CommentaryChallenges in using the ToxRefDB as a resource for toxicity prediction modeling

- Challenges were uncovered in applying ToxRefDB to a reproductive toxicity trigger.
- Limitations involved interpretation of toxicity effects and designation of toxicity endpoints.
- Toxicological interpretation of EPA DERs is just as critical as fidelity of input into ToxRefDB.
- ToxRefDB should be used cautiously in prediction model development and performance.
- Designating true positive and true negative reproductive toxicants requires deeper analysis.

Developing and evaluating toxicity prediction models requires selection and use of datasets of known positive and negative agents for the endpoint(s) of interest. EPA's Toxicity Reference Database (ToxRefDB) is a publicly available dataset containing detailed study and effect information on more than 400 chemicals, and it has been used by EPA researchers to develop toxicity prediction models. During an initial evaluation of reproductive toxicity, however, limitations were uncovered in applying data from ToxRefDB that involved interpretation of toxicity effects and designation of toxicity endpoints, core attributes of the database that are critical to its use. These limitations for reproductive toxicity were found to be related, at least in part, to challenges faced in (1) evaluating the source of the original study data (EPA Data Evaluation Records (DERs)) for input into ToxRefDB and (2) interpretation of the biological significance of responses. These limitations of the ToxRefDB have important implications for the wider use of the database as it currently exists. Our results point to a need for improvements to the existing ToxRefDB and/or for researchers to independently evaluate, assign and verify positive or negative designations to data from ToxRefDB before use in development or validation of prediction models or testing frameworks.