Knowledge graph analytics platform with LINCS and IDG for Parkinson's disease target illumination

Jeremy J. Yang, Christopher Gessner, Joel Duerksen, Daniel Biber, Jessica L. Binder, Murat Ozturk, Brian Foote, Robin McEntire, Kyle Stirling, Ying Ding, David J. WildBMC Bioinformatics 2022 · DOI

TL;DR

Researchers have developed the Knowledge Graph Analytics Platform (KGAP), a tool that integrates massive experimental and curated datasets to "illuminate" understudied protein targets. By applying KGAP to Parkinson’s Disease (PD), the team successfully validated known drug targets and identified SYNGR3 as a high-priority, novel candidate for future therapeutic research.

What Problem Does This Paper Address?

Drug discovery for complex neurodegenerative disorders like Parkinson’s Disease is notoriously difficult. A major bottleneck is the "dark genome"—thousands of proteins that are potentially druggable but remain understudied because researchers tend to cluster around well-known targets.

To find new targets, scientists need to bridge the gap between unbiased experimental data (how cells react to thousands of different compounds) and curated clinical knowledge (which drugs are already approved and what they are known to do). This paper introduces a way to link these disparate worlds using a Knowledge Graph (KG) to identify hidden biological connections that traditional research might miss.

Methods

The authors built KGAP using a graph database (Neo4j). The platform's power lies in its semantic structure, which connects four key entities:

  • Drugs: Approved medications and their clinical indications.
  • Signatures: Differential gene expression data (how a drug changes gene activity).
  • Genes: The biological targets, categorized by how well they are currently understood.
  • Concepts: Disease terms and classifications.

The Analytic Approach: Instead of using a "black box" machine learning model, the authors used path-based analytics. They started with known Parkinson’s drugs and traced paths through gene expression signatures to find associated genes. They used a modified Stouffer’s function to aggregate z-scores from these signatures, creating a "KGAP score" to rank potential targets. This "machine learning-free" approach prioritizes interpretability and avoids the data biases often found in trained models.

Datasets

The platform integrates two major NIH-funded initiatives:

  1. LINCS (Library of Integrated Network-based Cellular Signatures): Specifically the L1000 dataset, which contains over a million gene expression profiles of cells treated with various perturbagens (drugs/chemicals).
  2. IDG (Illuminating the Druggable Genome): Provides curated data via resources like DrugCentral (drug-target-disease links) and TCRD (Target Central Resource Database), which classifies genes by their "Target Development Level" (from "dark" proteins to well-studied clinical targets).
  3. TIN-X: A bibliometric tool used to rank targets based on "novelty" (rarity in literature) and "importance" (association with a disease).

Key Insights and Ideas

  • Synergy of "Unbiased" and "Curated": By combining LINCS (which shows what actually happens in a cell) with IDG (which contains what we know about drugs), the model can find evidence for targets that haven't been formally linked to a disease yet.
  • Target Illumination: The study specifically looks for "Tdark" (understudied) and "Tbio" (biologically known but not drugged) proteins. This shifts the focus away from the "usual suspects" in PD research.
  • Interpretability is King: Because the results are based on graph paths, a researcher can look at a suggested target and see exactly which drugs and which expression signatures led to that conclusion. This "provenance" is vital for building scientific trust.

Key Results

  • Validation: The KGAP scoring function was validated against a "gold standard" of known Parkinson’s drug targets. It achieved an Area Under the ROC Curve (AUC) of 0.74 for targets with a known mechanism of action, proving the method's predictive power.
  • The PD Hitlist: The search yielded 641 genes ranked by their association with PD.
  • Case Study (SYNGR3): The platform identified Synaptogyrin-3 (SYNGR3) as a top-tier "dark" target. While its exact function was previously unclear, KGAP linked it to PD through multiple evidence paths. Independent recent studies have since shown that SYNGR3 interacts with dopamine transporters, making it a highly plausible and exciting new target for Parkinson’s.

Limitations and Future Work

The authors acknowledge that the current definitions and "nosology" (classification) of Parkinson’s Disease are a limitation. PD is likely a collection of subtypes with different biological drivers, but most current data treats it as a single entity.

Future directions include: * Applying KGAP to other complex disease areas (cancers, metabolic diseases). * Refining the graph to include more granular data, such as disease subtypes and specific cell-line responses. * Further experimental validation of the "dark" targets identified by the platform.

Summary updated

BibTeX
@article{yang2022knowledge,
  author = {Jeremy J. Yang and Christopher Gessner and Joel Duerksen and Daniel Biber and Jessica L. Binder and Murat Ozturk and Brian Foote and Robin McEntire and Kyle Stirling and Ying Ding and David J. Wild},
  title = {{Knowledge graph analytics platform with LINCS and IDG for Parkinson's disease target illumination}},
  journal = {BMC Bioinformatics},
  year = {2022},
  doi = {10.1186/s12859-021-04530-9},
}
Figure