Accelerated variant curation from scientific literature using biomedical text mining

Abstract

Biological databases collect and standardize data through biocuration. Even though major model organism databases have adopted some automation of curation methods, a large portion of biocuration is still performed manually. To speed up the extraction of the genomic positions of variants, we have developed a hybrid approach that combines regular expressions, Named Entity Recognition based on BERT (Bidirectional Encoder Representations from Transformers) and bag-of-words to extract variant genomic locations from C. elegans papers for WormBase. Our model has a precision of 82.59% for the gene-mutation matches tested on extracted text from 100 papers, and even recovers some data not discovered during manual curation. Code at: https://github.com/WormBase/genomic-info-from-papers.

Figure 1. Overview of the data extraction process and key results

(A) Using database publication IDs as input, the sentences from the corresponding papers are extracted using wbtools or TextPresso, and then processed by the hybrid system where all mutation mentions are extracted. Additional information relevant for curation is then extracted using bag-of-words. The gene and mutation matches are verified by validating that the reference base/codon is possible in at least one transcript in the gene. The final matches and additional data (eg strain name) are then presented in a structured output for validation by curators before being used to update the database. (B) Metrics on sentence classification with mutation mention tested on IDP4+ corpus.