Exploring graph-based models for predicting active compounds against triple-negative breast cancer.

Area of Science:

• Oncology
• Computational Biology
• Drug Discovery

Background:

• Triple-negative breast cancer (TNBC) is an aggressive subtype with limited treatment options.
• The absence of HER2, progesterone, and estrogen receptors in TNBC necessitates novel therapeutic strategies.
• Artificial intelligence (AI) and machine learning (ML) accelerate biological data analysis and therapeutic outcome prediction.

Purpose of the Study:

• To develop and validate AI-driven computational models for predicting active compounds against TNBC.
• To identify key structural fragments responsible for the inhibitory activity of compounds against TNBC cells.
• To assess the models' robustness in identifying potential TNBC drug candidates.

Main Methods:

• Curated a dataset of 756 mutant-type compounds from three cell lines.
• Developed four graph-based AI/ML models for predicting TNBC-active compounds.
• Employed stratified nested tenfold cross-validation and the Optuna framework for model optimization and validation.
• Utilized explainability techniques to interpret model predictions and identify key structural features.

Main Results:

• Achieved predictive accuracy with AUC values ranging from 0.65 to 0.82, with the Message Passing Neural Network (MPNN) model showing superior performance.
• Identified critical structural fragments associated with cell inhibition and model predictions.
• External validation using FDA-approved drugs demonstrated prediction accuracies from 66% to 97%.

Conclusions:

• The developed AI models effectively predict compounds with potential inhibitory activity against TNBC cells.
• The MPNN model shows significant promise for accelerating the discovery of novel TNBC drug candidates.
• Explainability techniques provide insights into the structural basis of compound activity, aiding drug design.