Designing and Operating an In-House Cryogenic Electron Microscopy Facility with IT Infrastructure and On-the-Fly Data Processing Pipelines

Summary

This protocol provides a practical framework for designing and operating the IT infrastructure of an in-house Cryogenic electron microscopy (cryo-EM) facility. It covers facility planning, network architecture, storage, GPU computing, remote access, and real-time data processing, enabling efficient management of multi-terabyte datasets while ensuring scalable, sustainable, and high-performance cryo-EM operations.

Abstract

Cryo-EM has become a dominant method for high-resolution structural determination of biological macromolecules. As the number of institutional cryo-EM facilities grows worldwide, a persistent challenge has emerged: the data infrastructure required to support these facilities is routinely underplanned. Modern cryo-EM instruments paired with direct electron detectors (DEDs) generate between 2 and 5 TB of raw data per day, and a 20-user facility can accumulate approximately 2 PB of data within a single operational year. Processing this data demands dedicated GPU computing nodes, high-speed internal networking, and robust long-term archival strategies—all of which must be in place before data collection begins. This protocol presents an integrated framework for the IT and computational design of an in-house cryo-EM facility. We cover physical facility requirements (room design, power, cooling, and vibration isolation), equipment selection (microscopes, DEDs, and supporting instruments), and the full computational stack: network switch configuration, storage server selection and sizing, GPU compute node provisioning, and remote access architecture. We then provide a detailed protocol for implementing on-the-fly data processing pipelines—covering motion correction, contrast transfer function (CTF) estimation, particle picking, and 2D classification in real time during data collection—using industry-standard tools. Data transfer strategies for external users are also addressed. This guide fills a critical gap in the literature and will enable researchers and administrators to build cryo-EM facilities that are computationally ready from day one.

Introduction

The cryo-EM resolution revolution has transformed structural biology, enabling atomic-resolution visualization of proteins, nucleic acids, and macromolecular complexes without the need for crystallization^1,2. Three technical advances drove this transformation: the development of DEDs with high frame rates, improvements in electron optics and microscope mechanics, and the optimization of GPU-accelerated image processing software^3,4,5,6. Together, these advances have made cryo-EM the method of choice for an expanding community of structural biologists.

The demand for cryo-EM access has prompted significant investment in both large national centers and smaller institutional facilities. National centers—such as NCCAT, S2C2, PNCC, and MCCET in the USA, and eBIC, NeCEN, and SciLifeLab in Europe^3,5,6, provide access to the highest-end instruments for approved projects. In-house facilities complement these resources by supporting iterative, project-specific research with faster turnaround times and closer user-staff interaction. However, establishing a functional in-house facility requires far more than purchasing a microscope: the supporting IT infrastructure is equally critical, and its inadequacy is a leading cause of operational failure in new facilities⁶.

A single high-end cryo-EM session can generate 2–5 TB of raw movie data in 24 h. With multiple microscopes and users, the data burden rapidly reaches petabyte scale. Processing this data requires GPU-accelerated pipelines that must be operational concurrently with data collection^7,8. Delays in preprocessing mean delayed feedback on data quality and wasted instrument time. Despite these realities, most published guides on cryo-EM facility management devote minimal attention to the IT components, and none provide a step-by-step operational protocol for setting up and validating a complete data infrastructure.

This protocol addresses that gap. We describe the physical and technical requirements for facility construction, the selection criteria for microscopes and detectors, and—most importantly—the design and implementation of the IT backbone that makes a cryo-EM facility operationally viable⁶. Particular attention is given to on-the-fly data processing: the automated, concurrent preprocessing of data during collection that enables real-time quality control, guides collection decisions, and dramatically reduces post-session turnaround time^9,10,11. By following this protocol, facility developers will be equipped to build a complete, integrated cryo-EM operation that serves a broad scientific community sustainably and efficiently.

Protocol

Details of the equipment, software, and materials used throughout this protocol are provided in the Table of Materials.

1. Physical facility design, personnel, and infrastructure planning

1. Conduct a location assessment. Select a building site that provides: a vibration-isolated floor (ambient vibration levels, ideally < 1 µm/s RMS at 1–100 Hz); magnetic field stability (< 1 mG/m spatial gradient and < 0.1 mG/m temporal variation for 300 kV instruments); stable room temperature (18–22 °C with < 0.5 °C fluctuation); humidity control (40–60% RH); and low acoustic noise (<50 dB). Consult the microscope manufacturer's site requirements early in the planning phase.
   NOTE: Refurbishing an existing room can cost as much as new construction. As a rule of thumb, budget approximately 20% of the microscope purchase price for room preparation.
2. Design the facility layout with clear spatial separation between functional zones: (a) microscopy room(s) for cryogenic transmission electron microscopes (cryo-TEMs); (b) a sample preparation anteroom for grid handling and storage; (c) a controller room housing water chillers, UPS units, and microscope chiller systems; and (d) a separate server room or data center for storage and compute infrastructure.
   NOTE: Maintain the controller room below 26 °C at all times and provide dedicated backup power and cooling. Temperature excursions in the controller room are a common source of microscope downtime.
3. Plan the power infrastructure. Install dedicated three-phase power circuits for high-end cryo-TEM. Connect all microscopes, cameras, and server equipment to the designated protected power infrastructure. Coordinate with the institution's facilities team to include an emergency generator capable of sustaining microscope operation and cooling water flow during power outages.
   NOTE: Several instruments within a cryo-EM facility require more than one day for safe power cycling. Unplanned power loss can damage sensitive circuit boards and result in the loss of active data collection sessions.
4. Install cooling water systems for the electron microscope and associated hardware. Establish a backup cooling water circuit with automatic failover. Monitor water flow rate, temperature, and conductivity continuously using the microscope's integrated monitoring system or a dedicated sensor array.
5. Install fiber optic cable runs from the microscopy room(s) to the server room during construction or renovation. Specify at minimum OM4 multimode or OS2 single-mode fiber, with sufficient slack and labeled patch panels at both ends.
   NOTE: Pre-installing fiber is substantially cheaper than retrofitting.
6. Assign dedicated microscope operators or facility scientists to perform instrument maintenance, user training, scheduling, sample optimization, and data acquisition support. Assign dedicated IT personnel to manage infrastructure, network performance, cybersecurity, data backup, software deployment, GPU computing resources, and remote user access.
7. Determine staffing requirements according to the level of user support provided and the complexity of the computational infrastructure. In smaller facilities, such as the UCSC Biomolecular cryo-EM Facility, these responsibilities may be shared by a facility manager with computational expertise or centralized institutional IT services; however, as the number of microscopes, users, and annual data collection sessions increases, specialized personnel become increasingly necessary.
8. Provide dedicated IT support for facilities operating multiple cryo-TEMs to ensure uninterrupted data acquisition, efficient data management, and scalable computing resources. Invest in robust IT infrastructure and personnel early in facility development to minimize operational downtime, simplify future expansion, and provide a sustainable framework capable of supporting growing user communities and increasingly data-intensive cryo-EM workflows.

