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

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Mutagenicity is a significant concern due to its association with genetic mutations and cancer.
• Early detection of mutagenic compounds in drug development is vital for safety and cost-efficiency.
• Current computational mutagenicity prediction models often rely on single data modalities.

Purpose of the Study:

• To develop a novel stacked ensemble model for mutagenicity prediction.
• To integrate multiple molecular data modalities for improved prediction accuracy.
• To enhance the early identification of mutagenic drug candidates.

Main Methods:

• Utilized a stacked ensemble machine learning model.
• Integrated multiple molecular data modalities: SMILES (substructural, physicochemical, geometrical) and molecular graphs (topological features via Graph Attention Network).
• Employed SHAP (Shapley Additive Explanations) for feature and classifier significance analysis.

Main Results:

• The proposed multi-modal model demonstrated superior performance compared to state-of-the-art methods.
• Achieved an Area Under the Curve (AUC) of 95.21% on the Hansen benchmark dataset.
• Successfully identified significant features and classifiers contributing to mutagenicity prediction.

Conclusions:

• The novel stacked ensemble model effectively integrates diverse molecular information for accurate mutagenicity prediction.
• This approach offers a significant advancement in early drug safety assessment.
• The findings are relevant for clinicians and computational biologists in translational research.