Robustness in experimental design: A study on the reliability of selection approaches

Area of Science:

• Chemoinformatics
• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Experimental design in chemoinformatics requires high standards for model performance, reliability, and robustness.
• Existing methods like Kennard-Stone may not adequately address dataset variations or structural diversity.

Purpose of the Study:

• To introduce and evaluate a novel stepwise, adaptive approach (DescRep) for experimental design in chemoinformatics.
• To compare DescRep's performance against established selection methods using statistical evaluation.

Main Methods:

• Developed DescRep: an iterative descriptor selection and representative compound sampling strategy.
• Utilized three diverse datasets for rigorous statistical evaluation.
• Compared DescRep with Kennard-Stone and most descriptive compound selection algorithms.

Main Results:

• DescRep demonstrated superior adaptability to dataset changes.
• Models generated using DescRep exhibited improved error performance and stability.
• The stepwise and adaptive nature of DescRep enhances robustness.

Conclusions:

• Stepwise and adaptive chemoinformatics approaches, like DescRep, offer significant advantages.
• DescRep provides enhanced reliability, consistency, and robustness in predictive modeling.
• This method is crucial for developing high-quality chemoinformatics models.