Improved cryo-EM reconstruction of sub-50 kDa complexes using 2D template matching
- RNA Therapeutics Institute, University of Massachusetts Chan Medical School, United States
- Howard Hughes Medical Institute, United States
- Department of Biochemistry, University of Wisconsin–Madison, United States
- Morgridge Institute for Research, United States
Figures
Figure 1 with 4 supplements
3D reconstruction of protein kinase using 2D template matching (2DTM)-derived particle stack.
(a) From left to right: the single-particle reconstruction (EMDB:0409) (Herzik et al., 2019), the 2DTM template, and the 2DTM-derived reconstruction of the protein kinase. (b) Angular distribution plot and Fourier shell correlation (FSC) curve calculated using cisTEM. Note that these FSC curves are not gold-standard FSCs, as the reconstruction uses orientations determined by 2D template matching rather than independent half-set refinement. denotes the uncorrected FSC computed within a generous mask, which crosses the 0.143 threshold at 3.0 Å, and ‘particle FSC’ denotes the solvent-corrected FSC obtained using the mask-volume correction factor as described in the cisTEM/Frealign framework (Grant et al., 2018). (c) Densities at the ATP-binding site and deleted residues 222–227. The X-ray structural model (PDBID: 1ATP) (Zheng et al., 1993) is shown with atoms retained in the template colored gray and deleted atoms colored blue. The alpha-helical model for residues 222–227 is rendered transparent to better show the recovery of the helical turn in the reconstruction. The average Q-scores between the 2DTM reconstruction and the X-ray model were calculated using the MapQ command line tool (Pintilie et al., 2020; Pintilie et al., 2025). Q-scores for ATP, Mn2+, and deleted residues 222–227 were 0.60, 0.48/0.73, 0.63, 0.51, 0.57, 0.54, 0.61, and 0.53, respectively.
-
Figure 1—source data 1
Grouped occupancy refinement of omitted and control residues.
- https://cdn.elifesciences.org/articles/109790/elife-109790-fig1-data1-v1.xlsx
Figure 1—figure supplement 1
Image statistics of the untilted micrographs in EMPIAR-10252.
(a) Contrast transfer function (CTF) fitting scores for 2488 untilted images, calculated using ctffind5. Images with scores above 0.2 or below 0.05 were excluded from 2D template matching (2DTM) searches. (b) Mean defocus values of the 2314 images retained for 2DTM. (c) Sample thickness estimates from ctffind5, with a median thickness of 367 Å. Negative thickness estimates were excluded from the histogram. (d) Particle counts per thickness bin, based on 17,274 particles extracted from the 2314 images using extract-particles. (e) Particle counts per thickness bin after particle selection using filter-particles, showing the final stack of 7353 particles.
Figure 1—figure supplement 2
Examples of micrographs excluded from 2D template matching (2DTM) search.
(a) Very low contrast and ice contamination. (b) Extremely low contrast, likely drift or astigmatism. (c) Particle aggregation or contamination. (d) Cross-grating calibration grid.
Figure 1—figure supplement 3
Effect of template resolution on reconstruction quality.
Comparison of 3D reconstructions using particles aligned at bin2x (1.117 Å/pixel) versus bin4x (2.234 Å/pixel) resolution. For the bin4x experiment, 2D template matching (2DTM) was performed on bin4x images, and the detected particle coordinates were used to extract particles from the bin2x images for reconstruction. Densities of the full map (a), the ATP-binding site (b), and deleted residues 222-227 (c) are shown. The bin4x-aligned reconstruction shows loss of ATP density and degraded backbone density for residues 222–227, demonstrating that high-resolution signal in the template is critical for accurate particle alignment and recovery of omitted densities. Map contour .
Figure 1—figure supplement 4
Difference map between the template (gray, map contour ) and the reconstruction shown in Figure 1.
The difference map was generated using the diffmap.exe program (Grigorieff, 2021) with a protein soft mask. Positive (pink) and negative (light blue) difference densities are shown for the full map (a), the ATP-binding site (b), and residues 222-227 (c) at map contour . Positive densities indicate regions present in the reconstruction but absent from the template, with coherent features at the ATP-binding site and residues 222–227, which were deleted from the template. Residual noisy densities elsewhere reflect limitations of the forward model.
Figure 2
2D template matching (2DTM)-based reconstruction of the ATP-binding pocket.
Each row shows, from left to right: (1) the atomic model used to generate the template, with residues within a specified radius from ATP deleted (highlighted in blue); (2) the 2DTM reconstruction fitted with the full model, with labeled residues corresponding to those circled in red in the Q-score plot; (3) the recovered ATP density fitted with the full model; and (4) backbone Q-scores of all residues. Omitted residues are shown in blue (circles for chain E, diamonds for IP20); kept residues in gray. ATP and Mn2+ are shown in orange. Red circles indicate the omitted residues with the lowest backbone Q-scores, corresponding to the labeled residues in column 2. Q-scores were calculated using the MapQ command line tool (Pintilie et al., 2020; Pintilie et al., 2025). (a) Residues within 3 Å of ATP were deleted. (b) IP20 and residues within 3 Å of ATP were deleted. (c) Residues within 5.5 Å of ATP were deleted. Map contour .
