Example datasets

The walkthrough uses these public GEO series. They are also the examples shown under the GSE ID boxes in Step 1.

The four analysis types

Choose one under Select Analysis Platform in Step 1. The table shows which box to fill for each type, using the example IDs above.

For a single-study run, choose Single dataset (1 GSE) with RNA-seq only or microarray only (e.g. just GSE144119 or just GSE268456); Step 5 batch correction is then skipped.

Step 1 — Download Data

Sidebar: 1. Download Data

Goal: Download GEO series, map row IDs to gene symbols, and keep per-GSE matrices. Merged (Both) intersects genes after Step 2 normalization. Parallel DE keeps each platform’s genes until Step 7.

What to do:

1. Optionally enter a Disease / Condition label (used in reports and saved workspace names).
2. Under Select Analysis Platform, choose one of four types:
   • RNA-seq — counts; DESeq2 / edgeR / voom or limma on TMM.
   • Microarray — platform normalization + limma.
   • Merged (Both) — per-study normalize, common genes, one joint batch, one limma DE (original merged workflow).
   • Parallel DE, then merge — two separate pipelines through DE (RNA-seq engine + microarray limma). Step 7 keeps same-direction overlap. Step 8 is one WGCNA on one processed platform matrix.
3. Datasets: RNA-seq-only and microarray-only can use Single dataset (1 GSE) (skips Step 5) or Multiple datasets. Merged and Parallel always keep every GSE typed in each box (several RNA-seq studies and several microarray studies in one run).
4. Select the DE method for Step 6 (this also controls Step 2):
   • limma — microarray, merged, or normalized log-expression.
   • limma-voom, DESeq2, or edgeR — RNA-seq integer counts.
   • Parallel: RNA-seq method on the left; microarray is always limma.
5. Enter comma-separated GSE IDs. RNA-seq and microarray each have their own box, e.g. GSE144119, GSE100026 in RNA-seq Datasets and GSE268456, GSE47927 in Microarray Datasets (see the four analysis types for which boxes each type uses).
6. Click Start Processing and read the Download log.
7. Click Next: Normalize Data.

A GEO series measured on several platforms (e.g. GSE18123: GPL570 and GPL6244) is stored on GEO as one matrix file per platform; GExPipe downloads all of them and keeps the platform with the most samples.

For multiple microarray platforms, install org.Hs.eg.db and biomaRt (see Installation) so STEP 2b converts probe IDs to HGNC symbols so overlap and later merge use gene symbols.

Success check: Download log lists samples per GSE; STEP 2b runs when needed; row names are symbols (TP53, BRCA1, …). Merged reports common genes; Parallel keeps each platform’s full gene list until Step 7.

Step 1: Select one of four analysis types, DE method, and GSE IDs (several per box).

Step 2 — Normalize Data

Sidebar: 2. Normalize Data

Goal: Normalize each GSE on its native scale. Merged then keeps the common genes (and, by default, applies one global quantile so RNA-seq and microarray share one scale); Parallel keeps each platform separate. Step 5 removes the remaining between-study batch.

What to do:

Count-based DE (DESeq2 / edgeR / limma-voom from Step 1): the panel shows Normalization Auto-Handled — continue to Next: QC & Visualization. Each RNA-seq GSE is TMM-normalized (Merged intersects common genes; Parallel does not). DESeq2/edgeR/voom still use raw counts.

limma path (microarray or merged):

1. Choose Auto or Manual under Normalization Strategy.
2. Microarray: Quantile or RMA per GSE; RNA-seq: TMM + log2-CPM per GSE.
3. Merged only: Apply global quantile is on by default so the RNA-seq and microarray studies share one scale for the single limma DE. Parallel never applies it (each platform is normalized on its own).
4. Click Apply Normalization. Merged intersects common genes here; Parallel keeps two separate matrices (no global quantile).
5. Scroll to Normalization Quality Assessment (panels a and b below):
   • Overall density and Q–Q plot (panel a).
   • Median/range alignment and distribution overlap (panel b).
6. Optional: under Download Before / After Normalization Data, download a ZIP with the input matrices (before), the normalized matrices (after, per dataset and per platform or combined), per-sample and per-dataset check tables and a README; or the check table alone as CSV. After a correct normalization Likely_log_scale is TRUE and Median_SD_across_samples falls towards 0.
7. Click Next: QC & Visualization.

