Network-based piecewise linear regression for QSAR modelling.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• Quantitative Structure-Activity Relationship (QSAR) models are essential for drug discovery, aiding in lead optimization and virtual screening.
• There is a growing demand for QSAR models that are not only predictive but also interpretable.
• Current methods often lack interpretability, hindering a deeper understanding of structure-activity relationships.

Purpose of the Study:

• To propose a novel methodology for building interpretable QSAR models.
• To introduce the modSAR algorithm, which integrates network analysis and piecewise linear regression.
• To demonstrate the effectiveness of modSAR in drug discovery tasks.

Main Methods:

• The modSAR algorithm employs a two-step data splitting procedure.
• Compounds are represented as networks based on structural similarity to identify modules of related chemical properties.
• Each module is further divided into regions, with each region modeled by a separate linear equation.

Main Results:

• modSAR demonstrates comparable predictive accuracy to established algorithms like Random Forest and Support Vector Machine across five protein inhibitor datasets.
• The models generated by modSAR are interpretable, allowing for an evaluation of the compound applicability domain.
• modSAR effectively supports virtual screening and the development of new drug leads.

Conclusions:

• modSAR provides a powerful approach to developing interpretable QSAR models.
• The method enhances understanding of structure-activity relationships, crucial for drug design.
• modSAR is a valuable tool for virtual screening and accelerating the discovery of novel drug candidates.