Reference

Last updated on 2026-10-13 | Edit this page

Glossary

16S amplicon sequencing: Sequencing one marker gene (16S rRNA) to find out which bacteria are in a sample. Gives counts of taxa.
AUC (area under the ROC curve): A measure of how well a model separates two classes: 1 is perfect, 0.5 is no better than guessing.
Balanced accuracy: The average of the share of correctly predicted samples in each class. Unlike accuracy, it is not inflated when one class is much larger.
Batch effect: A technical difference between samples that were measured on different days, runs or plates.
Classification error rate: The share of samples that a model assigns to the wrong group. With two equally large groups, guessing gives 50 %.
Confounding: When two factors vary together (for example the batch and the study group), so that their effects cannot be told apart.
Cross-validation: Estimating how well a model predicts new samples by repeatedly training it on part of the data and testing it on the rest.
Detection limit: The smallest amount that an instrument can measure. Values below it are often reported as missing.
DIABLO: A supervised integration method in the mixOmics package (block.splsda()). It selects a few features from each data type that together separate the groups and are correlated between the data types.
Early fusion: Joining the tables of all data types side by side and training one model on them. Also called concatenation.
Experiment: One data type stored in a MultiAssayExperiment, for example the microbiota as a TreeSummarizedExperiment.
Factor (MOFA): A hidden pattern that varies between the samples and is found by MOFA. A factor can be shared by several data types or be specific to one.
Factor model: A model in which a few hidden variables (factors), each with a weight per feature, explain most of the variation of the measured features.
Intermediate fusion: Training a model on each data type, joining what the models learned, and training one more model on top.
Holobiont / hologenomics: The host and its microbes seen as one biological system / the study of their genomes and molecules together.
Joint RPCA: Robust principal component analysis of two or more data types together. It draws one map of the samples, and fills in missing values, such as the zeros of the microbiota, itself.
Functional redundancy: Different taxa carrying out the same function, so that communities with different taxa can work in the same way.
Kernel: A matrix of similarities between samples, for example calculated from microbiome distances. Kernel methods such as MiRKAT use it to relate a variable to the whole microbiome.
Late fusion: Training a separate model on each data type and combining their predictions, for example as a weighted average. Also called stacking.
Mantel test: A test of whether two dissimilarity matrices, for example from two data types, are correlated.
Mass spectrometry (LC-MS, GC-MS): A method that separates and weighs molecules. In metabolomics, it measures hundreds or thousands of metabolite features.
Metabolite: A small molecule produced or used in metabolism, such as a short-chain fatty acid.
MiRKAT: The microbiome regression-based kernel association test. Tests whether a single variable, such as a metabolite, is related to the microbiome as a whole.
MOFA (multi-omics factor analysis): An unsupervised integration method that finds factors shared by several data types, like a PCA for several tables.
Multi-omics: Analysing several data types measured from the same samples, such as taxa, metabolites and host biomarkers.
MultiAssayExperiment (MAE): The Bioconductor container for multi-omics data. It holds one experiment per data type, a shared table of sample information (colData) and a sample map that links them.
NMR spectroscopy: Nuclear magnetic resonance spectroscopy. In metabolomics, it measures the concentrations of tens of known metabolites.
Pairwise association: The correlation between every feature of one data type and every feature of another.
Permutation test: A test that shuffles the samples many times to see how often a result as large as the real one arises by chance.
Prebiotic: A fibre that the host cannot digest but some gut bacteria can, such as xylo-oligosaccharides (XOS).
primary / colname: The columns of sampleMap(): primary is the subject (a row of colData(mae)), and colname is the name of the column in an experiment.
rclr (robust centred log-ratio) transformation: A log-ratio transformation that ignores the zeros. By default mia fills them in by matrix completion; with impute = FALSE they stay missing.
Robust (in joint RPCA): The method fits only the observed values and fills in zeros or missing values by matrix completion, so that a few extreme values do not dominate.
Sample map: The table in a MultiAssayExperiment that links each column of each experiment to a subject in the sample information.
Short-chain fatty acids (SCFA): Metabolites such as butyrate, acetate and propionate that gut bacteria produce when they ferment fibre.
Shotgun metagenomics: Sequencing all DNA in a sample. Gives counts of taxa and of genes or pathways (functions).
Sparse partial least squares (sPLS): A method that finds directions in two or more tables whose sample scores are strongly correlated (and, in DIABLO, related to the outcome), using only a few features.
Stacking: Late fusion in which a second model (the meta-learner) learns weights to combine the cross-validated predictions of the models of each data type.
Standardization: Scaling each feature to mean 0 and standard deviation 1, so that features measured on different scales can be compared.
Supervised method: A method that is given an outcome, such as the diet group, and looks for the features that predict it.
Targeted / untargeted metabolomics: Measuring a chosen list of known metabolites / everything that can be detected, including unknown molecules.
TreeSummarizedExperiment (TreeSE): The Bioconductor container for microbiome data: abundance tables, taxonomy, sample information and a phylogenetic tree.
Unsupervised method: A method that looks for patterns in the data without knowing the study groups, such as PCA or MOFA.
UpSet plot: A plot that shows the sizes of the overlaps between several sets, here which samples have which data types.
Weight (MOFA): How strongly a feature drives a factor. Like a loading in PCA.