Integrative transcriptome-based drug repurposing in tuberculosis

This study presents a robust computational workflow that integrates diverse tuberculosis transcriptomic signatures with multiple connectivity scoring methods to systematically identify 64 FDA-approved drugs as promising host-directed therapeutics and 12 novel bridging genes for future experimental validation.

The Gist

The Big Picture: Finding a "Cheat Code" for Tuberculosis

Imagine Tuberculosis (TB) as a very stubborn, armored burglar breaking into your house (your body). For decades, we've tried to fight this burglar with "antibiotics," which are like throwing rocks at the burglar to knock them out. But the burglar is getting smarter; they are wearing better armor (antibiotic resistance), and the rocks aren't working as well as they used to.

The scientists in this paper had a new idea: Instead of just throwing rocks at the burglar, let's fix the house so the burglar can't get in or can't survive inside. This is called Host-Directed Therapy (HDT). It means using drugs to boost your own immune system (the house security) rather than just attacking the bacteria.

The problem? There are thousands of drugs already approved for other things (like high blood pressure or cancer). Finding the right one to fix the "house" for TB is like finding a needle in a haystack. Doing this with lab experiments takes years and costs millions of dollars.

So, these researchers built a super-smart computer program to scan through thousands of existing drugs and find the "cheat codes" that could stop TB.

---

The Problem: Too Much Noise, Too Many Confusing Clues

To find the right drug, you need to understand exactly what TB does to your body. Scientists have been studying this for years, but the data is messy.

• The Analogy: Imagine trying to figure out what a "typical" storm looks like by looking at photos from 28 different photographers. Some took photos in black and white (microarray data), some in color (RNAseq data). Some took pictures of a storm in the ocean, others in the mountains. Some photos were blurry, some were sharp.
• The Issue: If you only look at one photo, you might think a storm is just rain. If you look at another, you might think it's just wind. Previous computer programs tried to match drugs to these single, messy photos, which led to unreliable results.

The Solution: The "Consensus Chorus"

The researchers decided to stop looking at single photos and instead listen to the whole chorus.

1. Gathering the Data: They collected 28 different "snapshots" of how TB changes human cells.
2. The Weighted Average: They didn't just average them all together equally. They realized some snapshots were clearer and more similar to each other than others. So, they gave the "clearer" snapshots more weight (like giving a louder voice in a choir more importance) and the "noisy" ones less.
3. The Result: They created one Master Signature of TB. This signature represents the true, core way TB messes up the human body, ignoring the noise from different labs or equipment.

The Search: The "Reverse Engineering" Game

Now that they had the "Master Signature" of the disease, they needed a drug that could reverse it.

• The Analogy: Imagine the disease signature is a song playing in a minor key (sad, chaotic). You are looking for a drug that plays the exact same song but in a major key (happy, organized). If the drug plays the "reverse" song, it cancels out the disease.
• The Method: They used a massive library of drug effects (called LINCS) containing over a million "songs" (gene expression patterns) from different drugs.
• The Match: They ran their Master TB Signature against this library using six different mathematical scoring systems. Think of this as having six different judges taste a dish to see if it's spicy. If all six judges agree a drug is a "perfect match," it's a winner.

The Findings: The "Hall of Fame" Drugs

The computer found 64 promising drugs that are already approved by the FDA for other uses. This is great news because it means we don't have to wait 10 years to test their safety; we just need to test if they work for TB.

The Winners Included:

• Statins (like Rosuvastatin): Usually used for high cholesterol. The computer predicted they work for TB, and indeed, other studies show they help the immune system fight TB.
• Tamoxifen: Used for breast cancer. It turns out it helps immune cells digest the bacteria.
• Clonidine: A blood pressure drug. The computer flagged this as a new, exciting possibility for TB.

The "Secret Connectors": Finding the Weak Points

The researchers didn't just stop at the drugs. They wanted to know how these drugs work. They built a giant map (a network) connecting the disease genes to the drug targets.

