A cryo-EM structure is not a photograph of a molecule. It is a three-dimensional map reconstructed from very noisy two-dimensional projections of individual particles frozen in random orientations, followed by an atomic model that a person builds into that map with software. Two inferential steps separate the deposited coordinates from the specimen in the tube: the map is inferred from images, and the model is inferred from the map. Almost everything that goes wrong lives in those two steps rather than in the microscope.

This guide covers the single-particle workflow, why local and global resolution are different claims, how models are fitted and scored, and what a deposited peptide-bound G protein-coupled receptor structure does and does not establish.

The measurement, and its central artifact

Electrons scattered by a thin vitrified specimen form a phase-contrast image. Each particle image carries far more noise than signal at high spatial frequency, because electron dose must stay low enough that the specimen survives the exposure. Averaging across many particles recovers the signal — but only if the orientation assigned to each particle is right.

That conditional is where the field's characteristic artifact comes from. Chen and colleagues put it directly: because there is much more noise than signal at high resolution in the images, iterative refinement creates the possibility of noise reinforcement in the 3D map, which can give a false impression of the resolution attained [1]. Scheres and Chen showed the size of the effect on a simulated set of 20,212 GroEL particles processed with a single model determining the orientations of all particles: the reported resolution was 4.6 Å while the true resolution of the map was 7.8 Å, and they reported making convincing-looking figures of apparent side-chain density that in fact corresponded to overfitted noise [2]. Density that looks like a side chain is not evidence that a side chain was resolved.

The specimen decides more than the microscope does

Grid preparation — applying a few microliters to a holey support, blotting to a thin film, plunging into liquid ethane — is stated here as routine practice, not as a finding. What is not routine is what happens in the moments before vitrification. Glaeser and Han argued that the very high surface-to-volume ratio of a thin aqueous film creates an obvious risk that macromolecules adsorb to the air-water interface with a preferred orientation, or become partially or fully unfolded there, and that adsorption may occur even before the film is thinned [3]. That paper is explicitly an opinion proposing mitigations — surfactant monolayers, support films, affinity grids — not a measurement of how often this happens.

Preferred orientation is at least visible in its consequences. Tan and colleagues presented a strategy of collecting data with the stage tilted, together with a tool that quantifies directional resolution using 3D Fourier shell correlation volumes, and applied both to the influenza hemagglutinin trimer, which adopts a highly preferred orientation, and to ribosomal biogenesis intermediates, which adopt moderately preferred orientations [4]. The transferable point: a single resolution figure says nothing about whether a map is equally resolved in every direction.

The resolution number is a half-set statistic

The headline figure comes from a Fourier shell correlation between two reconstructions, each refined from one half of the particles. Scheres and Chen argued for refining the two models fully independently so the resulting curves are free from spurious correlations, against the then-current objection that withholding half the data would degrade orientation accuracy; their simulations at realistic signal-to-noise ratios indicated that orientation accuracy was not affected by using a model reconstructed from only half the particles [2]. Ramlaul and colleagues described independent refinement of the two sides of a split dataset as the most common of the strategies routinely used to prevent over-fitting [5].

The correlation threshold at which the curve is read — conventionally 0.143 — is a convention, stated here without citation. Its direction matters. Because correlation falls as spatial frequency rises, reading the crossing at 0.143 rather than the older 0.5 returns a crossing at higher frequency, and therefore a numerically smaller value in Å, from the same curve. Two maps both quoted at "3.0 Å" under different thresholds are not the same claim.

Half-set independence is necessary but not sufficient. Chen and colleagues described a software-independent test in which the Fourier components of the particle images beyond a chosen resolution are replaced either by components from an adjacent area of background or by phase-randomized versions, so the substituted noise carries the same spectral power distribution as the original data; comparing the FSC from real data with the FSC from substituted data measures the overfitting present [1]. Applied to two sets of images of beta-galactosidase across several processing packages, overfitting varied greatly depending on the procedure, and in the best case they reported no overfitting and a resolution of about 6 Å [1].

