Glossary of Terms

Adaptive boosting: reweights misclassified samples so each new weak learner focuses on the hardest cases. Simple, effective, but noise-sensitive.

Bottom-up clustering that merges the closest samples/clusters step by step into a tree. The linkage rule (ward, complete, average) controls how cluster distance is measured.

Quantifies within-sample diversity of a community matrix: Shannon and Simpson indices combine richness and evenness; observed features count non-empty categories.

One-way analysis of variance: tests whether the mean of a numeric column differs across two or more groups defined by a categorical column.

Bootstrap Aggregating: trains many models on resampled subsets and averages them, cutting variance without changing the base learner.

Naive Bayes for binary (0/1) features — models per-class presence probabilities of each indicator.

Quantifies between-sample compositional differences of a community matrix via a distance matrix and ordination; PERMANOVA tests whether groups differ in composition.

Finds pairs of linear combinations (canonical variates) — one per omics block — that are maximally correlated with each other.

Tests independence between two categorical columns using a contingency table. Reports the chi-square statistic, p-value, and expected counts.

Groups samples into k clusters without labels. K-Means partitions into compact groups; agglomerative (hierarchical) clustering builds a nested tree with a chosen linkage rule.

Computes pairwise association coefficients (Pearson, Spearman, or Kendall) between numeric columns and reports the correlation matrix and the strongest pairs.

Generalised linear model for count outcomes. Poisson assumes mean equals variance; Negative Binomial adds a dispersion parameter for overdispersed counts.

Cleans a dataset before modelling: imputes missing values, detects and handles outliers, and rescales feature magnitudes. Produces a derived dataset that can be chained into any other analysis.

A single tree of if/then splits on feature values. Highly interpretable but prone to overfitting; the building block of random forests and boosting.

Supervised multi-block integration (mixOmics block.splsda): finds correlated components across omics blocks that discriminate a categorical outcome.

A trivial predictor (mean or most-frequent class) used as the floor any real model must beat. Included in comparisons for reference.

Community-ecology methods for species and diversity data: alpha and beta diversity, Mantel tests, Moran's I spatial autocorrelation, and indicator species analysis.

Linear regression combining L1 (lasso) and L2 (ridge) penalties — selects features like lasso while staying stable on correlated predictors.

Extremely Randomised Trees: like a random forest but splits are chosen at random rather than optimised, trading a little accuracy for speed and variance reduction.

Creates new columns from existing ones — ratios, sums, transforms (log, sqrt, polynomial), longitudinal summaries, one-hot encoding, or a wide-to-long reshape. Produces a derived dataset.

Ranks and reduces the feature set before modelling using filter (variance, correlation, FDR), wrapper (RFE), or embedded (model importance) strategies.

End-to-end run: data preprocessing, optional univariate screening, and a tuned Random Forest model — combined in a single result with numbered sub-steps.

Probabilistic classifier assuming features are independent Gaussians per class. Very fast; a strong small-data baseline despite the naive assumption.

A Bayesian regressor that returns both a prediction and an uncertainty estimate per sample via a kernel-defined covariance over functions.

Bayesian classifier placing a Gaussian-process prior over a latent function squashed through a link — yields class probabilities with uncertainty.

Builds trees sequentially, each correcting the residual errors of the ensemble so far. Often top accuracy on tabular data but slower and easier to overfit than forests.

Ranked-list pathway method: instead of a cutoff, it tests whether pathway genes concentrate at the top or bottom of the full ranked feature list. Reports ES, NES, and FDR.

Genome-wide association study: QC filters (MAF, missingness, HWE, LD pruning), then per-SNP association testing with multiple-testing correction and Manhattan/QQ figures.

Agglomerative clustering that merges samples into a tree (dendrogram) of nested groups using a linkage criterion — no need to fix k upfront.

Gradient boosting on binned feature histograms — much faster on large datasets and natively tolerant of missing values.

