The Bigger Fish: A Comparison of Meta-Learning QSAR Models on Low-Resourced Aquatic Toxicity Regression Tasks

Area of Science:

• Environmental Toxicology
• Computational Chemistry
• Artificial Intelligence

Background:

• Sparse toxicological data necessitates alternative methods to animal testing.
• Quantitative structure-activity relationship (QSAR) models are crucial for predicting chemical toxicity.
• Aquatic toxicity datasets often present low-resource challenges due to limited compound data per species.

Purpose of the Study:

• To benchmark meta-learning techniques for building QSAR models.
• To evaluate knowledge-sharing strategies between different species for toxicity prediction.
• To identify optimal AI approaches for low-resource aquatic toxicity modeling.

Main Methods:

• Benchmarking of transformational machine learning, model-agnostic meta-learning, fine-tuning, and multi-task learning.
• Application of meta-learning for knowledge sharing across aquatic species.
• Comparison of single-task versus knowledge-sharing approaches in QSAR modeling.

Main Results:

• Established knowledge-sharing techniques significantly outperform single-task QSAR models.
• Multi-task random forest models demonstrated competitive or superior performance compared to other methods.
• Meta-learning approaches proved effective in low-resource settings for aquatic toxicity prediction.

Conclusions:

• Multi-task random forest models are recommended for aquatic toxicity modeling due to their robustness and performance.
• Meta-learning effectively facilitates knowledge transfer across species, enhancing QSAR model accuracy.
• The developed models support species-level toxicity prediction across diverse phyla and chemical domains.