PV_LOO-Based Training Set Selection Improves the External Predictability of QSAR/QSPR Models

Area of Science:

• Quantitative Structure-Activity Relationships (QSAR)
• Quantitative Structure-Property Relationships (QSPR)
• Cheminformatics
• Machine Learning

Background:

• Statistical external validation is crucial for assessing QSAR/QSPR model predictability.
• Training set selection significantly impacts model performance.
• Gaussian processes (GP) offer a probabilistic framework for modeling.

Purpose of the Study:

• Develop a novel training set division method for QSAR/QSPR.
• Improve the external predictability of predictive models.
• Investigate the utility of posterior variance of leave-one-out cross-validation (PV_LOO) in Gaussian processes for training set selection.

Main Methods:

• Developed a division algorithm based on GP's PV_LOO.
• Collected and analyzed four diverse, high-quality chemical data sets.
• Compared PV_LOO-based selection with Kennard-Stone and random division.
• Utilized squared exponential, rational quadratic, and neural network covariance functions.

Main Results:

• Training sets selected using higher PV_LOO values demonstrated statistically superior external predictability.
• PV_LOO-based selection outperformed Kennard-Stone and random division methods.
• Root Mean Squared Error (RMSE) of external validation was the primary comparison metric.

Conclusions:

• PV_LOO of Gaussian processes can effectively guide training set selection for enhanced QSAR/QSPR model predictability.
• Higher PV_LOO values may correlate with greater mechanism diversity within training compounds.
• This approach offers a statistically robust method for optimizing predictive model development.