A Data-Driven Approach for Interpretable and Efficient Predictive Modeling: A Case Study in SARS-CoV-2 Protease...

Area of Science:

• Chemoinformatics
• Computational Chemistry
• Drug Discovery

Background:

• State-of-the-art chemoinformatic models require feature selection methods that meet all evaluation criteria.
• Developing robust, interpretable, and computationally efficient methodologies is crucial for predictive modeling.

Purpose of the Study:

• To develop a robust methodology for interpretable and efficient predictive models.
• To apply this methodology to discover SARS-CoV-2 main protease inhibitors.
• To identify a transparent and reproducible descriptor selection approach.

Main Methods:

• Evaluation of various descriptor selection procedures.
• Training and testing models on CHEMBL database molecules.
• Validation on an external set of compounds.
• Utilizing the FeatureWiz algorithm combined with stepwise feature selection.

Main Results:

• The FeatureWiz and stepwise selection procedure uniquely satisfied all evaluation criteria for advanced chemoinformatic models.
• Models using 2D descriptors and Ordinary Least Squares regression yielded optimal performance.
• The developed framework demonstrated high interpretability and computational efficiency.

Conclusions:

• The proposed framework offers significant advantages for decision-making in drug discovery due to interpretability and efficiency.
• The derived models are effective, transparent tools for predicting biological activity.
• A validated framework for data-driven decisions in drug discovery and beyond was established.