From vitrified single particles to near-atomic 3D density maps via electron cryo-microscopy
Cryo-electron microscopy begins not with a crystal, but with a droplet. A few microliters of purified protein in aqueous buffer are applied to a holey carbon grid, blotted to a thin film, and plunged into liquid ethane cooled by liquid nitrogen. Cooling rates exceeding 10,000 K/second prevent water molecules from ordering into crystalline ice, instead trapping them — and the protein — in an amorphous, glass-like solid that preserves the native solution structure of every particle.
Water frozen slowly forms hexagonal crystalline ice — its regular lattice would diffract electrons strongly and physically distort or destroy delicate protein structure as ice crystals grow and exclude solutes. Jacques Dubochet's breakthrough in the early 1980s was showing that sufficiently fast cooling denies water molecules the time needed to nucleate a crystal lattice. Instead they freeze into vitreous ice: an amorphous solid with liquid-like short-range order but no long-range periodicity, optically and physically similar to a glass.
Because vitreous ice has no crystalline planes, it does not generate strong diffraction spots and only contributes weak, smoothly varying background scattering to electron micrographs. This lets the much weaker scattering signal from individual protein molecules be preserved and later extracted computationally, rather than being drowned out by the ice itself.
Cooling rates of roughly 10,000 to 100,000 K per second are required to vitrify a thin aqueous film — the sample is fully frozen in well under a millisecond, fast enough that no crystalline ice nucleus has time to form.
Standard cryo-EM grids are 3 mm copper or gold meshes coated with a holey carbon or gold film (e.g. Quantifoil, C-flat), perforated with regularly spaced circular holes 1–2 μm in diameter. A robotic plunge-freezer (Vitrobot, EM GP2, Leica GP) applies 3–4 μL of sample, blots away excess liquid with filter paper for 1–5 seconds leaving only a thin film tens of nanometers thick spanning each hole, and immediately plunges the grid into liquid ethane held near its freezing point by a liquid-nitrogen jacket.
Ethane, not liquid nitrogen directly, is used as the cryogen because nitrogen at its boiling point forms an insulating vapor layer around the grid (the Leidenfrost effect), slowing heat transfer. Liquid ethane has much higher thermal conductivity at the relevant temperature and achieves the extreme cooling rates vitrification requires.
Getting the ice thickness right is a delicate balance: too thick, and particles stack in multiple layers with increased background noise and beam-induced motion; too thin, and particles are forced into a narrow range of preferred orientations against the air-water interface, or are excluded from the hole altogether. Optimal ice is only slightly thicker than the particle's largest dimension.
Traditional X-ray crystallography requires coaxing a protein into a well-ordered 3D crystal lattice — a slow, often impossible process for flexible, membrane-embedded, or heterogeneous protein complexes. Vitrification sidesteps this bottleneck entirely: the protein never needs to crystallize. Instead, thousands to millions of individual copies are frozen in random orientations, each one a separate "single particle" snapshot of the same molecule viewed from a different angle.
This is the conceptual foundation of the entire single-particle reconstruction pipeline: rather than solving a phase problem for periodic diffraction from a crystal, cryo-EM instead solves an orientation-recovery problem across a huge ensemble of individually noisy 2D images of randomly oriented copies of the same 3D object, later combined computationally into a single high-resolution 3D density map.
The vitrified grid is transferred under liquid-nitrogen temperature into a transmission electron microscope. A coherent 200–300 kV electron beam is passed through the ice at extremely low dose to avoid destroying the very structure it is trying to image, and a direct electron detector records thousands of noisy movie frames that are later aligned and summed into micrographs.
Electrons are ionizing radiation. A dose sufficient to see individual atoms would also break covalent bonds, radiolyze the sample, and destroy the very structure being imaged — long before a clear picture could be recorded. Cryo-EM data collection is therefore a compromise dictated by radiation damage: the electron dose per micrograph is kept to roughly 40–60 electrons per square Ångström, orders of magnitude below what would be needed for a single image to show atomic detail directly.
