TL;DR
This paper presents a novel ontology-based framework for representing and reasoning over pharmacogenomic knowledge to enable genome-based clinical decision support. By leveraging OWL 2 DL ontologies and automated reasoning, the system can identify errors in knowledge bases, assign patient genotypes to clinical guidelines, and detect inconsistencies in recommendations, paving the way for more personalized medicine.
What Problem Does This Paper Address?
Every year, a significant number of patients experience treatment failure or adverse drug reactions (ADRs) that could be prevented with pharmacogenomic testing. The core challenge lies in the fragmented and often unstructured nature of pharmacogenomic knowledge, making it difficult to build reliable IT systems for clinical application. This scattered information, spread across disparate data structures and formalizations, leads to ambiguity and complexity. The authors aim to bridge this gap by creating a unified, logical representation of this knowledge that can be queried and reasoned upon to provide actionable clinical insights.
Methods
The authors developed a comprehensive framework using Web Ontology Language (OWL 2 DL) ontologies and automated reasoning. Key methodological steps include:
- Ontology Development: They created an OWL 2 ontology, the "Genomic CDS ontology," to represent pharmacogenomic knowledge. This involved:
- Manually defining top-level classes and automatically generating lower-level classes.
- Converting dbSNP records for SNPs into OWL axioms.
- Formalizing allele and haplotype definitions from sources like PharmGKB as OWL axioms, using "necessary and sufficient" conditions and qualified cardinality restrictions.
- Curating pharmacogenomic decision support rules and drug-response phenotype inference rules from clinical guidelines (CPIC, Dutch Pharmacogenomics Working Group) and FDA drug labels.
- Patient Data Representation: Individual patient genetic profiles were represented as OWL individuals, allowing them to be classified and reasoned upon within the ontology. Services were developed to automatically create these representations from common data formats like 23andMe files and VCF files.
- Automated Reasoning: The OWL reasoner was used to:
- Infer patient alleles and phenotypes from their genetic data.
- Match patient profiles to relevant clinical decision support (CDS) rules.
- Detect errors, inconsistencies, and overlaps within the pharmacogenomic knowledge base and between different guidelines.
Datasets
The paper utilized several data sources:
- dbSNP: Used for identifying and obtaining information on Single Nucleotide Polymorphisms (SNPs).
- PharmGKB (Pharmacogenomics Knowledge Base): Provided lists of "very important pharmacogenes," associated SNPs, and haplotype definitions.
- FDA Drug Labels: Served as a source for pharmacogenomic decision support rules.
- CPIC (Clinical Pharmacogenetics Implementation Consortium) and Dutch Pharmacogenomics Working Group Guidelines: Curated for pharmacogenomic decision support rules and drug-response phenotype inference rules.
- Publicly Available Genetic Profiles: A collection of 935 genetic profiles from the openSNP repository (derived from 23andMe and VCF formats) was used for evaluation. These profiles were manually altered to create a virtual patient dataset for testing.
Key Insights and Ideas
- Ontologies as a Unifying Framework: The core idea is that OWL ontologies can provide a formal, unambiguous, and machine-readable way to integrate disparate pharmacogenomic knowledge, moving beyond unstructured text and semi-structured formats.
- Automated Reasoning for Quality Assurance: The system's ability to automatically detect inconsistencies (e.g., overlapping patient populations with conflicting recommendations) and errors in knowledge bases is a significant advantage for building reliable CDS systems.
- Bridging the Gap from Raw Data to Clinical Action: The framework effectively maps raw genetic variants to higher-level clinical concepts like alleles, haplotypes, and phenotypes, ultimately linking them to actionable CDS messages.
- Performance Trade-offs in OWL Reasoners: The evaluation highlighted substantial performance differences among OWL 2 DL reasoners, with TrOWL showing significant speed advantages due to its syntactic approximation approach, though potential completeness concerns were noted.
Key Results
- The final Genomic CDS ontology represented 336 SNPs, 707 variants, 665 haplotypes related to 43 genes, 22 rules for drug-response phenotypes, and 308 clinical decision support rules.
- OWL reasoning successfully identified CDS rules with overlapping target populations but differing treatment recommendations.
- For a collection of 943 public genetic profiles, an average of 188 SNPs per profile matched ontology polymorphisms, and nearly 6 decision support rules were triggered on average.
- Significant performance differences were observed among OWL 2 reasoners, with TrOWL completing classification and realization in 4.62 seconds on average, while others failed to terminate within 3 hours.
Limitations and Future Work
- Data Completeness: A modest number of CDS rules were triggered because many public genetic profiles lacked information on specific SNPs required for inferring certain alleles/haplotypes.
- Phenotype Definition: There is a lack of clearly defined phenotypes for many pharmacogenes, with differing descriptions across guideline sources, which can lead to ambiguity.
- Ontology Scope: The current ontology does not model patient treatment regimes (specific drugs, dosages), requiring external applications for filtering and refinement of recommendations.
- Future Work:
- Integrating Genomic CDS and OWL reasoning into a clinical decision support application.
- Broadening collaboration with key organizations (PharmGKB, CPIC, etc.) to develop well-defined, formalized, and logically coherent representations.
- Exploring the possibility of mapping the Genomic CDS ontology to external genomic databases (e.g., VariO).
- Investigating the scalability of the design patterns for larger datasets.
- Modeling treatment regimes and integrating with electronic health record infrastructures.
- Developing best practices for integrating pharmacogenomic data into clinical workflows.
Summary updated