Figure 3 with 1 supplement
RELION processing of 2D template matching (2DTM)-derived particle stack.
(a) RELION 3D classification without angular refinement of the 2DTM-derived particle stack from Figure 1, using five classes (Tau_fudge = 4, resolution limit 7 Å). Classes are colored individually and labeled with particle count, percentage of the dataset, and estimated resolution. Classes 1–4 were merged for 3D reconstruction; Class 5 (156 particles, ∼2%, ∼20.5 Å) was excluded. Map contour . (b) Comparison of 3D reconstructions using 2DTM orientations directly (top row, green) and RELION auto-refine (bottom row, red). From left to right: full map, zoomed view of the ATP-binding pocket and deleted residues 222–227 with the atomic model (blue) overlaid (map contour ), orientation distribution, and Fourier shell correlation (FSC) curves. The 2DTM orientation reconstruction reaches 3.1 Å and the RELION auto-refined reconstruction 3.7 Å at the FSC = 0.143 threshold. In both maps, densities for ATP and the backbone of the deleted residues were recovered, but the 2DTM-derived orientations yielded sharper and more continuous density in the omitted regions. (c) RELION auto-refinement resolution as a function of the initial low-pass filter (--ini_high). The final row includes all five classes (i.e. keeping Class 5), yielding 3.7 Å, identical to the reconstruction without Class 5.
Figure 3—figure supplement 1
Fourier shell correlation (FSC) comparison between cisTEM and RELION on the same half-maps.
Both curves were computed from the same cisTEM half-maps (Figure 1 reconstruction). The cisTEM particle FSC (blue) uses a spherical mask with an analytical solvent-fraction correction, while the RELION masked FSC (orange) uses a tight 3D protein mask applied directly to the half-maps. Both cross the FSC = 0.143 threshold at ∼2.6–2.7 Å, confirming that the two packages give consistent resolution estimates when applied to the same data.
Figure 4 with 1 supplement
2D template matching (2DTM) reconstructions with varying particle selection parameters.
The first column shows the template model with the X-ray structure overlaid (deleted atoms in blue). The remaining columns show reconstructions using different selection thresholds (map contour ). The top and bottom rows show zoomed views of the ATP-binding pocket and deleted residues 222–227, respectively. Each column header lists the number of particles and the template bias , defined as , where and are reconstructions using orientations and particles derived from independent 2DTM searches with the full and omit templates, respectively, and the sum is restricted to the omission mask derived from the difference between the two templates. means the omit reconstruction recovers the same density (no bias from the template), while means all density disappears without the template. Q-scores for ATP and residues 222–227 are reported in Figure 4—source data 1. Template bias was calculated using the 2DTM_postprocess_tool Python package, adapted from Lucas et al., 2023. The omission mask used to compute is shown in Figure 4—figure supplement 1.
-
Figure 4—source data 1
Q-scores for omitted regions across particle selection conditions.
- https://cdn.elifesciences.org/articles/109790/elife-109790-fig4-data1-v1.xlsx
Figure 4—figure supplement 1
Central section of the omission mask used for template bias () calculation.
The mask was derived from the difference between the full and omit templates (threshold = average of top-100 voxels/5; 959 masked voxels). Left: full template. Center: omit template (residues 222–227, ATP, Mn, and H2O deleted). Right: binary omission mask, with the two omitted regions (ATP+Mn and residues 222–227) annotated. The mask isolates the omitted region for the calculation shown in Figure 4. Voxel size is 1.117 Å.
Figure 5 with 3 supplements
Site-specific composite omit map assembled from 36 partial-deletion reconstructions.
Each of the 36 omit templates deletes ∼10 non-overlapping residues. For each template, local density was extracted within 3 Å of the omitted atoms, with neighboring residues (i±1) excluded using a 2 Å cutoff on backbone Cɑ atoms. Local density patches were assigned uniquely and merged such that each voxel is contributed only by the reconstruction in which the corresponding region was omitted from the template. (a) Atomic model of the protein kinase (PDBID: 1ATP). (b) Left: composite omit map. Right: zoomed view of the ATP-binding pocket. (c) Left: composite omit map low-pass filtered to 4 Å. Right: zoomed view of the ATP-binding pocket. The composite demonstrates that density can be recovered at distributed locations across the protein, including peripheral and surface-exposed regions, although the quality of recovery varies across sites. Map contour .
-
Figure 5—source data 1
Radial profile of the recovered density around omitted residues, averaged across the 36 omit reconstructions.
- https://cdn.elifesciences.org/articles/109790/elife-109790-fig5-data1-v1.xlsx
Figure 5—figure supplement 1
Map–model Fourier shell correlation (FSC) of the composite omit map (red) and Figure 1 reconstruction (blue), each computed against a density map simulated from the full 1ATP model using cisTEM calculate_fsc with a protein envelope mask.