At this low dose, any single particle image is buried in shot noise — its signal-to-noise ratio (SNR) is often well below 0.1. No single image is directly interpretable. The entire single-particle approach exists to solve this problem: average the signal from many thousands of independent noisy observations of (nominally) the same structure to recover a map whose effective SNR is far higher than any individual exposure.
A typical single cryo-EM particle image has a signal-to-noise ratio far below 1 — the underlying protein signal is often outweighed several-fold by shot noise. High-resolution 3D reconstruction only becomes possible by averaging tens to hundreds of thousands of these images together.
Before roughly 2012, cryo-EM images were recorded on photographic film or CCD cameras via a scintillator, both of which degraded resolution and made low-dose imaging even noisier. The introduction of direct electron detectors (DEDs) — CMOS sensors that register electron impacts directly, without a lossy scintillation conversion step — triggered what the field calls the "resolution revolution," rapidly pushing achievable resolutions from ~10 Å blobs to sub-3 Å, near-atomic detail within just a few years.
DEDs also enable dose fractionation: instead of a single static exposure, the total dose is split into a "movie" of 20–50 short frames captured over a few seconds. Because the ice (and the particles embedded in it) physically shift and deform slightly under electron bombardment — beam-induced motion — each frame can be aligned to correct for this drift before summing, using motion-correction software such as MotionCor2 or RELION's own implementation. This single innovation alone recovered several Ångströms of resolution that beam-induced blurring had been silently discarding.
Electron micrographs are formed under phase-contrast imaging conditions with deliberate defocus, meaning the recorded image is not a simple projection of the specimen but is convolved with an oscillating contrast transfer function (CTF) that depends on defocus, spherical aberration, and electron wavelength. The CTF modulates and periodically inverts the contrast of different spatial frequencies (resolution ranges), and even passes zero contrast — no signal at all — at specific "zero crossing" frequencies.
Before particles can be combined into a 3D reconstruction, the CTF for every micrograph must be estimated (tools such as CTFFIND4 or Gctf fit Thon ring patterns in the power spectrum) and later computationally corrected, or "phase flipped," during reconstruction. Uncorrected CTF effects are one of the most common causes of blurry, uninterpretable cryo-EM maps.
Raw micrographs are dense fields of noisy particle silhouettes embedded in a background of vitreous ice and carbon support. Before 3D reconstruction can begin, every individual particle image must be located, extracted into a small "box," and sorted by apparent orientation — a process that both cleans up the dataset and gives the first visual confirmation that the sample looks like the expected molecule.
Particle picking identifies the pixel coordinates of every particle projection scattered across a micrograph. Classical approaches use template matching or Laplacian-of-Gaussian blob detection to score candidate locations against reference shapes (often generated from an initial, low-resolution 3D model). More recent tools such as Topaz and crYOLO use trained convolutional neural networks that learn to distinguish particles from ice contamination, carbon edges, and aggregates with far higher sensitivity and fewer false positives than classical correlation-based pickers, particularly for small or low-contrast particles.
Automated picking on a large dataset routinely yields hundreds of thousands to over a million candidate particle boxes, many of which are junk: ice contamination, overlapping particles, carbon film edges, or aggregates. This raw pool must be aggressively cleaned before it is useful.
2D classification (implemented in RELION, cryoSPARC, and similar packages) groups extracted particle images into a user-specified number of classes — typically 50 to 200 — based on similarity in apparent 2D projection. Particles that happen to have landed in the ice with similar orientation relative to the beam will, apart from noise, look alike; the algorithm iteratively aligns and averages particles within each class, using an expectation-maximization or Bayesian framework to refine both class assignment and in-plane alignment simultaneously.
Because averaging N independent noisy images improves signal-to-noise roughly as the square root of N, a class average built from a few thousand particles can go from an unrecognizable blob in any single raw image to a crisp, clearly interpretable 2D projection showing secondary structure features, domain boundaries, or even helical pitch. 2D classification therefore serves two purposes at once: it is a powerful particle-cleanup filter (junk particles form blurry, discardable classes) and the first genuine visual evidence that the sample is what it is expected to be.
