular environments. Key applications demonstrate its ability to filter junk particles, resolve detailed continuous motions, and identify distinct translational elongation states in ribosomes and rotational states in fatty acid synthase from yeast.
The second part details CryoDragon AI, a novel method for ab-initio reconstruction for both single-particle cryo-EM and cryo-ET data. It features a new two-stage pose estimation strategy combining robust hierarchical exhaustive search with fast gradient descent optimization. By replacing the encoder with an auto-decoder framework (lookup table for latents), it improves efficiency and generalization for large, unfiltered datasets. This approach is shown to robustly reconstruct structures from challenging datasets with high junk particle content and to discover rare or previously missed complex states. The method is also adapted for ab-initio subtomogram averaging, demonstrating effectiveness with fewer tilts.
The talk concludes with a discussion of future directions, including the reconstruction of complex mixtures using methods like Hydra, and ongoing benchmarking efforts (CryoBench) to establish standardized evaluation metrics for heterogeneous cryo-EM reconstructions. Updates on the open-source CryoDragon software (versions 3 and 4 beta) and its user interface are also provided.
Key Insights from the SP Grid Webinar: Advances in CryoDragon for Heterogeneity Analysis
0:00 Introduction to SP Grid and Webinar Series: The webinar series hosts software tutorials by developers and lectures by structural biologists. Future webinars are planned, with the next on June 10th featuring Jenkei Terasaki from the Kahara lab.
1:32 Speaker Introduction: Ellen Zong from Princeton University, known for her work on CryoDragon and CryoDragon AI, is introduced as the speaker.
1:56 CryoDragon Overview: CryoDragon is a software package for cryo-EM that uses deep generative networks to reconstruct 3D density maps from unlabeled 2D images. Its core goal is to reconstruct a continuous distribution of structures, moving beyond static or discrete models.
8:20 Neural Network Representation: A key idea is using a neural network to represent 3D volumes as a continuous function mapping 3D coordinates to density, rather than discrete voxels. This allows for elegant modeling of heterogeneity through latent variables.
10:11 Encoder-Decoder Architecture: CryoDragon employs an encoder-decoder architecture. An encoder predicts a low-dimensional "latent variable" for each particle's conformational state, which the decoder uses to output cryo-EM densities. This is trained with unsupervised learning, comparing rendered slices to input images.
11:04 Latent Space Analysis: After training, the generative model allows for visualizing and analyzing the "latent space" of millions of low-dimensional vectors, enabling continuous trajectories, reconstruction of representative samples, and filtering of "junk" particles.
12:43 Open-Source Software: CryoDragon is open-source, available at cryogragon.cs.princeton.edu, with extensive tutorials in its gitbook. It has been used to discover new states (e.g., E. coli ribosomal subunit) and continuous heterogeneity (pre-catalytic spliceosome).
14:19 CryoDragon ET for In-situ Data: An extension for cryo-ET, published in Nature Methods, addresses the challenges of lower signal-to-noise ratio and background noise in cellular environments. It handles multiview inputs (subtilts) corresponding to each subtomogram.
18:41 Reliance on Upstream Alignments: CryoDragon ET currently requires pre-aligned tilt series and particles, and alignments from a consensus reconstruction. Good initial alignments are crucial for accurate results.
19:14 Effective Particle Filtering: The method is highly effective for rapid filtering of junk particles, which is a recommended best practice to optimize representation capacity for conformational states.
20:12 In-situ Ribosome and Fatty Acid Synthase Analysis: CryoDragon ET was applied to Mycoplasma pneumoniae ribosomes, reproducing translation elongation states, and Saccharomyces cerevisiae fatty acid synthase, revealing rotational dynamics.
25:57 CryoDragon AI for Ab-Initio Reconstruction: The latest method, accepted at Nature Methods, aims for single-shot ab-initio reconstruction of large, unfiltered cryo-EM and cryo-ET datasets.
28:09 Joint Pose and Heterogeneity Estimation: CryoDragon AI seeks to jointly estimate unknown poses (viewing angles) and the distribution of heterogeneous structures, addressing limitations of traditional sequential approaches.
31:58 Two-Stage Pose Estimation: A key innovation is a two-stage pose estimation strategy: an initial, robust Hierarchical Pose Search (HPS) followed by faster, resolution-unlimited gradient descent (SGD) based optimization.
36:00 Auto-Decoder Framework: To improve efficiency for large, noisy datasets, CryoDragon AI replaces the encoder with an auto-decoder framework where latent variables are directly optimized via a lookup table for each particle.
38:33 Robustness to Junk Particles: CryoDragon AI demonstrated robust reconstruction of the DSL1 snare complex even when over 50% of the initial particles were junk.
39:34 Discovery of Rare States: It successfully identified a rare super-complex state of the ankryin one complex that was missed by traditional 3D classification methods.
41:09 Ab-Initio Subtomogram Averaging with AI: The two-stage pose estimation is adapted for tilt series data, with gradient descent refining individual subtilts. It was found that resolutions plateau beyond 11-13 tilts, making processing more efficient.
45:00 Future Direction: Complex Mixtures (Hydra): Future work includes extending the approach to ab-initio reconstruction of complex mixtures (e.g., cell lysates) using methods like Hydra, which employs a mixture of "expert" neural models.
47:49 Software Updates: CryoDragon ET is available in stable release (v3.4.4). CryoDragon AI is in a beta branch (v4.1.0) and is being merged into the main development branch, with a new user interface and job setup commands.
50:27 Benchmarking Efforts (CryoBench): The group is actively benchmarking heterogeneous reconstruction methods (CryoBench) using diverse synthetic datasets (conformational, compositional, simple, challenging mixtures) to develop objective evaluation metrics, as current methods often rely on subjective "benchmarking by eye."
55:58 Q&A Highlights: Discussion included total dose for tilts (standard, dose-symmetric, damage accumulates), reasons for latent space differences (optimized per dataset/model, arbitrary layout), and practical limits for CryoDragon (no theoretical limit, but smaller particles are harder due to less signal; hardware VRAM is a practical limit).