2. Microscope and detector selection

1. Define the facility's primary scientific mission before selecting a microscope. Consider: (a) whether the primary workflow will be single-particle analysis (SPA), cryo-electron tomography (cryo-ET), micro-electron diffraction (microED), or a combination; (b) the expected user base size and project diversity; (c) budget constraints for both acquisition and ongoing service contracts.
2. Evaluate available cryo-TEM platforms according to the intended applications. For high-resolution SPA and subtomogram averaging (STA), consider 300 kV instruments. For screening and medium-resolution data collection, consider 200 kV instruments. For a lower-cost entry point with screening capability, evaluate 100–120 kV (summarized in Table 1).
   NOTE: Specification sheets change frequently. Always request current configuration details and pricing directly from the manufacturer before finalizing equipment decisions.
3. Select a DED matched to the microscope and application. For high-resolution SPA on Thermo Fisher Scientific (TFS) instruments, the Falcon 4i is currently the primary option on Krios and Glacios series.
4. For JEOL CryoARM instruments, the Gatan K3 provides excellent sensitivity and frame rates. For facilities targeting both SPA and microED, confirm detector compatibility with electron diffraction data collection modes (Table 2).
   NOTE: The DED is one of the most expensive components of the facility and significantly impacts data quality and storage requirements. Electron Event Register (EER) mode, for example, produces substantially larger raw files than conventional compressed formats, requiring proportionally greater storage and network bandwidth.
5. Evaluate the role of an energy filter when selecting a cryo-EM instrument configuration, as it can significantly influence image quality, detector choice, and overall system performance. By removing inelastically scattered electrons, energy filters improve contrast and signal-to-noise ratio, which is particularly beneficial for challenging samples, high-resolution SPA (beneficial), cryo-ET (critical), and microED (beneficial) applications.
6. Evaluate available energy filter systems during instrument selection. Gatan Bioquantum and Biocontinuum or Thermo Fisher Scientific Selectris and SelectrisX systems typically cost approximately US$0.3–1.5M, depending on the configuration.
7. Evaluate the additional cost and complexity associated with energy filters in combination with detector capabilities, facility goals, and expected user applications before selecting the final instrument configuration.
8. Factor in service contracts when calculating total acquisition cost. Service contracts typically cover the microscope, energy filter, detector, and water-cooling system.
9. Negotiate service contract terms—including response time guarantees and loaner parts availability—at the time of instrument purchase. Budget 8–15% of the instrument purchase price per year for service coverage.

3. Network architecture design and implementation

1. Design the internal facility network around a core high-speed switch. Connect all data-generating and data-consuming devices directly to this switch at the highest supported speed: microscope controller, camera controller (often a separate workstation), storage servers, and GPU compute nodes.
2. Provide a minimum 10 Gbps connection to each device. Provision 25 Gbps or faster uplinks between the switch and the storage and compute nodes where budget allows.
   NOTE: At 1 Gbps, transferring 20 TB of raw data would take 2–3 days. At 10 Gbps, the same transfer takes 5–6 h. At 25 Gbps, it takes approximately 2 h. Network speed directly determines whether on-the-fly processing can keep pace with data acquisition.
3. Specify the core switch. Requirements: (a) non-blocking switching fabric at full line rate; (b) sufficient SFP+ or QSFP ports for all facility devices; (c) managed switch functionality (VLAN support, traffic monitoring, port mirroring for diagnostics); (d) redundant power supply modules. Select an enterprise-grade managed switch that provides a nonblocking switching fabric, sufficient high-speed ports, traffic-management functions, and redundant power supplies.
4. Segment the network using VLANs to separate: (a) the data acquisition VLAN (microscope and camera controllers, storage servers); (b) the compute VLAN (GPU workstations and cluster nodes); (c) the management VLAN (remote access, monitoring, administration).
   NOTE: Network segmentation reduces broadcast traffic, improves security, and allows independent troubleshooting of each subsystem.
5. Connect the facility network to the institutional WAN via a dedicated uplink port on the core switch. A 1–10 Gbps uplink to the campus network is sufficient for remote access and external data transfer.
6. Coordinate with institutional IT to provision static IP addresses, DNS entries, and appropriate firewall rules for all facility-accessible devices (Figure 1).
7. Install and label all network cable runs. Use direct attach cables or active optical cables for 10+ Gbps connections up to 100m (shielded Cat6A also works for 10 Gbps up to 100m); use pre-terminated fiber assemblies (LC-LC or MTP) for inter-rack 25+ Gbps links.
8. Document the complete cabling scheme with a network diagram specifying device names, IP addresses, port assignments, and VLAN membership.
   NOTE: Refer to Figure 2 for a schematic of the recommended minimal network topology for a small cryo-EM facility.

