Predicting the acute ecotoxicity of chemical substances by machine learning using graph theory

Area of Science:

• Environmental Chemistry
• Computational Toxicology
• cheminformatics

Background:

• Conventional ecotoxicity prediction models like ECOSAR rely on empirical clustering and log P linear regression, limiting accuracy improvements with new data.
• Existing methods struggle with inorganic and ionized compounds and present challenges in managing multiple Quantitative Structure-Activity Relationship (QSAR) formulas for single substances.
• Accurate in silico ecotoxicity assessment is crucial for environmental risk management.

Purpose of the Study:

• To develop a novel in silico method for predicting chemical substance acute ecotoxicity.
• To overcome limitations of conventional methods, including empirical classification and inability to handle diverse chemical structures.
• To improve prediction accuracy and applicability across a broader range of chemical compounds.

Main Methods:

• Application of unsupervised machine learning and graph theory for ecotoxicity prediction.
• Utilized the AIST-MeRAM ecotoxicity dataset and Molecular ACCess System (MACCS) keys for chemical structure vectorization (166-bit binary information).
• Developed a novel approach independent of log P and linear regression models.

Main Results:

• The new method accurately predicts the acute toxicity of fish, daphnids, and algae.
• Achieved good prediction accuracy without relying on log P values or linear regression.
• Cross-validation confirmed superior accuracy compared to ECOSAR predictions for a wider chemical spectrum.

Conclusions:

• The developed unsupervised machine learning and graph theory method offers a significant advancement in predicting chemical ecotoxicity.
• This novel approach demonstrates enhanced accuracy and broader applicability, including for inorganic and ionized compounds.
• The method provides a more robust and versatile tool for ecological risk assessment and management.