Predictive and Descriptive CoMFA Models: The Effect of Variable Selection

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Drug discovery

Background:

• Comparative Molecular Field Analysis (CoMFA) is a key technique in drug design.
• The predictive accuracy and stability of CoMFA models can be influenced by variable selection.
• Identifying optimal variable selection strategies is crucial for reliable CoMFA modeling.

Purpose of the Study:

• To evaluate the impact of eight variable selection approaches on CoMFA model performance.
• To determine which variable selection methods enhance the predictive power and stability of CoMFA models.
• To assess the efficiency and information preservation of different variable selection techniques.

Main Methods:

• CoMFA models were developed for three datasets: EPAC antagonists, CD38 inhibitors, and ATAD2 bromodomain inhibitors.
• Initially, CoMFA models were built using all available descriptors.
• Subsequently, new CoMFA models were generated by applying each of the eight variable selection methods to each dataset.

Main Results:

• Noisy and uninformative variables negatively impact CoMFA model outcomes.
• Five variable selection approaches—FFD, SRD-FFD, IVE-PLS, SRD-UVE-PLS, and SPA-jackknife—significantly boosted CoMFA model predictive power and stability.
• SPA-jackknife is highly effective at reducing variable numbers, while FFD retains more variables; SRD-FFD and SRD-UVE-PLS offer rapid computation.

Conclusions:

• Variable selection is essential for improving CoMFA model quality.
• FFD, SRD-FFD, IVE-PLS, SRD-UVE-PLS, and SPA-jackknife are recommended for enhancing CoMFA predictive ability and stability.
• Certain methods like FFD, SRD-FFD, IVE-PLS, and SRD-UVE-PLS effectively preserve crucial information in CoMFA contour maps.