4. Storage server design and deployment

1. Calculate storage requirements before procuring hardware. As a baseline: a single high-end cryo-EM session generates 2–5 TB of raw data per day.
2. Estimate long-term storage requirements based on anticipated facility usage. A 20-user facility running two microscopes can generate approximately 2 PB of data within 3–6 months of operation. Plan for both raw data storage (short-term, high-throughput) and processed data storage (medium-term, larger volume), as processed datasets are frequently 10–20x larger than the raw input.
3. Select a robust storage architecture. Standalone ZFS-based servers are the recommended starting point for most facilities, offering a favorable cost-to-capacity ratio, built-in data integrity verification (checksumming), flexible RAID-equivalent configurations (RAIDZ, RAIDZ2), and snapshot support. Clustered distributed file systems (Gluster, Ceph, GPFS/IBM Spectrum Scale) can offer higher performance and scalability but require multiple physical servers and more complex administration.
4. Configure ZFS storage pools. Recommended starting configuration: two RAIDZ2 vdevs (each comprising 6–8 high-capacity SAS HDDs, 16–20 TB each), providing approximately 60–70% usable capacity with protection against two simultaneous drive failures.
5. Add a dedicated SSD write-cache (L2ARC) vdev using 2–4 NVMe SSDs to accelerate random read performance for active processing jobs.
   NOTE: Do not use consumer-grade SATA drives in production storage arrays. Enterprise SAS or NL-SAS drives with rated workload capacity (e.g., 550 TB/year) are essential for the sustained random I/O patterns of cryo-EM workflows.
6. Configure NFS exports (for Linux clients) and SMB shares (for Windows clients) on the ZFS storage server. Mount the primary data share on all GPU compute nodes and on the camera controller workstation to enable direct data access without intermediate copies.
7. Establish a data retention and lifecycle policy. Define: (a) raw movie retention period (typically 3–6 months post-collection); (b) motion-corrected micrograph retention; (c) final processed datasets retention (indefinite until publication); (d) archival procedure for completed projects.
8. Implement automated alerts when storage utilization exceeds 70% capacity to allow timely procurement of additional storage.
   NOTE: Storage is the most commonly underestimated cost in cryo-EM facility planning. Do not conflate microscope acquisition approval with adequate storage funding. Storage procurement must be planned in parallel with instrument installation, not after data collection begins.

5. GPU compute infrastructure

1. Determine compute requirements based on expected throughput. For a single microscope generating 3,000–5,000 micrographs per 24 h session, a minimum of two compute nodes, each equipped with 4x NVIDIA GPU cards (Maxwell or newer, 11+ GB VRAM, e.g., RTX 3090, RTX 4090, or A100), is required to sustain on-the-fly preprocessing at collection speed.
2. For larger facilities or cryo-ET workflows with high tilt-series volume, provision additional nodes.
3. Specify GPU compute node hardware. Each node should include: (a) 4x NVIDIA GPUs with CUDA support (minimum 24 GB VRAM per GPU for CryoSPARC and RELION 3D refinements); (b) a high-core-count CPU (AMD EPYC or Intel Xeon, ≥32 cores) to support parallel CPU-bound steps (CTF estimation, file I/O); (c) ≥256 GB system RAM; (d) 2–4 TB NVMe SSD local scratch space for active job intermediates; (e) a 10/25 Gbps network interface card (NIC); (f) sufficient local storage for working data sets (typically a RAID volume).NOTE: Consumer-grade GPUs (RTX 4090) can offer comparable performance to enterprise-grade alternatives (A100) at significantly lower cost for cryo-EM workloads. Evaluate benchmarks for CryoSPARC and RELION specifically, as their memory access patterns favor high-memory-bandwidth GPUs.
4. Install the NVIDIA driver stack, CUDA toolkit (version compatible with CryoSPARC and RELION), and cuDNN on each compute node. Use a configuration management tool to standardize software environments across all nodes and simplify future updates.
5. Deploy CryoSPARC in a standalone or cluster configuration. For a multi-node facility, install CryoSPARC-master on a dedicated head node and CryoSPARC-worker on each GPU compute node.
6. Configure job scheduling to allow multiple users to submit concurrent processing jobs. Follow the official CryoSPARC architecture documentation for minimum system requirements and network configuration.
7. Install RELION alongside the primary processing platform to provide users with both processing ecosystems. Configure the software to use the local compute nodes via a scheduler (SLURM or PBS/Torque).
8. Install accessory tools: MotionCor2 (motion correction), CTFFIND4 and Gctf (CTF estimation), TOPAZ (machine-learning-based particle picking), crYOLO (neural-network particle picker), and cryoDRGN (heterogeneity analysis). Refer to Table 3 for an overview of the available software packages for SPA, cryo-ET, STA, and microED. 
9. Use SBGrid consortium deployment or container-based software management (Singularity/Apptainer for HPC environments; Docker for workstations) to maintain reproducible, versioned software environments.
   NOTE: SBGrid provides pre-compiled, validated builds of all major cryo-EM software tools with automated update management for academic institutions.

6. Remote access infrastructure

1. Establish VPN-based remote access as the primary secure entry point for all off-site users and staff. Work with institutional IT to provision a facility-specific VPN profile or tunnel that grants access to the facility's compute and storage network without exposing devices to the broader public internet.
2. Implement two-factor authentication (2FA) on all remote access points (Figure 2).
3. Deploy remote desktop software for graphical access to microscope controllers and GPU workstations. Recommended options: (a) TeamViewer (commercial, cross-platform, NAT-traversing—suitable for microscope controller access); (b) NoMachine (commercial, cross-platform, supports OpenGL 3D graphics forwarding—recommended for GPU workstations running visualization tools); (c) TurboVNC with VirtualGL (open-source, Linux-only, supports remote 3D rendering).
   NOTE: Many cryo-EM visualization tools use OpenGL for 3D rendering. Standard X11 forwarding does not support hardware-accelerated OpenGL. Use NoMachine or VirtualGL/TurboVNC for these applications.
4. Configure remote access to the microscope data acquisition software. Use Leginon or SerialEM for remote operation over the institutional VPN. The microscope acquisition software can be accessed remotely through an approved remote desktop application.
5. Test remote connectivity from representative off-site locations (home networks, collaborating institutions) before going live with user operations (Table 4).
6. Set up a web-based facility monitoring dashboard to allow staff and users to monitor data collection status, preprocessing outputs, storage utilization, and compute queue status remotely. Leginon includes a built-in web monitoring interface.
7. For the CryoSPARC workflow, configure Grafana or a custom web dashboard fed by pipeline log outputs and system monitoring tools (Prometheus, node_exporter).
8. Establish a data delivery mechanism for external users. Configure an authenticated data-transfer service, such as Globus Connect Server, on the facility storage.
9. For users without Globus access, prepare USB hard drive shipping procedures with documented checksums. A 10 Gbps internet uplink is necessary for practical direct download of multi-TB datasets by external users.

