Periodic Reporting for period 1 - kANNa (Knowledge graph completion using Artificial Neural Networks for Herb-Drug Interaction discovery)
Período documentado: 2019-04-01 hasta 2021-03-31
Herbal medicinal products are commonly used in combination with conventional drugs, therefore there is an increased risk of serious interactions between them. In Eastern Europe 51.8% of women use herbal medicines in pregnancy and in Germany 70% of the population reported using “natural medicines”, often as a complement to conventional forms of therapy. This raises concerns especially in pharmacovigilance, a field of study concerned with the identification, evaluation and prevention of adverse drug reactions. While the effects of certain herbal drugs such as St John’s Wort are well understood, there is still little information on possible interactions for many drugs and herbal products. For this reason, the European Medicines Agency (EMA) is actively monitoring medical literature to identify adverse reactions for a predefined list of herbs to populate its pharmacovigilance database. But their list is far from being exhaustive and manual curation does not scale up in the face of an increasingly frequent publication rate. Take for example the case of kanna (sceletium tortuosum), a plant used for its effects against anxiety. Because of its neuro-receptor activities there are plausible concerns that it might interact with psychiatric drugs or cardiac medication. Although clinical evidence is still limited, understanding interaction mechanisms by comparing with similar plants can inform risk assessment.
The main objective of the kANNa project is to automatically extract facts about herb-drug interactions from unstructured text through information extraction, combining and linking extracted information with rich information already available in general-purpose and domain-specific knowledge bases. To achieve this, the kANNa research program pursues the following measurable objectives: (i) Integrate information extraction into the process of monitoring herb-drug interactions from medical literature; (ii) Enhance additional knowledge acquisition from sparse, incomplete, and unreliable evidence; (iii) Provide support for clinical decision making and promote collaboration and reuse over the acquired knowledge base.
The main objective of the kANNa project is to automatically extract facts about herb-drug interactions from unstructured text through information extraction, combining and linking extracted information with rich information already available in general-purpose and domain-specific knowledge bases. To achieve this, the kANNa research program pursues the following measurable objectives: (i) Integrate information extraction into the process of monitoring herb-drug interactions from medical literature; (ii) Enhance additional knowledge acquisition from sparse, incomplete, and unreliable evidence; (iii) Provide support for clinical decision making and promote collaboration and reuse over the acquired knowledge base.
FIDEO, a knowledge graph of food-drug and herb-drug interactions was populated based on herb-drug interactions described in the Stockley’s Herbal Medicines Interactions compendium, the French HEDRINE database, the DrugBank knowledge base, and Thériaque database. FIDEO describes information about drugs, foods, plants, interactions and interaction mechanisms. We proposed a feature selection approach for automatically identifying appropriate queries to construct a domain-specific corpus. In addition, we proposed an evaluation dataset and methodology for automatically extracting taxonomies from the Wikipedia category structure.
An interactive visual interface was developed showing side by side two hierarchies and matching elements between those hierarchies. This prototype allows users to search and explore interactions between foods or plants and drugs.
An interactive visual interface was developed showing side by side two hierarchies and matching elements between those hierarchies. This prototype allows users to search and explore interactions between foods or plants and drugs.
kANNa goes beyond the state of the art in herb-drug interactions discovery by extracting and representing supporting evidence in the form of pharmacokinetic and pharmacodynamic mechanisms. kANNa fosters data integration by linking biological and biomedical entities represented in existing ontologies and terminologies from the Linked Open Data cloud and contributes a new multilingual Linked Data resource on herb-drug interactions. The knowledge base covers English and French, as a first step for covering other languages. A novel aspect of this work is the application of knowledge graph completion to the biomedical domain, that has specific challenges related to data integration, knowledge base size, and a need for high accuracy and confidence level. Compared to previous work on extracting drug-drug interactions that is limited to inferring interactions, kANNa will additionally classify interactions based on clinical significance (e.g. major, moderate, minor) and documentation level. To accomplish this kANNa brings together multidisciplinary expertise from Pharmacognosy, Pharmacovigilance, Artificial Intelligence, and Human Computer Interaction.