Identifies columns (species/features) that are characteristic of a group by combining their specificity (fidelity) and sensitivity (frequency) per group.

Partitions samples into k clusters by minimising within-cluster distance to each cluster centroid. Fast and effective for compact, roughly spherical clusters.

Predicts each sample from the k most similar rows in feature space — no fitted model. Simple and effective on smooth data, sensitive to scaling and irrelevant features.

Kendall's tau rank correlation based on concordant/discordant pairs. Most robust of the three coefficients; slower on large datasets.

Ridge regression in kernel space: L2-regularised fits that can follow non-linear surfaces via RBF or polynomial kernels.

Non-parametric multi-group comparison (rank-based ANOVA). Tests whether any of two or more groups differ in median without assuming normality.

Linear regression with L1 regularisation: uninformative coefficients are shrunk exactly to zero, giving built-in feature selection.

Classifier that models each class as a Gaussian with shared covariance, giving linear boundaries; also usable for supervised dimensionality reduction.

Fits a straight-line relationship between features and a continuous target by minimising squared error. Fast and interpretable, but sensitive to outliers and unable to capture non-linear patterns.

A linear classifier that models class probability through the logistic function. Fast, calibrated, and interpretable — the standard classification baseline.

Predictive modelling methods — chiefly random forest — that learn a mapping from feature columns to a target. Evaluated with cross-validated metrics and interpreted through feature importance.

Non-parametric two-group comparison based on ranks. Does not assume normality — preferred over the t-test for skewed or ordinal data.

Correlates two distance matrices (e.g. community dissimilarity vs environmental distance) and assesses significance by permuting labels.

Linear mixed model with a fixed effect of primary interest plus a random grouping factor (e.g. pasture within ranch) that accounts for non-independence between observations.

Benchmarks several regression algorithms under identical cross-validation splits and reports comparable metrics so the best-suited model can be chosen.

Multi-Omics Factor Analysis: decomposes multi-block data into a small set of latent factors that capture the dominant shared and block-specific sources of variation.

Native MOFA2 engine (mofapy2) fitting a Bayesian factor model across omics views. Tolerates missing values and supports convergence tuning.

Measures spatial autocorrelation of a numeric column given sample coordinates: positive values mean nearby plots have similar values; significance is permutation-based.

A feed-forward neural network of stacked non-linear layers. Flexible on complex patterns but needs scaling, tuning, and more data than tree ensembles.

Joint analysis of two or more omics blocks measured on the same samples. Methods include SNF, PCA, MOFA/MOFA2, CCA, NMF, and DIABLO for supervised integration.

Naive Bayes for count features — classic text-classification model that also fits species or event counts.

Assigns each sample to the class whose mean feature vector is closest — the simplest possible parametric classifier.

Rank-based ordination that preserves the ordering of pairwise distances rather than their exact values. Standard for ecological community data; reports a stress value.

Factorises a non-negative data matrix into basis (W) and coefficient (H) matrices, revealing parts-based patterns shared across samples.

Reduces a samples-by-features table to a few axes for visualisation. PCA maximises variance on standardised data; PCoA and NMDS embed a distance/dissimilarity matrix.

Pathway method that tests whether genes from your list appear in a pathway more often than expected by chance (hypergeometric test, reported as an odds ratio).

Maps a gene/feature list onto GO and KEGG pathways and tests over-representation (ORA) or ranked enrichment (GSEA) with FDR control.

Classical multidimensional scaling: embeds a distance/dissimilarity matrix (e.g. Bray-Curtis) into a few axes so sample differences can be plotted.

Linear correlation coefficient from -1 to +1. Measures straight-line association; assumes roughly continuous, normally distributed data.

Permutational multivariate ANOVA: tests whether groups differ in community composition by permuting the distance matrix. Reported alongside beta-diversity ordinations.

Partial Least Squares: projects features onto latent components maximally covariant with the target — strong for wide, collinear data like spectra.