Success check: Within each dataset, sample medians align. Mixed-platform blocks remaining after Step 2 are expected; Step 5 removes study/platform batch. Parallel shows two columns (RNA-seq left, microarray right).

Step 2a: Overall density and Q-Q plots after normalization.

Step 2b: Median/range alignment and distribution overlap.

Step 3 — QC & Visualization

Sidebar: 3. QC & Visualization

Goal: Check gene overlap, inspect expression per platform after Step 2, and detect outlier samples within each GSE. Parallel DE does not require a gene intersection here.

What to do:

1. Review Venn and UpSet plots (figure below, panel a). In Parallel mode RNA-seq (left, blue) and microarray (right, orange) have separate boxes; two datasets with identical gene sets are shown as one circle marked “100% overlap”.
2. Open Quality Control Plots — Boxplot and Density tabs on raw expression (panel b).
3. In Sample Outlier Detection, click Run Outlier Detection (panel c). PCA (Mahalanobis distance) and sample connectivity are computed on the normalized matrix — the combined matrix in Merged mode, each platform separately in Parallel mode (at least 5 samples per platform). The plots and outlier tables appear after the run; each table has a CSV download.
4. Exclude flagged samples if needed; keep ≥ 3 samples per group after Step 4.
5. Review Data Summary (datasets, samples, genes).
6. Proceed to Step 4 (Select Groups).

Success check: Overlap and QC plots look reasonable; outliers reviewed.

Step 3a: Venn and UpSet gene overlap.

Step 3b: Expression boxplot.

Step 3c: PCA and connectivity outlier detection.

Step 4 — Select Groups

Sidebar: 4. Select Groups

Goal: Assign each sample to Normal, Disease, or Exclude.

What to do:

1. Browse Phenodata Browser for each GSE.
2. Under Select Phenotype Columns (panel a), pick the condition column per dataset (e.g. characteristics_ch1, or title when treatment is only in sample titles).
3. Click Extract Groups from Selected Columns.
4. Categorize Groups — map labels to reference (Normal) or comparison (Disease), or None.
5. Click Apply Groups.
6. In Group Summary (panel b), optionally rename groups (e.g. Control / Treated); click Update, or leave as Normal / Disease and continue.
7. Confirm sample counts, then click Next: Batch Correction.

Success check: Group Summary counts match your study design.

Step 4a: Select phenotype column per GSE.

Step 4b: Group Summary after Apply Categorization.

Step 5 — Batch Correction

Sidebar: 5. Batch Correction

Goal: Remove technical variation between datasets while keeping biological Normal vs Disease signal.

What to do:

1. Read Dataset × Condition confounding — if dataset and condition are confounded, use limma or SVA instead of ComBat.
2. Review Gene variance distribution.
3. Set Variance percentile cutoff (panel a, default 25%).
4. Choose Batch correction method (panel a, default ComBat-ref): ComBat-ref, SVA, limma, ComBat, Quantile + limma, or Hybrid.
5. Click Apply Batch Correction.
6. Compare before vs after PCA by dataset (panel b), plus PCA by condition, hierarchical clustering, and PVCA on the same panel.
7. Check Batch effect check (before vs after) in the download box. For each corrected matrix it reports the % of top-PC variance explained by batch (dataset) and by condition, PC1 batch R², silhouette by batch and the % of genes with a significant batch effect, with a verdict (OK below 10% batch variance; warning above 25%; over-correction warning if the condition signal collapses). Download the check table (CSV) or a ZIP with matched before/after matrices, sample metadata and a README.
8. Click Next: Differential Expression.

Note: With only one GSE loaded in Step 1, there is no between-study batch to remove, so cross-study correction is skipped by default - but a single dataset can still carry a technical batch effect from sample processing (extraction date, sequencing lane, processing day). Step 5 shows a diagnostic panel for this case: pick a candidate technical covariate from that GSE’s own phenodata, compare PCA colored by the covariate versus by Condition, and either confirm there is no batch effect or apply limma::removeBatchEffect() using that covariate before continuing.

Success check: After-correction PCA (panel b) shows datasets intermingled, and the batch-effect check reports a falling batch share with the condition signal kept.

Step 5a: Variance filtering and batch correction method.

