You are here:
Biological Interpretation Model for Developmental Neurotoxicity
Citation:
Carstens, K. Biological Interpretation Model for Developmental Neurotoxicity. DNT5, Konstanz, GERMANY, April 07 - 10, 2024. https://doi.org/10.23645/epacomptox.25743081
Impact/Purpose:
This invited presentation will be given at the DNT5 conference in Konstanz, Germany in an effort to share research and network with an international group of researchers and stakeholders. This research falls under EPA's Chemical Safety for Sustainability research program, addressed the need to develop and evaluate new approach methods (NAMs) to assess neurotoxicity potential in a human-relevant model.
Description:
New approach methodologies (NAMs) are being evaluated for their utility to inform developmental neurotoxicity (DNT) hazard instead of using animals. This work compares predictive models for DNT using DNT NAMs data. The dataset comprised 200 chemicals screened in 17 DNT assays with 56 endpoints that measure neurodevelopmental processes: proliferation, differentiation, migration, apoptosis, neurite outgrowth, synaptogenesis and neural network function. The dataset included a chemical training set of 39 chemicals with evidence of in vivo DNT and 13 putative negatives. An unsupervised machine learning (ML) model using K-means clustering, identified distinct bioactivity fingerprints informing differential patterns of DNT and classified DNT reference chemicals with 77% sensitivity and 64% specificity, with 9 false negatives and 5 false positives. No clear trends between chemicals with a shared mode of action and bioactivity fingerprints were observed. A supervised ML model using Random Forest classified putative DNT reference chemicals with 100% sensitivity and 36% specificity, suggesting a lack of specificity for DNT and overfitting. Recursive feature elimination revealed that data from a single endpoint resulted in the same model accuracy as 56 endpoints, suggesting that the suite of 17 assays does not outperform a single informative endpoint. However, these two classification approaches demonstrate poor balanced accuracy, likely explained in part by a small, imbalanced reference chemical set. To address this limitation, we explored the performance of Bayesian modeling which uses a probability-based approach for hypothesis testing. Another supervised ML model using a naïve Bayes classifier model demonstrated 77% sensitivity and 100% specificity, a preliminary result suggesting that a probabilistic approach optimized with techniques to reduce overfitting may improve specificity for DNT. This work highlights current obstacles in modeling with DNT NAMs data, including a limited set of reference negatives, overfitting, and endpoints prone to outliers. Additional screening of DNT reference chemicals, particularly reference negatives, and consideration of additional bioactivity descriptors such as physicochemical properties and toxicokinetic information will likely improve the biological interpretation of a predictive model for DNT. This abstract does not reflect EPA policy.
URLs/Downloads:
DOI: Biological Interpretation Model for Developmental Neurotoxicity
CARSTENS_DNT5_TALK_APRIL2024_V2.PDF (PDF, NA pp, 5384.41 KB, about PDF)