Background Human leukocyte antigen (HLA) genes facilitate communication between tumor cells and the immune system through the cell surface presentation of a diverse set of peptides. HLA loss of heterozygosity (LOH) has been associated with reduced immune pressure on neoantigens and impaired response to checkpoint blockade immunotherapy. Although HLA LOH is emerging as a key biomarker for response to immunotherapy, few tools exist to detect HLA LOH. Moreover, the accuracy of these tools is not well understood due to lack of orthogonal validation approaches. Here, we briefly describe DASH (Deletion of Allele-Specific HLAs), an algorithm to detect HLA LOH from exome sequencing data, and present a three-pronged validation approach to assess its performance.

Methods In-silico evaluation of the limit of detection (LOD) of DASH was performed by deeply sequencing a tumor-normal paired cell line with HLA LOH and mixing reads at different proportions to simulate variable tumor purity and clonality. Direct genomic validation was performed using digital PCR (dPCR) with allele-specific primers targeting both predicted kept and lost alleles in ten patient samples and one cell line. Quantitative immunopeptidomics was performed to compare peptides presented by HLA alleles in tumor cells and adjacent normal cells. The relative increase or decrease of peptide presentation per allele was estimated by predicting the binding of each peptide to the patient-specific alleles.

Results DASH is a machine learning model built upon the HLA-enhanced ImmunoID NeXT Platform®. We validated the performance of DASH using three orthogonal approaches to better understand the factors driving sensitivity and specificity of the algorithm. Evaluation using cell line mixtures that simulate LOH at various dilutions helped establish the LOD of DASH. For fully clonal tumors, DASH had 100% sensitivity at all tumor purity levels above 8% and 100% specificity at tumor purity levels higher than 24%. Patient-specific and allele-specific dPCR assays provided sensitive, direct evidence of HLA LOH. All samples predicted to have HLA LOH by DASH with high confidence were confirmed by dPCR. Finally, a quantitative immunopeptidomics experiment in one patient with HLA LOH revealed a large decrease in the peptides presented by deleted alleles, revealing the functional implications of HLA LOH.

Conclusions HLA LOH detection methods need to be rigorously validated in order to be used as a clinical biomarker. Here, we introduced three methods to assess performance, demonstrated the strong predictive power of DASH, and highlighted the need to consider tumor purity in such assessments.