WebCalEM: a browser-based tool for routine and accurate pixel-size calibration in cryo-EM

Accurate determination of the image pixel size is critical for quantitative cryo-electron microscopy analyses, yet existing calibration methods remain under-utilized because installation barriers and workflow complexity discourage routine adoption. To fill in this gap, a web-based application, WebCalEM, was developed to transform specialized calibration procedures into an accessible routine practice. Micrographs of any specimen with a known crystalline lattice, such as gold or graphene oxide, are uploaded through a standard browser, processed entirely client-side and analyzed with real-time visualization and downloadable statistical outputs. The application is delivered as a single self-contained HTML file that runs in any modern web browser without server-side computation, a configuration that is well suited to isolated core-facility microscope workstations. Cross-standard consistency between gold and graphene oxide measurements across two microscopes and ten magnification settings yields a Bland–Altman bias of −0.005% of nominal with 95% limits of agreement of [−0.30%, +0.29%]. By delivering this workflow with no local installation, WebCalEM lowers the practical barrier to documented per-dataset magnification calibration in routine cryo-EM operation.

1. Introduction

Magnification calibration to accurately determine the pixel size of a cryo-electron microscopy (cryo-EM) image dataset is critical for downstream biological structure determination (Tiwari et al., 2020; Dickerson et al., 2024; Wasilewski et al., 2012). At the high resolutions now standard for modern cryo-EM, errors of even ∼1% in pixel size compromise CTF phase fitting (Rohou & Grigorieff, 2015; Heimowitz et al., 2020), leading to incorrect spatial frequency determination and loss of information at high frequencies, ultimately limiting the achievable resolution of the final structure (Dickerson et al., 2024). This issue is particularly important for amyloid reconstructions, where pixel-size errors cause the inferred cross-β spacing and helical rise to deviate from the expected 4.75 Å periodicity (Lövestam & Scheres, 2022; Scheres, 2020). More importantly, accurate calibration directly affects atomic modeling: even small scaling errors translate into incorrect atomic positions in the final structure (Tiwari et al., 2020; Dickerson et al., 2024; Wang et al., 2022). Accurate calibration is also required to place structures on a common absolute scale across instruments, time points and facilities, and to report resolution in Å without overstating or understating the true information content of a map (Penczek, 2010, 2020).

Three classes of methods are currently used for pixel-size calibration. External standards such as diffraction-grating replicas and MAG*I*CAL grids cover only narrow magnification ranges and can be biased by handling-induced distortion (McCaffrey & Baribeau, 1995; MAG*I*CAL Calibration Standard for TEM). Post hoc calibration by atomic-model fitting against reference proteins such as apoferritin requires a full reconstruction pipeline and yields a single value per dataset, precluding real-time use or per-region statistics (Tiwari et al., 2020; Wasilewski et al., 2012; Wang et al., 2022). Internal crystalline references – gold {111} at 2.35 Å on UltrAuFoil/HexAuFoil supports (Russo & Passmore, 2014; Naydenova et al., 2020; Naydenova & Russo, 2022) and graphene/graphene oxide reflections at 2.13 Å on standard supports (Han et al., 2020; Palovcak et al., 2018; Meyer et al., 2007; Sader et al., 2013) – provide directly measurable lattice spacings on every micrograph and span the magnification range relevant to single-particle and tomography cryo-EM.

While crystalline reference calibration is the most promising route, existing implementations (Dickerson et al., 2024) are still largely standalone or offline tools rather than integrated components of routine cryo-EM workflows. Users typically install software locally, transfer data between packages, compile results from multiple measurements and perform statistical analyses on the result separately. This fragmented workflow increases complexity, creating a substantial gap between available precise methods and their practical adoption. To close this practical gap, we present WebCalEM, a browser-based application that supports routine, SOP-compliant pixel-size calibration. Users upload micrographs, interactively select regions of interest and receive subpixel-precision pixel sizes from gold or graphene oxide reflections. Optional ellipse fitting is exposed as an interactive diagnostic for users investigating projector-lens anisotropy or specimen tilt. WebCalEM supports manual and automated workflows, aggregates results across measurements with mean and standard deviation, and runs entirely in a standard web browser through client-side computation, lowering the practical barrier to routine, documented calibration on any computer and operating system.

