AdapTor: Adaptive Topological Regression for quantitative structure-activity relationship modeling.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery and development

Background:

• Quantitative Structure-Activity Relationship (QSAR) modeling is crucial for drug design.
• Topological Regression (TR) offers efficiency and interpretability but has limitations in anchor selection and reconstruction.
• Existing QSAR models, including deep learning approaches, face challenges in balancing predictive power and interpretability.

Purpose of the Study:

• To introduce Adaptive Topological Regression (AdapToR), an enhanced QSAR model.
• To overcome the limitations of standard TR by implementing adaptive anchor selection and optimization-based reconstruction.
• To improve the accuracy, interpretability, and computational efficiency of drug response prediction models.

Main Methods:

• Developed AdapToR with novel adaptive anchor selection strategies.
• Implemented an optimization-based approach for response reconstruction.
• Evaluated AdapToR on the NCI60 GI50 dataset, comprising over 50,000 drug responses across 60 cancer cell lines.
• Compared AdapToR against Transformer CNN, Graph Transformer, TR, and other baseline QSAR models.

Main Results:

• AdapToR demonstrated superior performance in predicting drug responses compared to existing QSAR models.
• The proposed method achieved significantly lower computational costs than deep learning-based models.
• AdapToR offers greater interpretability in QSAR modeling compared to complex deep learning architectures.

Conclusions:

• AdapToR represents a significant advancement in QSAR modeling for drug response prediction.
• The model effectively balances predictive accuracy, interpretability, and computational efficiency.
• AdapToR holds promise for accelerating drug discovery and development processes.