Adverse Drug Event Prediction Using Noisy Literature-Derived Knowledge Graphs: Algorithm Development and Validation (Preprint)

Abstract

BACKGROUND Adverse Drug Events (ADEs) are unintended side-effects of drugs that cause substantial clinical and economic burden globally. Not all ADEs are discovered during clinical trials and so, post-marketing surveillance, called pharmacovigilance, is routinely conducted to find unknown ADEs. A wealth of information, that facilitates ADE discovery, lies in the enormous and continuously growing body of biomedical literature. Knowledge graphs (KG) encode information from the literature, where vertices and edges represent clinical concepts and their relations respectively. The scale and unstructured form of the literature necessitates the use of natural language processing (NLP) to automatically create such KGs. Previous studies have demonstrated the utility of such literature-derived KGs in ADE prediction. Through unsupervised learning of representations (features) of clinical concepts from the KG, that are used in machine learning models, state-of-the-art results for ADE prediction were obtained on benchmark datasets. OBJECTIVE In literature-derived KGs there is `noise’ in the form of false positive (erroneous) and false negative (absent) nodes and edges due to limitations of the NLP techniques used to infer the KGs. Previous representation learning methods do not account for such inaccuracies in the graph. NLP algorithms can quantify the confidence in their inference of extracted concepts and relations from the literature. Our hypothesis that motivates this work is that by utilizing such confidence scores during representation learning, the learnt embeddings would yield better features for ADE prediction models. METHODS We develop methods to utilize these confidence scores on two well-known representation learning methods – Deepwalk and TransE – to develop their `weighted’ versions – Weighted Deepwalk and Weighted TransE. These methods are used to learn representations from a large literature-derived KG, SemMedDB, containing more than 93 million clinical relations. They are compared with Embeddings of Sematic Predictions (ESP), that, to our knowledge, is the best reported representation learning method on SemMedDB with state-of-the-art results for ADE prediction. Representations learnt from different methods are used (separately) as features of drugs and diseases to build classification models for ADE prediction using benchmark datasets. The classification performance of all the methods is compared rigorously over multiple cross-validation settings. RESULTS The `weighted’ versions we design are able to learn representations that yield more accurate predictive models compared to both the corresponding unweighted versions of Deepwalk and TransE, as well as ESP, in our experiments. Performance improvements are up to 5.75% in F1 score and 8.4% in AUC, thus advancing the state-of-the-art in ADE prediction from literature-derived KGs. Implementation of our new methods and all experiments are available at https://bitbucket.org/cdal/kb_embeddings. CONCLUSIONS Our classification models can be used to aid pharmacovigilance teams in detecting potentially new ADEs. Our experiments demonstrate the importance of modelling inaccuracies in the inferred KGs for representation learning, which may also be useful in other predictive models that utilize literature-derived KGs.