STDrug enables spatially informed personalized drug repurposing from spatial transcriptomics

STDrug is a novel computational framework that leverages spatial transcriptomics, graph-based modeling, and multimodal learning to overcome the limitations of single-cell approaches by capturing tissue microenvironment context, thereby enabling more accurate, patient-specific drug repurposing for cancers like hepatocellular and prostate carcinoma.

The Gist

Imagine your body is a bustling, complex city. When a disease like cancer strikes, it's not just a few bad apples causing trouble; it's entire neighborhoods (tissues) that have gone rogue, changing their architecture, their traffic patterns, and their relationships with the surrounding areas.

For a long time, doctors and scientists trying to find new uses for old drugs (a process called drug repurposing) were looking at this city through a blurry, black-and-white photo. They could see the population statistics (which genes were active), but they couldn't see where those people were living or how the neighborhoods were interacting. They missed the most important clue: context.

Enter STDrug, a new, high-tech "smart city planner" that changes the game. Here is how it works, broken down into simple steps:

1. The Problem: The "Blurred Map"

Previous methods used single-cell data, which is like having a list of every citizen in the city and what they are doing, but without knowing their addresses. You know a "rebel" cell exists, but you don't know if it's in the wealthy district, the industrial zone, or right next to the police station. This lack of location data made it hard to predict which drugs would actually work, because drugs often need to target specific neighborhoods to be effective.

2. The Solution: STDrug's "3D Holographic Map"

STDrug uses a technology called Spatial Transcriptomics. Think of this as upgrading from a flat, blurry photo to a 3D, high-definition hologram of the city. It sees exactly where every cell is located and how they are talking to their neighbors.

The tool has three main superpowers:

• The Matchmaker (Spatial Domain Identification):
  Imagine you have two maps of the same city: one from when it was healthy and one from when it was sick. STDrug uses a sophisticated "matchmaker" algorithm to line them up perfectly. It finds the "healthy neighborhood" that corresponds to the "sick neighborhood." This allows it to see exactly what changed in that specific spot, rather than just looking at the city as a whole.

• The Drug Detective (The Repurposing Engine):
  Once it knows exactly what's wrong in the specific neighborhood, it goes to a massive library of drug information (like a giant pharmacology database). It asks: "Which existing drug can reverse these specific changes?"

  But here's the clever part: It doesn't just guess. It uses an AI "Librarian" (a Large Language Model) to read millions of medical research papers. This AI helps the tool understand which genes are the "villains" causing the most trouble and assigns them a "weight" or importance score. It then calculates a "Drug Score" for every candidate, factoring in:

  • How well the drug fixes the specific neighborhood.
  • How the neighborhoods talk to each other (if you fix one, does it help the neighbor?).
  • Safety (will this drug cause a traffic jam or a riot elsewhere in the body?).

• The Personalized Prescription:
  Because every patient's "city" is slightly different, STDrug creates a custom drug ranking for that specific person. It's not a "one-size-fits-all" approach; it's a tailored plan.

3. The Proof: Does it Work?

The researchers tested STDrug on two tough cities: Liver Cancer (HCC) and Prostate Cancer (PCa).

• The Results: In a head-to-head race against older methods, STDrug won easily. It predicted the right drugs with much higher accuracy (about 81-82% success rate vs. 50-60% for others).
• The Surprise Winners: It didn't just find cancer drugs. It suggested repurposing common drugs like:
  • Statins (heart medicine) and Digoxin (heart failure medicine) for liver cancer.
  • Bortezomib (a drug usually for blood cancer) and Vorinostat for prostate cancer.
• Real-World Verification:
  • The "Time Machine" Test: They looked at millions of real patient records. They found that people who happened to be taking these "repurposed" drugs (like statins) before getting cancer developed the disease much later than those who didn't.
  • The "Lab Test": They took prostate cancer cells in a petri dish and treated them with the top drugs STDrug suggested. The drugs killed the cancer cells effectively, even the ones that were resistant to standard chemotherapy.

The Big Picture

Think of STDrug as a GPS for drug discovery. Before, we were driving blind, hoping a drug would hit the right target. Now, STDrug gives us a live, high-definition map of the disease, showing us exactly which "streets" to block and which "buildings" to repair.

By combining the power of spatial location, AI reading, and real-world data, STDrug turns the slow, expensive process of finding new cancer cures into a faster, smarter, and more personalized journey. It proves that sometimes, the best new medicine isn't a new invention at all—it's an old drug used in a new, smarter way.

Technical Summary

1. Problem Statement

Drug repurposing is a promising strategy to accelerate therapeutic discovery, but existing computational methods face significant limitations:

• Lack of Spatial Context: Most current approaches rely on bulk RNA-seq or single-cell RNA-seq (scRNA-seq). While scRNA-seq captures cellular heterogeneity, it loses the critical spatial organization of cells within the tissue microenvironment.
• Microenvironment Dependency: Drug responses are often dependent on cell-cell interactions and spatial domain architecture (e.g., tumor-stroma interfaces), which scRNA-seq cannot fully capture.
• Gap in Methodology: There is a lack of frameworks that specifically leverage Spatial Transcriptomics (ST) data to identify patient-specific drug candidates that account for spatial domain interactions and tissue architecture.

