PFClust

PFClust performs parameter-free clustering by automatically determining an optimal number of clusters for a dataset using a partitioning-based approach for applications such as protein domain classification and general biological data analysis.

Key Features:

• Automatic Cluster Number Determination: Automatically identifies a suitable number of clusters using intrinsic data characteristics such as minimum expectation value and intra-cluster similarity variance.
• Versatile Clustering Approach: Can partition datasets into any number of clusters from one to n objects and accommodates hierarchical, partitional, agglomerative, and divisive approaches.
• Heuristic Method: Operates heuristically by evaluating multiple potential cluster configurations rather than optimizing a single global metric.
• Validation and Performance: Validated on synthetic 2D vectors and real-world data from the CATH structural bioinformatics database, showing performance at least equivalent to six other methods even when those methods were given the number of clusters.
• Application in Protein Domain Classification: Applied to classify three-dimensional structures of protein domains with strong agreement to manually curated CATH classifications.

Scientific Applications:

• Structural Bioinformatics: Classification of protein domain 3D structures and comparison to CATH manual classifications.
• General Data Analysis: Clustering of synthetic datasets (e.g., 2D vectors) and diverse real-world biological data without prior knowledge of cluster number.

Methodology:

Partitioning-based, parameter-free clustering that heuristically evaluates multiple candidate cluster configurations and uses intrinsic data characteristics (minimum expectation value and intra-cluster similarity variance), supporting hierarchical, partitional, agglomerative and divisive partitioning modes.