DEMO DATA: simulated. These counts were generated by a computer (negative binomial simulation) to test the dashboard. They are not real biological measurements.
1. Data import
Load an expression matrix (genes in rows, samples in columns) and a sample metadata table. Files are read by this page only; nothing is uploaded.
Your files
First column: gene ID or name. Other columns: one per sample. See the format.
One row per sample: sample name, group/condition, and any other columns (batch, sex, time...).
The demo is simulated data: 2,000 genes, 3 groups with 4 samples each, with known differentially expressed genes.
Validation report
Choose your files and press "Validate and load files", or try the demo.
2. Quality control
Check that every sample looks reasonable before any comparison: sequencing depth, detected genes, value distributions and missing values.
Load data first.
3. Filtering and normalization
Remove genes with too little signal to be tested reliably, then make samples comparable by correcting for sequencing depth and composition.
Load data first.
4. Exploratory analysis
Do samples of the same group look alike? Are there outliers or batch effects? PCA, correlation and clustering answer these questions before any statistical test.
Load data first.
5. Differential expression
Find genes whose expression differs between two groups.
Load data first.
6. Gene explorer
Look at any gene sample by sample, see its statistics, and compare several genes.
Load data first.
7. Enrichment and Venn diagram (optional)
Compare the gene lists of several comparisons, and test which biological pathways are over-represented in a gene list.
Load data first.
8. Export and report
Download tables, and a one-file HTML report with the main plots and every parameter used.
Load data first.
Help: how to use, methods and references
Everything this dashboard calculates, explained in simple words.
How to use the dashboard
Data import. Upload your expression matrix and your sample metadata, or press Load demo data. Read the validation report: red messages must be fixed, orange ones are warnings.
Quality control. Check library sizes, detected genes and distributions. A sample that looks very different from the others may be a failed library.
Filter & normalize. Keep genes with enough reads, then choose a normalization (TMM is a good default for raw counts).
Exploratory analysis. Look at the PCA, the correlation heatmap and the clustering. Replicates of the same group should be close together.
Differential expression. Choose two groups, set the fold change and FDR cut-offs, and read the volcano plot, MA plot and table. Click a gene to open it.
Gene explorer. Search any gene, see its values per sample, and open NCBI, Ensembl or GeneCards.
Enrichment & Venn (optional). Compare gene lists of several comparisons. Enrichment sends gene names to the Enrichr service only after you allow it.
Export. Every chart has PNG and SVG buttons, every table a CSV button, and the report page builds a full HTML summary.
Expected file formats
Expression matrix
One row per gene, one column per sample. The first row is the header. The first column holds the gene ID (Ensembl ID or gene symbol). Separators: comma, tab, semicolon or spaces are detected automatically.
The column called group, condition or treatment is used for comparisons (you can choose another one). Example files are in the sample_data folder of the project.
Methods, step by step
These steps follow the standard RNA-seq workflow described by Love et al. (DESeq2 workflow), Law et al. (limma, Glimma, edgeR workflow) and Conesa et al. (best practices). The statistics here are simplified so they can run in a browser.
Library size and detected genes
The library size of a sample is the sum of all its counts (its sequencing depth). Detected genes are genes with a value above zero. Big differences between samples are normal and are corrected by normalization, but an extreme sample deserves a closer look.
Counts per million (CPM)
CPM = count / library size x 1,000,000. It puts samples with different depth on the same scale. log2(CPM + 1) adds 1 to avoid log(0), and makes the data more symmetric so that a few very high genes do not dominate.
Filtering lowly expressed genes
A gene is kept if its CPM is at least the threshold in at least N samples. The default follows the idea of edgeR's filterByExpr: the threshold corresponds to about 10 reads in a sample of median library size (CPM = 10 / median library size in millions), and N is the size of the smallest group. Library sizes are recalculated on the kept genes. Filtering removes genes that cannot be tested reliably and makes the multiple testing correction less severe.
TMM normalization
Trimmed Mean of M-values (Robinson and Oshlack, 2010), re-implemented following edgeR's defaults. One sample is chosen as reference (the one whose upper quartile is closest to the average). For every other sample, the log ratios (M) of genes against the reference are computed; the 30% most extreme ratios and the 5% most extreme average expressions are trimmed, and a weighted mean of the remaining ratios gives a factor. Factors are scaled so that their product is 1. The normalized value is log2(count / (library size x factor) x 10^6 + 1). TMM corrects for composition: when a few genes take a large share of the reads in one sample, the other genes look artificially lower.
Median of ratios
The DESeq/DESeq2 method (Anders and Huber, 2010). For each gene, a geometric mean across samples is the pseudo-reference (genes with a zero in any sample are skipped). The size factor of a sample is the median of its ratios count / pseudo-reference. Normalized counts = count / size factor, then log2(x + 1).
DESeq2 itself uses variance stabilizing transformations (VST, rlog) for PCA and heatmaps. Here log2(x + 1) is used instead because it is simple and fast; low-count genes can look more variable than with VST.
PCA (principal component analysis)
Computed on the log2 values of the most variable genes (500 by default, adjustable). Each gene is centered (its mean is subtracted) but not scaled. PC1 is the direction with the largest variation between samples, PC2 the second, and so on. The percent shown on each axis is the share of the total variance it explains.
