An end‑to‑end, automated workflow for single‑particle cryo‑EM that achieves state‑of‑the‑art consensus maps suitable for model building—demonstrated across 21 GPCR datasets, meeting or exceeding published resolutions and map quality in the majority of cases.
A reusable workflow that encodes best practices—micrograph curation, robust picking, rigorous 3D curation via decoy classification and Ab‑Initio reconstruction, followed by Non‑Uniform and Local Refinement—with optional Reference‑Based Motion Correction.
Reliable, scalable repeat‑target cryo-EM structure determination that can be adopted immediately and extended to other target classes in structure-based drug design and life science research contexts.
We tested the automated workflow on 21 GPCR datasets (both active and inactive states, across a range of species and sample conditions). In 17 of 21 cases, the automated workflow attained equal or better resolution and map quality compared to manual depositions. Furthermore, in 10 cases, we were able to obtain improved interpretability in the ligand binding site.
Automated workflows rely on new developments in CryoSPARC, notably the Workflows UI, Micrograph Junk Detector for automatic segmentation of junk regions, Micrograph Denoiser for rapid, dataset-specific denoiser training and inference, as well as auto-clustering in Inspect Particle Picks and Reference Based Auto Select 2D and 3D for unattended particle and volume selection.
Average end‑to‑end runtime ~34 hours on a 2‑GPU workstation across all 21 datasets; 2-3x speedups on pre-processing stages using an 8-GPU server and 1.7x speedup for the entire workflow. On‑the‑fly pre‑processing in CryoSPARC Live can reduce time-to-structure by ~12.5 hours on average. Float16 storage yields project data approximately 52% of raw size.
Complete workflow for end-to-end automation of repeat-target structure determination. A. Workflow flowchart from import to final refinements, detailing required inputs at each stage. Dataset level inputs must be provided for each new dataset. Class level inputs (C) only need to be set once, and are reused for all datasets in the target class. For example, the low resolution reference map and corresponding mask for each target class in this work are shown in insets. Workflow level inputs (W) can be reused across multiple classes. B. 2D classes for template picking are automatically selected using Reference Based Auto Select 2D (RBAS 2D). C. Micrograph Denoiser improves particle picking, and particles close to junk and contaminants are automatically rejected by Micrograph Junk Detector. D. Decoy classification curates particles in 3D without introducing orientation bias. E. Further curation using Ab-initio Reconstruction and Heterogeneous refinement yields final high-quality particles. F. Non-Uniform Refinement (left) produces an optimal final global refinement, and Local Refinement (right) using the input mask improves map quality in the receptor region.
To explore how automation can accelerate your own structure determination workflows, consult the accompanying research paper and follow the step-by-step guide to reproduce on your data.
