Generalizability Improvement of Interpretable Symbolic Regression Models for Quantitative Structure-Activity...

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Quantitative Structure-Activity Relationship (QSAR) modeling requires both predictive robustness and model interpretability.
• Symbolic Regression (SR) offers globally interpretable models by deriving explicit mathematical expressions.
• Previous SR methods provided human-readable expressions but could be enhanced for broader applicability.

Purpose of the Study:

• To introduce an enhanced symbolic regression method, Filter-Induced Genetic Programming 2 (FIGP2).
• To improve the generalizability and predictive performance of SR models, especially for datasets with complex descriptors.
• To ensure the interpretability and robustness of generated mathematical models.

Main Methods:

• FIGP2 extends a previously developed SR method.
• Incorporates a modified domain filter to eliminate diverging mathematical expressions.
• Introduces a stability metric to prevent overfitting and enhance model generalization.

Main Results:

• FIGP2 demonstrated superior predictive performance compared to the prior SR method and conventional techniques (SVR, MLR) across 12 datasets.
• Generated mathematical expressions were simple, interpretable, and domain-stable.
• The method is applicable to datasets utilizing cost-intensive descriptors.

Conclusions:

• FIGP2 represents a significant advancement in symbolic regression for QSAR modeling.
• The method effectively balances predictive accuracy with model interpretability.
• FIGP2 provides a powerful tool for developing reliable and understandable regression models.