• The Analogy: Imagine the disease is a fortress and the drug is a key. The researchers found the bridges (genes) that connect the fortress to the key.
• The Discovery: They found 12 "bridge genes" (like IL-8 and CXCR2) that are crucial for the infection. If we can target these bridges, we might be able to stop the infection even better.

The Team-Up: Synergy vs. Antagonism

Finally, they asked: "If we use these new drugs with the old antibiotics, will they help each other or fight each other?"

• Synergy (The Power Couple): Some drugs work great together. For example, Statins + certain cancer drugs might supercharge the immune system.
• Antagonism (The Bad Mix): Some drugs fight each other. For example, if a drug slows down the bacteria too much, the old antibiotics (which need the bacteria to be growing fast to work) might stop working. The computer warned them to avoid these bad combinations.

The Bottom Line

This paper is like a GPS for drug discovery.

1. It cleaned up the messy map of TB data.
2. It found the best "reverse" drugs from the existing pharmacy.
3. It identified the specific biological bridges to target.
4. It told us which drug combinations to try and which to avoid.

The scientists are now saying, "We have the list. Now, let's go into the lab and test these 64 drugs to see if they can save lives." This approach could be used for other diseases too, turning the slow process of drug discovery into a fast, smart search.

Technical Summary

1. Problem Statement

Tuberculosis (TB) remains a leading cause of infectious disease mortality, exacerbated by rising antibiotic resistance. Host-Directed Therapies (HDTs), which modulate the host immune response rather than directly killing the pathogen, offer a promising alternative. However, developing new HDTs is costly and time-consuming. Drug repurposing—identifying new uses for FDA-approved drugs—is a viable strategy, but current computational approaches face significant limitations:

• Data Heterogeneity: Existing TB transcriptomic datasets vary widely in profiling platforms (microarray vs. RNA-seq), biological contexts (cell lines vs. primary tissues, blood vs. lung), and infection conditions.
• Methodological Bias: Previous studies often rely on a single connectivity scoring method or a narrow subset of data, leading to unstable predictions that may not represent robust, population-wide TB signals.
• Lack of Robust Signatures: Individual disease signatures often reflect context-specific responses rather than core host perturbations, making it difficult to identify consistent drug candidates.

2. Methodology

The authors developed a comprehensive computational workflow to integrate heterogeneous data and multiple scoring methods to identify robust HDT candidates.

A. Data Collection and Preprocessing

• Disease Signatures: Aggregated 28 TB gene expression datasets from GEO (10 microarray, 23 RNA-seq studies), covering diverse tissues, cell types, and time points.
• Drug Signatures: Utilized the LINCS L1000 database (GSE92742), restricting analysis to 728 FDA-approved drugs tested at 10 µM concentration for 24 hours across 14 cell lines to ensure consistency.

B. Signature Aggregation Strategy

• Weighted Consensus: Instead of analyzing individual signatures in isolation, the authors constructed a "consensus" TB signature.
• Weighting Mechanism: Individual signatures were weighted based on their Jaccard similarity to other signatures. Signatures with higher overlap (indicating shared core host responses) received higher weights, while outlier signatures were down-weighted. This approach aimed to capture dominant biological signals while filtering out noise from specific experimental conditions.

C. Multi-Method Connectivity Scoring
To mitigate method-specific biases, the study employed six distinct connectivity scoring methods categorized into two groups:

1. Enrichment-Based (ES):
   • CMAP 1.0 Connectivity Score (KS test).
   • CMAP 2.0 (LINCS) Scores: Weighted Connectivity Score (WCS), Normalized Connectivity Score (NCS), and Tau score ($\tau$).
2. Correlation-Based (Pairwise Similarity):
   • XCor: Pearson correlation between ranked gene expression values.
   • XSpe: Spearman correlation between ranked gene expression values.

