Predicting chemical ecotoxicity by learning latent space chemical representations

Area of Science:

• Computational toxicology
• Machine learning in chemistry

Background:

• In silico prediction of chemical ecotoxicity (HC50) complements in vivo and in vitro methods for assessing manufactured chemicals.
• Existing machine learning models for HC50 prediction show variable performance due to challenges in learning effective chemical representations from high-dimensional data.

Purpose of the Study:

• To develop an improved method for predicting chemical ecotoxicity (HC50) using machine learning.
• To enhance the learning of chemical representations for more accurate toxicological assessments.

Main Methods:

• Developed an autoencoder model to learn latent space chemical embeddings for HC50 prediction.
• Compared the autoencoder model's performance against other dimension reduction techniques (PCA, kernel PCA, UMAP) and various regression models (Random Forest, FCNN, LASSO, Ridge).

Main Results:

• The autoencoder model achieved state-of-the-art HC50 prediction performance with R² = 0.668 ± 0.003 and MAE = 0.572 ± 0.001.
• The autoencoder-based approach outperformed PCA, kernel PCA, and UMAP in dimension reduction for HC50 prediction.
• A simple linear layer using autoencoder-learned embeddings surpassed Random Forest, FCNN, LASSO, and Ridge regression models that used raw input features.

Conclusions:

• Learning latent chemical representations is crucial for improving HC50 prediction accuracy.
• The developed autoencoder model offers a robust and effective alternative for predicting chemical ecotoxicity (HC50).