QuantitativeTox

QuantitativeTox: Meta-ensemble multitask deep learning for quantitative toxicity prediction

QuantitativeTox predicts quantitative chemical toxicity using a meta-ensembling approach over multitask deep learning models that integrate multiple feature representations and base neural network architectures.

Key Features:

• Multi-feature representation: Uses five base deep learning models, each with a distinct feature representation, to capture complementary information relevant to toxicity prediction.
• Meta-ensemble aggregation: Combines outputs from the base models using a separate meta-ensemble model to integrate diverse model predictions.
• Weighted multitask learning: Trains across four quantitative toxicity datasets (LD50, IGC50, LC50, LC50-DM) with weighting to minimize root-mean-square error.
• Model evaluation metrics: Reports root-mean-square error, mean absolute error, and coefficient of determination; includes comparison to TopTox with stated relative improvements across three datasets.

Scientific Applications:

• Toxicology prediction: Supports early identification of potentially hazardous compounds by generating quantitative toxicity estimates for chemical safety evaluation.
• Drug development: Prioritizes compounds by predicted toxicity to inform screening and reduce experimental testing burden.
• Environmental safety assessment: Assists chemical hazard assessment workflows by estimating quantitative toxicity endpoints.

Methodology:

Trains multiple deep learning models on quantitative structure-activity relationship data using diverse feature representations, then applies a separate meta-ensemble model to synthesize base-model outputs into final quantitative toxicity predictions across LD50, IGC50, LC50, and LC50-DM endpoints, optimizing prediction error via weighted multitask learning.