A benchmark dataset for machine learning in ecotoxicology.

Area of Science:

• Ecotoxicology
• Environmental Science
• Computational Biology

Background:

• Machine learning (ML) holds promise for predicting ecotoxicological outcomes but is underutilized due to data curation challenges.
• A lack of standardized datasets and methodologies hinders model performance comparison across studies.
• Expertise in both ML and ecotoxicology is often required, creating a barrier to entry.

Purpose of the Study:

• To provide an extensive, well-described dataset (ADORE) for acute aquatic toxicity research.
• To facilitate standardized model development and performance evaluation in ecotoxicology.
• To challenge researchers to develop novel ML models for predicting ecotoxicity.

Main Methods:

• Curated an extensive dataset on acute aquatic toxicity, including experimental data.
• Integrated phylogenetic, species-specific, chemical, and molecular data.
• Developed standardized datasets and train-test splits for reproducible research.

Main Results:

• The ADORE dataset encompasses acute aquatic toxicity data for fish, crustaceans, and algae.
• Includes detailed species and chemical properties, alongside molecular representations.
• Provides specific data subsets and splits for defined research challenges.

Conclusions:

• ADORE addresses the need for standardized, high-quality data in ecotoxicological ML research.
• Facilitates reproducible model development and benchmarking.
• Promotes advancements in predicting environmental risks of chemicals.