Science Inventory

Development of computationally predicted Adverse Outcome Pathway (AOP) networks through data mining and integration of publicly available in vivo, in vitro, phenotype, and biological pathway data


Oki, N., S. Bell, R. Wang, M. Nelms, AND S. Edwards. Development of computationally predicted Adverse Outcome Pathway (AOP) networks through data mining and integration of publicly available in vivo, in vitro, phenotype, and biological pathway data. Pacific Symposium on Biocomputing (PSB), Kohala Coast, Big Island, HI, January 04 - 08, 2016.


Adverse Outcome Pathways can expand and enhance the use of ToxCast and other in vitro toxicity information by providing a mechanistic link to adverse outcomes of regulatory concern, but the expert-driven development of AOPs is labor-intensive and time-consuming. This work is intended to complement that process by creating a broad array of computationally-predicted AOPs (cpAOPs) by data mining of publicly available data. This increases the coverage of AOPs and provides minimal information in many cases where nothing is available otherwise. It also provides a starting point for expert-driven development of AOPs and thereby potentially accelerating that process.


The Adverse Outcome Pathway (AOP) framework is increasingly being adopted as a tool for organizing and summarizing the mechanistic information connecting molecular perturbations by environmental stressors with adverse outcomes relevant for ecological and human health outcomes. However, the conventional process for assembly of these AOPs is time and resource intensive, and has been a rate limiting step for AOP use and development. Therefore computational approaches to accelerate the process need to be developed. We previously developed a method for generating computationally predicted AOPs (cpAOPs) by association mining and integration of data from publicly available databases. In this work, a cpAOP network of ~21,000 associations was established between 105 phenotypes from TG-GATEs rat liver data from different time points (including microarray, pathological effects and clinical chemistry data), 994 REACTOME pathways, 688 High-throughput assays from ToxCast and 194 chemicals. A second network of 128,536 associations was generated by connecting 255 biological target genes from ToxCast to 4,980 diseases from CTD using either HT screening activity from ToxCast for 286 chemicals or CTD gene expression changes in response to 2,330 chemicals. Both networks were separately evaluated through manual extraction of disease-specific cpAOPs and comparison with expert curation of the relevant literature. By employing data integration strategies that involve the weighting of network edges for prioritization, the two networks can be merged into a global network of cpAOPs. Automated extraction techniques can then be used to identify individual cpAOPs that connect molecular perturbations with adverse outcomes by traversing the network for the most probable paths based on the weighted edges. This will result in a more comprehensive hypothetical AOP list than is possible by expert evaluation alone. Our workflow allows for additional datasets to be integrated into the global network regardless of the methods used to generate the individual networks, which means that the value of this resource will continue to grow as additional datasets are added. These methods highlight the value and utility of data mining and integration strategies in the development and assembly of AOPs. When prioritized, a rapid review of the resulting list of putative AOPs found can be performed by domain experts given an adverse outcome of interest.The views expressed in this abstract are those of the authors and do not necessarily reflect the views or policies of the U.S. EPA

Record Details:

Product Published Date: 01/18/2016
Record Last Revised: 03/29/2016
OMB Category: Other
Record ID: 311584