7. On-the-fly data processing workflow: implementation

1. Pre-configure the data flow path before each data collection session. Confirm that: (a) the camera controller writes raw movies directly to the facility's NFS-mounted high-performance storage; (b) the NFS share is mounted and accessible on all GPU compute nodes; (c) sufficient free storage capacity exists for the planned session (minimum 5 TB free for a 24 h SPA session); (d) the GPU compute nodes are available and no long-running jobs will compete for resources.
2. Set up a gain reference correction. Prior to each session, acquire a gain reference image as specified by the detector vendor (typically a blank beam exposure on an empty area of the grid).
3. Place the gain reference file in the designated preprocessing input directory. All motion correction software (MotionCor2, RELION's own implementation) will apply this reference automatically if correctly configured.
4. Launch the on-the-fly preprocessing pipeline using CryoSPARC-Live (preferred for real-time visualization). The pipeline should be initiated once the first movies appear in the output directory—typically within the first 5–10 min of automated data collection.
   NOTE: CryoSPARC-Live continues to monitor a designated directory for new movie files and processes them as they arrive. The key distinction is that CryoSPARC-Live provides a real-time interactive web interface for monitoring.
5. CryoSPARC-Live workflow
   1. In the web interface, navigate to a new or existing project and create a new live session^11,36.
   2. Configure the input source: specify the directory path being written to by the camera controller (the 'watch folder'), the file pattern matching raw movie files (e.g., *.tif, *.mrc, or *.eer for EER-format Falcon 4i data), and the gain reference file path.
   3. Set motion correction parameters: number of frames to use, patch-based motion correction (recommended: 5x5 patches), output binning factor (typically 1 for data collection near Nyquist; 2 for initial screening), and electron dose per frame (calculated from total dose and number of frames).
   4. Set CTF estimation parameters: minimum and maximum resolution for fitting (typically 30 Å and 4 Å), defocus search range (0.5–5 µm recommended for SPA), and maximum acceptable astigmatism. Enable per-tilt CTF estimation for cryo-ET datasets.
   5. Configure particle picking: select the picking method (blob picker for initial sessions; template picker after first good 2D classes are available; TOPAZ picker for difficult samples). Set expected particle diameter and minimum/maximum particle spacing.
   6. In order to substantially reduce storage requirements, data transfer load, and computational cost, Fourier cropping during particle extraction by a factor of 4 (resulting in 4x the physical pixel size) is recommended. For routine data quality assessment, intermediate resolutions of 3–5 Å are typically sufficient during on-the-fly processing.
   7. Enable 2D classification within the live session. Set the number of 2D classes (typically 50–200), the batch size for incremental updates (200–500 particles), and the resolution limit for classification (6–8 Å for initial assessment).
   8. Start the live session. Monitor the real-time dashboard for: incoming movie count and rate, motion correction quality (total motion per movie—reject movies with >30 Å total motion as a rule of thumb), CTF fit quality (reject micrographs with CTF fit resolution >6–8 Å), particle yield per micrograph, and evolving 2D class average quality.
   9. Set acceptance thresholds for automated curation. In the Filters panel, define rejection criteria for motion (total motion cutoff), CTF quality (fit resolution cutoff), and defocus range (exclude extremely under- or over-focused micrographs). Curated particles are automatically routed to the downstream 2D classification job.

8. Real-time quality control and feedback during data collecti

1. Monitor CTF statistics continuously throughout the session. Plot defocus values for each micrograph: the distribution should cluster within the intended defocus range (typically −0.5 to −3.0 µm for SPA).
2. Systematic drift toward higher defocus values may indicate stage drift or eucentric height miscalibration. Alert facility staff immediately if CTF fit resolution degrades below 6 Å, as this typically indicates ice contamination, beam instability, or a focus drift requiring intervention.
3. Monitor motion correction outputs. Total in-frame motion per movie should be less than 10–30 Å for high-quality data. Systematic high motion in early frames is expected and is corrected; however, high motion persisting throughout the movie (>40 Å total) may indicate mechanical stage instability or sample charging and should be investigated.
4. Inspect 2D class averages as they update during the session (every 30–60 min). Indicators of a high-quality dataset: class averages showing secondary structure features, multiple distinct views of the particle, and absence of stacking or aggregation artifacts.
5. Monitor for indicators requiring intervention: all particles sorting into featureless 'junk' classes (suggesting poor sample or ice quality), complete absence of certain views (suggesting preferred orientation—may require grid optimization), or extremely low particle yields per micrograph.
6. Log all observations in a structured session log. Record: session start/end times, microscope and camera settings, on-the-fly statistics summary (total movies collected, fraction passing CTF filter, fraction passing motion filter, total particles picked, 2D class assessment), and any interventions made. This log is essential for retrospective quality assessment and facility performance reporting.
7. At session end, generate a summary data quality report. The live-processing platform exports a session summary automatically.
   NOTE: Another option for generating summaries and reports is to use AI tools such as Vitreum. In the UCSC Biomolecular Cryo-EM Facility, the Vitreum platform is used to manage laboratory workflows and experimental documentation. Vitreum records experimental steps as they occur, enforces protocol-level constraints, verifies workflow execution against predefined quality criteria, and provides an auditable record of laboratory activities. By integrating laboratory operations with the facility's computational and data management infrastructure, the platform minimizes manual record keeping, improves reproducibility, facilitates communication among facility personnel and users, and enables consistent tracking of experiments from sample preparation through data collection and downstream computational analysis.