Averaging independent noisy observations improves signal-to-noise proportional to the square root of the number of particles combined — a 2D class average built from 5,000 raw particle images has roughly 70 times better SNR than any single constituent image.
Real samples are rarely perfectly homogeneous. Multiple 2D classification rounds are typically used to progressively discard particles belonging to obviously broken, denatured, aggregated, or otherwise uninterpretable classes, while retaining particles that fall into sharp, structurally sensible class averages. Datasets are commonly reduced from an initial million-plus raw picks down to a curated set of a few hundred thousand "good" particles carried forward into 3D processing — the quality of this curation step has an outsized effect on the achievable final resolution.
Every clean 2D particle image is a projection of the same 3D molecule, but at an unknown viewing angle. Recovering that angle for every particle — and then mathematically combining all the resulting projections — is what actually builds the 3D density map. This is accomplished in Fourier space using the projection-slice (central slice) theorem, refined iteratively via Bayesian or projection-matching algorithms as implemented in RELION and cryoSPARC.
The mathematical foundation of 3D reconstruction from 2D projections is the central slice (projection-slice) theorem: the 2D Fourier transform of a projection image of a 3D object is identical to a central 2D slice through the 3D Fourier transform of that object, oriented perpendicular to the projection (viewing) direction.
This means that if the orientation of every particle image can be determined, each particle's Fourier transform contributes one slice through the full 3D Fourier volume of the molecule. Enough slices at enough different, well-distributed angles — ideally hundreds of thousands of them — fill out 3D Fourier space densely enough that an inverse Fourier transform recovers the full 3D density map. Insufficient angular coverage (preferred orientation) instead leaves gaps in Fourier space and produces a directionally distorted, anisotropic map.
Before any orientations are known, an initial 3D model is typically built ab initio from the cleaned 2D class averages themselves — using common-lines approaches (each pair of 2D projections theoretically shares a common line of intersection in Fourier space, constraining relative angle) or stochastic gradient descent methods that build a rough, low-resolution starting volume, usually to ~15–30 Å resolution.
That crude initial model is then used as a reference for projection matching: synthetic 2D projections are generated from the current 3D model at a dense grid of candidate orientations, and each real particle image is compared against all candidates to find its best-matching angle and in-plane shift. RELION formalizes this within a Bayesian (maximum a posteriori) framework, treating orientation and class assignment as probability distributions rather than single hard choices, which is markedly more robust to noise than classical maximum-likelihood or cross-correlation matching. cryoSPARC instead uses a GPU-accelerated stochastic gradient descent / branch-and-bound approach that is dramatically faster for large datasets while achieving comparable accuracy.
Orientation determination and 3D reconstruction are then repeated iteratively: current best orientations refine the 3D model, and the improved 3D model in turn refines the orientation estimates, converging over several to dozens of iterations toward self-consistent angles and a sharper volume.
Because every particle contributes only a single 2D projection at an unknown angle, gold-standard cryo-EM refinement independently reconstructs two half-maps from randomly split halves of the particle set — their Fourier Shell Correlation (FSC) at the 0.143 threshold is the field's standard, bias-resistant resolution estimate.
Many biological samples are not one single rigid structure but an ensemble of related conformational or compositional states — a ribosome mid-translocation, a receptor with a flexible extracellular domain, a complex with sub-stoichiometric ligand occupancy. 3D classification extends the same probabilistic framework used in orientation search to also sort particles into discrete structural classes, while heterogeneous refinement and newer neural approaches (3D variability analysis, cryoDRGN) can even model continuous conformational motion rather than discrete states.
Properly separating heterogeneous particles before final refinement is often the single biggest determinant of whether a reconstruction reaches near-atomic resolution or stalls at a blurry intermediate resolution — averaging together particles in genuinely different conformations smears the resulting density just as surely as random noise does.