A GLM for count targets (0,1,2,…) that models the log of the expected count. Appropriate when the target is event counts rather than a continuous measure.

Variance-maximising ordination: finds orthogonal axes (principal components) that capture the most variation in the data. Used both as a single-dataset ordination method and across concatenated omics blocks.

Like LDA but each class gets its own covariance, producing curved boundaries — more flexible, needs more data per class.

Predicts a chosen quantile (e.g. the median) instead of the mean — robust to outliers and able to describe the whole conditional distribution.

Ensemble of decision trees trained on bootstrap samples for regression or classification (auto-detected from the target). Reports cross-validated metrics, feature importances, and an exportable model.

Analyses measurements taken on the same subjects at multiple timepoints (wide table: one column per timepoint). Tests the within-subject time effect and optionally a between-subject factor.

Wrapper feature selection: trains a model, drops the weakest features, and repeats until the target count remains. Slower than filters but captures feature interactions.

Ridge regression repurposed for classification by fitting class indicator variables — fast linear baseline with L2 stability.

Linear regression with L2 regularisation: coefficients are shrunk toward zero, which stabilises fits when features are correlated or numerous.

Linear model trained by stochastic gradient descent — scales to very large data and supports different losses/penalties, at the cost of tuning.

Fuses per-omics sample similarity networks into one fused network, then clusters samples into groups that are consistent across all omics layers.

Rank-based correlation coefficient for monotonic (not necessarily linear) relationships. Robust to outliers and non-normal data.

Hypothesis tests that compare groups or associate variables: t-tests, ANOVA, chi-square, Mann-Whitney, Kruskal-Wallis, and correlation screening. Each returns a test statistic, p-value, and effect size.

Finds the maximum-margin boundary between classes; kernels extend it to non-linear decision surfaces.

SVR fits a tube around the data where errors inside the margin are ignored; kernel functions let it model non-linear targets.

Compares the mean of a numeric column between exactly two groups. Student's test assumes independent samples; the paired option handles matched before/after measurements.

A GLM for zero-inflated, right-skewed continuous targets — mass at zero plus a positive tail (e.g. rainfall or biomass with many zeros).

Runs Pearson, Spearman, and ANOVA F tests of every feature against the target before modeling, with FDR correction. Significant features are saved as a derived dataset for downstream analyses.

Select the statistical or machine learning methodology for your inquiry. Options range from simple univariate correlations to comprehensive multi-omics and path-discovery pipelines.

Compares feature means across groups to flag variables that differ significantly between classes or conditions.

Optionally attach this dataset to an existing project so related data and analysis results stay grouped together.

The regression framework for SNP-trait testing: linear for quantitative traits, logistic for binary case/control outcomes.

The reference result against which the comparison analysis is evaluated. Metrics are reported relative to this baseline.

Optional grouping column (e.g. treatment) whose effect is tested on subject means alongside the within-subject time effect.

The species context for the gene identifiers. This selects the appropriate genomic backend (e.g., Human vs. Custom) for pathway mapping.

How columns are assigned to blocks in a single wide table. Prefix splits on the first underscore; manual lets you assign columns explicitly.

The set of regression algorithms benchmarked under identical cross-validation splits in a model comparison run.

How non-numeric columns are converted before modeling: keep as-is, map to integers, or expand into one-hot dummy columns.

How non-numeric feature columns are converted before modeling: Exclude drops them, Integer maps each level to a number (1, 2, 3…), One-hot expands each level into its own 0/1 column.

Algorithm used to group rows. K-Means partitions into k compact clusters; hierarchical (agglomerative) clustering merges rows bottom-up and supports different linkage rules.

Optional column whose values colour the ordination scatter points — typically a grouping factor such as ranch or treatment.

The result whose performance is compared against the baseline, e.g. a differently tuned model or alternative dataset.

