Topological regression as an interpretable and efficient tool for quantitative structure-activity relationship...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Quantitative structure-activity relationship (QSAR) models are vital in drug discovery but often lack interpretability, limiting their use in molecular design.
• Current interpretable QSAR methods may not match the predictive power of complex models like deep learning.

Purpose of the Study:

• To introduce Topological Regression (TR), a novel similarity-based regression framework for drug response prediction.
• To evaluate the performance and interpretability of TR against deep learning QSAR models.
• To demonstrate TR's utility in extracting meaningful relationships between chemical space and biological activity.

Main Methods:

• Developed a similarity-based regression framework named Topological Regression (TR).
• Compared TR's predictive performance against deep learning QSAR models using 530 ChEMBL human target activity datasets.
• Analyzed the interpretability of TR by examining the extracted mapping between chemical and activity spaces.

Main Results:

• The sparse TR model achieved predictive performance comparable to, or exceeding, that of deep learning-based QSAR models.
• TR provided more intuitive interpretations by revealing an approximate isometry between drug chemical space and activity space.
• TR demonstrated statistical grounding and computational efficiency.

Conclusions:

• Topological Regression (TR) presents a powerful, interpretable, and efficient alternative for quantitative structure-activity relationship (QSAR) modeling in drug discovery.
• TR enhances molecular design by offering interpretable insights into the relationship between chemical structures and biological activity.
• The framework facilitates a deeper understanding of drug response prediction beyond purely predictive accuracy.