CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference
Authors: Shayan Shekarforoush, David Lindell, Marcus A. Brubaker, David J. Fleet
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through evaluation on synthetic datasets, we demonstrate that cryo SPIN is able to handle multi-modal pose distributions during the amortized inference stage, while the later, more flexible stage of direct pose optimization yields faster and more accurate convergence of poses compared to baselines. On experimental data, we show that cryo SPIN outperforms the state-of-the-art cryo AI in speed and reconstruction quality. |
| Researcher Affiliation | Collaboration | Shayan Shekarforoush1, 2 shayan@cs.toronto.edu David B. Lindell1, 2 lindell@cs.toronto.edu Marcus A. Brubaker1, 2, 3, 4 mab@eecs.yorku.ca David J. Fleet1, 2, 4 fleet@cs.toronto.edu 1University of Toronto 2Vector Institute 3York University 4Google Deep Mind |
| Pseudocode | No | The paper describes the method using diagrams and text but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions "Project Webpage." in the abstract but does not provide a direct link to a source-code repository for cryo SPIN or explicitly state that the code is provided in supplementary materials. |
| Open Datasets | Yes | We generate synthetic datasets by simulating the image formation process formalized in Sec. 3 using atomic models deposited in the Protein Data Bank (PDB). ...As a widely adopted experimental benchmark, we use the 80S ribosome dataset (EMPIAR-10028 [48]) |
| Dataset Splits | No | The paper mentions splitting data into two halves for Fourier Shell Correlation calculation but does not specify explicit train/validation/test splits with percentages or sample counts for the training process. |
| Hardware Specification | Yes | Experiments are run on a single NVIDIA A40 GPU. |
| Software Dependencies | Yes | Methods are implemented in Pytorch [50]. ...run cryo SPARC v4.4.0 [8] with default settings... |
| Experiment Setup | Yes | During auto-encoding, we use encoders with M = 7 and M = 15 heads for reconstruction on synthetic and real datasets, respectively, with Adam [49] to optimize encoder and decoder with learning rates 0.0001 and 0.05. Once switched to direct optimization (after 7 epochs for synthetic and 15 epochs for real data), we reduce the decoder learning rate to 0.02 and allocate a new optimizer for pose parameters with learning rate 0.05. To be consistent with cryo AI, we use a batch size of 64 and train for the same number of epochs (20 for synthetic, 30 for real data). |