High-dimensional QSAR modelling using penalized linear regression model with L1/2-norm.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Medicinal Chemistry

Background:

• High-dimensional quantitative structure-activity relationship (QSAR) modeling often requires methods for simultaneous descriptor selection and model estimation.
• Penalization techniques are widely adopted for addressing these challenges in QSAR.

Purpose of the Study:

• To propose a novel penalized linear regression model incorporating an L1/2-norm penalty.
• To evaluate the efficacy of this new method in high-dimensional QSAR modeling.

Main Methods:

• A penalized linear regression model with an L1/2-norm penalty was developed.
• The local linear approximation algorithm was employed to handle the non-convexity of the L1/2-norm penalty.
• The proposed method was tested on benchmark datasets.

Main Results:

• The proposed L1/2-norm penalized method demonstrated superior predictive ability compared to existing penalized methods.
• The model provided enhanced interpretability, yielding an easily understandable QSAR model.
• Applicability domain and Y-randomization tests confirmed the model's efficiency and robustness.

Conclusions:

• The proposed L1/2-norm penalized regression model is a promising approach for computational chemistry research.
• It is particularly effective in QSAR scenarios with a high number of molecular descriptors relative to the number of compounds.
• The method offers a robust and interpretable solution for complex QSAR problems.