2. Materials and methods

2.1. Grid preparation and EM data acquisition

UltrAuFoil gold grids were used as test samples to evaluate crystalline internal calibration standards. To introduce an additional crystalline reference, grids were coated with graphene oxide (GO) using a Langmuir–Blodgett-style deposition method (Palovcak et al., 2018). Data were collected on two microscopes at The Huck Cryo-EM Facility (RRID SCR_0244456) at Penn State University: a Thermo Fisher 200 kV Talos Arctica G2 and a 300 kV Titan Krios G3. Micrographs were acquired at five nominal magnifications on each instrument (Arctica, 120 000–310 000×, 0.36–0.97 Å per pixel; Krios, 105 000–270 000×, 0.37–0.96 Å per pixel; defocus ranges −0.5 to −2.0 µm). Exposure doses were 50–65 e⁻ Å⁻².

2.2. Image analysis and pixel-size calculation

From 4000 × 4000 micrographs, regions of 1000 × 1000 to 2000 × 2000 pixels were transformed to reciprocal space by 2D fast Fourier transform (FFT). Calibration materials produce reflections at known spatial frequencies: graphene/GO at 1/d_ref = 0.469 Å⁻¹ (d_ref = 2.13 Å, in-plane hexagonal spacing) and gold at 1/d_ref = 0.426 Å⁻¹ (d_ref = 2.35 Å, {111}). The peak position r (in Fourier pixels) was taken from the maximum of the resulting 1D radial intensity profile. For an N × N image region with pixel size p (Å per pixel), a peak at radial distance r corresponds to spatial frequency r/(N × p) Å⁻¹. Equating this with 1/d_ref gives

where d_ref is the known lattice spacing (Å), d is the measured radial peak position (Fourier pixels) and N is the image dimension (pixels).

2.3. Correction for elliptical magnification distortion

Anisotropic magnification or specimen tilt displaces Fourier-space lattice peaks from a circular arrangement to an elliptic arrangement; the resulting axis ratio carries direct information about either projector-lens anisotropy or stage tilt. WebCalEM exposes the ellipse-fitting algorithm of Yu et al. (2016) as an interactive analysis option to recover the undistorted pixel size. Quantitative validation of the recovered axis ratio against an independently characterized tilt or anisotropy is beyond the scope of this short communication; users with such datasets are encouraged to verify the recovery on their own systems.

2.4. Cross-standard consistency and measurement precision

No ground-truth microscope pixel size was available for direct accuracy assessment, so cross-standard consistency was used as a proxy. The agreement between pixel sizes independently calibrated from gold and GO under identical microscope settings and the deviation of calibrated values from manufacturer-reported nominal values were both reported. Cross-standard consistency between the gold and GO measurements provides a check against sample-dependent systematic errors such as inaccurate reference lattice values, ring-fitting biases that vary with peak sharpness, or radial detector distortions that act differently on rings at different radii. The consistency was quantified using the Bland–Altman comparison framework (Bland & Altman, 1986; Giavarina, 2015), reporting the per-condition difference Δ = mean(GO) − mean(AU) with its 95% confidence interval and summarizing across the ten (microscope, magnification) conditions as bias (mean Δ) and 95% limits of agreement (LoA) [mean Δ ± 1.96 × SD(Δ)]. Given the modest number of condition-level pairs (n = 10), a small-sample-corrected multiplier was also computed to evaluate confidence bounds for the LoA (Carkeet, 2015). Precision was assessed by calibrating pixel sizes from 15 to 50 independent image regions at each magnification setting and computing the standard deviation of these measurements across different grid squares and imaging sessions for both calibration standards.

2.5. Web application implementation

WebCalEM is distributed in two interchangeable forms. The primary deliverable is a client-side written in JavaScript, delivered as a single self-contained HTML file. The app can be run directly in a browser from a local copy with no internet connection: a configuration well suited to isolated core-facility microscope workstations. A functionally equivalent server-side reference implementation, built in Python, is also distributed in the codebase and was used to benchmark the browser numerical core (The Shiny Development Team; Plotly, 2015). Detailed user interface implementation is discussed in Section 3.3.

