Scholar Presentation: Identifying HLA Amino Acid Mismatches that Predict Kidney Graft Failure with Cross Validation using FIBERS 2.0 Machine Learning Binning

Aim: Discover HLA amino acid mismatches (AA-MMs) that are predictive of kidney graft failure (GF) risk using the recently developed FIBERS 2.0 machine learning (ML) algorithm (Fig.1) and a rigorous cross-validation (CV) analysis.

Methods: Data on 166932 adult deceased donor kidney transplants from 2010-2023 were obtained from the Scientific Registry of Transplant Recipients. We assigned AA-MMs polymorphisms for a high-resolution imputation of all classical HLA class I and class II loci from serologic antigen (Ag) specificities using haplotype frequencies. FIBERS 2.0 ML with 10-fold CV was applied to identify bins maximizing separation of high/low-risk survivorship and deviance residuals adjusting for standard covariates (e.g. demographics, comorbidities) while relying on a burden of 0 MMs = low-risk and >0 MMs = high-risk. To identify all potential AA-MM risk factors, Ag-MM covariates were excluded. Top bins were evaluated on testing CV partitions with standard covariate adjusted Cox regression.

Results: Top FIBERS 2.0 bins for each CV partition yielded Cox adjusted hazard ratios (HRs) (1.037-1.171) with 5/10 partitions yielding p<0.05 (Fig 2). Across CV partitions, AA-MM positions in HLA class I and class II were identified with variable CV consistency, attributed to underlying data heterogeneity, linkage disequilibrium, and CV sample partitioning variability, i.e. sample bias (Fig 2). AA-MM positions identified in >5 CV partitions included several positions that are strongly correlated with AA-MMs that line P4 pocket and peptide-contact sites in DRB1 (30, 73, 33, and 16) and DQ molecules (25 and 26 in DQA1; 14 in DQB1). In addition, 3 high-risk AA-MM positions were identified in HLA-A (56, 67, and 158). Notably, AA substitutions at positions 67 and 158 separate A1 and A36 alleles from most of other HLA-A lineages, while AA substitutions at position 56 distinguish A30 and A31 allele lineages from others.

Conclusion: Combining FIBERS 2.0 with a CV evaluation analysis framework was effective in identifying AA-MM bins
associated with covariate-adjusted GF risk, yielding both consistent and new AA-MM positions in contrast with previous
analyses without CV. CV offers a rigorous predictive evaluation, but introduces challenges, i.e. multiple evaluated models
and smaller datasets for final Cox regression testing analysis of HR.