Multi-Task Cascade Forest Framework for Predicting Acute Toxicity across Species.

Area of Science:

• Computational toxicology
• cheminformatics
• bioinformatics

Background:

• Chemical toxicity evaluation is vital for human health and environmental safety.
• Multi-species acute toxicity prediction (MSATP) traditionally relies on animal testing, raising ethical concerns.
• Existing AI methods using deep neural networks struggle with the small, high-dimensional, and sparse nature of MSATP tabular data.

Purpose of the Study:

• To develop an efficient and accurate artificial intelligence-based framework for MSATP, addressing limitations of current deep learning approaches on tabular data.
• To enhance feature and sample representation for improved multi-task learning in toxicity prediction.
• To validate the framework's performance, generalizability, and interpretability in cross-species toxicity assessment.

Main Methods:

• Proposed a multi-task cascade forest framework integrating feature enhancement via knowledge transfer and sample enhancement using a greedy search with covariance distance.
• Implemented strategies for accommodating varying task sizes and designed the framework specifically for tabular data.
• Conducted single-view (ablation, external dataset) and multi-view (feature fusion, consensus ensemble) experiments to assess performance and generalization.

Main Results:

• Achieved a 12% performance improvement over state-of-the-art methods, with R² = 0.64 and RMSE = 0.57.
• Data enhancement strategies were validated for effectiveness.
• Demonstrated strong generalization capability for cross-species prediction and further performance gains through multi-view learning.

Conclusions:

• The proposed multi-task cascade forest framework effectively addresses MSATP challenges with tabular data.
• The method offers a promising, ethical alternative to animal testing for toxicity prediction.
• The framework shows substantial potential for broad application in chemical safety and risk assessment domains.