The coefficient used to measure the strength and direction of the linear or monotonic relationship between variables.

The number of folds for K-Fold Cross-Validation. Data is split into N subsets, with the model training on N-1 and validating on the remainder to ensure robust performance metrics.

How the omics blocks are organised. Single wide table splits columns by prefix; multiple files uses one file per block; transposed has samples in columns and features in rows.

Provide a unique, descriptive name for the dataset to identify it across the platform dashboards.

The primary data source for the current analysis. All calculations, preprocessing, and modeling will be executed against the records in this file.

Number of latent components for the multi-block PLS-DA model. More components capture more discriminative structure but risk overfitting.

Controls how blocks are linked. Full links all blocks to each other; null treats blocks independently; custom lets you specify pairwise weights.

Number of features to retain per block per component (sparse PLS). Use the same value for all blocks or specify per-block values.

Number of cross-validation folds for the perf() error-rate estimation.

Number of times the cross-validation is repeated for BER stability.

Integer seed for reproducible DIABLO results.

Opt-in tuning of keepX via tune.block.splsda. Shows an estimated cost derived from grid_size x ncomp x nrepeat x folds.

Pairwise distance or dissimilarity measure used by PCoA and NMDS. Bray-Curtis suits compositional abundance data; Euclidean and Manhattan suit measured variables.

How genotypes are converted to numeric values. Additive counts effect alleles; dominant and recessive collapse genotypes by inheritance model.

Numeric columns forming the second distance matrix in a Mantel test — typically measured environmental or trait variables.

Distance metric applied to the environmental columns in a Mantel test.

The significance level for the False Discovery Rate (Benjamini-Hochberg) correction. Controls the expected proportion of false positives among the reported significant findings.

Comma-separated list of numeric columns to ordinate. Leave blank to use every numeric column in the dataset.

In transposed layout, the column that identifies features (rows). This is never auto-detected — you must specify it explicitly.

Strategy for reducing the feature set before modeling: variance filter, correlation, RFE, K-Best scores, or model-based selection.

The number of top-ranking genes to include in the Over-Representation Analysis (ORA). Selecting more genes increases coverage but may introduce noise.

The first categorical variable in the chi-square test of independence. Its distinct values form the rows of the contingency table.

The predictor of primary interest in a mixed model — e.g. a treatment or slope position. Categorical columns are treated as factors automatically.

Determines the input for enrichment analysis. You can utilize high-ranking features from a previous Random Forest run or provide a manual list of identifiers.

Applies an inflation-factor correction to p-values to account for population stratification and reduce spurious associations.

Error distribution for count data. Poisson assumes mean equals variance; Negative Binomial adds a dispersion parameter for overdispersed counts.

The Gene Ontology (GO) partition to query: Biological Process (BP), Molecular Function (MF), or Cellular Component (CC).

The categorical column that defines the groups to compare. t-test and Mann-Whitney require exactly two distinct values; ANOVA and Kruskal-Wallis accept two or more.

Hardy-Weinberg Equilibrium p-value cutoff. SNPs below this value deviate from expected genotype frequencies and may indicate genotyping error.

Searches a parameter grid to find better model settings. Slower, but often improves predictive performance.

Identifier columns kept per row in a wide-to-long reshape (e.g. subject, plot). Blank generates a row id.

Optional column used to label each row in the prediction output so results can be traced back to samples.

Mean decrease in impurity importance computed inside tree-based models. Fast and built-in, but biased toward high-cardinality features.

When checked, the polynomial operation emits only cross-products (a x b), skipping squared/higher-power terms.

The shared variable used to align samples across different omics datasets. This column must exist in all selected datasets with matching entry IDs.

The shared column used to match rows between the two datasets. Auto-detect picks the first column name present in both files.

How rows are combined: Inner keeps only matching samples, Left keeps all primary rows, Outer keeps every row from both datasets.

The number of top-ranked features retained by the selection method.

