Integrative pharmacogenomics to infer large-scale drug taxonomy

Nehmé El-Hachem, Deena M.a. Gendoo, Laleh Soltan Ghoraie, Zhaleh Safikhani, Petr Smirnov, Christina Chung, Kenan Deng, Ailsa Fang, Erin Birkwood, Chantal Ho, Ruth Isserlin, Gary D. Bader, Anna Goldenberg, Benjamin Haibe‐KainsbioRxiv 2016 · DOI

Abstract

Identification of drug targets and mechanism of action (MoA) for new and uncharacterized drugs is essential for optimizing therapeutic efficacy. Traditional MoA prediction approaches often rely on prior clinical information, such as side effects or therapeutic indications, which are unavailable for newly identified small molecules. We present Drug Network Fusion (DNF), an integrative computational framework that infers scalable drug taxonomies by fusing basic drug characteristics: molecular structure, high-throughput drug perturbation (transcriptomic signatures), and drug sensitivity profiles. DNF enables the classification of experimental compounds with minimal prior information. Our results demonstrate that the DNF taxonomy successfully identifies novel drug-drug relationships and outperforms single-data-layer models in predicting drug targets and therapeutic classifications. This framework serves as a valuable resource for the cancer research community, providing new hypotheses for compound MoA and drug repurposing.

The Challenge of Uncharacterized Molecules

Identifying the mechanism of action (MoA) and specific protein targets of a new drug remains one of the most significant hurdles in pharmacology. Most existing prediction tools suffer from inherent biases, as they depend on "prior information"—known side effects, established therapeutic indications, or chemical similarities to well-documented drugs.

This dependency creates a bottleneck for uncharacterized small molecules. Newly discovered compounds lack the clinical history required by traditional models, making them difficult to classify. To address this, the Drug Network Fusion (DNF) framework was developed as an unbiased approach that utilizes high-throughput experimental data to place any compound into a global drug taxonomy, regardless of its clinical maturity.

The Drug Network Fusion (DNF) Framework

DNF utilizes a three-layered integration strategy to represent drug-drug relationships through diverse biological and chemical lenses:

  1. Drug Structure: Molecular similarity is calculated using SMILES representations and the Tanimoto index.
  2. Drug Perturbation: Transcriptomic signatures are analyzed to see how a drug alters gene expression in cancer cells, utilizing the massive LINCS L1000 dataset.
  3. Drug Sensitivity: Growth inhibition patterns are measured across hundreds of cancer cell lines (using CTRPv2 and NCI60 datasets).

These three distinct similarity matrices are integrated using Similarity Network Fusion (SNF). This iterative algorithm strengthens relationships that are consistently supported across multiple data types while effectively filtering out noise inherent in individual datasets. Finally, Affinity Propagation is applied to the fused network to identify "drug communities"—clusters of molecules that likely share a common MoA.

Key Insights: The Power of Integration

One of the primary findings of this research is the non-redundancy of pharmacogenomic data. The authors observed that chemical structure, cellular perturbation, and drug sensitivity are only weakly correlated. This suggests that each data layer captures a unique facet of a drug's biological profile. By fusing them, DNF creates a "sum greater than its parts" effect, uncovering biological nuances that a single data source would overlook.

Beyond "Guilt by Association"

DNF identifies more than just pairs of similar drugs; it reveals "super-communities." For instance, the framework successfully linked CDK inhibitors with topoisomerase inhibitors. While these drugs target different primary proteins, they share a systems-level MoA by perturbing the same underlying cell cycle pathways.

Opportunities for Drug Repurposing

The taxonomy also provides immediate leads for drug repurposing. A notable example is the anti-inflammatory natural compound parthenolide, which clustered closely with statins (cholesterol-lowering drugs). This suggests that parthenolide may share the statins' ability to inhibit the NF-Kb pathway, offering a potential new avenue for cancer treatment.

Performance and Validation

The DNF framework was rigorously validated against gold-standard benchmarks, including DrugBank (for protein targets) and ChEMBL (for Anatomical Therapeutic Chemical [ATC] classifications).

  • Target Prediction: DNF achieved an AUROC of 0.89, significantly higher than models using sensitivity alone (0.83) or perturbation alone (0.64).
  • Therapeutic Classification: In predicting ATC codes, DNF reached an AUROC of 0.80–0.85, consistently outperforming individual data layers.
  • Scalability: The framework proved robust across different data scales, maintaining high accuracy even when using the smaller NCI60 dataset compared to the much larger CTRPv2.

Future Directions

While DNF represents a significant leap forward, its current scope is limited by the "overlap bottleneck"—the requirement that drugs be tested across all three data layers. As high-throughput screening projects like LINCS L1000 continue to expand, the DNF taxonomy will grow in resolution and coverage.

Future iterations of the model may incorporate additional "omic" layers, such as proteomics or methylation data, and explore tissue-specific taxonomies. By mapping the drug universe with increasing precision, DNF provides a flexible tool for clinical applications in precision medicine and the discovery of next-generation therapeutics.

Summary updated

Abstract

ABSTRACT Identification of drug targets and mechanism of action (MoA) for new and uncharacterlzed drugs is important for optimization of drug efficacy. Current MoA prediction approaches largely rely on prior information including side effects, therapeutic indication and/or chemo-informatics. Such information is not transferable or applicable for newly identified, previously uncharacterlzed small molecules. Therefore, a shift in the paradigm of MoA predictions is necessary towards development of unbiased approaches that can elucidate drug relationships and efficiently classify new compounds with basic input data. We propose a new integrative computational pharmacogenomlc approach, referred to as Drug Network Fusion (DNF), to infer scalable drug taxonomies that relies only on basic drug characteristics towards elucidating drug-drug relationships. DNF is the first framework to integrate drug structural information, high-throughput drug perturbation and drug sensitivity profiles, enabling drug classification of new experimental compounds with minimal prior information. We demonstrate that the DNF taxonomy succeeds in identifying pertinent and novel drug-drug relationships, making it suitable for investigating experimental drugs with potential new targets or MoA. We highlight how the scalability of DNF facilitates identification of key drug relationships across different drug categories, and poses as a flexible tool for potential clinical applications in precision medicine. Our results support DNF as a valuable resource to the cancer research community by providing new hypotheses on the compound MoA and potential insights for drug repurposlng.
BibTeX
@misc{elhachem2016integrative,
  author = {Nehmé El-Hachem and Deena M.A. Gendoo and Laleh Soltan Ghoraie and Zhaleh Safikhani and Petr Smirnov and Christina Chung and Kenan Deng and Ailsa Fang and Erin Birkwood and Chantal Ho and Ruth Isserlin and Gary D. Bader and Anna Goldenberg and Benjamin Haibe‐Kains},
  title = {{Integrative pharmacogenomics to infer large-scale drug taxonomy}},
  year = {2016},
  doi = {10.1101/046219},
  howpublished = {bioRxiv},
}
Figure