Crossings at FSC = 0.5 (dashed) and FSC = 0.143 (dotted) are labeled. The composite omit map crosses FSC = 0.5 at 3.5 Å and FSC = 0.143 at 3.0 Å; Figure 1 reconstruction crosses at 3.0 Å and 2.4 Å, respectively. Because each voxel of the composite map derives from a reconstruction in which the corresponding residues were absent from the template, the high-resolution agreement reflects genuine signal recovery rather than reproduction of the template. The composite curve falls below Figure 1 curve because its FSC is dominated by locally omitted density, whereas Figure 1 curve is dominated by template-included density. The composite curve’s negative correlation at low resolution is examined in Figure 5—source data 1.
Figure 5—figure supplement 2
Uncorrected half-map Fourier shell correlation (FSC) of the 36 individual omit reconstructions (thin gray), with the median (blue) and interquartile range (shaded).
Figure 5—figure supplement 3
Phase-randomization control for the composite half-map Fourier shell correlation (FSC).
Shown are the observed FSC between the two composite half-maps (blue), the randomized null (red dashed; the parent half-maps’ phases randomized independently beyond 4 Å and the maps reassembled with the same voxel-assignment masks), and the phase-randomization-corrected FSC (orange). The randomized null remains high beyond the randomization cutoff, indicating that the shared voxel-assignment masks introduce substantial support-induced correlation. Therefore, the observed composite half-map FSC cannot be interpreted as a gold-standard resolution estimate.
Figure 6 with 1 supplement
2D template matching (2DTM) reconstruction using the AlphaFold model as the search template.
(a) Structural comparison between the X-ray model (PDBID: 1ATP, gray) and the AlphaFold3 predicted model (tan) (Abramson et al., 2024). Residues within 3 Å of ATP were deleted from both templates and not shown. RMSD between the undeleted structures is 0.45 Å. (b) 2DTM-derived maps using the AlphaFold3 template (tan) and the X-ray template (gray). (c) Reconstruction at the ATP-binding site using the X-ray model (top) and the AlphaFold3 model (bottom) as the template (map contour ). (d) Reconstruction at residue 18 (ARG) on IP20 using the X-ray model (top) and the AlphaFold3 model (bottom) as the template (map contour ).
Figure 6—figure supplement 1
Map–model Fourier shell correlation (FSC) for X-ray and AlphaFold3 templates in Figure 6.
Map–model FSC was computed between each atomic model-derived template and the corresponding 2D template matching (2DTM) reconstruction using cisTEM calculate_fsc, with a protein envelope mask. Left: X-ray-derived template versus X-ray-derived 2DTM reconstruction. Right: AlphaFold3-derived template versus AlphaFold3-derived 2DTM reconstruction. Each panel shows the uncorrected FSC (blue) and the solvent-corrected particle FSC (green); the uncorrected FSC crosses FSC = 0.143 at ∼2.4 Å and FSC = 0.5 at 3.0 Å (X-ray) and 3.1 Å (AlphaFold3).
Figure 7
Theoretical lower molecular weight limit.
A constrained search restricts the and dimensions to a 5-by-5 pixel window. At the minimal molecular weight, the signal-to-noise ratio (SNR) calculated from alignment noise and phase contrast are equal.
Tables
Table 1
Single-particle cryo-electron microscopy (cryo-EM) lower molecular weight limits under different assumptions.
A constrained search restricts the and dimensions to a 5-by-5 pixel window.
| Parameters | Values | Full search | Constrained search |
|---|---|---|---|
| Acceleration voltage | 300 keV | 14.8 | 11.8 |
| Defocus | 1500 nm | ||
| Total exposure | 45 e−/Å2 (M = 45) | ||
| Resolution limits | 2–120 Å | ||
| Perfect beam | – | ||
| +Inelastic scattering | Ice, 30 nm | 16.3 | 13.0 |
| +Particle envelope | Ice thickness = particle diameter | 15.0 | 12.0 |
| +Amplitude contrast | 0.1 | 14.8 | 11.8 |
| +Phase plate | 90° | 9.2 | 7.4 |
| +Liquid helium | 1.8× | 7.1 | 5.7 |
Key resources table
| Reagent type (species) or resource | Designation | Source or reference | Identifiers | Additional information |
|---|---|---|---|---|
| Software, algorithm | cisTEM | Grant et al., 2018; Himes and Grigorieff, 2021; Zhang et al., 2025 | RRID:SCR_016502 | https://cistem.org/ |
| Software, algorithm | 2DTM_postprocess_tool | This paper | https://github.com/kekexinz/2DTM_postprocess_tool, copy archived at Zhang, 2026; also available in the cisTEM repository (Grant, 2025) |
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Improved cryo-EM reconstruction of sub-50 kDa complexes using 2D template matching
eLife 15:RP109790.
https://doi.org/10.7554/eLife.109790.4
