On the Best Way to Cluster NCI-60 Molecules

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Machine learning models are crucial in early drug design.
• Cross-validation, often using molecular clustering, validates these models.
• Ineffective clustering compromises validation, especially with dissimilar test molecules.

Purpose of the Study:

• To identify the optimal molecular clustering method for the National Cancer Institute (NCI)-60 dataset.
• To compare hierarchical, Taylor-Butina, and Uniform Manifold Approximation and Projection (UMAP) clustering algorithms.
• To assess the impact of outlier removal on clustering quality.

Main Methods:

• Comparison of hierarchical, Taylor-Butina, and UMAP clustering techniques.
• Evaluation of clustering quality using three standard metrics.
• Assessment of cluster quality via an average similarity matrix.
• Analysis of clustering performance with and without outlier removal.

Main Results:

• Clustering quality varied significantly across the tested methods.
• Hierarchical and Taylor-Butina methods demonstrated poor clustering performance and high computational cost.
• Uniform Manifold Approximation and Projection (UMAP) yielded superior clustering quality.

Conclusions:

• UMAP is the recommended method for analyzing the NCI-60 dataset due to its high clustering quality.
• Effective molecular clustering is essential for robust validation of machine learning models in drug design.
• UMAP offers a promising approach for enhancing cross-validation strategies in cheminformatics.