Research map / 3D vision / Human pose estimation
CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference
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 cryoSPIN, a semi-amortized ab-initio cryo-EM reconstruction method that uses a multi-head encoder with winner-takes-all loss for multi-modal pose inference, then switches to per-particle SGD pose refinement with an explicit Hartley-space volume decoder.
Problem
Ab-initio cryo-EM reconstruction requires jointly estimating a 3D density and per-image poses; the pose posterior is highly multi-modal early on, and amortized (encoder-based) pose inference can get stuck in wrong modes and converge slowly.
Approach
Combine a multi-head pose encoder trained with a winner-takes-all (multi-choice learning) loss through an explicit volume decoder, then switch from amortized auto-encoding to per-image pose optimization via SGD (auto-decoding) once the pose posterior is closer to unimodal.
Key results
- On synthetic HSP, mean rotation error drops to 3.27 degrees (vs 6.23 for cryoSPARC and 45.83 for cryoAI).
- cryoSPIN is reported as about 6x faster and about 5x less GPU memory than cryoAI.
- On EMPIAR-10028 cryoSPIN gives higher resolution than cryoAI and matches cryoSPARC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| switch from auto-encoding to direct per-particle pose optimization (ablation) | CryoSPIN · Spike and Spliceosome synthetic | pose error during training | Helped |
| use a multi-head pose encoder to return multiple candidates | CryoSPIN · HSP synthetic | reconstruction resolution | Helped |
| replace the implicit decoder with an explicit decoder (ablation) | CryoAI · Reconstruction (decoder ablation) | time per epoch: 622.39 -> 142.40 | Helped |
| replace the implicit decoder with an explicit decoder (ablation) | CryoAI · Reconstruction (decoder ablation) | GPU memory decoding: 14.05 -> 0.75 | Helped |
| continue auto-encoding instead of switching to auto-decoding (ablation) | CryoSPIN · Reconstruction (decoder ablation) | time per epoch: 99.76, 81.61; 99.75 | No clear effect |
| use explicit decoder with CryoAI's input-augmentation encoder (ablation) | CryoAI · Reconstruction (decoder ablation) | time per epoch: 99.76, 81.61; 142.40 | Hurt |
| use explicit decoder with CryoAI's input-augmentation encoder (ablation) | CryoAI · Reconstruction (decoder ablation) | GPU memory encoding: 1.17; 2.32 | Hurt |
| add the multi-head encoder | CryoSPIN · HSP synthetic | rotation error (mean / median, degrees): 45.83 / 61.86 -> 3.27 / 0.97 | Helped |
| add semi-amortized pose inference | CryoSPIN · CryoAI comparison | reconstruction speed: ~6x faster | Helped |
Surprising findings
- cryoAI gets stuck in local minima on HSP and on the 80S ribosome when particles are off-centered, requiring pre-centering of images for it to work.
Limitations
- Assumes a rigid homogeneous structure; heterogeneity not addressed.
- No principled criterion for when to switch from amortized to auto-decoding stage.
- cryoAI required preprocessing (centering) to work on off-centered experimental data.
Experimental setup
Datasets: EMPIAR-10028 Protein Data Bank atomic models (HSP, pre-catalytic spliceosome, SARS-CoV-2 spike)
Benchmarks: synthetic HSP, Spliceosome, Spike datasets EMPIAR-10028 80S ribosome
Metrics: Fourier Shell Correlation resolution mean/median rotation error resolution vs. time training time per epoch GPU memory
Baselines: cryoSPARC cryoAI cryoDRGN cryoAI-explicit fully-amortized variant
Builds on: cryoAI cryoDRGN cryoDRGNv2 multi-choice learning cryoSPARC
Closest papers by meaning
- cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.91)
- CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM (NeurIPS 2024 · similarity 0.90)
- Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery (NeurIPS 2025 · similarity 0.89)
- Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data (NeurIPS 2025 · similarity 0.89)
- Mixture of neural fields for heterogeneous reconstruction in cryo-EM (NeurIPS 2024 · similarity 0.89)
- CryoFM: A Flow-based Foundation Model for Cryo-EM Densities (ICLR 2025 · 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.