Hybrid reduced graph for SAR studies.

Area of Science:

• * Computational chemistry
• * Chemoinformatics
• * Graph theory

Background:

• * Structure-activity relationship (SAR) studies are crucial for drug discovery.
• * Existing descriptors may have limitations in capturing complex molecular information.
• * Novel descriptors are needed to improve predictive accuracy in SAR.

Purpose of the Study:

• * To define and evaluate new atomic and local hybrid indices.
• * To apply these indices in structure-activity relationship (SAR) studies.
• * To assess the predictive performance of different machine learning classifiers.

Main Methods:

• * Development of new atomic indices inspired by the Refractotopological State Index for Atoms.
• * Definition of local indices, Descriptor Centres (DCs), derived from atomic values.
• * Utilization of various classifiers: multilayer perceptron (MLP), support vector machines (SVM), and meta-classifiers (bagging, decorate).

Main Results:

• * New hybrid descriptors exhibit low mutual correlation coefficients.
• * Meta-classifiers achieved over 92% prediction accuracy, significantly outperforming MLP and SVM (around 60%).
• * Incorporating the distance between Descriptor Centres (DCs) further improved results with meta-classifiers.

Conclusions:

• * The novel hybrid indices and Descriptor Centres are effective for SAR studies.
• * Meta-classifiers demonstrate superior performance for predicting molecular properties using these descriptors.
• * The developed descriptors show promise for enhancing chemoinformatics and drug design efforts.