Peripheral Blood Transcriptomics Identifies a Six-Gene Diagnostics Signature with Network-Aware Biomarker Prioritization and Uncovers Transcriptomic Overlap with Motor Neuron Disease Mimics

This repository contains the complete R pipeline for the manuscript:

"Peripheral Blood Transcriptomics Identifies a Six-Gene Diagnostics Signature with Network-Aware Biomarker Prioritization and Uncovers Transcriptomic Overlap with Motor Neuron Disease Mimics"

---

Overview

This project presents a systems-integrated, leakage-free bioinformatics framework for peripheral blood biomarker discovery in ALS. Using GSE112676 (n = 702) as the sole discovery cohort, the pipeline applies differential expression analysis, WGCNA, and a triple-consensus machine learning framework (Random Forest, LASSO, SVM-RFE) to derive a six-gene diagnostic signature — BRI3, ABCA1, QPCT, PPP2R5A, ETFRF1, and SLC37A3. A novel Biomarker Priority Score (BPS) integrating cross-cohort diagnostic stability with network centrality is introduced for principled biomarker ranking. The signature is validated across five independent arms spanning three external cohorts, including a dedicated ALS-versus-mimic discrimination analysis.

---

Data Availability

All datasets are publicly available through the NCBI Gene Expression Omnibus (GEO):

Dataset	Platform	Role	n	Groups
GSE112676	Illumina HumanHT-12 V3.0	Discovery	702	ALS, CON
GSE112680	Illumina HumanHT-12 V4.0	Validation (held-out)	376	ALS, CON, MIM
GSE28253	Agilent 4×44K GPL4133	Validation (external)	22	ALS, CON
GSE234297	Illumina RNA-seq	Validation (cross-platform)	144	ALS, CON

Note: GSE112680 was held out entirely and never accessed during discovery, feature selection, or model training.

---

Prerequisites

The analysis was performed in R version 4.5.2 with a global seed (set.seed(1122)). To install all required packages, run:

source("00_Install_Dependencies.R")

---

Pipeline Structure

Run scripts in sequential order. Each script loads intermediate objects saved by the preceding script from the Processed_Data/ directory.

Script	Description
00_Install_Dependencies.R	Installs all required CRAN and Bioconductor packages
01_Data_Prep_and_DEG.R	Data preprocessing, quality control, and differential expression analysis
02_WGCNA_Analysis.R	Co-expression network construction, module-trait correlation, and hub gene extraction
03_Machine_Learning_Consensus.R	Triple-consensus feature selection and logistic regression model training; discovery-cohort gene-wise scaling parameters (mean and SD) saved alongside the trained model for deployment-ready normalization
04_Clinical_Validation_and_BPS.R	Five-arm independent validation using discovery-locked z-scoring, BPS computation, and ALS-mimic discrimination analysis
05_Immune_Analysis.R	EPIC, MCP-counter, and ssGSEA immune deconvolution and ALS immune subtype stratification
06_Drug_Repurposing.R	DSigDB drug repurposing via Enrichr
07_Cohort_Table_and_Covariate_Analysis.R	Cohort characteristics table summarising sex, age at onset, site of onset, and survival across all groups in GSE112676 and GSE112680; sex-adjusted limma sensitivity analysis
08_Threshold_CI_Analysis.R	Youden-optimal decision threshold with 2000-bootstrap 95% confidence intervals for sensitivity and specificity across GSE112680 Arms 1–3
09_Immune_Subtype_Validation.R	Clinical anchoring of ALS immune subtypes against survival (Kaplan-Meier, log-rank), age at onset (Mann-Whitney U), site of onset, and sex (Fisher's exact); external validation by EPIC deconvolution and centroid projection in GSE112680 (n = 164 ALS)

Note on Functional Enrichment and PPI Analysis

Functional enrichment (GO Biological Process, GO Cellular Component, Reactome) and PPI network construction were performed using the STRING database web server (v12.0) (https://string-db.org/). The 240 ML hub candidates were used as enrichment input and the six consensus genes with first-degree interactors were used for the PPI network. These steps are not included as R scripts as they use the STRING graphical interface.

---

Reproducibility Notes

• Deterministic Seed Control: A global seed (set.seed(1122)) is strictly enforced across all scripts. This guarantees that all stochastic processes — Random Forest tree generation, LASSO cross-validation folds, SVM recursive feature elimination, and WGCNA module clustering — yield identical, perfectly reproducible results on every run.

• Sequential Execution: The pipeline relies on a continuous computational environment. Scripts 04 through 09 explicitly load filtered matrices, trained models, and signature lists saved to the Processed_Data/ directory by the preceding scripts. Please run scripts in exact numerical order.

• Discovery-Locked Normalization: Gene-wise means and standard deviations are computed once from the GSE112676 discovery cohort (n = 702) and stored in Step4_ML_Results.rds. All validation cohorts are scaled using these fixed parameters (z = (x − μ_discovery) / σ_discovery) rather than their own cohort statistics, making the model applicable to a single future patient without requiring cohort-level data.

• Automated Directories: The scripts automatically create Processed_Data/ and Manuscript_Figures/ directories in the working folder to store intermediate data objects (.rds, .csv) and high-resolution figures (.pdf, .png).

• GEO Data Retrieval (Timeouts): Scripts 01, 04, 07, and 09 download transcriptomic matrices and phenotype records directly from NCBI GEO via GEOquery. If you encounter timeout errors, increase the R timeout limit before running: options(timeout = 600). Local series matrix caches (e.g., GSE112676_series_matrix.txt.gz) are detected automatically and loaded in preference to live downloads where available.

• Hardware Recommendations: WGCNA soft-threshold topology fitting and triple-consensus machine learning cross-validation are memory-intensive. A workstation with at least 8–16 GB of RAM is recommended for seamless execution.

---

Citation

If you use this pipeline in your research, please cite:

[Citation to be added upon publication]

---

Contact

• Academic & Collaborative Inquiries: For questions regarding the manuscript, methodology, or potential collaborations, please contact the corresponding authors via email as listed in the publication.
• Code & Pipeline Issues: For technical questions, troubleshooting, or code-related bugs, please open an Issue directly in this GitHub repository.