Building of Robust and Interpretable QSAR Classification Models by Means of the Rivality Index.

Area of Science:

• * Quantitative Structure-Activity Relationship (QSAR) modeling
• * Cheminformatics and computational toxicology
• * Machine learning in drug discovery

Background:

• * QSAR models require unambiguous algorithms, defined applicability domains, and robust validation for regulatory acceptance.
• * Existing QSAR classification methods may lack sufficient reliability and clear applicability domain assessments.
• * There is a need for advanced algorithms that provide reliable predictions and robust applicability domain measures.

Purpose of the Study:

• * To introduce a novel algorithm, RINH (Rivalry Index for Neighbor analysis), for constructing QSAR classification models.
• * To evaluate the RINH algorithm's performance against established machine learning methods.
• * To demonstrate the algorithm's capability in assessing prediction reliability and defining the applicability domain.

Main Methods:

• * Development of the RINH algorithm based on the rivality index, measuring neighbor class distinctions.
• * Application of RINH to four diverse benchmark QSAR datasets.
• * Comparative analysis of RINH against 12 different machine learning algorithms.
• * Validation using 20 additional datasets of varying sizes and balance.

Main Results:

• * The RINH algorithm generated highly accurate QSAR classification models across diverse datasets.
• * RINH provided robust measurements of prediction reliability and model applicability domain.
• * Performance was consistently superior or comparable to 12 other machine learning algorithms.

Conclusions:

• * The RINH algorithm offers a reliable and robust approach for QSAR classification modeling.
• * RINH enhances the interpretability and trustworthiness of QSAR models for regulatory applications.
• * The algorithm effectively addresses the need for clear applicability domain assessment in QSAR studies.