Local resolution versus the global figure

Kucukelbir and colleagues proposed a definition of local resolution based on local sinusoidal features, with no free parameters, and on evaluating single-particle reconstructions and subtomogram averages for four example datasets reported variable resolution across a 4- to 40-Å range [6]. That range spans the four maps they tested, which ran from near-atomic single-particle reconstructions to subtomogram averages; it is not the spread inside any one map. Their within-map estimates are narrower and specimen-specific: in one 80S ribosome reconstruction they fell between 4.5 and 5.5 Å in the 60S subunit and between 4.5 and 9 Å in the 40S subunit [6]. That kind of variation, not the 4- to 40-Å span, is what a global number summarizes away.

For a peptide-receptor complex the spread is structural rather than incidental. A transmembrane bundle held in detergent or a nanodisc, a G protein heterotrimer, an extracellular domain and a bound peptide are not equally ordered, and the regions a reader most wants to interpret — the peptide's amino-terminal residues inside the pocket, the extracellular loops — are often among the least well resolved in the map.

Independent half-sets do not repair this. Ramlaul and colleagues showed that over-fitting remains an issue within regions of low local signal-to-noise despite independent refinement of half datasets, and proposed applying a local signal-to-noise filter during refinement; they reported improved density and noted that in extreme cases prone to severe over-fitting, such as small membrane proteins, the filter can also lead to improvement of the final resolution [5]. A locally overfitted region can sit inside a globally honest number.

One further convention sits between map and figure, stated here without citation: maps are routinely sharpened before display by boosting high-frequency amplitudes, usually with a negative B-factor. Sharpening makes features easier to see. It does not add information, and past a point it renders noise with the same crispness as signal — which is why the sharpening applied, and whether an unsharpened map was deposited, belongs in the methods.

Heterogeneity: one map, many molecules

Every particle image is a different molecule. Where the population occupies more than one conformation, an unrestricted average is a map of nothing in particular. Two families of method respond. Focused classification and refinement restrict the analysis to a masked region so a mobile part can be treated separately; Barchet and colleagues, working on the human ribosome and on the cannabinoid G protein-coupled receptor, reported that these can improve local map resolution from an essentially disordered region into the 3-4 Å range and finally into the 2 Å range, without assigning either endpoint to a particular system; a systematic analysis with variable spherical masks showed that choosing an optimal mask size helps refinement reach high resolution, and across their four examples mask size and shape, and the inclusion of neighboring structural elements, mattered to the outcome [7]. Continuous methods model the motion instead of partitioning it: Punjani and Fleet's 3D variability analysis fits a linear subspace model of conformational change, and among their experimental results they reported resolving multiple flexible motions of alpha-helices in the sub-50 kDa transmembrane domain of a GPCR complex [8].

A published conformer count is a property of the classification that was run: it depends on the mask, the number of classes requested and the particle count, and nothing in the method establishes that states you did not ask for are absent.

Map-to-model fitting

The atomic model is fitted into the density in real space under restraints from known covalent geometry, supplemented at lower resolution by further prior information. Afonine and colleagues described real-space refinement in PHENIX, which uses a simplified target function and can draw on secondary-structure and rotamer-specific restraints and on restraints or constraints from internal molecular symmetry in addition to standard geometry restraints; they re-refined 385 cryo-EM-derived models available in the Protein Data Bank at resolutions of 6 Å or better and reported significant improvement both of the models and of the fit of those models to the target maps [9].

Read that result in the right direction. It says deposited coordinates were improvable by refinement, not that they were wrong, and not that the re-refined versions are correct. What it does establish is that map and model are separable objects that can disagree, and that their agreement must be measured rather than assumed.

Scoring a model against its map

Geometry statistics — bond lengths and angles, rotamer and Ramachandran outliers, steric clashes — report whether a model is chemically reasonable, not whether it matches the density: restraints can hold geometry tidy over a poor fit. Fit-to-map metrics are the other half of the assessment.