9. Post-collection data processing pipeline

1. After on-the-fly preprocessing is complete, the curated binned particle stack is exported along with the curated preprocessed movies.
2. Detach the CryoSPARC project from the main workstation and then reattach it to another shared, remotely accessible workstation. This allows the quick offload and turnover to start the next data acquisition session.
   NOTE: A typical data collection may vary between 1–2 days, easily monitored by CryoSPARC-Live. However, 3D processing steps may take weeks or months, therefore justifying the project transfer in between workstations and storage servers.
3. In CryoSPARC, export curated binned particles from the live session into a standard project for offline 3D processing³⁶. 
4. Perform ab initio 3D reconstruction. In CryoSPARC, use the Ab Initio Reconstruction job (3–5 initial models recommended for heterogeneous samples). Generate 1–4 initial models to assess structural heterogeneity.
5. Once the reconstructed values are satisfactory, re-extract the unbinned particles to enhance the overall resolution.
6. Perform 3D refinement using heterogeneous refinement (if multiple conformations or classes are suspected), followed by homogeneous refinement for the best class. In CryoSPARC, use heterogeneous refinement followed by non-uniform refinement. 
7. Apply Bayesian polishing in RELION¹³ or use per-particle motion correction and CTF refinement (CryoSPARC: CTF refinement job) to improve per-particle signal by accounting for beam-induced motion and per-particle defocus, depending on the user’s preference. This step routinely improves final map resolution by 0.2–0.5 Å. Additionally, reference-based motion correction (RBMC) in CryoSPARC can also be an option to improve resolution and map quality.
8. Validate the final map using the gold-standard Fourier shell correlation (FSC) procedure. Split the particle stack into two independent half-sets, refine independently, and calculate the FSC between the two half-maps. Report the final resolution at FSC = 0.143.
9. Compute the map-to-model FSC (FSC-Q) and Q-scores after atomic model fitting as independent validation metrics³⁷.
10. Perform atomic model building and validation. Use ModelAngelo for automated de novo model building from the density map²⁰. Refine the model using PHENIX real-space refinement and validate with MolProbity.
11. Use COOT³⁵ after the initial model has been generated to visualize the density map, fit or adjust atomic coordinates, correct modeling errors, build missing regions, and optimize protein, nucleic acid, or ligand structures. It enables interactive editing of residues, side-chain conformations, secondary structures, and sequence assignments, followed by real-space refinement to improve the agreement between the atomic model and experimental density.
12. Finally, deposit the final atomic model in the Protein Data Bank (PDB), the density map in the Electron Microscopy Data Bank (EMDB), and raw or preprocessed data in the Electron Microscopy Public Image Archive (EMPIAR).

Results

A successfully implemented facility IT infrastructure and on-the-fly data processing pipeline will produce the following verifiable outcomes at each stage:

Network and storage
Internal data transfer between the camera controller and storage server should sustain ≥800 MB/s (approaching theoretical 10 Gbps throughput) when measured with tools such as iperf3 or fio. This throughput is sufficient to transfer a 10–12 GB movie stack from the detector in under 15 seconds, ensuring that storage writing does not create a bottleneck during high-speed EER data collection. ZFS storage pool utilization and per-pool IOPS should be monitored daily; healthy pools show no degraded or faulted VDEVs and consistent read speeds of ≥500 MB/s for sequential reads.

On-the-fly preprocessing
The on-the-fly workflow should process each new movie—including motion correction, CTF estimation, and particle picking—within seconds of it being written to disk, thereby keeping pace with modern detector frame rates, such as the TFS Falcon 4i, operating at 5–15 movies per minute (considering 10–100 Gbps network connections and switches). For a 24 h SPA session generating ~8,000–12,000 micrographs at 150,000x magnification with the selected detector, the pipeline should deliver approximately 200,000–800,000 picked particles, a median CTF fit resolution of 3.0–5 Å, and an initial set of 2D class averages showing clear secondary structural features (alpha-helices, beta-strands) by 4–6 h into the session.

2D and 3D processing benchmarks
With 4x NVIDIA RTX 4090 GPUs per compute node and a dataset of 500,000 particles, CryoSPARC 2D classification (200 classes, 25 iterations) should complete in approximately 30–60 min. Ab initio 3D reconstruction from 100,000 particles should complete in 15–30 min. Full 3D refinement of a 300 kDa complex to near-atomic resolution typically requires 2–8 h, depending on particle count and target resolution. These benchmarks serve as a baseline for detecting performance regressions caused by hardware failures or resource contention.

Facility performance metrics
A well-functioning facility should track the following key performance indicators (KPIs) monthly: (a) instrument uptime percentage (target: >90%); (b) fraction of sessions with on-the-fly pipeline active from session start (target: 100%); (c) storage utilization and growth rate; (d) number of active users and projects; (e) number of structures deposited to EMDB/PDB; (f) fraction of sessions producing publishable-quality 2D classes. These metrics provide objective evidence of operational success and support facility reporting to funding agencies.

Figure 1: Overview of the cryo-EM facility data flow architecture. Schematic illustration of the complete data path from electron detector to final deposited structure. Raw movie frames are written from the camera controller to high-performance NFS storage via the 10 Gbps facility network switch. GPU compute nodes mount the same storage share and begin on-the-fly preprocessing (motion correction → CTF estimation → particle picking → 2D classification) concurrently with data collection. Preprocessed outputs and final 3D reconstructions are stored on the same or secondary storage pool. Remote users access compute resources and storage through an institutional VPN and approved remote desktop software. External data delivery is performed through an authenticated transfer service or by shipping physical storage media. Please click here to view a larger version of this figure.

Figure 2: Minimal network topology for a small cryo-EM facility. All data acquisition and processing devices—camera controller, microscope controller, storage servers (>1 PB capacity), and GPU compute servers (N nodes)—are connected via a central 10 Gbps (or faster) managed network switch. The switch uplinks to the institutional internet/WAN for remote access. Remote data processing and remote monitoring are accessible over the public internet via VPN. Storage capacity shown is representative for a single-microscope, 20-user facility operating for 3–6 months. Please click here to view a larger version of this figure.

