Frozen functional HLA classification from pocket physicochemistry: 159 alleles in 88 groups across five loci

This deposit contains a functional classification of HLA molecules built from the physicochemistry of the residues lining the pockets of the peptide binding groove, together with the tooling required to verify its integrity. Scope. 159 alleles were retained on an explicit and symmetric criterion, a frequency of at least 1 % in at least one of ten reference cohorts, six African and four European, so that an allele frequent only in Africa enters the panel. Five loci are covered: HLA-A, -B, -C, -DRB1 and -DQB1. Protein alignments are from IPD-IMGT/HLA release 3.65.0, read in mature protein numbering. Each molecule is described by five physicochemical properties per pocket-lining residue, z-scored across the panel. Clustering uses euclidean distance and Ward linkage under a single global cophenetic threshold selected to maximise mean silhouette across loci, and yields 88 functional groups. Sealing protocol. The classification was built from sequence data alone, with no clinical data or disease association loaded, and the resulting groups were serialised, timestamped and digested with SHA-256 before any validation was performed. The purpose is to make the subsequent validation verifiable rather than asserted: the groups cannot have been adjusted to the expected answer. Frozen at 2026-07-29T16:22:13Z. SHA-256 of the sealed file: 2bebec56b385bb1940e5aed47bcb458287dacf944858550d573b527bc1a3b7f9 The included verification script recomputes the digest and compares it with the stored value, returning a non-zero exit code on divergence. Contents. phaseA_frozen_groups.json, the sealed classification with its method block, group assignments and consensus motifs. verify_freeze.py, the integrity verification script. notebook_hla_reclassification.py, the analysis notebook, which runs offline and reproduces the reported results. A human-readable table of the 88 groups with, for each allele, its maximum frequency in the African and European cohorts and the pocket motif of its group. What this deposit does and does not establish. It establishes that the classification is computable, reproducible, and that its integrity can be checked independently. It does not establish clinical validity. No patient data was used and no clinical claim is made. Limitations. The frequency threshold mechanically excludes alleles specific to the least studied populations, so the panel reflects the state of the reference databases as much as the biology. A single clustering threshold applied to five loci is not optimal for any locus taken alone. Class II molecules are described here by their beta chain only, which is immaterial for DR but not for DQ, where DQA1 is polymorphic. Related work. This resumes an approach first proposed in Busson M, Djoulah S, Karsenty E, Bleux H, Bouteiller AM, Charron D, Transplantation Proceedings 1998;30:2855-2856, and developed in Djoulah S, Busson M, Sasazuki T, Maillère B, Yasunaga S, Kimura A, Charron D, Hors J, Tissue Antigens 1999;54:341-348. Acknowledgement of tooling. The analysis pipeline, sealing protocol and verification tooling were developed with the assistance of Claude, an AI assistant developed by Anthropic. Study design, inclusion criteria, the sealing protocol and the interpretation of results remain the responsibility of the authors, who assume full responsibility for the content.