Buckets continuous features into discrete intervals, turning numeric variables into categorical bins.

Specific to factor analytic models like MOFA+. Represents the underlying biological drivers or technical variance components spanning multiple omics views.

Numeric column with each sample latitude (or northing) coordinate, used to build the spatial neighbour graph.

Removes highly correlated SNPs in linkage disequilibrium to reduce redundancy and multicollinearity before association testing.

Pairwise correlation cutoff used during LD pruning. SNP pairs above this r² are reduced to a single representative.

How distances between clusters are measured in hierarchical clustering. Ward minimises within-cluster variance and is the usual default.

Applies log(x) to numeric features, compressing right-skewed distributions common in count and concentration data.

Numeric column with each sample longitude (or easting) coordinate.

Gene or feature identifiers entered directly for enrichment analysis. Provide one symbol per line or a comma-separated list.

Maximum fraction of missing genotypes allowed per sample before the sample is excluded.

Maximum fraction of missing genotype calls allowed per SNP before it is dropped from the analysis.

The name assigned to the new combined dataset created by the merge.

Minimum Minor Allele Frequency. SNPs rarer than this threshold are filtered out as statistically uninformative.

For IQR and Z-Score outlier detection, a row is only flagged as an outlier when it is extreme in at least this fraction of numeric columns. Prevents mass row deletion on wide datasets.

The minimum number of genes from your list that must be present in a pathway for it to be tested for enrichment.

How empty cells are handled before modeling: keep them, or impute each column with its mean, median, or mode.

The machine learning algorithm to train. Random Forest is a robust default for tabular omics data; linear models suit well-understood linear relationships.

Controls the convergence criteria for the variational EM algorithm. Fast stops early; slow runs more iterations for tighter convergence.

Minimum R-squared contribution for a factor to be retained. Factors explaining less variance than this threshold are pruned.

Integer seed for reproducible MOFA2 factor estimation.

Whether to scale each view to unit variance before fitting. Useful when views have very different magnitudes.

Controls false positives across many SNP tests. Bonferroni is strict; Benjamini-Hochberg FDR balances discovery and error.

Dispersion parameter of the negative-binomial family. Larger alpha means more overdispersion; alpha near 0 approaches Poisson.

How many nearest neighbours each point connects to in the spatial weight matrix. Four to eight is typical for plot grids.

Optional name for the created column; a descriptive name is generated automatically when blank.

Rescales feature magnitudes: min-max maps each column to 0-1; z-score standardises to zero mean and unit variance.

Scales feature values before analysis so variables measured on large ranges do not dominate the model. Recommended when columns have very different units.

How many ordination axes to compute. Two axes are enough for a scatter plot; more axes capture additional variance.

How many groups to split the rows into. For management zones, choose a small interpretable number (2–5).

The number of decision trees in the Random Forest ensemble. Increasing this generally improves stability and performance at the cost of significantly higher computational time.

The feature-engineering transform to apply: arithmetic (ratio/sum/difference/product/mean), log1p, polynomial interactions, longitudinal summaries, one-hot encoding, or a wide-to-long reshape.

Algorithm used to reduce the samples-by-features table to a few axes. PCA maximises variance on standardised data; PCoA (classical MDS) and NMDS embed a chosen distance/dissimilarity matrix.

The categorical variable DIABLO discriminates. Lists class counts; warns when any class has fewer than ~10 members. Required for DIABLO.

The algorithm used to flag anomalous samples: IQR (simple), IsolationForest, LOF, EllipticEnvelope, OneClassSVM, or DBSCAN.

What to do with detected outliers: remove rows, keep them, winsorize extreme values, or impute with median/mean.

When outlier handling is applied relative to the train/test split. Training-only avoids leaking test information.

Prefix for longitudinal summary columns: <prefix>_mean, _delta, _slope, _min, _max.

