Developing Predictive Models to Facilitate Interpretation of Toxicology Study Results

A computational pipeline to build models to predict target organs of toxicity from SEND datasets has been developed and published on GitHub under PHUSE. Project team members will evaluate the feasibility and performance of this pipeline when run on data from within their organisations. The pipeline will be updated to improve compatibility with different database systems, and efforts will be made to improve its performance across disparate data sources. Additional study interpretations – e.g. adversity of findings, NOAEL determination, clinical translatability, structure activity relationship – will be explored for development of predictive models. Successful modeling approaches will be published in peer-reviewed scientific journal articles.

Expert toxicologists spend hours interpreting the results of multiple studies to support the safety of new drug products. SEND data has enabled the construction of large databases of toxicology study results for building predictive models. These models can be trained using expert interpretations of old study data to make predictions that will streamline the interpretation of future studies.

Many hours are spent by both industry toxicologists and regulatory authority reviewers attempting to properly integrate and interpret the results of toxicology studies conducted to support drug safety. SEND data can be used to train models to use prior interpretations to predict likely interpretations of current studies. This approach will decrease time spent reviewing study results and increase study interpretation across toxicologists because the models will be trained on their interpretive conclusions. The models may also provide valuable insight into which toxicology study endpoints are most relevant for prediction of specific interpretations, e.g. hepatotoxicity, nephrotoxicity.