Rethinking HLA compatibility with protein language models

The Problem

Haematopoietic stem cell transplantation (HSCT) outcome depends critically on HLA compatibility between donor and recipient. The standard approach encodes this as a binary match/mismatch count — treating all mismatches as equally dangerous regardless of the underlying protein sequence difference.

A mismatch between two structurally similar alleles and two completely divergent ones are treated identically. This discards most of the immunologically relevant information.

Our Approach

CAPA encodes every HLA allele as a 1 280-dimensional vector using ESM-2, a large protein language model pre-trained on 250 million protein sequences. Alleles with similar binding-groove conformations cluster together in embedding space; immunologically distant alleles are far apart.

A cross-attention network then models the interaction between donor and recipient allele sets, learning which locus pairs are most predictive. The resulting representation is fed into a DeepHit survival head that jointly models GvHD, relapse, and transplant-related mortality as competing risks.

Data

Primary validation uses the UCI Bone Marrow Transplant Children Dataset — 187 paediatric patients with allogeneic HSCT, donor/recipient HLA typing at antigen and allele resolution, and time-to-event outcomes for GvHD, relapse, and cause of death. HLA protein sequences are sourced from the IPD-IMGT/HLA database.