Cryo-Electron Microscopy

De Broglie Wavelength with Relativistic Correction

The kinetic energy of an electron accelerated through potential $V$ is$E_k = eV$. The relativistic energy-momentum relation gives:

$$E^2 = (pc)^2 + (m_0 c^2)^2$$

where the total energy is $E = E_k + m_0 c^2 = eV + m_0 c^2$. Solving for momentum:

$$p = \frac{1}{c}\sqrt{E^2 - (m_0 c^2)^2} = \frac{1}{c}\sqrt{(eV + m_0 c^2)^2 - (m_0 c^2)^2}$$

Expanding the square:

$$p = \frac{1}{c}\sqrt{e^2V^2 + 2eVm_0c^2} = \frac{1}{c}\sqrt{2m_0 c^2 eV\left(1 + \frac{eV}{2m_0c^2}\right)}$$

The de Broglie wavelength $\lambda = h/p$ becomes:

$$\boxed{\lambda = \frac{h}{\sqrt{2m_0 eV\left(1 + \frac{eV}{2m_0c^2}\right)}}}$$

At 300 kV, this gives $\lambda \approx 0.0197$ Å — far smaller than interatomic distances. The resolution in cryo-EM is therefore not limited by diffraction but by radiation damage, specimen movement, and the contrast transfer function.

The Contrast Transfer Function (CTF)

In the weak phase object approximation (WPOA), the specimen acts as a thin phase plate. The image formed by the objective lens is modulated by the contrast transfer function, which arises from the combination of defocus and spherical aberration:

The electron wave after the objective lens acquires a phase shift $\chi(k)$ that depends on the spatial frequency $k$:

$$\chi(k) = \pi \lambda \Delta f\, k^2 - \frac{\pi}{2} C_s \lambda^3 k^4$$

where $\Delta f$ is the defocus (positive for underfocus),$C_s$ is the spherical aberration coefficient, and$k = |\mathbf{k}|$ is the spatial frequency. The CTF for a weak phase object is:

$$\boxed{\text{CTF}(k) = \sin\!\left(\pi \lambda \Delta f\, k^2 - \frac{\pi}{2} C_s \lambda^3 k^4\right)}$$

The CTF oscillates between +1 and -1 as a function of spatial frequency, passing through zero at specific frequencies. At these zeros, information is lost entirely. The first zero occurs at:

$$k_1 = \frac{1}{\sqrt{\lambda \Delta f}}$$

(neglecting the $C_s$ term for moderate spatial frequencies). This is why cryo-EM images are typically collected at a series of defocus values: information lost at the CTF zeros of one defocus value can be recovered from images at other defocus values.

The Fourier Slice Theorem

Let $\rho(\mathbf{r})$ be the 3D electron density of the specimen, and let$P_\theta(x)$ be a 1D projection along direction $\theta$ (for the 2D case, which generalizes straightforwardly to 3D):

$$P_\theta(x) = \int_{-\infty}^{\infty} \rho(x\cos\theta - y\sin\theta,\; x\sin\theta + y\cos\theta)\, dy$$

Taking the 1D Fourier transform of the projection:

$$\hat{P}_\theta(k) = \int_{-\infty}^{\infty} P_\theta(x)\, e^{-2\pi i k x}\, dx$$

Substituting the projection integral and changing variables, one can show that:

$$\boxed{\hat{P}_\theta(k) = \hat{\rho}(k\cos\theta,\; k\sin\theta)}$$

Interpretation: The 1D Fourier transform of a projection at angle $\theta$ gives a central slice through the 2D Fourier transform of the object, oriented at the same angle $\theta$. In 3D, each 2D projection image provides a central plane through the 3D Fourier transform. By collecting projections at many orientations (which cryo-EM achieves through random particle orientations on the grid), one can fill in the 3D Fourier space and reconstruct $\rho(\mathbf{r})$ by inverse Fourier transform.