2. Methodology: The STDrug Framework

STDrug is a computational framework designed to integrate spatial transcriptomics, graph-based modeling, and multimodal learning for personalized drug repurposing. It operates in three primary stages:

A. Data Preprocessing and Spatial Domain Identification

• Input: Paired spatial transcriptomics data from diseased (tumor) and adjacent normal tissues from the same patient.
• Embedding & Alignment:
  • Uses Harmony for batch effect correction on gene expression embeddings.
  • Constructs a Graph Convolutional Network (GCN) to learn spatially-aware low-dimensional embeddings, integrating gene expression with spatial coordinates.
  • Disease-Control Balanced Clustering: Unlike standard clustering, STDrug employs a Coherent Point Drift (CPD) algorithm to align spatial domains between diseased and normal tissues. This ensures that corresponding biological regions (e.g., specific tumor niches vs. normal equivalents) are matched within the same patient.
• Output: Paired spatial domains representing matched regions across disease and control conditions.

B. Drug Repurposing Module

• Gene Signature Identification: Identifies Differentially Expressed Genes (DEGs) between the paired spatial domains (Tumor vs. Normal).
• Perturbation Matching: Searches large-scale pharmacogenomics databases (L1000 and Tahoe-100M) for compounds that can reverse the disease-associated gene signatures.
• LLM-Enhanced Weighting:
  • Utilizes a pre-trained Large Language Model (GPT-4o) to extract drug-disease relationships from literature.
  • Trains a machine learning classifier (XGBoost) to assign weights to reversible genes based on their importance, refining the prioritization of candidate drugs.
• Comprehensive Scoring: Calculates a patient-specific drug score by integrating:
  1. Reversal Score: Ability to reverse disease signatures.
  2. Spatial Domain Interactions (DDI): Uses CellChat v2 to quantify cell-cell communication strength changes between domains in disease vs. control.
  3. Efficacy & Safety: Incorporates drug sensitivity data from GDSC and side-effect profiles from SIDER.
• Output: A ranked list of personalized drug candidates.

C. Validation Strategy

The framework employs a multi-tiered validation approach:

1. In-silico Benchmarking: Comparison against existing methods (ASGARD, Beyondcell) using AUC metrics against clinical trial and literature-derived "ground truth."
2. Real-World EHR Validation: Analysis of the MarketScan database (264M patients) to assess if prior exposure to predicted drugs delays cancer onset (using Cox proportional hazards models and Kaplan-Meier curves).
3. In-vitro Validation: Cell viability assays (WST-1) on prostate cancer cell lines (including taxane-resistant variants) to test top candidates.

3. Key Contributions

• First Spatially Informed Repurposing Tool: STDrug is the first framework to explicitly model spatial domain interactions and tissue architecture for drug repurposing, moving beyond single-cell resolution to spatial context.
• Novel Alignment Algorithm: The integration of GCN with Coherent Point Drift (CPD) allows for precise, patient-specific alignment of diseased and normal spatial domains, overcoming the lack of ground truth in paired tissue analysis.
• LLM-Guided Prioritization: Uniquely leverages GPT-4o to curate knowledge-guided gene weights, enhancing the biological relevance of the drug scoring mechanism.
• Multimodal Integration: Seamlessly combines spatial omics, pharmacogenomics, clinical side-effect data, and machine learning into a unified scoring system.

4. Results

The framework was validated on Hepatocellular Carcinoma (HCC) and Prostate Cancer (PCa) datasets.

• Benchmarking Performance:

  • Spatial Domain Identification: STDrug achieved superior accuracy (NMI = 0.93, ARI = 0.94) compared to state-of-the-art tools like STAGATE, STAligner, SPACEL, and GraphST.
  • Drug Prediction Accuracy: STDrug significantly outperformed scRNA-seq-based methods (ASGARD and Beyondcell).
    • HCC: AUC = 0.81 (vs. 0.61 for ASGARD, 0.59 for Beyondcell).
    • PCa: AUC = 0.82 (vs. 0.57 for ASGARD, 0.48 for Beyondcell).

• Top Candidate Discoveries:

  • HCC: Correctly identified Sorafenib (standard of care) and Bosutinib as top candidates. Also predicted non-oncology agents like Niacin, Caffeine, Atorvastatin, and Digoxin, all of which have literature support for anti-tumor effects in liver cancer.
  • PCa: Prioritized Bortezomib and Vorinostat. Also identified Digoxin, Colchicine, and Sirolimus.

• Validation Outcomes:

  • EHR Analysis: Patients with prior exposure to predicted drugs (e.g., Atorvastatin, Digoxin for HCC; Sirolimus, Colchicine for PCa) showed statistically significant delays in cancer onset compared to controls.
  • In-vitro Assays: Bortezomib and Vorinostat demonstrated potent cytotoxicity against multiple PCa cell lines, including taxane-resistant strains, at clinically achievable concentrations.

• Mechanistic Insights:

  • HCC: Bosutinib was found to target the MAPK/ERK and Src-associated signaling pathways, specifically downregulating oncogenic kinases and cytoskeletal regulators in specific spatial domains (e.g., transitioning immunoactive domains).
  • PCa: Bortezomib was shown to restore pro-apoptotic signaling (TRAIL/TNFSF10) and tumor suppressors (DCN, WIF1) while inhibiting the PI3K-Akt and cGMP-PKG pathways across multiple spatial domains.

5. Significance

• Precision Medicine: STDrug enables truly patient-specific therapeutic strategies by accounting for the unique spatial architecture of an individual's tumor microenvironment.
• Translational Potential: By successfully identifying repurposable drugs that are validated by real-world clinical data and wet-lab experiments, STDrug bridges the gap between computational prediction and clinical application.
• Generalizability: The framework is designed to be adaptable to other diseases and spatial omics technologies, establishing a new standard for integrating spatial biology into drug discovery pipelines.
• Efficiency: It offers a scalable route to accelerate drug discovery by leveraging existing safety data and repurposing approved compounds, potentially reducing the time and cost of bringing new treatments to market.