Prototype-based clustering represents each compound cluster with a prototype, such as a compound or centroid, then assigns compounds to their nearest prototype. The standard k-means workflow initializes centroids, assigns compounds, recomputes means, and repeats until convergence. section 7.2.3
Apply the method carefully to molecular representations:
- Avoid default k-means for binary fingerprints with Tanimoto similarity; it is better suited to normalized continuous descriptors and Euclidean distance. section 7.2.3
- Consider Taylor-Butina clustering for cheminformatics applications, varying neighbor cutoffs and evaluating cluster quality. section 7.2.3
