Diagnostic tools to determine the quality of "transparent" regression-based QSARs: the "modelling power" plot

Area of Science:

• * Quantitative Structure-Activity Relationship (QSAR) modeling
• * Cheminformatics and computational chemistry
• * Statistical modeling in drug discovery

Background:

• * Traditional methods for comparing QSAR models can be complex and lack intuitive interpretation.
• * There is a need for robust metrics to assess both the explanatory and predictive capabilities of QSAR models.
• * Evaluating the global importance of descriptors and the potential for SAR vs. QSAR estimations is crucial.

Purpose of the Study:

• * To introduce two novel statistics, descriptive power (Dp) and predictive power (Pp), for QSAR model evaluation.
• * To develop an algorithm for calculating Dp and Pp using multiple linear regression and partial least squares.
• * To facilitate visual comparison and selection of optimal QSAR models.

Main Methods:

• * Development of an algorithm for calculating descriptive power (Dp) and predictive power (Pp).
• * Estimation of Dp using the relative uncertainty of model coefficients.
• * Estimation of Pp using fitted and cross-validated explained variance of the response variable.
• * Validation of the algorithm against commercial QSAR software.

Main Results:

• * Dp and Pp statistics range from 0 to 100%, providing an intuitive measure of model quality.
• * A bivariate plot of Dp versus Pp allows for easy visual comparison of multiple QSAR models.
• * The developed method provides results comparable to established commercial software.

Conclusions:

• * Dp and Pp offer a powerful and intuitive approach to assessing QSAR model quality.
• * These statistics can potentially replace or supplement traditional model evaluation metrics.
• * The Dp-Pp plot serves as a valuable tool for model selection and comparison in QSAR studies.