Glossary

The Cryo-ET and statistical-machine-learning terms used across this site — bilingual, each linked to the article that develops it.

Electron Tomography

Cryo-ET / 冷冻电子断层成像: Imaging a unique, frozen-hydrated object from many tilt angles and reconstructing it in 3-D.
Cryo-EM / 冷冻电镜: Imaging vitrified specimens with electrons; single-particle analysis and Cryo-ET are its two main branches.
Vitrification / 玻璃化: Freezing a sample fast enough that water solidifies into amorphous, glass-like ice, preserving near-native structure.
Tilt series / 倾转序列: The set of 2-D projections recorded as the specimen stage is tilted through a range of angles.
Missing wedge / 缺失楔形: The unmeasured region of Fourier space left by the limited tilt range; it blurs the reconstruction along the beam axis.
Reconstruction / 重构（断层体）: Combining the tilt-series projections into a 3-D volume (a tomogram).
In situ / 原位: Imaging molecules inside their native cellular context rather than purified in isolation.

CryoGEN-I / II: Self-supervised Cryo-ET reconstruction: a MAP / energy-prior generation (I) and an optimal-transport generation (II).
CryoWGEN-I / II: Entropic-OT reconstruction giving a Boltzmann posterior, sampled by Monte-Carlo (I) or Langevin / SGLD (II).

Subtomogram averaging / 子断层平均: Aligning and averaging many copies of the same structure to raise SNR and fill the missing wedge.
Particle picking / 颗粒挑选: Locating instances of a target structure within a crowded tomogram.
Segmentation / 分割: Partitioning a tomogram into membranes, organelles, and other components.

Fourier transform / 傅里叶变换: Re-expressing a signal as a sum of frequencies — the natural language of imaging and reconstruction.
Central-slice theorem / 中心切片定理: A projection's Fourier transform equals a central slice of the object's 3-D transform, linking tilt angle to frequency coverage.
CTF / 衬度传递函数: The microscope's frequency response, which modulates and partly inverts contrast across spatial frequencies.
Point-spread function / 点扩散函数: The image of a single point; convolving the object with it models blur.
Sampling & Nyquist / 采样与奈奎斯特: How finely a signal must be sampled — the Nyquist limit — to capture a given frequency without aliasing.
SNR / 信噪比: The ratio of signal power to noise power; cryo imaging operates at very low SNR.
Filtering / 滤波: Suppressing or emphasizing chosen frequency bands to reduce noise or sharpen features.

Bayesian inference / 贝叶斯推断: Updating a prior belief into a posterior using observed data via Bayes' rule.
Prior / 先验: The distribution encoding what is plausible before any data is seen.
Posterior / 后验: The distribution over the unknown after combining prior and data.
MAP estimate / 最大后验估计: The single most probable value under the posterior — its mode.
Maximum likelihood / 最大似然: Choosing the parameters that make the observed data most probable.
EM algorithm / EM 算法: An iterative scheme alternating estimation of hidden variables (E-step) and parameter updates (M-step).
Entropy / 熵: A measure of the uncertainty or spread of a distribution.
KL divergence / KL 散度: An asymmetric measure of how one distribution differs from a reference; +∞ when their supports disagree.

Variational inference / 变分推断: Approximating an intractable posterior by the closest member of a tractable family.
ELBO / 证据下界: The evidence lower bound; maximizing it is equivalent to minimizing KL to the true posterior.
Energy-based model / 能量模型: A distribution p(x) ∝ e^{−E(x)} defined by an energy function, with no tractable normalizer.
Langevin dynamics / SGLD / 朗之万动力学: Sampling from a density using its gradient plus injected noise; SGLD is its stochastic-gradient form.

Optimal transport / 最优传输: The least-cost way to move one distribution's mass onto another.
Wasserstein distance / Wasserstein 距离: The optimal-transport cost read as a distance between distributions; finite even for disjoint supports.
Entropic OT / Sinkhorn / 熵正则最优传输: Adding an entropy penalty to the transport cost, giving a smooth, everywhere-positive coupling solved by matrix scaling.

GAN / WGAN / 生成对抗网络: A generator trained against a discriminator; WGAN swaps the discriminator for a Wasserstein critic for stability.
VAE / 变分自编码器: An autoencoder trained by maximizing the ELBO, with a probabilistic latent space.
WAE / Wasserstein 自编码器: An autoencoder that matches its aggregated latent posterior to the prior by an optimal-transport cost.
AAE / 对抗自编码器: An autoencoder that matches latent codes to a prior using an adversarial discriminator.
EVIA / 熵变分推断自编码: Entropic variational inference autoencoder — the entropic-OT framework that CryoWGEN instantiates.
Self-supervised learning / 自监督学习: Learning from a supervisory signal derived from the data itself, with no external labels.