Use a paired t-test when the two groups contain matched observations (e.g. the same samples measured before and after treatment). Requires equal group sizes in the same row order.

Creates ratio features between column pairs, exposing proportional relationships that absolute values hide.

Univariate screen measuring the linear association between each feature and the target. Fast and interpretable; assumes roughly normal data.

Shuffles each feature and measures the accuracy drop, giving a model-agnostic estimate of feature contribution.

The number of times to shuffle the feature values to calculate empirical p-values for feature importance. Higher counts provide more precise significance estimates.

Number of label-shuffling permutations used to build the null distribution for the test p-value. More permutations give finer p-value resolution at the cost of runtime.

Highest power/interaction degree generated by the polynomial operation.

The maximum power of generated polynomial terms. Higher degrees grow the feature count rapidly.

Generates interaction and power terms between features up to the chosen degree, allowing the model to capture non-linear relationships.

Estimates an optimal power transform per feature to approximate normality; supports zero and negative values.

Runs Pearson, Spearman, and ANOVA screens before modeling to summarise each feature relationship with the target.

One or more independent variables in the model. Categorical predictors are dummy-coded automatically.

The base dataset in the merge. With a Left join, all of its rows are preserved in the result.

Maps each feature to a uniform or normal distribution via its quantiles. Robust but destroys relative distances.

An integer used to fix the random number generator, ensuring that the results of the analysis are reproducible across multiple runs.

Applies 1/x, strongly shrinking large values and spreading small ones. Useful for highly skewed positive data.

The column identifying samples across blocks. Used to align rows when joining multiple omics layers.

The name given to the edited dataset copy produced by the dataset editor. The original dataset is never modified.

The metric used to rank features during selection: R-squared, F-regression, or mutual information.

Grid search evaluates every parameter combination; random search samples combinations and is faster on large grids.

The second categorical variable in the chi-square test of independence. Its distinct values form the columns of the contingency table.

The dataset joined onto the primary one. Matching columns besides the join key may be suffixed to avoid collisions.

Threshold used by filter-type feature selection methods: minimum variance (Variance), maximum inter-feature correlation (Correlation), or FDR alpha (FDR).

The model whose internal feature weights drive SelectFromModel feature selection.

Enable SHAP (SHapley Additive exPlanations) to interpret model predictions. This provides local explanations for how each feature contributes to individual predictions.

The p-value threshold for declaring a test result statistically significant.

The genome-wide p-value cutoff for calling a SNP-trait association significant after multiple-testing correction.

Columns the operation reads. Ratio and difference take exactly two (first/second); sum, product, mean and longitudinal take two or more; log1p and one-hot apply per selected column.

Select a local data file to upload. Supported formats are CSV, TSV (tab-separated), and Excel (XLSX). Ensure headers are clearly defined in the first row.

Rank-based univariate screen for monotonic (not necessarily linear) relationships. More robust to outliers than Pearson.

Expands features into piecewise polynomial basis functions so linear models can fit smooth non-linear trends.

Applies the square-root transform, a milder variance-stabilising correction than log for count data.

The outcome variable (Y) that you intend to predict or correlate against other features. For regression models, this must be a continuous numeric variable.

Regression predicts a continuous numeric outcome; classification predicts discrete categories.

Numeric columns holding the repeated measurements in a wide table — one column per timepoint, one row per subject. Select them in time order.

A pickled scikit-learn compatible model exported from a previous analysis result. Works for both classification and regression.

Number of cross-validation folds used inside the hyperparameter search loop.

The proportion of the dataset to reserve for the final model evaluation (unseen data). Expressed as a decimal between 0.0 and 1.0.

The numeric column whose distribution is compared across groups by the statistical test.

Fraction of classification predictions that are correct. Can mislead on imbalanced classes — check F1 and the confusion matrix too.

DIABLO's cross-validated error measure: average of per-class error rates, robust to imbalanced classes. Lower is better.

