This video presents an in-depth webinar on the latest advancements in the CryoDragon software package, focusing on its applications for heterogeneous cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET) data. The speaker, Ellen Zong from Princeton University, details CryoDragon ET for heterogeneous subtomogram averaging in situ, addressing challenges like low signal-to-noise ratio in cellular environments. She then introduces CryoDragon AI, a novel ab initio reconstruction method for both single particle and cryo-ET data. CryoDragon AI leverages a two-stage pose estimation strategy (hierarchical search followed by gradient descent) and an auto-decoder framework to jointly estimate particle poses and reconstruct continuous distributions of structures, proving robust for large, noisy, and unfiltered datasets. The presentation highlights its ability to discover rare conformational states and effectively filter junk particles. Future directions include tackling complex mixtures and ongoing benchmarking efforts for heterogeneity analysis.
Exploring CryoDragon: Advanced Methods for Heterogeneity in Cryo-EM and Cryo-ET
0:00 Welcome & Upcoming Webinars: Introduction to the SP Grid YouTube channel and upcoming webinars, including Jenkei Terasaki on June 10th. Call for webinar ideas for the next academic season.
1:32 Speaker Introduction: Ellen Zong from Princeton University, known for her work on CryoDragon and Dragon AI, is introduced as the speaker.
1:53 CryoDragon Overview: Ellen expresses excitement to share updates on CryoDragon, specifically covering CryoDragon ET for heterogeneous subtomogram averaging and CryoDragon AI for ab initio reconstruction in single particle and cryo-ET data.
2:24 Team Acknowledgement: The speaker acknowledges key collaborators and team members instrumental in developing these methods, including Axel, Rich, Rama, Ryan, and Michael (research software engineer).
3:25 Biological Motivation: Heterogeneity: Proteins are dynamic macromolecular machines, and cryo-EM aims to study their structures, dynamics, and conformational states in situ.
3:59 The Heterogeneity Problem in Cryo-EM: Each image can contain unique molecules in different conformational states. Traditional methods often discard heterogeneous data or use discrete 3D classification, which requires pre-specification and struggles with continuous heterogeneity.
5:30 Workflow Limitations: The typical cryo-EM workflow is user-guided and ad-hoc, often resulting in significant data loss (e.g., 90% in an example).
6:16 Research Question: Modern ML for Heterogeneous Cryo-EM: The core research question is how to design modern machine learning methods for heterogeneous cryo-EM reconstruction.
6:40 CryoDragon Introduction: CryoDragon (Deep Reconstructing Generative Networks) reconstructs 3D density maps from unlabeled 2D images, focusing on continuous distributions of structures using deep generative models and neural fields.
8:22 Neural Network for 3D Volumes: CryoDragon uses a neural network to represent 3D volumes as a continuous function, mapping Cartesian coordinates to density. This differs from traditional voxel arrays and allows for elegant modeling of heterogeneity by extending the neural network input with a "latent variable vector."
10:11 Encoder-Decoder Architecture: CryoDragon uses an encoder-decoder architecture. The encoder predicts a low-dimensional latent variable describing a particle's conformational state, and the decoder outputs densities based on this latent vector and spatial frequencies. It's trained with unsupervised learning via the 4D slice theorem.
11:04 Analyzing Generative Models: After training, the neural network evaluates latent embeddings for each image, transforming millions of noisy images into millions of low-dimensional, visualizable vectors. This "latent space" can reveal clusters, continuous trajectories, and enable filtering of junk particles.
12:36 CryoDragon Software & Applications: CryoDragon is open-source (cryodragon.cs.princeton.edu) with extensive tutorials. It has been used to discover new compositional and continuous heterogeneity, like an E.coli ribosomal subunit assembly state and continuous motion in a spliceosome dataset.
14:13 CryoDragon ET for In Situ Data: CryoDragon ET (published in Nature Methods) extends CryoDragon to cryo-ET data, specifically for heterogeneity analysis among subtomograms, addressing lower SNR and background noise in cellular environments.