Manufacturer	Model	kVmax	Electron Source	Compatible DEDs	Primary Applications	Strengths	Limitations/Notes
Thermo Fisher Scientific	Krios G4/Krios 5	300 kV	XFEG/CFEG	Falcon 4i, Falcon 3EC, Gatan K3/K2	SPA, cryo-ET/STA, MicroED	Gold-standard platform, full automation (EPU), highest throughput, best for high-res (<2.5 Å)	Highest cost, infrastructure demanding
	Glacios 2	200 kV	XFEG	Falcon 4i, Falcon 4, Gatan K3	SPA, cryo-ET, MicroED	Strong performance at lower cost, compact footprint, high accessibility	Lower ultimate resolution vs 300 kV
	Talos Arctica	200 kV	XFEG	Falcon 3, Falcon 4, Gatan K2/K3	SPA, cryo-ET	Proven, robust, widely adopted	Older generation, less automation
	Talos L120C	120 kV	LaB6	Ceta CMOS (limited DED)	Screening, MicroED (limited)	Cost-effective, training/screening	Not suitable for high-resolution SPA
JEOL	CRYO ARM 300 II (JEM-Z300FSC)	300 kV	CFEG	Gatan K3/K2, Direct Electron DE64	SPA, cryo-ET/STA, MicroED	Cold FEG (high coherence), top-tier resolution, strong optics	Less integrated ecosystem
	CRYO ARM 200	200 kV	CFEG	Gatan K3/K2, DE64	SPA, cryo-ET	Excellent performance-cost ratio, high stability	Fewer turnkey automation tools
	JEM-3200FSC	300 kV	FEG	Gatan K2/K3, DE20	SPA, cryo-ET	Large installed base, proven	Legacy system, aging support
Hitachi High-Tech	HF8300	300 kV	FEG	Gatan K3/K2, DE64 (custom integration)	SPA, cryo-ET	High-voltage capability, flexible integration	Limited adoption in structural biology
	HT7800	120 kV	LaB6	CMOS, limited DED	Screening, MicroED	Reliable screening scope	Not suitable for high-resolution cryoEM
	SU9000 (TEM/STEM)	200 kV	CFEG	Limited cryo-compatible DEDs	Niche cryo, materials	Advanced optics	Not widely used for SPA/ET

Table 1: Comparison of current cryo-TEM platforms and recommended DEDs. Summary of commercially available cryo-TEMs organized by manufacturer (ThermoFisher Scientific, JEOL, Hitachi), including model name, maximum accelerating voltage (kV_max), compatible DEDs for cryogenic operation, and primary imaging applications (SPA, STA/cryo-ET, microED). Intended as a planning reference; consult manufacturers for current specifications. Abbreviations: DED = direct electron detector; XFEG = extreme field emission gun; CFEG = cold field emission gun; FEG = field emission gun; SPA = single-particle analysis; cryo-ET = cryo-electron tomography; STA = subtomogram averaging; MicroED = micro-electron diffraction; TEM = transmission electron microscope; STEM = scanning transmission electron microscope.

Detector	Vendor	Key Features	Typical Pairing	Best Use Case
Falcon 4i	Thermo Fisher Scientific	EER mode, very high frame rate, high DQE	Krios, Glacios	High-throughput SPA, cryo-ET
Gatan K3	Gatan (AMETEK)	Counting mode, large FOV, super-resolution	JEOL + Thermo	High-resolution SPA (standard)
Gatan K2	Gatan	Previous-gen counting detector	Legacy systems	Established pipelines
DE64	Direct Electron	Large sensor, high throughput	JEOL, selected Thermo	Cryo-ET, large field-of-view
Apollo	Direct Electron	High frame rate, high throughput	Krios, Glacios, JEOLs	SPA and microED

Table 2: Current available DEDs in the market. Summary of commercially available DEDs organized by manufacturer. Abbreviations: EER = Electron Event Register; DQE = detective quantum efficiency; FOV = field of view; SPA = single-particle analysis; cryo-ET = cryo-electron tomography; microED = micro-electron diffraction.

Software	Main Function	Key Features	Reference
RELION	End-to-end SPA processing	Bayesian refinement, classification, CTF refinement	7, 12, 13
cryoSPARC	End-to-end SPA	Fast GPU pipelines, ab initio reconstruction	11
EMAN2	Image processing suite	Reconstruction, heterogeneity analysis (e2gmm)	14
cisTEM	SPA processing	User-friendly GUI, refinement and validation	15
SPHIRE	SPA (SPARX-based)	High-resolution refinement workflows	16
cryoDRGN	Heterogeneity analysis	Deep learning latent space reconstruction	17
Topaz	Particle picking	Deep learning-based picking	18
crYOLO	Particle picking	Real-time neural network picking	16
Warp	Preprocessing	Motion correction, CTF estimation, real-time processing	19
Model Angelo	Model building	Automated model building from maps	20
cryoET + STA
Software	Main Function	Key Features	Reference
IMOD	Tilt-series alignment & reconstruction	Gold standard for tomogram reconstruction	21
Dynamo	STA	Geometry-based particle picking, flexible workflows	22
emClarity	STA refinement	High-resolution STA, template matching	23
RELION (ET module)	STA refinement	Bayesian STA, integration with SPA workflows	24
EMAN2 (SPT)	STA	Subtomogram refinement tools	14
Warp/M	Preprocessing + STA	Real-time tomographic preprocessing	19
AreTomo	Tomogram reconstruction	Fast alignment and reconstruction	25
PEET	STA	Subtomogram averaging (IMOD-based)	26
PyTom	STA	Template matching, classification	27
Tomo3D	Reconstruction	Fast GPU tomogram reconstruction	28
crYOLO (3D)	Particle picking	ML-based picking in tomograms	16
AITom	AI-based analysis	Deep learning for segmentation and classification	29
MicroED
Software	Main Function	Key Features	Reference
DIALS	Diffraction processing	Integration of diffraction data	30
XDS	Data processing	Indexing and integration of diffraction data	31
MOSFLM	Data processing	Crystallography indexing	32
SHELX	Structure solution	Phasing and refinement	33
Phenix	Structure refinement	Automated model building	34
COOT	Model building	Manual model correction	35

Table 3: Examples of commonly used software packages for SPA, cryo-ET, STA, and microED. Summary of commercially and publicly available software packages. Abbreviations: SPA = single-particle analysis; cryo-ET = cryo-electron tomography; STA = subtomogram averaging; MicroED = micro-electron diffraction; CTF = contrast transfer function; GUI = graphical user interface; ML = machine learning.

