Glossary
Terms and concepts used throughout the GARAGE documentation and codebase.
A
ARI (Adjusted Rand Index)
A metric that measures the similarity between two clusterings, adjusting for chance. Range: \([-1, 1]\). Used in GARAGE to compare Leiden clusters of generated data to ground-truth cell type labels.
Attention mechanism
A neural network component that assigns different weights to inputs based on their relevance. In GAT, attention scores determine how much each neighbor contributes to a node’s update. In GARAGE, attention weights from the trained GAT are used to select seed cells.
B
Batch effect
Systematic technical variation between samples processed in different batches, labs, or protocols. Can confound biological signal in scRNA-seq data.
C
CV² (Coefficient of Variation Squared)
\(\sigma^2 / \mu^2\) — a measure of relative variability for each gene. Used as a feature selection criterion: genes with high CV² are selected for clustering.
Cell type
A biologically defined category of cells sharing a common identity and function (e.g., “T cell CD4+”, “B cell”, “neuron”).
D
DDPM (Denoising Diffusion Probabilistic Model)
A generative model that learns to reverse a gradual noising process. Used in benchmarking/scrna_seq_specific/scdiffusion.py.
Discriminator
The neural network in a GAN that tries to distinguish real data from generated data.
E
Embedding
A low-dimensional vector representation of a cell (or gene) learned by a neural network.
Expression matrix
A matrix \(X \in \mathbb{R}^{n \times g}\) where \(X_{ij}\) is the expression level of gene \(j\) in cell \(i\).
F
Fano index
\(\sigma^2 / \mu\) — variance-to-mean ratio for each gene. Used as an alternative to CV² for feature selection.
Feature selection
The process of selecting a subset of informative genes from the full gene set before clustering or classification. GARAGE supports CV², Fano, and PCA loading methods.
G
GAN (Generative Adversarial Network)
A framework of two competing neural networks: a Generator that produces synthetic data and a Discriminator that tries to distinguish real from synthetic.
GAT (Graph Attention Network)
A graph neural network that uses self-attention to weight the importance of neighboring nodes. Used in GARAGE to identify seed cells.
Generator
The neural network in a GAN that produces synthetic data from input noise (or a hybrid batch in GARAGE).
H
Hybrid input batch
In GARAGE, the generator receives a mixed batch: \((1-\lambda) \cdot z_{\text{noise}} \oplus \lambda \cdot x_{\text{seed}}\), where a fraction \(\lambda\) of the input is real seed cells from the GAT and the rest is random noise.
L
Leakage fraction (\(\lambda\))
The proportion of the GAN generator’s input batch that consists of real seed cells (selected by the GAT) rather than random noise. Default: 0.2.
Leiden clustering
A community detection algorithm that partitions cells into clusters by optimising modularity. Used in GARAGE for quantitative evaluation.
M
Macro-F1
The average of per-class F1 scores, weighted equally. Used in GARAGE to evaluate rare-cell preservation — a low macro-F1 indicates that rare cell types are poorly captured.
MMD (Maximum Mean Discrepancy)
A kernel-based metric for comparing two distributions. Computed in analysis/distribution_metrics.py.
Mode collapse
A failure mode of GANs where the generator produces only a few types of output repeatedly, rather than the full diversity of the training data.
N
NMI (Normalised Mutual Information)
A normalised measure of mutual information between two clusterings. Range: \([0, 1]\). Less sensitive to cluster size imbalance than ARI.
P
PCA (Principal Component Analysis)
A linear dimensionality reduction method that projects data onto axes of maximum variance. Used in GARAGE as a preprocessing step before UMAP and as a feature selection criterion.
PCA loading
The contribution (weight) of each original feature to a principal component. Genes with high aggregate loading on PC1–PC3 are selected as informative features.
Priority node / Priority cell
A cell belonging to a rare cell type (with cell count below rare_threshold). These cells receive extra attention weight during GAT training.
Priority weight
A scalar multiplier applied to rare-cell features in the GAT to boost their influence. Default: 2.0 (effective factor: \(1 + 2.0 = 3.0\)).
R
Rare cell type
A cell type with fewer than rare_threshold cells in the dataset. GARAGE is designed to explicitly preserve these populations in synthetic data.
Resolution (Leiden)
A parameter controlling the granularity of Leiden clustering. Higher resolution → more clusters. Swept over a range to find the optimal partition for ARI/NMI/F1.
S
scRNA-seq (single-cell RNA sequencing)
A technology that measures gene expression in individual cells, revealing cellular heterogeneity that is masked in bulk RNA-seq.
Seed cell
A real cell selected by the GAT as highly representative. Seed cells are mixed into the generator’s input batch to guide and stabilise GAN training.
SOTA (State of the Art)
Refers to the collection of baseline generative models implemented in benchmarking/sota/ for comparison with GARAGE.
SWD (Sliced Wasserstein Distance)
A computationally cheaper variant of the Wasserstein distance, computed by projecting data onto random 1D lines and averaging the resulting 1D Wasserstein distances.
U
UMAP (Uniform Manifold Approximation and Projection)
A non-linear dimensionality reduction technique for visualising high-dimensional data in 2D or 3D. Used in GARAGE for qualitative comparison of real and generated data clusters.
V
VAE (Variational Autoencoder)
A generative model that learns a latent representation of data via an encoder-decoder architecture with a KL divergence regularisation term.
W
Wasserstein distance (WD)
Also known as Earth Mover’s Distance. Measures the minimum cost to transform one probability distribution into another. Used in GARAGE as a primary distributional fidelity metric. Computed via Optimal Transport.