Step 5b: PCA by dataset before and after batch correction.

Step 6 — Differential Expression Analysis

Sidebar: 6. Differential Expression Analysis

Goal: Identify genes differing between Normal and Disease.

What to do:

1. Confirm the DE method banner matches your Step 1 choice (limma, limma-voom, DESeq2, or edgeR).
2. Set DE Parameters (panel a):
   • LogFC cutoff (default 0.5; e.g. 1.0 for 2-fold change).
   • Adj. P-value (default 0.05, Benjamini–Hochberg FDR).
   • Heatmap Genes (number of top DEGs for the heatmap, default 50).
3. Click Run DE Analysis.
4. Read Statistical model and Pipeline verification — design formula, filterByExpr gene counts, samples used, and batch-in-model note.
5. Review results (panel b):
   • Volcano plot (up / down / not significant).
   • Top DEGs table and heatmap below.
6. Optionally download Analysis report (TXT).
7. Click Next: WGCNA Analysis (RNA-seq only, microarray only, or Merged) or Next: RNA-seq ∩ microarray (Parallel DE).

Parallel DE runs two engines with one click (RNA-seq on the left, microarray limma on the right). Merged (Both) still uses one joint limma on the combined matrix.

Success check: Volcano shows expected direction; DEG count is plausible for your contrast; pipeline verification lists the correct method and samples.

Step 6a: DE parameters, statistical model, and pipeline verification.

Step 6b: Volcano plot and top DEG table.

Step 7 — RNA-seq ∩ microarray (Parallel DE only)

Sidebar: 7. RNA-seq ∩ microarray

Goal: Keep genes that are significant on both platforms with the same fold-change direction. This is the first merge after Parallel DE. RNA-seq only, microarray only, and Merged (Both) skip this tab.

What to do:

1. Confirm both Step 6 DE tables exist (RNA-seq left, microarray right).
2. Click Apply to store the same-direction overlap (Consensus DEGs).
3. Review the Venn: Common is the two-circle overlap; Consensus is the list stored for Step 9.
4. Click Next: WGCNA Analysis.

Success check: Consensus gene count is greater than 0 and matches the same-direction overlap, not the union of all DEGs.

Step 8 — WGCNA Network Analysis

Sidebar: 8. WGCNA Analysis

Goal: Find co-expression modules associated with disease and extract hub genes from significant modules.

What to do:

The WGCNA tab has internal sub-steps; complete them in order:

1. Data Preparation & QC (panel a):
   • Choose All genes or Top variable genes (e.g. 5000).
   • Set Min samples per gene fraction (default 0.5).
   • Click Prepare Data.
   • Check the sample dendrogram; exclude outliers or click No outliers — proceed to Step 2 (within this tab).
2. Pick Soft Threshold Power (panel b):
   • Click Calculate Power; choose power where scale-free R² ≥ 0.80.
3. Network Construction & Module Detection (panel c):
   • Set minimum module size and merge cut height.
   • Click Build Network; inspect the module dendrogram and module colours.
4. Module–trait relationships (panel c):
   • Click Calculate Correlations & GS/MM; review the heatmap of module–condition correlations.
5. Identify significant modules (panel d):
   • Click Identify Significant Modules; note hub genes (GS/MM filters).
   • Optionally run GS vs MM plots for all modules.
6. Click Next: Common Genes (DEG & WGCNA) (Step 9).

Success check: Soft-threshold R² ≥ 0.80; at least one module correlates with Disease; hub genes are listed for significant modules.

Step 8a: WGCNA data preparation and gene selection.

Step 8b: Soft-threshold power selection.

Step 8c: Module dendrogram and module-trait heatmap.

Step 8d: Significant modules and hub gene summary.

Step 9 — Common Genes (DEG & WGCNA) & Enrichment

Sidebar: 9. Common Genes (DEG & WGCNA)

Goal: Intersect DEGs (Step 6, or Step 7 consensus in Parallel) with genes in significant WGCNA modules (Step 8), then run GO/KEGG enrichment on the overlap.

What to do:

1. Click Compute Common Genes (panel a).
2. Review the summary counts and Venn diagram: DEG ∩ WGCNA.
3. Download the common-gene table (CSV) if needed.
4. Click Run GO Enrichment (panel b) on the common genes — dot plots for BP, MF, and CC.
5. Click Run KEGG Enrichment (panel c) — bar plot and chord diagram.
6. Choose next path at the bottom of the tab:
   • Path 1: PPI Interaction → then ML, or
   • Path 2: Direct to Machine Learning.