Back-Projection Reconstruction

The simplest reconstruction method is direct back-projection. Given a set of projections $\{P_{\theta_i}\}$, the back-projection estimate of the density is:

$$\rho_{\text{BP}}(\mathbf{r}) = \sum_{i=1}^{N} P_{\theta_i}(\mathbf{r} \cdot \hat{n}_i)$$

where $\hat{n}_i$ is the projection direction. This simple sum produces a blurred reconstruction because low spatial frequencies are overrepresented (each projection contributes a central slice, and slices are denser near the origin). The correction is weighted back-projection:

$$\rho(\mathbf{r}) = \mathcal{F}^{-1}\!\left[\frac{\sum_i \hat{P}_{\theta_i}(k) \cdot W_i(k)}{\sum_i W_i(k)}\right]$$

where $W_i(k)$ are appropriate weighting functions. In practice, modern cryo-EM software (RELION, cryoSPARC) uses iterative maximum-likelihood refinement to simultaneously determine particle orientations and reconstruct the 3D map.

Shot Noise and the SNR of Individual Images

For a pixel receiving $n$ electrons on average, electron counting follows Poisson statistics. The signal is $n$ and the noise (standard deviation) is$\sqrt{n}$, giving a per-pixel SNR:

$$\text{SNR}_{\text{pixel}} = \frac{n}{\sqrt{n}} = \sqrt{n}$$

For a typical cryo-EM exposure of 20 e$^-$/Ų on a detector with 1 Špixel size, $n \approx 20$, giving SNR $\approx 4.5$per pixel. However, the contrast of unstained biological specimens is only a few percent, so the structural SNR is much lower:

$$\text{SNR}_{\text{structural}} \approx \frac{\Delta n}{\sqrt{n}} \ll 1$$

where $\Delta n$ is the difference in electron counts due to the specimen contrast. For a typical protein, $\Delta n / n \sim 0.02$, making individual particle images essentially invisible by eye.

SNR Improvement by Averaging

The key to cryo-EM is that averaging $N$ independent, aligned images of identical particles improves the SNR. If each image has signal $s$ and independent noise with variance $\sigma^2$:

The average signal is unchanged: $\bar{s} = s$. The noise variance of the average decreases as:

$$\text{Var}(\bar{x}) = \frac{\sigma^2}{N}$$

Therefore the SNR after averaging $N$ particles is:

$$\boxed{\text{SNR}_N = \sqrt{N} \cdot \text{SNR}_1}$$

To achieve SNR $\sim 10$ from images with SNR $\sim 0.1$, one needs$N \sim 10{,}000$ particles. For higher resolution (smaller features, lower contrast), the required number grows. Modern high-resolution reconstructions routinely use$10^5$ to $10^6$ particles.

This analysis assumes perfect alignment. In practice, alignment errors introduce an additional noise source that limits resolution. The required number of particles to achieve resolution$d$ scales approximately as $N \propto 1/d^2$ (more particles are needed for higher resolution, since the signal at high spatial frequencies is weaker).

Image Formation Model

In the weak phase object approximation, the observed image in Fourier space is:

$$\hat{I}(k) = \text{CTF}(k) \cdot \hat{\rho}_{\text{proj}}(k) + \hat{N}(k)$$

where $\hat{\rho}_{\text{proj}}$ is the Fourier transform of the projected density and $\hat{N}$ is additive noise. The goal is to recover$\hat{\rho}_{\text{proj}}$ from the observed $\hat{I}$.

Naive division by the CTF diverges at zeros. The Wiener filter provides the optimal linear estimate (in the minimum mean-square error sense):

$$\boxed{H_{\text{opt}}(k) = \frac{\text{CTF}(k)}{\text{CTF}^2(k) + \frac{1}{\text{SNR}(k)}}}$$

The corrected estimate is $\hat{\rho}_{\text{est}} = H_{\text{opt}} \cdot \hat{I}$.

