Colours
groups up / down heatmaps continuous scales
group colours (click to choose any colour):
Applies to every plot, SVG export and composite figure, and is saved with the session. Colour-blind-safe choices are marked. Choosing a new group palette clears individual group colours. Text colours in on-screen tables are not changed.
Licenses
benchsiDE application code is released under the MIT License. The application embeds the following third-party material. Where marked, the material is a modified subset of the source files: only the columns needed for identifier mapping, links and orthology were kept, and the data were re-encoded to reduce file size. The source data are provided as is, without warranty. Citations for each resource appear in the auto-generated methods text.
| Resource | Use in benchsiDE | License | Modification |
|---|---|---|---|
| Plotly.js 2.32.0 | Plotting library (loaded from the Plotly CDN; embedded in the offline build) | MIT License | Unmodified |
| MSigDB v2024.1: Hallmark, GO BP, GO CC, GO MF, Reactome (human and mouse) | Built-in gene-set libraries | CC BY 4.0 | Re-encoded; set membership unchanged |
| MGI: MRK_ENSEMBL.rpt, MGI_EntrezGene.rpt, HOM_MouseHumanSequence.rpt (The Jackson Laboratory) | Mouse identifier-to-symbol mapping, accession links, mouse-human orthologs, homology families | CC BY 4.0 | Subset: selected columns; re-encoded |
| RGD: GENES_RAT (Medical College of Wisconsin) | Rat identifier-to-symbol mapping, accession links | CC BY 4.0 | Subset: selected columns; re-encoded |
| HGNC: complete set and gene groups (HUGO Gene Nomenclature Committee at the University of Cambridge) | Human identifier-to-symbol mapping; curated discovery modules | CC0 1.0 | Subset: selected columns; re-encoded |
MGI data © The Jackson Laboratory (terms: informatics.jax.org/mgihome/other/copyright.shtml). RGD data © Medical College of Wisconsin (terms: rgd.mcw.edu/wg/disclaimer/). HGNC data: genenames.org/about/license/. MSigDB © Broad Institute, Inc., Massachusetts Institute of Technology, and Regents of the University of California.
1 · Expression matrix (required)
Drop a TSV/CSV here or click to browse.
Rows = genes, columns = samples. Annotation columns (gene id / symbol / biotype) are auto-detected. Raw counts, CPM/TPM, or log-transformed values all accepted.
Rows = genes, columns = samples. Annotation columns (gene id / symbol / biotype) are auto-detected. Raw counts, CPM/TPM, or log-transformed values all accepted.
2 · Experimental design (optional)
Drop a design file here or click to browse.
Either samples as rows (
Either samples as rows (
sample group, one column per factor) or samples as columns
(one row per factor). Without a design file, groups are guessed from sample names and can be edited below.2b · Species & gene symbols (optional)
Species:
Drop a gene-annotation file here to convert Ensembl (or other) gene IDs to official symbols.
3 · Review & analyze
Group by factor:
Normalization:
Design type:
Values are:
Filter:
CPM ≥
in ≥ samples
Edit the design: sample groups and other columns (for example a patient column for paired samples)
Samples included in all statistics (toggle to recompute everything live)
Library size (column totals)
Detected genes per sample (CPM ≥ 1, among filtered genes)
PCA: top 2,000 variable genes, log2 scale (recomputed live) display batch-adjusted:
Library complexity & sample identity
per-sample MA:
Sample dendrogram: correlation distance, average linkage (top 2,000 variable genes)
colour leaves by
cut into clusters
RLE: relative log expression per sample (composition / normalization QC)
expression density per sample (normalization QC)
Each box: distribution of log₂(value) − gene-wise median across samples. Well-normalized samples center on 0 with similar spread; a shifted or wide box flags a normalization or composition problem.
MDS: leading log₂FC distance (limma plotMDS, top 500 genes per pair) (recomputed live)
Power & design guidance
Biotype composition of expressed genes
Sample–sample correlation (Pearson, log₂, all kept genes)
Auto-generated methods text
Group summary
Fold change vs baseline (first group)
ANOVA and Kruskal–Wallis on log₂ values across included samples; groups need ≥2 included samples.
Trend = Spearman correlation with numeric group value (shown only when group labels are numeric, e.g. time or dose).
This gene in the comparison datasets
gene:
Each panel uses that dataset's own normalization, groups and statistics; values are never merged across datasets. Genes are matched by symbol, by ortholog for another species, or through a custom ID map. Type a symbol to look up a gene that is not in the session.