Requirements: Step 6 DE (and Step 7 if Parallel) plus WGCNA Identify Significant Modules must be complete.

Success check: Venn overlap > 0; GO/KEGG plots show interpretable terms.

Step 9a: Common genes Venn diagram (DEG intersect WGCNA).

Step 9b: GO enrichment of common genes.

Step 9c: KEGG enrichment of common genes.

Step 10 — PPI Interaction

Sidebar: 10. PPI Interaction

Goal: Map common genes onto a STRING protein–protein interaction network and identify hub proteins.

What to do:

1. Set STRING score threshold (default 400; 150–900).
2. Click Run PPI Analysis (panel a).
3. Review Interactive vs non-interactive gene counts and table.
4. Choose genes for network view: Hub genes only, Top N by degree, or Select genes manually; click Run.
5. Inspect PPI network layouts (panel b) and the hub score and consensus hub tables. Hub Scores (CSV) saves PPI_Hub_Gene_Scores.csv (all nodes, six centrality measures); Consensus Hubs (CSV) saves Consensus_Hub_Genes.csv (genes in ≥ 3 methods, which methods, and their scores).
6. Click Extract Data for ML (required before Step 11).
7. Click Next: Machine Learning Process.

Requirements: Step 9 common genes computed.

Success check: Network builds with mapped edges; hub genes listed; ML extract succeeds.

Step 10a: Run PPI and interactive vs non-interactive gene table.

Step 10b: PPI network graph and hub gene metrics.

Step 11 — Machine Learning Process

Sidebar: 11. Machine Learning Process

Goal: Rank genes by predictive importance using multiple ML algorithms and find genes selected by ≥ 2 methods.

What to do:

1. Confirm Extract Data for ML was run on the PPI tab (Step 10).
2. Under Run ML Analysis (panel a), select one or more methods: LASSO, Elastic Net, Ridge, Random Forest, SVM-RFE, Boruta, sPLS-DA, XGBoost+SHAP.
3. Set Top N per method where shown.
4. Click Run ML Analysis.
5. Review Venn diagram (selected methods) and the common ML genes table (panel b) — genes in ≥ 2 methods form the candidate panel.
6. Inspect per-method importance plots (RF, Elastic Net, etc.).
7. Click Continue to Validation Setup.

Note: If glmnet fails after install, restart R and relaunch the app.

Success check: At least one common ML gene; Venn overlap across methods is non-empty.

Step 11a: Select ML methods and run analysis.

Step 11b: Method overlap Venn diagram and common ML genes.

Step 12 — Validation

Sidebar: 12. Validation Setup

Goal: Test generalization on an external GEO cohort or an internal 70/30 train–test split. Target accuracy ≥ 70%.

What to do:

1. Choose Validation Strategy (panel a):
   • External Validation — download an independent GSE, categorize groups, and run DE on the validation cohort.
   • Internal Validation — use a 70/30 stratified split of your current data (no extra download).
2. External path only:
   • Enter validation GSE IDs (e.g. GSE162462 for RNA-seq or GSE13159 for microarray), select Platform Type and DE Method (limma, or DESeq2/edgeR when the platform is RNA-seq with raw counts).
   • Click Download - this normalizes the validation data with the same per-dataset method used in Step 2 (quantile for microarray, TMM for RNA-seq) and applies the same gene ID -> symbol conversion as Step 1, so validation gene identifiers and counts match the training pipeline for the same GSE.
   • Browse phenodata and select the group column.
   • A Batch Correction panel appears next: with 2+ validation datasets it runs the same auto-selected ComBat-ref/limma correction as Step 5; with a single dataset it offers the same optional technical-covariate correction described in Step 5.
   • Categorize labels as Normal / Disease, set LogFC cutoff and Adj. P-value for this validation-set DE (defaults 0.5 / 0.05, independent of Step 6’s thresholds), and click Run Validation DE
     • this uses the same DE engines (gexpipe_run_count_de / gexpipe_run_limma_on_subset) as Step 6, including a Dataset covariate when 2+ validation datasets are combined.
3. Review validation status and DE results (panel b).
4. Click Continue to ROC Curve Analysis.