D. Prioritization and Filtering Pipeline

• Dual Approach: Drugs were ranked against both individual signatures and the aggregated consensus signature.
• Rank Aggregation: Results from microarray and RNA-seq pipelines were consolidated using the BiG (Bayesian Latent Variable Approach) method to handle heterogeneity across scoring metrics.
• High-Confidence Selection: Candidates were selected if they appeared in the top rankings across multiple methods and signature types. The final list was refined using a combined mean-rank z-score ($z_{combined}$), requiring positive scores for both individual ($z_{indiv}$) and aggregated ($z_{aggr}$) predictions.

E. Network and Synergy Analysis

• Network Analysis: Constructed disease-drug networks using STRING (PPI) and DGIdb to identify "bridging genes" (in-path genes) connecting TB disease genes to drug targets.
• Synergy Prediction: Evaluated potential synergistic or antagonistic interactions between predicted HDTs and standard TB antibiotics using DrugCombDB metrics (ZIP, Bliss, Loewe, HSA).

3. Key Contributions

1. Integrative Framework: Developed a novel workflow that successfully reconciles technical (platform) and biological (tissue/cell type) heterogeneity in transcriptomic data for drug repurposing.
2. Weighted Aggregation: Introduced a Jaccard-similarity-weighted aggregation method to derive a robust consensus disease signature, demonstrating that this approach captures dominant biological signals better than individual signatures.
3. Multi-Method Robustness: Demonstrated that integrating six different connectivity scoring methods significantly improves the stability and reliability of drug prioritization compared to single-method approaches.
4. Mechanistic Insight: Moved beyond simple drug ranking to provide mechanistic insights via network analysis, identifying specific bridging genes and pathway convergences (cholesterol and cytokine signaling).

4. Key Results

• Candidate Identification: The workflow identified 64 high-confidence FDA-approved drug candidates for TB HDT.
  • Validation: The list successfully recovered known HDTs with experimental validation, including statins (rosuvastatin, fluvastatin, lovastatin) and tamoxifen.
  • Novel Predictions: Identified novel candidates such as clonidine (an antihypertensive), fostamatinib, and nabumetone, which lack prior published evidence for anti-TB activity but showed strong reversal scores.
• Biological Mechanisms:
  • Top candidates were enriched for mechanisms involving cholesterol metabolism inhibition (HMGCR inhibitors) and immune modulation (cytokine signaling, PDGFR inhibition).
  • Network analysis revealed HMGCS1 as a shared node linking disease and drug perturbations in cholesterol and cytokine pathways.
• Bridging Genes: Identified 12 key "in-path" genes (e.g., IL-8, CXCR2, MCL-1, BAX, ITGB1) that serve as mechanistic bridges between TB pathology and drug targets. These represent potential novel druggable targets for HDT.
• Synergy and Antagonism:
  • Predicted synergy between HMGCR inhibitors (statins) and HDAC inhibitors, as well as between bacterial cell wall synthesis inhibitors and tubulin polymerization inhibitors.
  • Predicted antagonism between bacterial cell wall synthesis inhibitors and fatty acid synthesis inhibitors, suggesting caution in combining these classes.
• Heterogeneity Analysis: Confirmed that TB host responses vary significantly by tissue type (circulating vs. local) and technology (microarray vs. RNA-seq), validating the necessity of the integrative aggregation approach.

5. Significance

• Generalizability: This study establishes a robust, generalizable framework for transcriptome-based drug repurposing that can be applied to other infectious diseases with heterogeneous data landscapes.
• Accelerated Discovery: By prioritizing FDA-approved drugs with established safety profiles, the identified candidates (e.g., statins, clonidine) offer immediate opportunities for experimental validation and potential clinical repurposing, significantly reducing development time and cost.
• Mechanistic Understanding: The identification of specific bridging genes and synergistic MOA combinations provides a rational basis for designing combination therapies that enhance host immunity while avoiding antagonistic interactions.
• Resource Availability: The authors have made all processed data and code publicly available, facilitating reproducibility and further research in the field of infectious disease therapeutics.

In conclusion, this work demonstrates that by systematically integrating heterogeneous transcriptomic data and multiple scoring algorithms, it is possible to overcome the limitations of traditional connectivity mapping to identify robust, mechanistically grounded host-directed therapeutics for tuberculosis.