A study on large-scale disease causality discovery from biomedical literature.

Abstract

Background: Biomedical semantic relationship extraction could reveal important biomedical entities and the semantic relationships between them, providing a crucial foundation for the biomedical knowledge discovery, clinical decision making and other artificial intelligence applications. Identifying the causal relationships between diseases is a significant research field, since it expedites the identification of underlying disease pathogenesis mechanisms and promote better disease prevention and treatment. SemRep is an effective tool for semantic relationship extraction in the biomedical field, but it is not accurate enough for disease causality extraction, bringing challenges for downstream tasks. In this study, we proposed an optimization strategy for SemRep to enhance its accuracy in disease causality extraction.

Methods: This study aims to optimize disease causality extraction of SemRep tool by constructing a semantic predicate vocabulary that precisely expresses disease causality to support the automatic extraction of disease causality knowledge from biomedical literature. The proposed method invloves the following four steps: Firstly, we obtained a collection of semantic feature words expressing disease causality based on current causality predicate studies and the disease causality pairs extracted from SemMedDB. Then, we constructed a disease causality semantic predicate vocabulary by filtering and evaluating the clue words using quantitative comparisons. Following that, we extracted disease causality pairs from the biomedical literature using 36 semantic predicates with an accuracy greater than 80% for more meaningful knowledge discovery. Finally, we conducted knowledge discovery based on the extracted disease causality triples, which primarily includes unidirectional disease causality, bidirectional disease causality, as well as two specific types of disease causality: primary disease causality and rare disease causality.

Results: We obtained a disease causality semantic predicate vocabulary containing 50 textual predicates with an accuracy of above 40%. 36 semantic predicates from the 60% accuracy group were used for disease causality extraction, yielding 259,434 disease causality pairs for subsequent knowledge discovery. Among them, 92,557 types with 176,010 unidirectional disease causality triples, and 6084 types with 83,424 bidirectional disease causality triples were found eventually. Two other types of disease causality, primary disease causality and rare disease causality, were also discovered.

Conclusions: The novelty of this research is that the proposed method enhanced the disease causality extraction of SemRep tool, resulting a more accurate and comprehensive disease causality extraction. It also facilitates an automatic disease causality extraction from large-scale biomedical literature. Additionally, a customized extraction of disease causality for its accuracy and comprehensiveness can be made possible by leveraging the quantified causality predicate vocabulary, allowing for flexible extraction of disease causality according to the actual circumstance.