The final stage of the pipeline iteratively sharpens the 3D reconstruction, corrects remaining optical aberrations, and validates resolution using rigorous statistical criteria — often reaching 2–4 Ångströms, sufficient to resolve individual amino acid side chains. An atomic model, built de novo or predicted by tools like AlphaFold, is then fitted into the density to produce the final interpretable structure.
Once an initial converged 3D reconstruction is available, several further rounds of refinement squeeze out remaining resolution. Per-particle CTF refinement re-estimates defocus and higher-order optical aberrations (beam tilt, astigmatism, spherical aberration) for each particle individually rather than per-micrograph, correcting subtle optical distortions that blur high-resolution features. Bayesian polishing (RELION) or equivalent motion-refinement tools re-estimate each particle's trajectory across the original movie frames using the now much-better 3D model as a reference, squeezing additional signal out of frames that were previously only coarsely corrected.
Masking and post-processing (dividing out the modulation transfer function, applying negative B-factor sharpening) further correct the amplitude falloff of high-resolution Fourier components, which is otherwise attenuated by detector response and beam-induced motion — sharpening transforms an initially "mushy" density into one with crisp, well-separated secondary structure.
Cryo-EM resolution is not a single number that can be read off a map by eye; it is a statistical estimate of how much the map can be trusted at each spatial frequency. The field-standard "gold-standard" procedure splits the full particle set randomly into two halves at the very start of refinement, processes each half completely independently (independent orientation search, independent 3D reconstruction), and then computes the Fourier Shell Correlation (FSC) between the two resulting half-maps as a function of spatial frequency.
The reported resolution is conventionally the spatial frequency at which the FSC curve crosses 0.143 — a threshold chosen to correspond roughly to the point where signal and noise are equally likely to explain the observed correlation. Because the two half-maps never shared any information during processing, this procedure is much more resistant to overfitting or model bias than any single-map metric, and is now essentially mandatory for publication and PDB/EMDB deposition.
The 2013 structure of the TRPV1 ion channel at 3.4 Å, solved by Yifan Cheng and David Julius using a direct electron detector, is widely credited as the opening shot of the "resolution revolution" — proof that membrane proteins impossible to crystallize could now be solved to near-atomic detail by cryo-EM alone.
Once density reaches roughly 3–4 Å or better, individual β-strands and α-helices are clearly separated and bulky side chains become visible as distinct protrusions; below about 2.5 Å even smaller side chains and ordered water molecules can be placed with confidence. Model building tools such as Coot, ISOLDE, and Phenix real-space refinement let structural biologists trace the polypeptide backbone and place side chains directly into the density, iteratively refining atomic coordinates to maximize agreement between the model and the experimental map.
The emergence of highly accurate structure predictors — most prominently AlphaFold2 and AlphaFold3 — has transformed this step: a predicted model for the protein's sequence can now be rigid-body or flexibly fitted into the cryo-EM density as a very strong starting point, dramatically accelerating model building and helping resolve ambiguous density in a way that was impossible before 2020. The final deposited model — coordinates plus the experimental density map — becomes a permanent, reusable structural biology resource in the PDB and EMDB.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| Cryo-EM Single Particle | Soluble & membrane complexes, no crystal needed | Thousands–millions of frozen particles imaged, computationally aligned & averaged in 3D | No crystallization; captures flexible/heterogeneous states; now reaches <2 Å |
| X-ray Crystallography | Well-ordered 3D protein crystals | X-ray diffraction from periodic lattice; phases solved by MR, MAD/SAD, or Patterson methods | Extremely high resolution (<1 Å) when crystals are excellent; mature, fast pipelines |
| Cryo-Electron Tomography | Cells, organelles, in-situ complexes | Tilt-series imaging of a single specimen reconstructed into a 3D tomogram; sub-tomogram averaging for resolution | Native cellular context; visualizes macromolecular complexes in situ |
| NMR Spectroscopy | Small, soluble, stable proteins in solution | Nuclear spin coupling measured in solution; distance restraints solved into ensemble structures | No crystal or freezing needed; captures dynamics and solution-state motion directly |