Optimal Piecewise Linear Regression Algorithm for QSAR Modelling.

Area of Science:

• * Computational chemistry and cheminformatics.
• * Development of predictive modeling techniques for drug discovery.

Background:

• * Quantitative Structure-Activity Relationship (QSAR) models are crucial in drug discovery for lead optimization and virtual screening.
• * Recent advancements prioritize predictive power, often at the expense of model interpretability.
• * A gap exists for QSAR models that are both predictive and transparent.

Purpose of the Study:

• * To introduce and evaluate OPLRAreg, a piecewise linear regression algorithm for creating interpretable QSAR models.
• * To demonstrate that OPLRAreg offers a balance between predictive accuracy and model transparency.
• * To provide a flexible and customizable alternative to existing machine learning algorithms in QSAR.

Main Methods:

• * Application of a piecewise linear regression algorithm (OPLRAreg) to QSAR modeling.
• * Algorithm identifies optimal data partitioning features and regional linear prediction equations.
• * Incorporation of a regularization term for overfitting prevention and implicit feature selection.
• * Leveraging mathematical programming for customizable constraint integration.

Main Results:

• * OPLRAreg successfully develops QSAR models that are both predictive and interpretable.
• * The algorithm demonstrates comparable predictive accuracy to established methods like Random Forest and Support Vector Machine.
• * Tests on five QSAR datasets from the ChEMBL database validate the model's performance.
• * Regularization effectively prevents overfitting and selects key features.

Conclusions:

• * OPLRAreg presents a viable, interpretable alternative to current black-box machine learning algorithms in QSAR.
• * The algorithm's foundation in mathematical programming allows for adaptable model constraints.
• * This approach facilitates a deeper understanding of structure-activity relationships in drug discovery.