Sample correlation and clustering
Pearson correlation between every pair of samples, on log2 values of all kept genes. Hierarchical clustering joins the two closest samples (or groups of samples) step by step; the dendrogram shows the order and the distance at which they were joined. Distance: Euclidean (default) or 1 minus correlation. Linkage: average (default), complete or single.
Heatmap of the most variable genes
The N genes with the highest variance of log2 values are selected. Each row is converted to a z-score (value minus the gene mean, divided by the gene standard deviation), so colors show whether a sample is above (red) or below (blue) the gene's own average. Genes and samples are clustered (Euclidean distance on z-scores).
Differential expression (simplified)
For each gene, on the log2 normalized values: mean of each group, log2 fold change = mean(test group) minus mean(reference group), and a Welch two-sample t-test (does not assume equal variances, Welch 1947). Two-sided p values come from the t distribution. P values are corrected for testing thousands of genes with the Benjamini-Hochberg method; the adjusted value (FDR) estimates the share of false positives among genes called significant. A gene is "up" when log2FC is at least the cut-off and FDR is below its cut-off, "down" when log2FC is at most minus the cut-off.
This is not DESeq2, edgeR or limma. Those methods model counts with the negative binomial distribution or with precision weights, and share information between genes (empirical Bayes), which gives much better results with few replicates. A Welch t-test with 3 to 4 replicates per group has low power and is less stable. Use this dashboard to explore; use DESeq2, edgeR or limma in R for results you publish.
Enrichment (optional)
The gene list is sent to Enrichr (Chen et al., 2013; Kuleshov et al., 2016), which tests over-representation of gene sets with Fisher's exact test and reports adjusted p values and a combined score. It only works with real gene symbols (for example TP53), not Ensembl IDs. Conesa et al. note that over-representation tests can be biased by gene length in RNA-seq.
References
Love MI, Anders S, Kim V, Huber W. RNA-Seq workflow: gene-level exploratory analysis and differential expression. F1000Research 2015, 4:1070 (version 2, 2016). doi:10.12688/f1000research.7035.2
Law CW, Alhamdoosh M, Su S, Dong X, Tian L, Smyth GK, Ritchie ME. RNA-seq analysis is easy as 1-2-3 with limma, Glimma and edgeR. F1000Research 2016, 5:1408 (version 3, 2018). doi:10.12688/f1000research.9005.3
Conesa A, Madrigal P, Tarazona S, et al. A survey of best practices for RNA-seq data analysis. Genome Biology 2016, 17:13. doi:10.1186/s13059-016-0881-8
Robinson MD, Oshlack A. A scaling normalization method for differential expression analysis of RNA-seq data. Genome Biology 2010, 11:R25. doi:10.1186/gb-2010-11-3-r25
Anders S, Huber W. Differential expression analysis for sequence count data. Genome Biology 2010, 11:R106. doi:10.1186/gb-2010-11-10-r106
Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society B 1995, 57(1):289-300. doi:10.1111/j.2517-6161.1995.tb02031.x
Welch BL. The generalization of "Student's" problem when several different population variances are involved. Biometrika 1947, 34:28-35. doi:10.1093/biomet/34.1-2.28
Su S, Law CW, Ah-Cann C, Asselin-Labat ML, Blewitt ME, Ritchie ME. Glimma: interactive graphics for gene expression analysis. Bioinformatics 2017, 33(13):2050-2052. doi:10.1093/bioinformatics/btx094
Ge SX, Son EW, Yao R. iDEP: an integrated web application for differential expression and pathway analysis of RNA-Seq data. BMC Bioinformatics 2018, 19:534. doi:10.1186/s12859-018-2486-6
Chen EY, Tan CM, Kou Y, et al. Enrichr: interactive and collaborative HTML5 gene list enrichment analysis tool. BMC Bioinformatics 2013, 14:128. doi:10.1186/1471-2105-14-128
Kuleshov MV, Jones MR, Rouillard AD, et al. Enrichr: a comprehensive gene set enrichment analysis web server 2016 update. Nucleic Acids Research 2016, 44:W90-W97. doi:10.1093/nar/gkw377
A static web page (HTML, CSS, JavaScript) that runs entirely in your browser. There is no server, no account and no database. Your files are read locally and never uploaded. The only exception is the optional enrichment step, which sends gene names (and nothing else) to the public Enrichr service, and only after you tick the consent box and press the button.
Libraries (stored in the vendor folder, work offline)
Plotly.js 4.3.0 (MIT license) for interactive charts.
PapaParse 5.7.1 (MIT license) for reading CSV/TSV files.
Inter font (SIL Open Font License).
Limitations
Differential expression uses a Welch t-test on log2 values, not a count model with information sharing between genes. Results are less reliable than DESeq2, edgeR or limma, especially with few replicates.
Only two-group comparisons. No design with covariates (for example batch + treatment), no paired tests.
log2(x + 1) is used instead of VST or rlog for exploratory plots.
No gene ID conversion: enrichment needs gene symbols.
Tested with 20,000 genes x 30 samples. Much larger data may be slow, depending on the computer.