15:18 Cryo-ET Workflow Overview: Briefly outlines the stages: vitrified sample, FIB milling, tomography (tilt series), 3D tomogram reconstruction, subtomogram averaging, and high-resolution reconstruction (in situ).
17:09 CryoDragon ET Architecture & Improvements: It's an encoder-decoder architecture adapted for multiview (subtilts) inputs. It requires pre-aligned tilt series and refined particles. Software improvements were made for large-scale datasets, and it inherits CryoDragon's analysis tools for filtering and trajectories.
19:00 Filtering Junk Particles: Emphasizes the importance of removing junk particles, as they waste the neural network's representation capacity.
20:11 CryoDragon ET Applications: Applied to Mycoplasma pneumoniae ribosome and Saccharomyces cerevisiae (yeast) data (260 tomograms, 119,000 ribosomes). It successfully resolved rotated/unrotated ribosomal states and mapped out translation elongation states. Orthogonal validation via back-projection is crucial.
23:06 Fatty Acid Synthase Example: Demonstrated effective filtering of non-complex particles (from 35,000 to 5,000 bonafide FAS particles) and revealed rotational heterogeneity.
24:09 CryoDragon ET Future Directions: Relating conformational and spatial heterogeneity, integration with upstream/downstream processing, and ab initio subtomogram averaging of intrinsically mixed samples.
24:59 Particle Picking Bottleneck: Highlights particle picking as a main bottleneck, especially in in situ data, leading to multiple particle types or junk. Calls for ab initio methods to jointly separate and reconstruct multiple particle types.
25:57 CryoDragon AI: Ab Initio Reconstruction: Introduces CryoDragon AI, accepted at Nature Methods, spearheaded by Axel, Rich, Ryan, and Michael. Its goal is single-shot ab initio reconstruction for large, unfiltered cryo-EM and cryo-ET datasets without multiple filtering stages.
26:17 Ab Initio Reconstruction Task: In cryo-EM, the task is to find optimal poses and volumes from noisy 2D images, usually splitting pose estimation from conformational estimation. CryoDragon AI aims to jointly estimate poses and structure distribution.
28:29 New Pose Search Strategy: CryoDragon AI employs a new pose search strategy that scales to large datasets, redesigns architecture to avoid memorization/generalization issues, and discovers new states (rare or challenging).
29:05 Pose Estimation Paradigms: Discusses exhaustive search (grid-based, behind most tools), differentiable gradient descent (fast, needs initialization), and amortized inference (neural models to predict pose). CryoDragon AI uses the first two.
29:55 Optimal Pose & Projection Matching: Explains projection matching: comparing rendered projections from candidate poses against a query image to find the lowest error. This is computationally intensive due to the vast number of possible poses.
31:58 Speeding Up Pose Estimation: Two key tricks: hierarchical search (coarse-to-fine subdivision of pose grid) and frequency marching (gradually increasing frequency cutoff for image evaluation, aligning on low-res first).
33:28 Translations: Considers translational search as well, which is faster than orientation search.
33:59 Hierarchical Pose Search (HPS): Detailed explanation of HPS, taking top K poses and subdividing.
34:36 Two-Stage Pose Estimation in CryoDragon AI: Combines robust but slow HPS for initial good poses, then switches to much faster gradient descent (SGD) for refinement, leveraging differentiability of the loss function w.r.t. pose variables.
35:26 Hyperparameters: Mentions various hyperparameters (lart/ln, iterations, kept poses, translational extents, base resolution grid) that users can tune, hoping defaults are reasonable.
36:03 CryoDragon AI Latent Estimation (Auto-Decoder): Replaces the encoder with an auto-decoder framework (lookup table) where each particle is directly assigned a latent, which is then optimized with gradient descent. This addresses efficiency and capacity issues of encoders on very large/noisy datasets.
37:03 Overall Architecture of CryoDragon AI: A single neural network (decoder/neural field) takes a conformational latent and a pose to render an image. Loss is used to update neural network weights, pose variables, or conformational latents.