Range that contains the true parameter value with the stated confidence (usually 95%). Narrow intervals mean precise estimates.

Cross-table of true vs predicted classes. Diagonal cells are correct predictions; off-diagonals reveal which classes are confused.

Magnitude of a difference or association independent of sample size (e.g. Cohen's d, eta-squared). Complements the p-value with practical significance.

Amount of variance captured by an ordination axis. The scree plot ranks eigenvalues; axes with large eigenvalues carry most structure.

The objective function MOFA2 maximises during variational inference. A rising, plateauing ELBO trace indicates the factor model has converged.

GSEA's running-sum statistic measuring how concentrated a pathway's genes are at the top or bottom of the ranked feature list.

How uniformly abundance is distributed among features in a community — 1.0 means all categories are equally abundant. Complements richness in diversity indices.

Harmonic mean of precision and recall — a balanced accuracy measure robust to class imbalance. Weighted F1 averages per-class scores by support.

False Discovery Rate adjusted p-value (Benjamini-Hochberg): the expected proportion of false positives among results called significant at that level.

How much each feature contributes to model predictions. Impurity importance is computed inside trees; permutation importance measures the accuracy drop when the feature is shuffled.

Expected genotype frequencies under random mating. SNPs deviating strongly may reflect genotyping error and are filtered by the HWE threshold.

Indicator-species statistic combining specificity (how exclusive a feature is to a group) and fidelity (how consistently it occurs there). Ranges 0–1.

Non-random correlation between nearby SNPs. High-LD SNP pairs are redundant, so pruning keeps one representative per correlated group.

Weights showing how strongly each original feature contributes to an ordination axis or factor. Large absolute loadings identify driving variables.

Average absolute difference between predictions and observations, in target units. More robust to outliers than RMSE.

GWAS figure plotting −log10(p) per SNP across chromosomes; peaks mark candidate loci above the significance threshold line.

Pearson correlation between the entries of two distance matrices. Positive values mean samples close in one space are close in the other.

Frequency of the less common allele at a SNP. Rare variants (low MAF) are filtered because they carry little statistical power.

Average of squared prediction errors. Lower is better; squaring emphasises outliers. RMSE is its square root in target units.

Enrichment score adjusted for pathway size, allowing comparison across pathways; accompanied by a permutation-based FDR q-value.

ORA measure of enrichment: the odds of a pathway gene appearing in your list vs outside it. Values above 1 indicate over-representation.

Probability of observing a statistic at least this extreme under the null hypothesis. Small values (below alpha) are declared significant.

Model-agnostic importance: shuffle one feature, measure the performance drop. Unbiased compared to impurity importance but slower to compute.

Of the samples predicted as a class, the fraction that truly belong to it. High precision means few false alarms.

Observed vs expected −log10(p) quantiles. Early deviation suggests inflation (population structure); deviation at the tail suggests true associations.

Of the samples truly in a class, the fraction the model finds. High recall means few missed positives.

Square root of the mean squared prediction error, in the same units as the target. Lower is better; large errors are penalised more than small ones.

Area under the receiver-operating-characteristic curve: probability the model ranks a random positive above a random negative. 0.5 is chance, 1.0 perfect.

Plot of true-positive rate vs false-positive rate across classification thresholds; the AUC summarises it in one number.

Fraction of outcome variance explained by the model. 1.0 is a perfect fit, 0 means the model is no better than predicting the mean, negative means worse.

Coordinates of each sample on ordination axes or latent factors — the values plotted in the ordination scatter.

Alpha-diversity metric combining richness and evenness; higher values mean more diverse communities. Sensitive to rare features.

Per-prediction feature contribution derived from Shapley values: how much each feature pushed this sample's prediction above or below the average.

Clustering quality from -1 to 1: how similar a sample is to its own cluster vs the nearest other cluster. Above ~0.5 indicates well-separated clusters.

