Modern protein databases grow far faster than experts can annotate them, which leaves a large share of sequences and structures without a known function. Two unsupervised methods, DPCfam and DPCstruct, help close this gap by applying the Density Peak Clustering algorithm to raw similarity data and grouping protein domains into putative families called metaclusters, with no manual curation. Together they produced 81,384 sequence-based and 28,246 structure-based metaclusters, many of which have no counterpart in Pfam, CATH, or SCOP. Both datasets are public on Zenodo, yet they remain hard to use in practice: exploring them means downloading and parsing very large archives, including an 81 GB XML file for DPCfam.
This thesis presents DPCexplorer, a unified Django web application that makes these datasets easy to explore in a browser. The raw files were cleaned and consolidated on the ORFEO HPC cluster, then loaded into a PostgreSQL database indexed for fast lookup across millions of rows. The platform answers four kinds of queries (DPCfam MCID, DPCstruct MCID, Pfam ID, and UniProt accession), returns paginated tables and per-protein domain-architecture diagrams, and lets users download biological files (FASTA, MSA, HMM, PDB). For structural metaclusters it embeds the PDBe-Molstar (Mol*) viewer to display AlphaFold2 models in 3D, coloured by per-residue confidence. By turning static repositories into an interactive tool, the work supports the FAIR principles and widens access to unsupervised protein domain classifications. The application is available at https://dpcexplorer.areasciencepark.it/ and is fully reproducible from its public GitHub repository and Zenodo deposit.