Multi-gene comparison group by statistics
Group A vs Group B
FDR ≤
|log₂FC| ≥
variance from
adjust for
label top up/down by
drag to
p-value histogram (diagnostic)
voom mean–variance trend
Significant genes
Moderated t: pooled gene-wise variance shrunk by empirical Bayes toward a global prior (limma-style eBayes, Smyth 2004); recommended for small groups. voom: per-observation precision weights from the count-level mean–variance trend (Law et al. 2014), then weighted moderated t. Best when counts are low or library sizes vary widely; needs raw counts. Welch t: unequal-variance t-test, no moderation. Both on log₂ values, Benjamini–Hochberg FDR over all kept genes. log₂FC = mean(B) − mean(A). Recomputed live on any change.
Two-factor design: one factor compared at each level of the other
compare vs
All pairwise contrasts
Venn:
Compare two contrasts
X: vs
Y: vs
Interaction: does a difference change between conditions?
Tests the difference of differences (B2 − A2) − (B1 − A1), for example whether the treatment response differs between genotypes or time points. For a two-factor design, combine the two factors into one grouping under Review → group by (for example Treatment · Genotype), so that each group is one combination. Linear model ~ 0 + group (+ covariates), as limma.
reference condition: →
compared with: →
test
variance from
adjust for
FDR ≤ |log₂FC| ≥
Query gene:
show top per direction
Most correlated genes across included samples (click a gene to plot the pair; click a column header to sort)
Pearson r on log₂ values between the query gene and all kept genes. Correlation across a designed experiment reflects shared response, not necessarily direct regulation.
Network hub genes
top variable genes ·
soft power β ·
Network view
top hubs ·
partners per hub ·
edges |r| ≥ ,
strongest per gene ·
Hub neighbourhood
hub
· partners
· correlation
· partner–partner edges |r| ≥
What is in it: gene-set enrichment of the neighbourhood
Trajectory heatmap (numeric designs)
Genome-wide moderated F (any group differs)
FDR ≤
Cluster top
variable genes into
patterns (k-means on per-group mean z-profiles)
Click a pattern line to list its genes
Gene-set module score
1 · Gene-set library (GMT)
Built-in library: MSigDB v2024.1, CC BY 4.0; citations are added to the methods text. KEGG is excluded for license reasons. Your own GMT can be dropped below.
Drop a
Gene symbols or IDs; matching is case-insensitive against this dataset.
.gmt file here or click to browse (e.g. MSigDB Hallmark, GO BP, KEGG; download from MSigDB or Enrichr for your species).Gene symbols or IDs; matching is case-insensitive against this dataset.
2 · Query set & over-representation test
Query:
set size –
permute
Enrichment plot: top sets by p (or click "curve" in the table to choose)
Click a ridge for its barcode and cumulative distribution.
Click a cell for the genes shared by the two sets.
One-sided hypergeometric (Fisher exact) test of query–set overlap; universe = the kept genes annotated in the library; Benjamini–Hochberg FDR across tested sets. DE-based queries use the current Differential Expression contrast, method, and thresholds.
Automated discovery: coordinated modules the per-gene tests miss
How discovery works and how to read the results
Per-gene tests require each gene to reach significance on its own. A set of related genes that all shift modestly in the same direction can therefore be missed entirely, even when the shift is consistent across the whole set.
This screen builds candidate modules from two sources: gene families (shared symbol prefix; Skint1…Skint9 → Skint*) and the loaded gene-set library. Each module is summarized per sample as the mean z-scored expression of its members (the module score). For two-group designs the group effect on that score is tested with CAMERA (Wu & Smyth 2012, as in limma), a competitive test whose variance-inflation factor accounts for correlation between module members; random-set permutation nulls understate that correlation and over-call co-regulated families. For designs with more than two groups a permutation test against random same-size gene sets is used instead, with each module’s inter-gene correlation displayed as a caution. BH correction is applied across all modules screened; because modules overlap (families nest inside GO sets), the BH assumption holds under positive dependence (PRDS) rather than independence.
In each result card: Δz is the module-score difference between groups (ANOVA F for >2 groups); z is the number of null standard deviations above random same-size modules; “n individually significant” counts members that pass per-gene FDR on their own. The hidden module tag marks modules significant as a module (FDR ≤ 0.05) while fewer than 25% of members are individually significant; when no module survives FDR (common for small group sizes), the tag falls back to nominal p ≤ 0.01 and is labeled “(nominal)”. A leave-one-out note appears when a module’s direction depends on a single sample.
This is a screening tool, not a confirmatory test: treat hits as ranked hypotheses and follow up in the Gene Explorer and the heatmap export. Family membership is based on symbol structure alone; a shared prefix (Ddx*, for example) can be structural rather than functional.
