TL;DR
This paper introduces PGxMine, an automated text-mining system designed to help curators of the Pharmacogenomics Knowledgebase (PharmGKB) identify new, relevant scientific literature. By extracting pharmacogenomic associations between genetic variants and drugs from published papers, PGxMine significantly aids the time-consuming curation process, identifying a substantial number of potentially curatable papers.
What Problem Does This Paper Address?
Precision medicine aims to tailor treatments based on an individual's genetic makeup. The Pharmacogenomics Knowledgebase (PharmGKB) is a crucial resource that curates information on how genetic variations affect drug responses and side effects. However, manually reviewing and annotating the ever-increasing volume of scientific literature for these associations is a monumental and time-consuming task for human curators. This bottleneck limits the speed at which new pharmacogenomic knowledge can be incorporated into PharmGKB, potentially delaying the adoption of precision medicine.
Methods
The authors developed PGxMine, a supervised machine learning pipeline to extract pharmacogenomic associations. The core of their approach involves:
- Data Acquisition: Leveraging the PubTator Central resource, which provides pre-annotated biomedical entities (chemicals, genes, variants) from PubMed abstracts and full-text articles.
- Text Alignment: A novel method to precisely map the identified entities from PubTator Central to their exact locations within the text, even when PubTator Central only provides the text string.
- Star Allele Identification: A specialized module to identify and normalize "star alleles" (e.g., CYP2D6*2), which are common in pharmacogenomics literature.
- Sentence Filtering: A keyword-based filtering step to enrich sentences for those likely discussing pharmacogenomic associations, thereby reducing noise for the classifier.
- Relation Classification: Training two "Kindred" relation classifiers using dependency-path based features. These classifiers identify sentences that describe a pharmacogenomic association between a chemical and a variant. A high-precision, low-recall strategy was adopted to prioritize accurate findings.
- Output Generation: Producing three output files: an unfiltered list of associations, a filtered list of high-probability associations, and a collated list that ranks associations by the number of papers discussing them.
Datasets
- PubTator Central: Used as the primary source for pre-annotated biomedical entities (chemicals, variants, genes) from PubMed abstracts and PubMed Central full-text articles.
- DrugBank: Used to map PubTator Central chemical identifiers to PharmGKB chemical IDs and to filter out non-relevant biological molecules.
- Annotated Sentences: 500 sentences were manually annotated for each of the two relation classifiers (one for rs IDs/star alleles, one for DNA/protein modifications) to serve as training and testing data.
Key Insights and Ideas
- Leveraging Existing Resources: The project effectively builds upon existing, large-scale biomedical text-mining resources like PubTator Central, demonstrating a pragmatic approach to tackling complex biological data.
- The Challenge of Star Alleles: The explicit effort to identify and normalize star alleles highlights their importance in pharmacogenomics and the need for specialized handling beyond standard variant annotation.
- High-Precision, Low-Recall Strategy: The choice to prioritize precision over recall in the relation classifier is a strategic decision to minimize false positives for curators, acknowledging that some true positives might be missed but that the identified ones are highly reliable.
- The "Long Tail" of Associations: The finding that a large percentage of extracted associations are discussed in only a single paper is significant. While curators often focus on well-established associations, these "singletons" represent a valuable trove of emerging knowledge that PGxMine can help uncover.
Key Results
- PGxMine identified 19,930 mentions of pharmacogenomic associations across 15,228 sentences from 7,170 papers.
- 41.1% of these associations were found in the title or abstract of the papers.
- The system identified 6,099 unique gene/variant associations.
- Of all variant mentions, 39.9% could be normalized to an rs ID and 39.6% were star alleles.
- A significant portion (75.6%) of the unique associations were discussed in only a single paper.
- In an evaluation by PharmGKB curators, 57 out of 100 associations not previously in PharmGKB led directly to 83 curatable papers.
- An additional 24 associations were likely to lead to curatable papers through citations.
- The most frequently discussed drug in extracted associations was Warfarin, and many common variants belonged to the CYP family.
Limitations and Future Work
- Reliance on PubTator Central: PGxMine's performance is inherently limited by the accuracy and coverage of the PubTator Central resource and its underlying NER tools. Errors in PubTator Central will propagate to PGxMine.
- Limited Full-Text Access: A significant limitation is the poor access to full-text articles for text mining. Only a fraction of curated papers in PharmGKB have full-text available, meaning PGxMine misses many previously curated associations.
- Sentence-Level Extraction: The current system is limited to extracting relations within a single sentence. This can lead to false positives in complex sentences with multiple drugs and variants.
- Handling Complex Haplotypes: Capturing diplotypes or more complex haplotype arrangements remains a challenge.
Future work could involve expanding the system to handle cross-sentence relations, improving the handling of complex genetic variations, and integrating with more comprehensive full-text article access. The authors also note that the resource can be updated regularly as PubTator Central is updated.
Summary updated