Learning from biomedical linked data to suggest valid pharmacogenes.

Area of Science:

• Pharmacogenomics
• Computational Biology
• Bioinformatics

Background:

• Identifying pharmacogenes is crucial for understanding drug response variability.
• Genomic studies often yield false positives, necessitating computational approaches.
• Existing methods primarily use molecular networks or literature, overlooking other data sources.

Purpose of the Study:

• To develop a computational approach for identifying and prioritizing pharmacogenes.
• To leverage diverse linked data resources beyond traditional sources.
• To compare the performance of machine learning methods in pharmacogene identification.

Main Methods:

• Integrated diverse linked data sources (e.g., DisGeNET, Clinvar) related to genes, drugs, and diseases.
• Employed machine learning, specifically random forest and graph kernel methods, for classification.
• Classified gene-drug pairs as pharmacogenomically associated or not based on linked data.

Main Results:

• Assembled a comprehensive linked dataset of 2,610,793 triples from six resources.
• Random forest achieved a F-measure of 0.73 in identifying pharmacogenes.
• Graph kernel method demonstrated superior performance with a F-measure of 0.81.

Conclusions:

• Machine learning, particularly graph kernel methods, effectively identifies valid pharmacogenes using diverse linked data.
• The proposed approach enhances pharmacogene discovery by integrating multiple data types.
• The study provides a prioritized list of candidate pharmacogenes for further research.