Research and development of an adverse outcome pathway-focused mechanistic inference tool for 'omics data using semanticknowledge graphs

Project Summary1 Adverse outcome pathways (AOPs) are risk assessment tools that provide a transparent, mechanistic description of a2 stressor resulting in an adverse outcome. Each AOP describes a logical sequence of measurable, causally linked events at3 varying levels of a biological hierarchy (e.g., molecular, cell, organ, and population). Starting with exposure to a stressor,4 and proceeding through a series of key events, each AOP terminates with an adverse outcome for the health of an5 organism or population as a whole. Despite the great popularity of AOPs in risk assessment, there are few tools that can6 exploit the rich mechanistic information they provide, especially quantitatively and in an automated setting. The lack of a7 systematic framework for representing AOPs in a way that is amenable to quantitative analysis is an important obstacle8 that limits their regulatory applications. We have developed a unique system, TOXGRAPH, to address the aforementioned9 limitations. TOXGRAPH integrates biological and biochemical knowledge across a heterogeneous collection of open 10 biomedical ontologies and public databases: a massive, tissue-specific semantic knowledge graph (KG) curated entirely in- 11 house. We have also developed a deep learning-based model that maps unstructured textual AOP descriptions to specific 12 concepts in this KG. To take advantage of these semantic AOP representations, we have developed a framework that 13 allows us to turn AOPs into hypothesis validation tools, by generating enrichment statistics of AOPs an their individual 14 events. This approach can significantly simplify downstream analysis by calculating enrichment statistics at varying levels 15 of granularity, using the results of experimental datasets, such as the results of differential gene expression experiments 16 to quantify AOP enrichment. 17 The research we propose encompasses three specific aims: (1) improve our AOP enrichment statistics; (2) enhance our 18 NLP mapping of AOP event descriptions to biomedical knowledge graph concepts; (3) software engineering for improved 19 user experience with interactive visualizations. In the first aim, we will improve our model to account for multiple 20 experimental doses and time points concurrently for the enrichment of AOP events. In our second aim, we will improve 21 our semantic similarity model to map arbitrary AOP event descriptions to a controlled vocabulary. For this research, we 22 will employ state-of-the-art machine learning, NLP and text mining methodologies. In our third aim, we will develop a 23 web interface to the current command line tool which besides producing the same enrichment results and mechanistic 24 hypotheses will allow users to navigate the results in a lightweight, interactive 3D space. We envision this as a powerful 25 exploratory tool to generate mechanistic hypotheses, with the added ability to show enriched relations across time 26 points in this 3D space. 27 Our overarching goal is to provide an easy, intuitive, and unbiased (data-driven) framework for querying AOPs and 28 researching mechanism of action. TOXGRAPH will directly contribute towards the development of novel non-animal 29 testing strategies and streamline regulatory decision making. Fundamentally, it will greatly facilitate the intended use of 30 AOPs for regulatory applications: to help minimize the uncertainty in decision making.