The BioAssay network and its implications to future therapeutic discovery

Jintao Zhang, Gerald H. Lushington, Jun HuanBMC Bioinformatics 2011 · DOI

Abstract

The pharmaceutical industry faces a significant "productivity gap" where R&D investment increases while the output of new drugs remains stagnant. This paper introduces a network-based framework to bridge high-throughput chemical screening and therapeutic discovery. By constructing a "BioAssay network" from PubChem data and mapping it onto human protein interaction maps, the authors developed a physics-inspired model to prioritize novel drug targets. Their model successfully prioritized targets with approximately 70% accuracy, demonstrating that bioassay targets are strategically positioned near essential and disease-related genes within the human interactome.

Introduction

Traditional drug discovery often follows a "one gene, one drug, one disease" philosophy, which fails to account for the inherent complexity of biological systems. To address this, the NIH's Molecular Libraries Initiative (MLI) was launched to explore the "undruggable" parts of the human genome. However, the resulting data in repositories like PubChem is massive, noisy, and high-dimensional. This study addresses the challenge of systematically translating these millions of data points into a prioritized list of viable, biologically relevant drug targets.

Methods

The researchers employed a multi-step network biology approach to analyze chemical screening data:

  1. Network Construction: They started with a bipartite network connecting compounds to the assays where they showed activity. This was projected into a BioAssay Network, where nodes represent assays. Two assays are linked if they share a significant portion (at least 10% Jaccard coefficient) of active compounds, suggesting a shared biological mechanism.
  2. Topological Analysis: They characterized the network’s structure, finding it to be "scale-free" (a few hubs have many connections, while most have few) but notably non-hierarchical.
  3. Cross-Database Integration: Bioassay targets were mapped onto the UniHI (Unified Human Interactome) to analyze their position within the broader web of human protein-protein interactions (PPI).
  4. The "Characteristic Distance" Model: Inspired by Coulomb’s Law in electrostatics, the authors treated essential genes as "unit charges" and calculated a "field strength" to determine how closely a bioassay target clusters around essential or disease-related genes.
Figure 1: Topological analysis of the BioAssay network, showing its scale-free structure and clustering coefficients.

Datasets

The study integrated several major biological and chemical databases, including PubChem for screening data, DrugBank for approved drug targets, UniHI for PPI data, OMIM for disease-gene associations, and Mouse Genome Informatics (MGI) for defining essential genes.

Key Insights and Results

The Proximity Principle and Hub Characteristics

  • Bioassay Targets as Hubs: Surprisingly, bioassay targets have a higher median degree (more connections) in the human PPI network than even currently approved drug targets. This suggests that the MLI is successfully targeting "central" biological players.
  • The Proximity Principle: Bioassay targets are not randomly distributed; they cluster significantly closer to essential genes and known disease genes than random proteins do. This proximity validates the target selection strategy of the MLI.
  • Non-Hierarchical Nature: Unlike many biological networks that show hierarchical modularity, the BioAssay network is more uniform in its clustering, likely reflecting the diverse nature of chemical probes used in screenings.
Figure 2: Comparison of BioAssay targets, drug targets, and random proteins across network degree and distance to essential/disease genes.

Quantitative Performance

  • Network Scale: The final BioAssay network consisted of 899 nodes and 6,080 edges.
  • High Accuracy: Using their physics-based model, the authors predicted the top 10 most promising bioassay targets. A literature survey confirmed that 7 out of 10 (70%) were already being independently investigated as promising drug targets (e.g., WEE1, RGS7, and PAK7).
  • Statistical Significance: Bioassay targets were found to be statistically closer to existing drug targets ($P < 10^{-24}$) and disease genes ($P < 10^{-12}$) than random proteins.

Discussion and Future Work

While the study demonstrates the power of network biology in drug discovery, it also highlights challenges regarding data quality. PubChem data is inherently noisy, and the lack of universal normalization for "activity scores" across different assays can introduce bias.

Future directions include moving beyond target-based assays to better incorporate cell-based assay data. The authors suggest that future models should integrate more diverse data types, such as gene expression profiles and metabolic pathways, to further refine the "druggability" score and improve the prioritization of therapeutic targets.

Summary updated

BibTeX
@article{zhang2011bioassay,
  author = {Jintao Zhang and Gerald H. Lushington and Jun Huan},
  title = {{The BioAssay network and its implications to future therapeutic discovery}},
  journal = {BMC Bioinformatics},
  year = {2011},
  doi = {10.1186/1471-2105-12-s5-s1},
}
Figure