3. Results

3.1. Gold and graphene oxide yield consistent calibrations across magnifications and instruments

A key question for any internal-standard approach is whether different crystalline materials produce consistent results under identical imaging conditions. We addressed this by independently calibrating pixel sizes from gold and GO regions of the same grid (15–50 per magnification, depending on the distribution of test specimens on collected micrographs) at five magnification settings on both Krios and Arctica microscopes (Fig. 1). Across all ten conditions, mean pixel sizes from the two standards differed by less than 0.3% of the nominal value, with absolute measurement uncertainties on the order of 10⁻³ Å per pixel for both materials. The largest discrepancy between standards occurred on the Arctica at the 0.970 Å per pixel nominal setting (150 000×), where GO measurements ran ∼0.26% below gold; the two means still agreed within their combined measurement uncertainties under this condition.

Figure 1 Centered pixel-size distributions by nominal magnification and specimen type. Box plots show the distribution of calibrated pixel sizes (after subtracting the nominal value; dashed line) for gold (blue) and GO (red) across five magnification settings on (a) Titan Krios G3 and (b) Talos Arctica G2 microscopes. (c) Bland–Altman comparison of GO and gold calibrations expressed as percentage of nominal. Bias is −0.005% of nominal; 95% LoA are [−0.30%, +0.29%].

A Bland–Altman analysis across the ten conditions yielded an essentially zero overall bias of −0.005% of nominal with 95% LoA of [−0.30%, +0.29%]. Applying a t-based small-sample correction (Carkeet, 2015) for N = 10 gave confidence bounds for the LoA of [−0.58%, +0.56%] of nominal, indicating subpercent cross-standard discrepancy between gold and GO across the tested conditions. On both instruments, calibrated pixel sizes deviated systematically from nominal. The Arctica read uniformly low: about −2.6% at the lowest magnifications, shrinking to −1% at the highest. The Krios was nonmonotonic: near-nominal at the lowest magnification, peaking at +1.3% mid-range and crossing to −0.6% at the highest. Both standards reproduced these instrument-specific deviations to within ∼0.3% under every condition, indicating they are not calibration-method artifacts and reinforcing the need for empirical, per-instrument calibration.

3.2. Retrospective calibration of archived datasets

Internal crystalline standards present on sample grids enable the recalibration of previously collected data without recollection. We demonstrated this using archived micrographs from two amyloid-β fibril datasets (unpublished work): an ex vivo amyloid dataset collected at 105 000× on the Titan Krios G3 and an in vitro amyloid data collected at 190 000× on the Talos Arctica G2. For these datasets, GO-coated grids had been chosen to increase sample concentration rather than for calibration purposes. By selecting background regions of these micrographs, we were able to retrospectively calibrate the pixel size. The recovered values showed excellent agreement with those obtained from dedicated test micrographs at the corresponding magnification settings (Fig. 2). These examples demonstrate the feasibility of retrospective calibration when archived micrographs contain sufficiently visible crystalline support reflections in analyzable background regions.

Figure 2 Retrospective per-region calibration on archived amyloid-β fibril datasets compared with calibrations from dedicated test micrographs at matching magnification settings. (a) Titan Krios G3 at 105 000× (nominal 0.96 Å per pixel): new GO calibration versus the ex vivo amyloid dataset. (b) Talos Arctica G2 at 190 000× (nominal 0.76 Å per pixel): new GO calibration versus the in vitro amyloid dataset. p-values are from independent-samples t-tests on the centered pixel-size distributions.