Alpha-diversity metric weighted toward dominant features; reported as 1 − D so higher means more diverse. Robust to rare-feature noise.

Estimated standard deviation of a statistic's sampling distribution — how much the estimate would vary across repeated samples.

NMDS goodness-of-fit: how faithfully the 2-D plot preserves the original distance ordering. Below 0.1 is good, 0.1–0.2 usable, above 0.2 poor.

Percentage of total data variance captured by an ordination axis or latent factor — the axis's eigenvalue divided by the sum of all eigenvalues.

A 128-bit token minted per result row. Token URLs (/results/by-token/<hex>) open exactly that result without a session — safe to share or bookmark.

The inferred type of each dataset column (e.g. float64 for decimal numbers, int64 for whole numbers, object for text). Types decide which analyses a column can serve: numeric for features and targets, categorical for grouping.

A samples-by-species table where each cell is an abundance or cover value. Required by diversity, ordination, and indicator-species methods.

Model evaluation by rotating train/test folds: train on N−1 folds, validate on the held-out fold, average the metrics. Reduces dependence on one lucky split.

A samples-by-features table (CSV, TSV, or Excel) uploaded to the platform. Rows are observations (plots, samples); columns are variables.

Density-based clustering that finds arbitrarily shaped clusters and labels points in low-density regions as noise — usable for outlier detection.

A new dataset produced by preprocessing, feature engineering, merging, or the dataset editor. The original is never modified — chain derived datasets into downstream analyses.

Transforming numeric features to comparable magnitudes — z-score standardisation, min-max, or robust scaling. Required by distance- and gradient-based methods (k-NN, SVM, neural nets, ordination).

Standard vocabulary describing gene function in three namespaces: Biological Process, Molecular Function, and Cellular Component.

Framework where a null hypothesis (no effect) is rejected when the observed statistic is unlikely under it. The p-value quantifies that unlikeliness; significance thresholds and multiple-testing corrections control error rates.

Filling in missing values so analyses can run on complete data. Common strategies: column mean, median, mode, or k-nearest-neighbour estimates.

Outlier detector that isolates samples with random splits — anomalies separate in few splits. Robust default for multivariate outlier detection.

Kyoto Encyclopedia of Genes and Genomes pathway database — curated metabolic and signalling pathway maps used for enrichment testing.

Outlier detector comparing each sample's local density to its neighbours — points in sparse regions score as outliers.

Preprocessing step that fills NaN cells with a column statistic (mean, median, mode) or k-NN estimates so downstream analyses see complete data.

Empty cells (NaN) in a dataset column. Most analyses either drop affected rows or require imputation first; the Missing Values strategy control chooses the behaviour per run.

Testing many hypotheses at once inflates false positives: at alpha 0.05, ~5% of null tests pass by chance. Corrections (Bonferroni, FDR) control the error rate.

Preprocessing step that rescales numeric columns to a common range or distribution so features with large units don't dominate distance-based methods.

One molecular layer in a multi-omics run (e.g. metabolomics, proteomics). Blocks are aligned on a shared sample column before integration.

When a model memorises training noise and fails on new data. Held-out test sets, cross-validation, and tuned hyperparameters guard against it.

A session-independent result URL carrying the access token. Works cookieless, survives logout, and unlocks the result's figures, exports, and reports.

Significance testing by shuffling labels many times to build a null distribution; the p-value is the fraction of shuffles as extreme as the observed statistic.

Applying a previously trained and exported model to new data. Rows are scored through the saved pipeline; when a ground-truth target column is present, evaluation metrics are computed as well.

Open-source pathway database of biological reactions and pathways, complementary to GO and KEGG for enrichment analysis.

Wide tables hold one column per variable/timepoint (COVER_JUN, COVER_JUL); long tables stack measurements in one column with a label column. Feature engineering can reshape between them.

Outlier handling that clips extreme values to a chosen percentile instead of deleting the row — preserves sample count while taming extremes.