Requirements: Step 11 ML analysis complete with at least one common ML gene.

Success check: External mode shows downloaded samples and validation DE results; internal mode confirms the 70/30 split is ready for ROC and nomogram.

Step 12a: Validation mode and external validation settings.

Step 12b: External validation DE results.

Step 13 — ROC Analysis

Sidebar: 13. ROC Curve Analysis

Goal: Per-gene AUC with 95% CI (pROC). Retain genes with AUC ≥ 0.80 in training and validation.

What to do:

1. Review AUC Summary — Training Data (panel a): table and bar plot for common ML genes with AUC ≥ 0.80.
2. Inspect ROC Curves — Training Data and per-gene expression boxplots.
3. If external validation was run in Step 12, compare training vs validation AUC in the summary and external ROC panels (panel b).
4. Review Multi-variable ROC (combined biomarker panel vs best single gene).
5. Under Select Final Biomarker Genes, choose genes to carry forward to nomogram and GSEA (defaults to all common ML genes if none selected).
6. Click Continue to Diagnostic Nomogram, Continue to GSEA Analysis, or View Results Summary.

Requirements: Step 11 common ML genes; Step 12 validation mode selected.

Success check: At least one gene with training AUC ≥ 0.80; validation AUC reported when external data are loaded.

Step 13a: Training ROC curves and AUC summary.

Step 13b: Training vs validation AUC comparison.

Step 14 — Nomogram

Sidebar: 14. Diagnostic Nomogram Model

Goal: Build a points-based clinical score (rms::lrm()), calibration curve, and decision curve analysis.

What to do:

1. Confirm the validation mode banner (external or internal 70/30 split).
2. Click Run Nomogram Analysis.
3. Review outputs in order:
   • Panel A: nomogram (points scale).
   • Panel B: training and validation ROC curves.
   • Panel C: calibration curves.
   • Panel D: decision curve analysis (DCA).
   • Panel E: clinical impact curves.
4. Inspect Model Diagnostics (VIF, coefficients) and Performance Comparison (training vs validation).
5. Click Continue to GSEA Analysis or View Results Summary.

Requirements: Batch correction (Step 5), group metadata, and common ML genes from Step 11 (or common DEG ∩ WGCNA genes from Step 9).

Success check: Nomogram renders; training AUC is reasonable; validation metrics match your chosen validation strategy.

Step 14: Nomogram, calibration, and decision curves.

Step 15 — GSEA

Sidebar: 15. GSEA Analysis (Signature Genes)

Goal: Pathway-level enrichment (GO, KEGG, MSigDB Hallmark/C2/C5) stratified by final gene expression.

What to do:

1. Select one or more Gene Set Collection(s) (Hallmark, GO BP/MF, KEGG, Reactome, etc.).
2. Optionally enter Custom target genes; leave blank to use common ML genes (or genes selected in Step 13).
3. Click Run GSEA.
4. For each target gene, review the enrichment plot and pathway list in the per-gene panels below.
5. Click Continue to Results Summary.

Requirements: Batch-corrected expression and common ML genes from Step 11 (or Step 9 overlap).

Success check: GSEA completes for each target gene; enrichment plots and pathway tables are interpretable.

Step 15: GSEA enrichment plots per signature gene.

Step 16 — Summary Report

Sidebar: 16. Results Summary

Goal: Review a plain-text overview of the whole pipeline and its key result counts, in order.

What to do:

1. Read the Pipeline summary narrative at the top - one paragraph covering genes/samples processed, normalization, batch correction, DE (significant/up/down genes), WGCNA modules, common genes, GO/KEGG/PPI counts, ML overlap, nomogram AUCs, and GSEA.
2. Scroll through the step cards below it for the same information broken out per step (normalization, batch correction, DE, WGCNA, common genes, GO/KEGG, PPI, ML, nomogram, GSEA, and input/pipeline size).

This tab is text and tables only - it does not duplicate plots or offer JPG/PDF downloads. For figures, use the download buttons on each step’s own tab (e.g. Step 6 for the volcano plot, Step 9 for the PPI network).

Success check: The narrative and step cards reflect your completed steps’ actual numbers (e.g. DEG count matches Step 6, module count matches Step 7).

Step 16: Results summary overview.