Linking Semantic Fingerprints of Literature – from Simple Neural Embeddings Towards Contextualized Pharmaceutical Networks

Abstract

The exponential growth of publications in medical digital libraries requires new access paths that go beyond term-based searches, as these increasingly lead to thousands of results. An effective approach for this problem is to extract important pharmaceutical entities and their relations to each other in order to reveal the embedded knowledge in digital libraries. State-of-the-art approaches in the field of neural-language models (NLMs) enable progress in learning and predicting such relations in terms of semantic quality, scalability, and performance and already now make them valuable for important research tasks such as hypothesis generation. However, in the field of pharmacy a simple list of (predicted) associations is often challenging to interpret because, between typical pharmaceutical entities, such as active substances, diseases, and genes, complex associations will exist. A contextualized network of pharmaceutical entities can support the exploration of these associations and will help to assess and interpret predicted relationships. On the other hand, the prerequisite for building meaningful entity networks is an answer to the question: When is an NLM-learned entity relation meaningful? In this paper, we investigate this question for important pharmaceutical entity relations in the form of drug-disease associations (DDAs). To do so, we present a new methodology to determine entity-specific thresholds for the existence of associations. Such entity-specific thresholds open-up the possibility of automatically constructing (meaningful) embedded pharmaceutical networks, which can then be used to explore and to explain learned relationships between pharmaceutical entities.