37:43 Results: Two-Stage Pose Estimation: Demonstrates HPS (orange) and SGD (blue) on a synthetic dataset, showing SGD is 20 times faster and decreases pose error. Performance is on par with CryoSPARC.
38:34 DSL1 SNARE Complex Example: Applied to a challenging dataset with over 50% junk particles, CryoDragon AI robustly estimated structures where traditional methods struggled.
39:33 Ankrin One Complex Example: Found a rare "super complex" state of the ankrin one complex (700,000 particles) that was missed by traditional 3D classification, suggesting joint pose and heterogeneity estimation can reveal more states.
40:48 Ab Initio Subtomogram Averaging in CryoDragon AI: Adapts the two-stage pose estimation to tilt series data, constraining relative orientations between subtilts during HPS, and allowing individual subtilts to "jiggle" during SGD for refinement.
41:58 Tilt Selection: Finds that resolution plateaus beyond 11-13 tilts for both ribosome and FAS examples, saving computational time. SGD improves relative error for lower tilt angles.
43:05 Bacterial 70S Ribosome Example: Applied to a bacterial ribosome dataset (18,000 particles), CryoDragon AI's ab initio homogeneous reconstruction is qualitatively similar to traditional pipelines. Heterogeneous reconstruction identifies junk particles and known ribosomal translation states (AP, P, EF2 tRNA bound).
44:36 Spatial Heterogeneity: Enables visualization of spatial context for particles, like polyribosomes or ribosomes near membranes, without masking.
45:05 FAS Complex Ab Initio Reconstruction: Learned that for symmetric complexes, higher resolution information in frequency marching is needed during pose search to overcome local minima due to symmetry.
45:38 Next Steps & Future Directions: Evaluating on new datasets, method improvements (constraining pose search with existing alignments, refining CTF, deeper analysis of hyperparameters), and ab initio reconstruction of complex mixtures using Hydra (mixture of K expert neural models).
47:45 Software Update: CryoDragon ET is in stable release (v3.4.4). CryoDragon AI is in beta (v4.1.0 branch) and being merged into the main development branch.
48:48 New Interface in Version 4: Introduces a new two-command interface (cryodragon setup and cryodragon train) for clearer job setup and parameter inspection, replacing complex single-line commands. Legacy commands will still be supported.
50:26 Benchmarking Efforts (CryoBench): Emphasizes the importance of benchmarking heterogeneous reconstruction. Discusses the lack of ground truth for real data and the need for standardized evaluation.
51:44 CryoBench Design: Focuses on conformational and compositional heterogeneity, with simple diagnostic datasets (e.g., IGG1D's 1D circular motion) and challenging ones (random linkers, mixtures of 46,000 MD structures, 100 distinct structures).
52:21 Benchmarking Challenges: Shows results for spike MD and tommo twin 100 datasets, highlighting the difficulty in evaluating different methods due to varying latent spaces and volume distributions. Calls for more ab initio methods and better metrics.
54:19 Conclusion: Cryo-EM is pushing structural biology frontiers. CryoDragon's neural representation is expressive for heterogeneity. Recent work extended it to ab initio reconstruction, in situ data (CryoDragon ET), and will tackle complex mixtures. Benchmarking is ongoing, and advanced models/validation metrics are still needed.
55:33 Q&A: Total Dose for Tilts: Clarifies that dose is equal among tilts but damage accumulates.
56:43 Q&A: Latent Space Transferability: Explains that latent spaces are optimized per dataset/model and are not directly transferable due to arbitrary layouts and powerful, nonlinear decoders.
58:21 Q&A: Intuition for Hyperparameters & Overfitting: Discusses latent space dimensionality (low-dimensional like 8D usually best; too large can underfit) and architecture size (larger allows more powerful modeling but can overfit to noise; relates to box size).
1:01:18 Q&A: Size/Particle Limit: States there's no theoretical limit, but it's data-dependent. Smaller samples are harder due to less signal. Has worked with ~100 kDa proteins.