Category	Software	Type	Primary Function	Key Features / Role in Facility
Operating System	Linux (Ubuntu, CentOS/Rocky)	OS	Core compute environment	HPC standard, GPU support, scripting, cluster deployment
Operating System	Windows	OS	User interface / workstation OS	Compatibility with vendor software, remote desktop access
File System	ZFS	File system	Data storage & integrity	Snapshots, redundancy, data integrity (checksums), scalable storage
Network File Sharing	NFS	Protocol	Linux-based file sharing	High-performance sharing across compute nodes
Network File Sharing	SMB (CIFS)	Protocol	Cross-platform file sharing	Windows–Linux interoperability
Remote Access	Microsoft Remote Desktop (RDP)	Remote access	GUI-based remote control	Native Windows remote access
Remote Access	NoMachine	Remote access	High-performance remote desktop	Fast GPU-enabled visualization, widely used in cryoEM
Remote Access	TeamViewer	Remote access	Remote support/access	Easy setup, cross-platform support
Remote Access	VirtualGL + TurboVNC	Remote visualization	GPU-accelerated remote rendering	Essential for remote cryoSPARC/RELION visualization
GPU Computing	CUDA	Compute framework	GPU acceleration	Required for cryoSPARC, RELION, deep learning tools
Data Transfer	Globus	Data transfer	Secure large-scale transfer	High-speed, fault-tolerant data movement across institutions

Table 4: Software reference for cryo-EM facility IT operations and data processing. Reference list of key software tools covering operating systems (Linux, Windows), file systems (ZFS), network file sharing (NFS, SMB), remote access (Microsoft RDP, NoMachine, TeamViewer, VirtualGL/TurboVNC), GPU computing (CUDA), and data transfer (Globus). Abbreviations: OS = operating system; HPC = high-performance computing; NFS = Network File System; SMB = Server Message Block; CIFS = Common Internet File System; RDP = Remote Desktop Protocol; GUI = graphical user interface.

Discussion

The IT infrastructure of a cryo-EM facility is not a supporting afterthought—it is an operational prerequisite. Facilities that begin data collection without adequate storage, networking, and processing capacity invariably face a choice between discarding raw data, halting data collection, or accepting months-long delays in structure determination. The protocols described here are designed to prevent these outcomes by ensuring that the computational backbone is designed, tested, and validated before the first user session.

The single most important principle in cryo-EM IT design is that storage and network capacity must be provisioned to match the data output of the detector, not the historical intuition of the facility designer. In EER mode, the detector can write 8–12 GB per movie stack; at ~15 movies per minute over a 24 h session, this yields approximately 2–5 TB of raw data before any compression. While on-the-fly lossless compression (available in CryoSPARC-Live) can reduce this by 50–70%, the network and storage must be able to handle the uncompressed write rate in the event that compression is disabled or fails. Designing for peak capacity rather than average capacity is essential.

On-the-fly processing represents one of the most impactful advances in cryo-EM facility operations in recent years. Before real-time pipelines were available, facility staff had no objective way to assess data quality until hours or days after collection, by which time the session was over and the sample grid may have been discarded. With CryoSPARC-Live, 2D class averages showing recognizable secondary structure can be available within 2–4 h of session start, enabling informed decisions about grid selection, data collection parameter adjustments, and session extension or termination. This feedback loop is perhaps the most direct way in which IT infrastructure improves scientific outcomes.

A known limitation of on-the-fly workflow is its sensitivity to configuration errors. An incorrectly specified pixel size, wrong gain reference, or mismatched dose-per-frame value will propagate through the entire pipeline, producing misleading quality metrics and corrupted particle stacks. We therefore recommend a standardized pre-session checklist that explicitly verifies all pipeline configuration parameters before data collection begins. This checklist should be treated as a safety-critical document, analogous to pre-flight checklists in aviation, as described in the broader cryo-EM facility management literature.

The choice between CryoSPARC-Live and RELION depends on facility priorities. The live-processing platform offers a superior real-time web interface, tighter integration between the live and offline processing environments, and simpler configuration for non-expert users. RELION offers more granular control over processing parameters and is the preferred platform for users who require specific algorithmic options (e.g., Bayesian polishing, specific 3D classification strategies), in comparison to 3D classification strategies (e.g. 3DVAR or 3D-classification) present in CryoSPARC. Many facilities deploy both in parallel, using CryoSPARC-Live for real-time monitoring and RELION for final structure determination.

Remote access infrastructure deserves special emphasis. In our experience, the majority of user-reported problems during data collection stem from remote access failures rather than microscope or detector issues. Reliable, low-latency remote desktop access to the microscope and data collection workstations is not a luxury—it is the primary means by which most users interact with the facility. Investing in professionally configured remote desktop servers, testing connectivity from representative user locations, and documenting the connection procedure clearly will have a larger positive impact on user experience than many more expensive hardware investments.

