Research map / 3D vision / Gaussian splatting
Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CryoSPIRE, a hierarchical part-aware Gaussian mixture model for cryo-EM heterogeneous reconstruction, with a learned-feature part-discovery step that initializes anchors, and reports state-of-the-art on CryoBench.
Problem
Cryo-EM 3D reconstruction of biomolecules with conformational flexibility and compositional variation (missing parts) from extremely noisy 2D images, where existing methods are limited in expressiveness, interpretability, or efficiency.
Approach
Discover a part-based segmentation by optimizing a coarse Gaussian mixture with per-Gaussian learnable features and clustering those features; then build a hierarchical GMM with one anchor per part (rigid anchor motion, per-Gaussian displacements, and sigmoid amplitude modulation for compositional absence) conditioned on a latent code.
Key results
- On CryoBench IgG-1D, CryoSPIRE reaches mean Per-Conformation FSC AUC 0.402 vs 0.391 for RECOVAR (second best).
- On IgG-RL, CryoSPIRE achieves 0.386 vs 0.372 for RECOVAR in Per-Conformation AUC.
- On Ribosembly, CryoSPIRE (0.427) is not statistically different from RECOVAR (0.430).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CryoSPIRE instead of RECOVAR | CryoBench methods · IgG-1D | Per-Conformation FSC AUC: 0.391 (0.001) -> 0.402 (0.002) | Helped |
| Use CryoSPIRE instead of RECOVAR | CryoBench methods · IgG-RL | Per-Conformation FSC AUC: 0.372 (0.008) -> 0.386 (0.014) | Helped |
| Use CryoSPIRE instead of RECOVAR | CryoBench methods · Ribosembly | Per-Conformation FSC AUC: 0.430 (0.016) -> 0.427 (0.014) | No clear effect |
| Add anchor-based motion modeling (ablation) | CryoSPIRE · IgG-1D and IgG-RL | Per-Conformation FSC AUC | Helped |
| Use anchor-based motion modeling rather than direct per-Gaussian deformation (ablation) | CryoSPIRE · Ribosembly | Per-Conformation FSC AUC | No clear effect |
| Reduce over-segmentation from 64 anchors (ablation) | CryoSPIRE · CryoBench | Per-Conformation FSC AUC | Helped |
| Use learnable Gaussian features instead of positional encoding (ablation) | CryoSPIRE · CryoBench | Per-Conformation FSC AUC | Helped |
Surprising findings
- Learned Gaussian features cluster into coherent parts that move together or appear/disappear together without any direct supervision on features.
- Positional encodings instead of learnable Gaussian features give inferior results and fail to identify meaningful parts.
Limitations
- Requires an initial structure and image poses; inaccuracies may limit performance.
- Latent space interpretation and validation on experimental heterogeneous data remain open.
- Not fully ab initio.
Experimental setup
Datasets: CryoBench IgG-1D CryoBench IgG-RL CryoBench Ribosembly EMPIAR-10076 EMPIAR-10180
Benchmarks: CryoBench IgG-1D CryoBench IgG-RL CryoBench Ribosembly
Metrics: Per-Conformation FSC AUC Per-Image FSC AUC
Baselines: 3DVA 3DFlex RECOVAR CryoDRGN DRGN-AI 3D Classification
Builds on: Scaffold-GS 3D Gaussian Splatting DynaMight E2GMM CryoBench
Closest papers by meaning
- Mixture of neural fields for heterogeneous reconstruction in cryo-EM (NeurIPS 2024 · similarity 0.92)
- CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM (NeurIPS 2024 · similarity 0.92)
- cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.91)
- CryoFM: A Flow-based Foundation Model for Cryo-EM Densities (ICLR 2025 · similarity 0.90)
- Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data (NeurIPS 2025 · similarity 0.90)
- CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.