Code and figures for the paper "Analytical validation pipeline for generating highly specific functional gene expression signatures" (BostonGene).
Functional gene expression signatures (FGESs) are commonly used in tumor transcriptomic analyses, including in predictive and prognostic models. However, creating and validating FGES remain challenging due to technical and biological limitations that prevent the establishment of reliable quality assessment criteria for FGES. Here, we describe a comprehensive pipeline for developing and validating FGESs that includes 1) creating a gene set using both computational and analytical methods; 2) calculating the gene set score using a modified single-sample gene set enrichment analysis (ssGSEA) and assessing its quality according to a defined set of technical and biological criteria to ensure reproducibility, specificity, and robustness; and 3) refining the FGES followed by final validation. Nineteen cell type-specific and five process-describing FGESs created using this workflow showed higher biological relevance than most publicly available gene signatures. Thus, the proposed workflow provides an approach to control FGES quality, supporting their further use in the investigation of complex biological processes, including tumor microenvironment analytics.
Figure 1. FGES classification and the proposed validation pipeline. A candidate gene set is built from literature- or data-driven gene search, iteratively tuned on train datasets until it meets four pre-defined technical and biological quality criteria, and then validated on an independent test dataset.
| Path | Contents |
|---|---|
src/signature_validation/ |
Installable Python library with the analysis logic (see below) |
Paper_Code_and_Figures/ |
One folder per paper figure: notebooks that call the library and produce figure panels and supplementary tables — see its README for the figure → notebook map |
Data/ |
Small annotation / reference tables (TCGA annotation and deconvolution, gene lengths, sorted-cell annotation) |
License.md, License_Notice.md |
Software license and third-party open-source notices |
ssgsea_calc/— the modified ssGSEA used throughout the paper: a direct O(N log N) formulation of the original ssGSEA score (analytic gene-set term + arithmetic-progression approximation of the non-gene-set term), with tied genes assigned the same highest rank instead of alphabetical tie-breaking (Methods, ssGSEA score calculation). Also classifies gene sets by source (internal / MSigDB / published / random).noise_calc/— single-gene technical expression noise model (Methods, Gene expression noise calculation), used for the noise criterion (Figure 2).benchmark/— sorted-cell benchmark of cell type-specific FGESs against public gene sets (Figure 4E–I): GOI / control cell-type mapping, signature loading, ssGSEA scoring, bootstrap F-score and CV metrics, rare cell-type backfill from train cross-validation, plots and supplementary tables.plotting/,utils/— shared palettes, plotting helpers, expression / annotation readers.
python -m venv .venv
.venv/bin/pip install -e . # Python >= 3.10
.venv/bin/python -c "import signature_validation"The input expression data are not distributed with this repository. Public datasets used in the study are listed in Supplementary Tables S8.1–S8.3 of the paper; several datasets are subject to access restrictions of their original providers. Internally generated data (sorted-cell bulk RNA-seq and the NSCLC scRNA-seq dataset) are available from the corresponding author upon request.
Every notebook has a data-locations cell with <PATH_TO_...> placeholders — replace them with
your local paths before running. Generated plots, tables and cached pickles are regenerated by the
notebooks.
See License.md and License_Notice.md.