A Semi-Supervised Pattern-Learning Approach to Extract Pharmacogenomics-Specific Drug-Gene Pairs from Biomedical Literature

We develop a semi-supervised pattern learning method to extract drug-gene relationships from free text. Central to our approach is the observation that: the semantic relationship between a drug and a gene can be expressed in many different ways due to the flexibility and expressive nature of human natural language. However, these patterns are not randomly distributed and there are predominant patterns people use to describe specific types of drug-gene relationships. For example, pattern “DRUG is metabolized by GENE” is typically used to describe metabolism relationship between a drug and a gene. Example sentences include “Quetiapine is metabolized by CYP3A4 and sertindole by CYP2D6” (PMID 10422890), and “Cerivastatin is metabolized by CYP2C8 and CYP3A4, and fluvastatin is metabolized by CYP2C9” (PMID 17178259). On the other hand, pattern “GENE inhibitor DRUG” is typically used to express the inhibition relationships between a drug and a gene.Example sentences include “In addition, the effect of the CYP2C9 inhibitor fluvastatin was evaluated using S-warfarin as a probe” (PMID 16758259) and “The CYP2C8 inhibitor gemfibrozil does not increase the plasma concentrations of zopiclone” (PMID 16832679). In this paper, we use two seed patterns for two types of drug-gene relationship extraction: seed “DRUG is metabolized by GENE” for drug-gene metabolism relationship (i.e. quetiapine-CYP3A4, cerivastatin-CYP2C8) extraction and the seed “GENE inhibitor DRUG” for drug-gene target relationship (i.e. fluvastatin-CYP2C9, gemfibrozil-CYP2C8) extraction. First, we use the seed patterns to find their associated drug-gene pairs. Then we iteratively learn new patterns that are associated with the extracted drug-gene pairs and extract corresponding drug-gene relationships from the newly discovered patterns. The iterative process stops when no additional good patterns are found.