A Multitask Active Learning Framework with Probabilistic Modeling for Multi-Species Acute Toxicity Prediction.

Area of Science:

• Computational toxicology
• Pharmacology
• Machine learning in drug discovery

Background:

• Accurate prediction of acute toxicity across species is crucial for early drug safety assessment.
• Existing methods often struggle with inter-species toxicity mechanism variations and experimental data noise.
• There is a need for robust frameworks that address these challenges in multi-species toxicity prediction.

Purpose of the Study:

• To introduce a novel Probabilistic Multitask Active Learning (PMAL) framework for enhanced multi-species acute toxicity prediction.
• To address limitations in current predictive models, including inter-species divergence and data noise.
• To provide well-calibrated uncertainty estimates for small molecules in toxicity prediction.

Main Methods:

• Developed a Probabilistic Multitask Learning (PML) component to jointly model multiple toxicity endpoints probabilistically.
• Integrated an Uncertainty-based Active Learning (UAL) component to select compounds with high predictive uncertainty for annotation.
• Employed a framework designed to handle noisy, multi-task learning environments.

Main Results:

• The PMAL framework demonstrated superior performance compared to state-of-the-art methods in multi-species acute toxicity prediction.
• PMAL provided reliable, well-calibrated uncertainty estimates for small molecules across various toxicity endpoints.
• Empirical evaluations confirmed the framework's effectiveness in diverse toxicity prediction scenarios.

Conclusions:

• The PMAL framework offers a significant advancement in predicting acute toxicity across different species.
• The approach effectively manages inter-species variations and experimental data noise.
• The core principles of PMAL provide a generalizable paradigm for machine learning in noisy, multi-task settings.