Leveraging developmental neurotoxicity data in zebrafish embryos through the use of artificial intelligence methods

The zebrafish is a well-known whole organism model to study developmental effects of chemical exposure, including developmental neurotoxicity (DNT). One method used to screen for DNT effects is the zebrafish light-dark transition test (LDT), which measures locomotory behavior under alternating light and dark conditions. Because the LDT is relatively new, experimental protocols and data analysis tools for this assay are still evolving. To advance our knowledge on potential applications of the LDT, we explored using artificial intelligence (AI) methods to distinguish between zebrafish exposed to vehicle controls or pharmaceuticals (fluoxetine, paroxetine, carbamazepine, phenytoin) on 5, 10 or 14 days post-fertilization (dpf).

AI methods were trained to distinguish control versus exposed zebrafish behavior and then applied to assign fish to either group, using five-fold cross-validation. This was done with four different AI methods that differ in nature and complexity, namely Generalized Linear Model (GLM), Random Forest (RF), Gradient Boosting Machine (GBM), and Deep Learning (DL).

Average prediction accuracy increased from 67% at 5 dpf to 76% at 10 and 14 dpf upon continuous exposure. For fish analyzed at 14 dpf but exposed for shorter durations, we found DNT effects clearly persistent for the antidepressants fluoxetine and paroxetine, but less clearly for the anticonvulsants carbamazepine and phenytoin. The AI methods GLM, RF and DL showed comparable performance, whereas GBM accuracies were lower.

Compared to conventional univariate analysis, AI appears more sensitive in detecting DNT effects. Overall, this shows the potential of implementing AI methods in DNT screening of chemicals and further development of this approach.