Generation of QSAR sets with a self-organizing map

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• Quantitative structure-activity relationship (QSAR) models are crucial for drug discovery.
• Developing representative training, cross-validation, and prediction sets is vital for robust QSAR modeling.
• Traditional methods for dataset generation can be suboptimal.

Purpose of the Study:

• To evaluate the efficacy of Kohonen self-organizing maps (SOMs) for classifying dihydrofolate reductase inhibitor data.
• To compare SOM-based dataset generation with activity binning and sphere exclusion methods for QSAR modeling.
• To assess the quality and predictive performance of QSAR models derived from SOM-generated sets.

Main Methods:

• Application of Kohonen self-organizing maps (SOMs) for data clustering.
• Utilizing Dragon descriptors for chemical feature extraction.
• Generation of training, cross-validation (CV), and prediction sets using SOM classification.
• Development of QSAR models via the ADAPT methodology.
• Comparison with QSAR models from activity binning and sphere exclusion methods.

Main Results:

• SOMs successfully generated QSAR sets that accurately represent the overall dataset's similarity.
• QSAR models developed using SOM-generated sets were approximately half the size of previously published models.
• The resulting QSAR models exhibited comparable Root Mean Square (RMS) errors to existing models.
• Consistent RMS errors across different QSAR sets indicated good predictive capabilities and generalizability.

Conclusions:

• Kohonen self-organizing maps offer an effective approach for generating representative datasets in QSAR studies.
• SOM-based QSAR modeling can yield smaller, highly predictive models with enhanced generalizability.
• This methodology provides a valuable alternative for dataset partitioning in cheminformatics and drug discovery.