3.3. Web app implementation

WebCalEM provides an integrated interface for pixel-size calibration organized into three functional panels (Fig. 3). The image panel (blue) allows users to upload cryo-EM micrographs and interactively select regions of interest for analysis. The FFT analysis panel (yellow) offers both the 1D radial profile and the 2D power spectrum. Peak detection is performed automatically in both views: in the 1D profile the software identifies the radial maximum near the expected lattice frequency, while in the 2D spectrum lattice peak positions are detected and an ellipse is fitted automatically for distortion estimation. Users can manually adjust any of these results if needed. Radial and angular sampling frequencies are also adjustable, and a focused heatmap around the selected peak provides visual confirmation of the localization. The results panel (green) displays the calibrated pixel size for the current measurement and maintains a running table of all measurements within the session, automatically computing the group statistics for each nominal magnification. An interactive scatter plot visualizes the distribution of measurements centered by their nominal values, enabling users to assess calibration consistency immediately. Saved measurements can be reloaded and edited by clicking a row in the table, which restores the original image and region for refinement before being written back to the table. Results export contains a measurements CSV file, a grouped-statistics CSV file per nominal magnification and the statistics scatter plot as a PNG file, suitable for direct integration into downstream processing pipelines. An auto-sample mode is also provided for batch use: the user selects a region count and region size, points the app at one or more images, and the app draws regions automatically and records the per-region pixel size for each. A detailed walkthrough is released along with the codebase.

Figure 3 WebCalEM app user interface.

4. Discussion

Although both standards yielded statistically equivalent calibrations across the conditions that we tested, the two materials differ in practical respects that affect routine use. The GO hexagonal lattice produces a single, well isolated 2.13 Å reflection that is unambiguous to pick in the 1D radial profile, whereas UltrAuFoil grids expose two closely spaced gold reflections at 2.35 and 2.48 Å, separated by only ∼5% in reciprocal space. This is wide enough to resolve the two rings most of the time, but the two bands are easily confused at lower nominal magnifications, where the rings sit closer together in Fourier pixels. GO peak picking was therefore more reliable than gold peak picking across the magnification range we tested, an observation that is consistent with the slightly tighter per-condition GO uncertainties that we obtained (median ∼0.31% of nominal for GO versus ∼0.46% for gold; Fig. 1). The two materials also differ in the nature of their elliptical distortion. When gold is used, the apparent ellipticity of the 2D power spectrum is dominated by global effects – projector-lens magnification anisotropy, stage tilt and gradients in ice thickness – that are largely invariant across imaging regions on a single grid, so the ellipse fit can be evaluated once for the entire dataset and applied uniformly. GO flakes, however, are not always flat on the support film; physical out-of-plane tilt of the flake introduces a local ellipticity that depends on the specific region selected, so with GO the ellipse fit needs to be re-evaluated per region rather than per dataset. WebCalEM exposes interactive per-region ellipse fitting for exactly this reason.

In general, WebCalEM reaches the precision band of established command-line lattice-calibration tools (e.g. <0.5% of nominal for magCalEM; Dickerson et al., 2024) while removing the installation and command-line barriers that have historically kept routine calibration out of standard cryo-EM workflows. The auto-sample mode introduced here already extends the workflow to batch calibration of pre-acquired datasets; the natural next step is to aggregate per-instrument calibrations into a shared community database, turning pixel-size calibration from an occasional procedure into a routine, field-wide quality-assurance practice.

5. Conclusion

We have developed WebCalEM, a browser-based tool that makes lattice-based pixel-size calibration a routine step in the cryo-EM workflow. Delivered as a single self-contained HTML file that runs entirely client-side, with no server, installation or internet connection required after first load, it is deployable in any computational setting at a standard facility. The application supports multiple lattice standards, manual and auto-sampled per-region measurements and automated statistical aggregation, while matching the precision of established command-line calibration tools. Widespread adoption should improve the accuracy and cross-comparability of cryo-EM structures across the field.

6. Code and software availability

WebCalEM is free and open-source software developed by the Jiang Lab at Penn State University, with the source code on GitHub at https://github.com/jianglab/WebCalEM. The client-side single-file web app is hosted at https://jianglab.github.io/WebCalEM/. A detailed walkthrough is released as part of the codebase. No new structures, models or maps were deposited as part of this work; PDB, EMDB and EMPIAR deposition are therefore not applicable to this Methods Communication. Source micrographs supporting the cross-standard analysis are available from the corresponding author upon request.

Footnotes

‡These authors made equal contributions.