QSAR model as a random event: A case of rat toxicity

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Quantitative structure-property/activity relationships (QSPRs/QSARs) predict experimental physicochemical and biochemical properties of substances.
• Accurate predictions from QSPR/QSAR models are crucial for drug discovery and chemical research.
• Model validation is essential before practical application of QSPR/QSAR predictions.

Purpose of the Study:

• To propose an improved validation strategy for QSPR/QSAR models.
• To enhance the reliability assessment of predictive models.
• To establish a robust method for evaluating the domain of applicability.

Main Methods:

• Analyzing geometrical features of data point clusters in experimental vs. calculated value plots.
• Utilizing multiple splits of data into training and test sets for validation.
• Evaluating the probability of correct predictions on external validation sets.

Main Results:

• Geometrical criteria analysis across multiple data splits provides better model reproducibility insights.
• This multi-split approach allows for a more reliable evaluation of QSPR/QSAR model performance.
• The probability of correct prediction on external sets effectively defines the model's applicability domain.

Conclusions:

• Multiple data splits enhance QSPR/QSAR model validation beyond simple geometrical analysis.
• This method improves the estimation of model reproducibility and reliability.
• The proposed validation strategy offers a robust way to determine the domain of applicability for predictive models.