kludo

kludo applies diffusion kernels on protein graphs to segment protein structures into structural domains for protein domain assignment and structural analysis.

Key Features:

• Diffusion kernel affinity: Uses diffusion kernels applied to protein graphs as affinity measures between residues.
• Protein graph representation: Represents protein structures as graphs of residues/amino acids to capture structural relationships.
• Reproducing kernel Hilbert space rationale: Leverages the concept that a suitable kernel function in a reproducing kernel Hilbert space can separate structural domains.
• Graph node kernels evaluated: Evaluates combinations of four graph node kernels to assess suitability for domain decomposition.
• Clustering algorithms compared: Tests two clustering algorithms in combination with kernels to segment protein graphs into domains.
• Alternative partitionings: Produces alternative partitionings for the same protein structure to reflect ambiguous domain boundaries.
• Benchmark evaluation: Validated on five benchmark datasets for protein domain assignment and on a comprehensive non-redundant dataset.
• Simple selection criterion: Employs a relatively straightforward criterion to choose optimal decompositions.
• Competitive performance: Demonstrates competitive accuracy with leading automatic methods, particularly for specific diffusion kernels.
• Kernel extensibility: Adaptable to incorporate higher-performing kernels for future improvements.

Scientific Applications:

• Protein domain assignment: Automated assignment of structural domains in solved protein structures.
• Structural analysis: Parsing proteins into domains to support downstream structural analyses.
• Ambiguity handling: Generating alternative domain partitionings to represent subjective or ambiguous domain definitions.
• Method benchmarking: Benchmarking and comparison of domain decomposition approaches across datasets.

Methodology:

Applies diffusion kernels to residue-level protein graphs to compute affinities, evaluates combinations of four graph node kernels and two clustering algorithms, and selects domain decompositions using a kernel-separation criterion motivated by reproducing kernel Hilbert space theory.