Hepatotoxicity

Hepatotoxicity predicts drug-induced liver toxicity from toxicogenomics data using gene-selection and machine-learning methods to support safety assessment.

Key Features:

• BioCB gene selection algorithm: The BioCB algorithm prioritizes genes based on their participation in biological processes to select discriminative genes for hepatotoxicity prediction.
• Biological process feature pattern map: Constructs two-dimensional biological process feature pattern maps to represent each drug and capture complex biological interactions beyond raw gene expression levels.
• Dual-strategy prediction framework: Implements a Two-dim branch that uses two-dimensional maps with deep convolutional neural networks and a One-dim branch that vectorizes maps and uses LightGBM.
• Stacked vectorized gene matrix: Proposes a stacked vectorized gene matrix that has demonstrated superior predictive performance compared to traditional raw gene matrices.

Scientific Applications:

• Toxicogenomics validation: Validated using in vivo and in vitro datasets from TG-GATES and DrugMatrix to assess model performance on public toxicogenomics data.
• Comparative modeling: Enables comparison of Two-dim and One-dim strategies, with the One-dim branch reported to outperform the Two-dim deep framework in prediction accuracy.
• Early safety assessment: Applied for early identification of potential drug-induced liver toxicity during drug development.

Methodology:

Computational steps explicitly stated include BioCB gene selection; construction of two-dimensional biological process feature pattern maps; vectorization of maps for the One-dim branch; modeling with deep convolutional neural networks for the Two-dim branch and LightGBM for the One-dim branch; use of a stacked vectorized gene matrix; and validation on in vivo and in vitro datasets from TG-GATES and DrugMatrix.