EMRinger assesses the precise fitting of an atomic model into the map through side-chain geometry during refinement; Barad and colleagues also reported that it shows how radiation damage alters scattering from negatively charged amino acids [10]. Q-scores characterize the resolvability of individual atoms. Pintilie and colleagues reported that Q-scores can be calculated for atoms in proteins, nucleic acids, water, ligands and other solvent, that averaged over entire models they correlate very well with the estimated resolution of cryo-EM maps for both protein and RNA, and that this use of them assumes the models are already well fitted to the map [11]. Comparing maps of the same proteins from different laboratories, they reported reproducibility of structural features from side chains down to water and ion atoms [11].

The most useful anchor is the field's own audit. In the 2019 EMDataResource model challenge, 13 participating teams produced models from four benchmark maps, three of which formed a resolution series from 1.8 to 3.1 Å. Lawson and colleagues reported relatively high accuracy and reproducibility across those teams and recommended the adoption of multiple scoring parameters, rather than any single one, to give a full and objective assessment of a model reflective of the observed map density [12]. Multiple, because each metric is blind somewhere.

What a deposited GPCR-peptide complex establishes

Consider a peptide-bound receptor-G protein complex deposited with a map, a model and a global resolution. What it establishes is bounded, and worth stating exactly. A set of particle images was consistent with a reconstruction. An atomic model was built into that reconstruction and can be scored against it. Where local resolvability supports it, particular peptide-receptor residue contacts represent the modeler's best interpretation of the density there.

What is not established is a mechanism, for three reasons. The complex was assembled and stabilized to be imaged — fiducials such as a nanobody or a single-chain antibody fragment bound to the G protein, receptor truncations, thermostabilizing substitutions and fusion partners are all preparation, not biology, and belong in the reader's assessment of the construct. The particles that entered the final reconstruction are a heavily selected subset of those on the grid. And the map is an average over many molecules, each vitrified in whatever state it happened to occupy.

Cary and colleagues stated the epistemic limit plainly in the course of a GLP-1 receptor study, writing that the relationship between receptor function and static structures carries inherent uncertainty [13]. Their cryo-EM analysis of a complex involving an exendin-4 analog, the GLP-1 receptor and the Gs heterotrimer revealed two receptor conformers with distinct modes of peptide-receptor engagement [13]. One sample yielded more than one answer to the question of how the bound peptide sits. A single deposited conformer would not have been wrong; it would have been incomplete, and nothing in the deposition would have said so.

What to check before quoting a structure

  • the global resolution together with the FSC threshold it was read at, and whether the half-sets were refined independently from the start
  • a local resolution map, a directional FSC, or both — not a single number standing alone
  • particle counts at each stage: picked, surviving 2D classification, and in the final reconstruction
  • which regions were masked, and whether the mask was applied before the FSC was calculated
  • how many classes were requested, how many were kept, and on what basis
  • fit-to-map scores reported alongside geometry statistics, not instead of them
  • the construct in full: fiducials, fusions, truncations, substitutions, and the membrane mimetic
  • whether the specific residues you intend to cite fall in a well-resolved part of the map

Convention, and what is not known

Several statements above are convention or routine practice rather than findings of the studies cited here, and are flagged where they appear: the grid-preparation sequence, the choice of the 0.143 FSC threshold, and sharpening with a negative B-factor before display.

What remains genuinely unresolved is more basic. There is no general method for establishing how the particle population on a grid relates to the population in solution before blotting; orientation bias and interfacial damage are inferred from their consequences rather than observed directly. Local resolution values are themselves estimates, produced by algorithms with their own assumptions, and are not calibrated against a ground truth for experimental data. There is no accepted numerical criterion for the local resolution at which a modeled rotamer, an ordered water or a peptide side-chain contact becomes reliable — practice varies, and the resolution at which a feature was modeled is frequently not reported per residue. And for peptide-receptor complexes specifically, the bound peptide is often among the weakest density in the map, which means the part of the structure most often quoted is the part least well constrained by the data.

None of this makes cryo-EM structures unreliable. It makes the difference between a map, a model and a mechanism worth keeping in view, and it makes the deposited map — not the picture in the figure — the thing to look at.