Protein Language Model-Based Intra-Gene Disparity of Donor HLA Alleles Informs Prognostic Risk in Hematologic Malignancy Patients Undergoing Haplo-Identical HSCT: A Combined Xinqiao-CIBMTR Analysis

It has been suggested that the intra-gene disparity in human leukocyte antigen (HLA) within a donor or recipient may play an important immunological role in haplo-identical hematopoietic stem cell transplantation (haplo-HSCT). Current methods to estimate HLA disparity primarily rely on epitope analysis and amino acid index scores. Recently, AlphaFold and emerging protein language models (PLMs) have highlighted the potential of large-scale deep learning techniques for predicting protein structures and functions. In this study, we utilized the newly proposed Protein-Vec PLM (PMID: 38045331) to extract functional embedding vectors from the amino acid sequences encoded by HLA alleles. With these vectors, we defined a novel PLM-based measure of intra-gene disparity for any allele pair of the same HLA gene within an individual and explored its prognostic value in haplo-HSCT.

We collected data from 425 hematologic malignancy (HM) patients undergoing haplo-HSCT between 2015 and 2022 at Xinqiao Hospital in Chongqing, China. We analyzed the association of PLM-based intra-gene HLA disparity (HLA-A/B/C/DRB1/DQB1) in a donor or recipient with overall survival (OS), disease-free survival (DFS), non-relapse mortality (NRM), and relapse. We validated our results by using a publicly available dataset from CIBMTR IB20-02 (PMID: 34774819), which included 985 HM patients who underwent haplo-HSCT between 2013 and 2017.

Using Cox regression to calculate proportional hazard ratios (HR), with p<0.01 as the threshold for statistical significance, our analysis of the Xinqiao dataset showed that higher disparity of donor HLA-B alleles was significantly associated with worse OS (HR=2.0, 95%CI 1.34-3.0, p=0.0007), worse DFS (HR=1.78, 95%CI 1.27-2.49, p=0.0008), higher NRM (HR=1.77, 95%CI 1.1-2.86, p=0.019), and higher relapse risk (HR=1.78, 95%CI 1.11-2.86, p=0.017). The result was further verified by the CIBMTR dataset, where higher disparity of donor HLA-B alleles was associated with worse OS (HR=1.3, 95%CI 1.09-1.57, p=0.0046), worse DFS (HR=1.20, 95%CI 1.02-1.42, p=0.03), higher NRM (HR=1.33, 95%CI 1.0-1.78, p=0.053), higher relapse risk (HR=1.14, 95%CI 0.93-1.40, p=0.2). Multivariate analysis demonstrated that PLM-based disparity of donor HLA-B alleles was an independent prognostic factor for OS in both Xinqiao and CIBMTR dataset. In contrast, we did not find significant results for intra-gene disparity of HLA-A/C/DRB1/DQB1 alleles in either donors or recipients.

In summary, PLM-based disparity of donor HLA-B alleles might be a potential prognostic factor in haplo-HSCT, and might be used as a novel tool assisting donor selection.