In silico prediction of chemical ecotoxicity (HC50) is important to improve in vivo and in vitro toxicological assessment of ever-increasing manufactured chemicals. Machine learning models have recently been used to predict chemical HC50, but the prediction performance depends on effectively learning chemical representations from high-dimension data. We hence proposed an autoencoder model to predict chemical HC50 by learning latent space chemical embeddings. This novel approach achieved the state-of-the-art prediction performance of HC50 with R2 = 0.67 and mean absolute error (MAE) = 0.56, and outperformed other dimension reduction methods including principal component analysis method (R2 = 0.60 and MAE = 0.63) and uniform manifold approximation and projection dimensionality reduction method (R2 = 0.36 and MAE = 0.80). In addition, a simple linear layer with the learned chemical embeddings from the autoencoder performed better than random forest (R2 = 0.66 and MAE = 0.59) and fully connected neural network (R2 = 0.60 and MAE = 0.62) with raw input features. Our results highlighted the usefulness of learning latent chemical representation on HC50 prediction with the autoencoder model.

Predicting chemical ecotoxicity by learning latent space chemical representations

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.