Derivation of the Wiener Filter

We seek the filter $H(k)$ that minimizes the expected squared error:

$$\mathcal{E} = \left\langle \left| H(k)\hat{I}(k) - \hat{\rho}(k) \right|^2 \right\rangle$$

Substituting $\hat{I} = \text{CTF}\cdot\hat{\rho} + \hat{N}$ and expanding:

$$\mathcal{E} = |H|^2 \left(\text{CTF}^2 S_\rho + S_N\right) - 2\,\text{Re}\!\left[H \cdot \text{CTF} \cdot S_\rho\right] + S_\rho$$

where $S_\rho = \langle|\hat{\rho}|^2\rangle$ is the signal power spectrum and$S_N = \langle|\hat{N}|^2\rangle$ is the noise power spectrum. Setting$\partial \mathcal{E}/\partial H = 0$:

$$H_{\text{opt}} = \frac{\text{CTF} \cdot S_\rho}{\text{CTF}^2 S_\rho + S_N} = \frac{\text{CTF}}{\text{CTF}^2 + S_N/S_\rho}$$

Identifying $\text{SNR}(k) = S_\rho(k)/S_N(k)$ recovers the boxed result above. At CTF zeros, $H_{\text{opt}} \to 0$, gracefully suppressing the noisy frequencies rather than amplifying them. At spatial frequencies where $\text{SNR} \gg 1$,$H_{\text{opt}} \approx 1/\text{CTF}$, performing full phase correction.

DeRosier & Klug (1968): 3D Reconstruction from EM

David DeRosier and Aaron Klug published the first 3D reconstruction from electron micrographs in 1968, using images of the T4 bacteriophage tail. They demonstrated that a series of 2D projection images, taken at different tilt angles, could be computationally combined to reconstruct the 3D structure using the Fourier slice theorem. This seminal work established the mathematical framework that underpins all of modern cryo-EM reconstruction. Klug received the Nobel Prize in Chemistry in 1982 for this and related work on crystallographic electron microscopy.

Henderson (1975): First Atomic-Resolution EM Structure

Richard Henderson and Nigel Unwin determined the first structure of a membrane protein at near-atomic resolution using electron microscopy of 2D crystals of bacteriorhodopsin (1975). By combining images and electron diffraction from purple membrane patches, they obtained a 7 Å resolution map showing seven transmembrane alpha-helices. Henderson later improved the resolution to 3.5 Å (1990), and crucially demonstrated that radiation damage was the fundamental limiting factor, not the physics of electron optics.

Henderson's key insight was that the electron wavelength is short enough for atomic resolution, and that the resolution limit in biological EM is set entirely by the tolerable dose. He calculated that approximately 10,000 images of identical molecules should suffice for atomic-resolution reconstruction even at minimal dose, laying the theoretical groundwork for single-particle cryo-EM.

Joachim Frank: Single-Particle Analysis

Joachim Frank developed the computational methods for single-particle analysis beginning in the 1970s. His group created algorithms for aligning, classifying, and averaging heterogeneous sets of particle images without requiring crystals or regular arrays. The key innovations included multivariate statistical analysis for classification, correlation-based alignment, and iterative refinement of 3D reconstructions from 2D projections.

Frank's software package SPIDER became the standard tool for single-particle EM. His work transformed EM from a qualitative imaging technique into a quantitative structural method. The ribosome served as his primary test specimen, and his group produced increasingly detailed reconstructions through the 1990s and 2000s.

Jacques Dubochet: Vitrification

The critical breakthrough that made cryo-EM practical was Jacques Dubochet's development of vitrification in the early 1980s at the EMBL in Heidelberg. Previous EM methods required staining or dehydration, which distorted or destroyed biological structures. Dubochet discovered that by rapidly plunging a thin aqueous film into liquid ethane (cooled by liquid nitrogen), water could be frozen so quickly that it formed vitreous (amorphous) ice rather than crystalline ice.

Vitreous ice preserves the native hydrated structure of biological specimens, avoids staining artifacts, and provides a uniform background. Dubochet's method (published 1984) made it possible to image unstained, fully hydrated macromolecules in the electron microscope. Combined with Frank's single-particle methods and modern direct electron detectors with movie-mode data collection, vitrification enabled the resolution revolution.