Stacked ensemble\-based mutagenicity prediction model using multiple modalities with graph attention network

Early drug safety screening saves lives and money: The model identifies potentially mutagenic (cancer-causing) compounds earlier in drug development, which prevents unsafe drugs from progressing further and reduces costly late-stage failures.

Multiple perspectives improve accuracy: By analyzing molecules through different "lenses"—chemical structure (SMILES), molecular shape, and network relationships (graph attention)—the model captures more complete information than previous single-method approaches, achieving 95% accuracy on standard datasets.

Explainability builds trust in AI predictions: The use of SHAP analysis means researchers can understand why the model flagged a compound as mutagenic, identifying which specific molecular features drove each prediction rather than treating the model as a "black box."

Ensemble approach reduces individual model bias: By combining multiple machine learning classifiers rather than relying on one, the stacked ensemble method is more robust and less likely to make systematic errors that a single model might produce.

Responsible AI in pharmaceutical development: This work demonstrates how responsible AI practices—combining multiple data sources, using explainability techniques, and rigorous validation—can be applied to real-world drug safety challenges where prediction errors have serious health consequences.