Finally, data management policy must be established and communicated to users before facility operations begin. Users frequently underestimate the size of their datasets and overestimate the duration of their storage allocation. A clear written policy specifying raw data retention periods, user data quotas, archival procedures, and cost recovery for storage overruns protects both the facility and its users, and prevents the operational crises that arise when storage is exhausted mid-session.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Krios 5	ThermoFisher Scientific	https://www.thermofisher.com/us/en/home/electron-microscopy/products/transmission-electron-microscopes/krios-5-cryo-tem.html	Transmission Electron Microscope
Krios G4	ThermoFisher Scientific	https://www.thermofisher.com/us/en/home/electron-microscopy/products/transmission-electron-microscopes/krios-g4-cryo-tem.html	Transmission Electron Microscope
Glacios	ThermoFisher Scientific	https://www.thermofisher.com/us/en/home/electron-microscopy/products/transmission-electron-microscopes/glacios-cryo-tem.html	Transmission Electron Microscope
Glacios 2	ThermoFisher Scientific	https://www.thermofisher.com/us/en/home/electron-microscopy/products/transmission-electron-microscopes/glacios-2-cryo-tem.html	Transmission Electron Microscope
Glacios 3	ThermoFisher Scientific	https://www.thermofisher.com/ (newest generation; product page may vary by region)	Transmission Electron Microscope
Talos Arctica	ThermoFisher Scientific	https://www.thermofisher.com/us/en/home/electron-microscopy/products/transmission-electron-microscopes/talos-arctica.html	Transmission Electron Microscope
Talos L120C	ThermoFisher Scientific	https://www.thermofisher.com/us/en/home/electron-microscopy/products/transmission-electron-microscopes/talos-l120c.html	Transmission Electron Microscope
CRYO ARM 300 II (JEM-Z300FSC)	JEOL	https://www.jeol.com/products/scientific/tem/cryoarm300ii.php	Transmission Electron Microscope
CRYO ARM 200	JEOL	https://www.jeol.com/products/scientific/tem/cryoarm200.php	Transmission Electron Microscope
JEM-3200FSC	JEOL	https://www.jeol.com/products/scientific/tem/jem3200fsc.php	Transmission Electron Microscope
HF8300	Hitachi	https://www.hitachi-hightech.com/global/en/products/microscopes/tem/hf8300.html	Transmission Electron Microscope
HT7800	Hitachi	https://www.hitachi-hightech.com/global/en/products/microscopes/tem/ht7800.html	Transmission Electron Microscope
SU9000 (TEM/STEM)	Hitachi	https://www.hitachi-hightech.com/global/en/products/microscopes/sem-tem/su9000.html	Transmission Electron Microscope
Falcon 4i	Thermo Fisher Scientific	https://www.thermofisher.com/us/en/home/electron-microscopy/products/direct-electron-detectors/falcon-4i.html	Direct Electron Detector
Gatan K3	Gatan (AMETEK)	https://gatan.com/products/tem-cameras/k3-is-camera	Direct Electron Detector
Gatan K2	Gatan	https://gatan.com/products/tem-cameras/k2-camera	Direct Electron Detector
Apollo	Direct Electron	https://directelectron.com/apollo/	Direct Electron Detector
DE64	Direct Electron	https://directelectron.com/de-64/	Direct Electron Detector
RELION	MRC Laboratory of Molecular Biology (MRC-LMB), University of Cambridge	https://relion.readthedocs.io	SPA software package
cryoSPARC	Structura Biotechnology	https://cryosparc.com	SPA software package
EMAN2	Baylor College of Medicine (Developed by Steven Ludtke Lab)	https://blake.bcm.edu/emanwiki/EMAN2	SPA software package
cisTEM	National Research Council Canada (Developed by Tim Grant and colleagues)	https://cistem.org	SPA software package
SPHIRE	Max Planck Institute of Molecular Physiology and Baylor College of Medicine	https://sphire.mpg.de	SPA software package
cryoDRGN	Princeton University (Developed by Ellen Zhong Lab)	https://cryodrgn.cs.princeton.edu	SPA software package
Topaz	Harvard Medical School (Developed by Travis Bepler Lab)	https://github.com/tbepler/topaz	SPA software package
crYOLO	Max Planck Institute of Biophysics	https://cryolo.readthedocs.io	SPA software package
Warp	MRC Laboratory of Molecular Biology (Developed by Dmitry Tegunov)	http://www.warpem.com	cryo-ET and STA software package
ModelAngelo	MRC Laboratory of Molecular Biology	https://www.modelangelo.org	cryo-ET and STA software package
IMOD	University of Colorado Boulder	https://bio3d.colorado.edu/imod	cryo-ET and STA software package
Dynamo	Centro Nacional de Biotecnología (CNB-CSIC)	https://www.dynamo-em.org	cryo-ET and STA software package
emClarity	University of California, San Francisco	https://github.com/StochasticAnalytics/emClarity	cryo-ET and STA software package
RELION (ET module)	MRC Laboratory of Molecular Biology (MRC-LMB), University of Cambridge	https://relion.readthedocs.io	cryo-ET and STA software package
EMAN2 (SPT)	Baylor College of Medicine	https://blake.bcm.edu/emanwiki/EMAN2	cryo-ET and STA software package
Warp/M	MRC Laboratory of Molecular Biology	https://warpem.github.io	cryo-ET and STA software package
AreTomo	Stanford University (Developed by Fei Sun and Wah Chiu groups)	https://github.com/czimaginginstitute/AreTomo2	cryo-ET and STA software package
PEET	University of Colorado Boulder	https://bio3d.colorado.edu/PEET	cryo-ET and STA software package
PyTom	Max Planck Institute of Biochemistry	https://pytom.org	cryo-ET and STA software package
Tomo3D	Spanish National Centre for Biotechnology (CNB-CSIC)	https://sites.google.com/site/3demimageprocessing/tomo3d	cryo-ET and STA software package
crYOLO (3D)	Max Planck Institute of Biophysics	https://cryolo.readthedocs.io	cryo-ET and STA software package
AITom	Carnegie Mellon University	https://github.com/xulabs/aitom	cryo-ET and STA software package
DIALS	Diamond Light Source and collaborators	https://dials.github.io	microED software package
XDS	Max Planck Institute for Medical Research (Developed by Wolfgang Kabsch)	https://xds.mr.mpg.de	microED software package
MOSFLM	MRC Laboratory of Molecular Biology (MRC-LMB), University of Cambridge	https://www.mrc-lmb.cam.ac.uk/mosflm	microED software package
SHELX	University of Göttingen (Developed by George M. Sheldrick)	https://shelx.uni-goettingen.de	microED software package
Phenix	Lawrence Berkeley National Laboratory (PHENIX Consortium)	https://phenix-online.org	microED software package
Coot	MRC Laboratory of Molecular Biology (MRC-LMB), University of Cambridge	https://www2.mrc-lmb.cam.ac.uk/personal/pemsley/coot	microED software package
Linux (Ubuntu, CentOS/Rocky)	Ubuntu: Canonical Ltd.; CentOS: CentOS Project; Rocky Linux: Rocky Enterprise Software Foundation (RESF)	https://ubuntu.com / https://www.centos.org (legacy) /  https://rockylinux.org	Operational System
Windows	Microsoft	https://www.microsoft.com/windows	Operational System
ZFS	Oracle Corporation (originally Sun Microsystems); OpenZFS maintained by the OpenZFS community	https://openzfs.org	File System
NFS	Sun Microsystems (now maintained as an IETF standard)	https://www.rfc-editor.org/rfc/rfc8881	Protocol
SMB (CIFS)	Microsoft	https://learn.microsoft.com/windows-server/storage/file-server/file-server-smb-overview	Protocol
Microsoft Remote Desktop (RDP)	Microsoft	https://learn.microsoft.com/windows-server/remote/remote-desktop-services	Remote Access
NoMachine	NoMachine	https://www.nomachine.com	Remote Access
TeamViewer	TeamViewer SE	https://www.teamviewer.com	Remote Access
VirtualGL + TurboVNC	VirtualGL: The VirtualGL Project; TurboVNC: The TurboVNC Project	https://virtualgl.org and https://turbovnc.org	Remote Access
CUDA	NVIDIA	https://developer.nvidia.com/cuda-toolkit	Computer Framework