Novel approach to evolutionary neural network based descriptor selection and QSAR model development

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Machine Learning

Background:

• Quantitative structure-activity relationship (QSAR) models are crucial for drug discovery.
• Descriptor selection significantly impacts QSAR model stability and interpretability.
• Evolutionary neural networks (ENN) offer a powerful framework for QSAR modeling.

Purpose of the Study:

• To evaluate the capability of ENN in guiding descriptor selection towards stable subsets.
• To investigate the influence of descriptor stability on QSAR model generalization and interpretation.
• To compare a novel descriptor subset selection strategy with a standard frequency-based approach.

Main Methods:

• Utilized multiple random dataset partitions to assess descriptor subset (DS) stability and QSAR model generalization.
• Applied Golbraikh et al.'s acceptability criteria for selecting highly predictive QSAR models.
• Collected QSAR models passing the filter generated by ENN for each dataset partition.
• Compared a standard descriptor frequency-based selection principle with a novel approach searching for universal DS solutions.

Main Results:

• The novel approach identified stable descriptor subsets through subsampling, leading to universal solutions.
• Benzodiazepine QSAR models were developed and evaluated based on the proposed principles.
• The study demonstrated the ENN's capability to direct descriptor selection for stable DS composition.
• Performance was benchmarked against existing literature regarding DS composition and predictive accuracy.

Conclusions:

• ENN-based QSAR approaches can effectively guide descriptor selection towards stable subsets.
• Stable descriptor subsets enhance QSAR model generalization and interpretability.
• The novel approach of searching for universal descriptor subsets shows promise for robust QSAR modeling.