Predicting chemicals' toxicity pathway of female reproductive disorders using AOP7 and deep neural networks

Experimental evidence shows that certain chemicals, particularly endocrine disrupting chemicals, may negatively affect the female reproductive system, thereby lowering women's fertility. However, humans are constantly exposed to a number of different chemicals with limited or no experimental data regarding their effect and the mechanism of action in the female reproductive system. To predict chemical hazards to the female reproductive system, we used a previously defined adverse outcome pathway (AOP) that links activation of the peroxisome proliferator-activated receptor gamma to the reproductive toxicity in adult females (AOP7) and the Convolutional Deep Neural Network models that produce meaningful predictions when trained on a significant amount of data. The models trained using CompTox assays with intended molecular and biological targets corresponding to AOP7 achieved high performance (over 90% validation accuracy). The integration of AOP7 and Deep Neural Network identified chemicals that could negatively affect female reproduction through the mechanism described in AOP7. We provide a solution to quickly analyze the data and produce machine learning models to identify potentially active chemicals in the female reproductive system. Although we focused on the female reproductive system, this approach could be valid for a number of other chemicals and AOPs if the right data exist.

Automated summary

Experimental evidence suggests that certain chemicals, especially endocrine disruptors, can have negative effects on the female reproductive system and decrease fertility in women. However, there is limited or no experimental data available for many chemicals regarding their impact and mechanisms of action in the female reproductive system. To predict the hazards of chemicals to the female reproductive system, a previously defined adverse outcome pathway (AOP) and Convolutional Deep Neural Network models were used. The models, trained using CompTox assays with intended molecular and biological targets corresponding to AOP7, achieved high performance in identifying potentially harmful chemicals. This approach provides a solution for quickly analyzing data and generating machine learning models to identify active chemicals in the female reproductive system. It may also be applicable to other chemicals and adverse outcome pathways if sufficient data is available.