Current approaches for choosing feature selection and learning algorithms in quantitative structure-activity...

Area of Science:

• Computational chemistry
• Cheminformatics
• Statistical modeling

Background:

• Quantitative structure-activity/property relationships (QSAR/QSPR) are statistical models correlating chemical structure descriptors to biological activity, properties, or toxicity.
• Effective QSAR/QSPR model development requires robust strategies including dataset curation, variable selection, and appropriate validation measures.

Purpose of the Study:

• To provide an accessible overview of feature selection methods and statistical learning algorithms for QSAR modeling.
• To guide nonexpert readers in understanding the principles of QSAR model development.
• To highlight the importance of interpretable models and addressing data complexity.

Main Methods:

• Exploration of diverse feature selection techniques to identify relevant molecular descriptors.
• Review of various linear and nonlinear machine learning algorithms for QSAR model building.
• Discussion of strategies to reduce model overfitting and enhance interpretability.

Main Results:

• Different feature selection methods and learning algorithms are crucial for handling complex datasets in QSAR/QSPR.
• These methods aid in selecting important features, minimizing overfitting, and deriving understandable models.
• Novel descriptors necessitate improved feature selection tools for statistically sound and user-friendly models.

Conclusions:

• Advanced feature selection tools are needed to manage new descriptors and develop statistically meaningful, interpretable QSAR/QSPR models.
• Techniques like double cross-validation and consensus modeling are valuable for small datasets to mitigate bias.
• The review aims to empower nonexperts with fundamental knowledge of QSAR modeling principles.