hla2vec: Developing a HLA embedding space using probabilistic cross-reactive groups

The genes in the major histocompatibility complex (MHC) that code for the Human Leukocyte Antigen (HLA) protein complex are some of the most diverse genes in the genome. And while some of the alleles such as HLA-A*02:01 are seen in almost 25% of the population, databases such as IMGT/HLA report nearly 20 thousand variants. In cancer populations, this produces a long tail of individuals who are unique within a sampled population. HLA class I typing for The Cancer Genome Atlas (TCGA) covered almost 10K individuals across 33 cancer types, and resulted in over 300 HLA alleles. Within this cohort alleles such as HLA-A*03, HLA-A*33, and HLA-A*2 have been associated with prognosis and immunotherapy response in various cancers. Not all HLA alleles are different, with some alleles observing similar binding patterns that can place them in cross-reactive groups (CREGs). And while this concept has traditionally been documented as a discrete set of documented alleles, from large scale experiments we can start to think of this as a probabilistic space representing probabilistic CREGs (pCREGs). Two alleles in a pCREG may only bind to some fraction, for example 90%, of the same peptides. Using this as a basis we can develop an embedding space where the distance between different HLA alleles is relative to their pCREG association. We have developed a method called HLA2Vec, which has been trained on the 15 million HLA/peptide binding experiments accumulated in the BigMHC dataset. In addition, we developed a benchmark to describe a new model's ability to identify CREGs. We have compared this method to other Protein Modelling Languages (PML), such as Evolutionary Scale Modeling (ESM) in their ability to identify pairs of HLAs that are likely to be in the same pCREG, and found the HLA2Vec to be much more accurate, while using many fewer parameters. With HLA alleles placed in this embedding space, we are able to cluster together smaller groups of individuals with near unique HLA alleles into larger subtypes, based on pCREG status, bringing together patients likely to have similar immune HLA binding patterns for statistical analysis. This analysis has been applied to TCGA data to group rare alleles into larger response groups. The next steps of this research are to see if this method for HLA embedding may also provide insights for developing similar embedding spaces for T Cell Receptor (TCR), and improve HLA pCREG conditional binding prediction.