This screen builds candidate modules from two sources: gene families (shared symbol prefix; Skint1…Skint9 → Skint*) and the loaded gene-set library. Each module is summarized per sample as the mean z-scored expression of its members (the module score). For two-group designs the group effect on that score is tested with CAMERA (Wu & Smyth 2012, as in limma), a competitive test whose variance-inflation factor accounts for correlation between module members; random-set permutation nulls understate that correlation and over-call co-regulated families. For designs with more than two groups a permutation test against random same-size gene sets is used instead, with each module’s inter-gene correlation displayed as a caution. BH correction is applied across all modules screened; because modules overlap (families nest inside GO sets), the BH assumption holds under positive dependence (PRDS) rather than independence.
In each result card: Δz is the module-score difference between groups (ANOVA F for >2 groups); z is the number of null standard deviations above random same-size modules; “n individually significant” counts members that pass per-gene FDR on their own. The hidden module tag marks modules significant as a module (FDR ≤ 0.05) while fewer than 25% of members are individually significant; when no module survives FDR (common for small group sizes), the tag falls back to nominal p ≤ 0.01 and is labeled “(nominal)”. A leave-one-out note appears when a module’s direction depends on a single sample.
This is a screening tool, not a confirmatory test: treat hits as ranked hypotheses and follow up in the Gene Explorer and the heatmap export. Family membership is based on symbol structure alone; a shared prefix (Ddx*, for example) can be structural rather than functional.
Screen:
module size –
permutations
Top
· rows:
k
· columns:
Sample traits: genes, gene sets and modules associated with a per-sample measurement
What this does and how to set it up
A trait is a numeric value per sample that is the variable of interest rather than a nuisance: a clinical score (PASI), cell counts from immunohistochemistry of the same section, epidermal thickness, a serum level. Add it as a column of the design file (or with Edit the design → Add column); any numeric column with three or more distinct values is offered here. Missing values (blank or NA) leave only that sample out.
Model. Each gene's log₂ expression is regressed on the trait (limma with an intensity trend). Effects are reported per unit of the trait and per standard deviation (SD) of it. Adjust for group tests association within groups: use it whenever the trait differs between groups (PASI is 0 in healthy controls), otherwise the result largely repeats the group comparison. Alternatively, keep only one group (for example lesional samples).
Repeated samples. When a subject or animal contributes several samples (several biopsies, both lesional and non-lesional skin), select the column that identifies it. If the trait is constant within each subject (PASI, age), each subject is summarized by its mean (block means), which keeps the test valid; if it varies between a subject's samples (IHC of each section), a mixed model with limma's duplicateCorrelation is used. Ignoring repeated samples gives false positives.
Transformations. log₂(x + 1) for counts, logit for proportions or percentages, rank for skewed values or outliers (a few extreme samples can otherwise determine a slope; the table flags genes where one sample dominates, Cook's distance > 1).
Model. Each gene's log₂ expression is regressed on the trait (limma with an intensity trend). Effects are reported per unit of the trait and per standard deviation (SD) of it. Adjust for group tests association within groups: use it whenever the trait differs between groups (PASI is 0 in healthy controls), otherwise the result largely repeats the group comparison. Alternatively, keep only one group (for example lesional samples).
Repeated samples. When a subject or animal contributes several samples (several biopsies, both lesional and non-lesional skin), select the column that identifies it. If the trait is constant within each subject (PASI, age), each subject is summarized by its mean (block means), which keeps the test valid; if it varies between a subject's samples (IHC of each section), a mixed model with limma's duplicateCorrelation is used. Ignoring repeated samples gives false positives.
Transformations. log₂(x + 1) for counts, logit for proportions or percentages, rank for skewed values or outliers (a few extreme samples can otherwise determine a slope; the table flags genes where one sample dominates, Cook's distance > 1).
trait
transform
samples:
adjust for –
repeated samples per
Gene sets and modules
modules
Cross-dataset comparisons here operate on within-dataset statistics only (fold changes, ranks, module scores). Raw expression values are never merged, pooled, or renormalized across datasets, and no p-value is ever computed on expression pooled across datasets: merging two studies' counts requires cross-batch assumptions this tool will not silently make.
Comparison datasets (this session is dataset A; add as many as memory allows)
species of next dataset:
Pairwise comparison
X: · Y:
Multi-dataset summary
forest plot:
Meta-analysis: shared signatures across datasets
model:
test:
gene must be matched:
meta FDR ≤
|pooled log₂FC| ≥
I² ≤ %
pathway enrichment of
heatmap: genes, ranked by
multi-gene forest:
genes
table:
rows
Homology families: many-to-many orthologs, scored as modules
X: Y:
family dot plot:
families
Hub across datasets
1. Is the neighbourhood preserved?
2. Is it a hub in every dataset?
3. Does it rewire between conditions?
4. Members in the meta-analysis
Signature transfer and replication: score each comparison dataset with a signature from the session
Signature: