Conceptual Knowledge Discovery in Databases for Drug Combinations Predictions in Malignant Melanoma

Kelly Regan, Satyajeet Raje, Cartik Saravanamuthu, Philip R. O. PayneMedInfo 2015 · DOI

TL;DR

This paper introduces a computational approach, TOKEn, to predict synergistic drug combinations for malignant melanoma by leveraging conceptual knowledge discovery from databases. This method aims to accelerate the identification of promising therapeutic strategies, addressing the limitations of traditional drug discovery and the growing challenge of drug resistance in melanoma.

What Problem Does This Paper Address?

Malignant melanoma is a deadly skin cancer with a poor prognosis for metastatic disease. While targeted therapies like BRAF inhibitors have shown promise, acquired resistance mechanisms limit their long-term effectiveness. Developing synergistic drug combinations is crucial for improving patient outcomes. Traditional methods for discovering these combinations, such as high-throughput screening, are expensive, time-consuming, and inefficient, with a high failure rate in clinical trials. This paper addresses the need for more efficient and efficacious computational approaches to identify these synergistic drug combinations.

Methods

The core of this research is the application of the Translational Ontology-anchored Knowledge Discovery Engine (TOKEn) for in-silico knowledge discovery. TOKEn utilizes a conceptual knowledge engineering approach, specifically "constructive induction," to identify relationships between biomolecular entities and therapeutic agents.

Here's a breakdown of the key steps:

  • Data Acquisition: 42 FDA-approved and investigational melanoma drugs were sourced from the DrugBank database.
  • Concept Extraction: Free-text fields from DrugBank (description, mechanism of action, targets, pharmacodynamics) were mapped to concepts within the Unified Medical Language System (UMLS) using MetaMap. Biomedical ontologies (NCI, SNOMED-CT, MSH, GO) were used for broad coverage.
  • Concept Expansion: The UMLS Metathesaurus was used to expand the initial concepts by identifying parent, child, and semantic relationships, creating an extended concept graph.
  • Constructive Induction: This process generates "induced facts" by identifying transitive paths between drug concepts. The algorithm avoids cycles and constrains concepts to a certain depth. In this study, paths were generated between BRAF inhibitor drugs and non-BRAF inhibitor drugs.
  • Novel Ranking Metric (Drug Combination Score - DCS): To prioritize predicted drug combinations, a novel ranking method was developed. This metric combines two components:
  • Overlap Score (OS): Measures the number of concept unique identifiers that directly map to drug concepts or intermediate concepts within induced facts that overlap with melanoma-associated concepts.
  • Distance Score (DS): Measures the number of induced fact relationships between drug and melanoma concepts, weighted by the number of "hops" (proximal relationships receive higher weight).
  • Melanoma-Specific Prioritization: The TOKEn method was also applied to generate relationships between NCI melanoma-associated concepts (specifically biomolecular properties) and the 42 DrugBank melanoma-associated therapies.

Datasets

  • DrugBank (version 4.1): A comprehensive database containing chemical, pharmacological, and pharmaceutical drug information for 42 FDA-approved and investigational melanoma drugs.
  • Unified Medical Language System (UMLS): Used for mapping and expanding drug-related concepts.
  • Biomedical Ontologies: NCI, SNOMED-CT, MSH, and GO were utilized for concept coverage.
  • NCI Thesaurus (version 14.07d): Used to extract 221 melanoma-associated concepts related to biomolecular properties of the disease.

No new benchmarks were introduced; the study leverages existing structured and semi-structured data.

Key Insights and Ideas

  • Computational Synergy Prediction: The study demonstrates that conceptual knowledge discovery from databases can effectively generate hypotheses for synergistic drug combinations in a complex disease like melanoma.
  • TOKEn's Efficacy: The TOKEn pipeline, by explicitly linking therapeutic modalities to their biomolecular underpinnings, yields informed relationships that can guide combination therapy strategies.
  • Complementary Ranking Metrics: The Overlap Score and Distance Score, while distinct, were found to be significantly correlated, suggesting they offer complementary perspectives for ranking drug combination predictions.
  • Potential for Drug Repurposing: The distance-based metrics used in the ranking method show promise for drug repurposing applications by identifying drug similarities based on conceptual knowledge.

Key Results

  • Extensive Relationship Discovery: The study generated 202,968 induced facts linking BRAF inhibitor drugs to 30 out of 39 non-BRAF inhibitor melanoma drugs.
  • Top-Ranked Candidates: AGRO100 was identified as the highest-ranked non-BRAF inhibitor drug for combination with BRAF inhibitors, followed by Peginterferon-alfa-2a and Interferon alfacon-1.
  • Validation of Known Combinations: The method successfully identified the known FDA-approved combination of trametinib and dabrafenib, linking them through the MAP2K1 and BRAF proteins in the MAPK signaling pathway.
  • Identification of Novel Hypotheses: The study also highlighted potential combinations with investigational drugs like PI-88, suggesting its potential to support inhibition of tumor angiogenesis.
  • Interferon-Based Therapies: Several interferon-based therapies (excluding natural alpha interferon) were highly ranked by their DCS, indicating potential synergistic roles.

Limitations and Future Work

  • Limited Search Space: The study acknowledges that constraining TOKEn to canonical knowledge within DrugBank may have limited the recovery of relationships for some non-BRAF inhibitor drugs.
  • Exclusion of Non-Melanoma Drugs: The study focused only on known melanoma drugs within DrugBank, excluding the possibility of evaluating other existing drugs or small molecules not currently indicated for melanoma.
  • Future Work:
    • Increasing the search space to include other databases and directly extracting data from the literature.
  • Validating findings and the approach through semantic similarity measures and formal subject matter expert review.
  • Assessing the individual contributions of the Overlap and Distance metrics in validated drug combination relationships.
  • Exploring drug repurposing applications more broadly.

Summary updated

BibTeX
@inproceedings{regan2015conceptual,
  author = {Kelly Regan and Satyajeet Raje and Cartik Saravanamuthu and Philip R. O. Payne},
  title = {{Conceptual Knowledge Discovery in Databases for Drug Combinations Predictions in Malignant Melanoma}},
  booktitle = {MedInfo},
  year = {2015},
  doi = {10.3233/978-1-61499-564-7-663},
}
Figure