Background A hallmark of cancer is the ability for the tumor cells to evade the immune system. One such immune-evasion mechanism is the loss of heterozygosity (LOH) of the classical human leukocyte antigen class I (HLA I) genes in the tumor cells. HLA-I LOH has been shown in many instances to be associated with worse cancer immunotherapy outcomes and attenuated cytotoxicity. Consequently, many studies have specifically focused on developing tools to detect HLA-I losses in order to better infer HLA-I LOH.

Methods We develop an in silico workflow, ASHLI/LEAHLA, for Allele-Specific HLA Loss Inference (ASHLI) that can perform statistical Loss Estimation of Alleles in HLA (LEAHLA), in HLA class I genes. The workflow uses matched normal and tumor next-generation sequencing read data to quantify allelic tumor and normal reads that map to the respective HLA alleles. To aid in normalizing these read counts, the workflow also maps reads to the non-HLA genomic regions.

Results We have applied ASHLI/LEAHLA to detect HLA-I losses in a set of 24 samples from a bladder clinical trial, which has been orthogonally assayed to detect HLA-I LOH using a combination of SNP array genotyping and ASCAT, to infer copy number in the HLA-I genes. We found that ASHLI/LEAHLA is also highly tunable. At a more relaxed threshold, it gives a performant sensitivity = 1 and specificity = 0.93. We can also apply a more stringent threshold to obtain more confident LOH calls, trading sensitivity for specificity (sensitivity = 0.86 and specificity = 0.96). Moreover, for all the genes that both SNP array and ASHLI/LEAHLA are estimated to have LOH, ASHLI/LEAHLA was able to identify the alleles correctly 100% of the time. We compared our pipeline on the same dataset to SpecHLA, which is another tool that is recently published; the latter’s sensitivity = 0.82, specificity = 0.79. Finally, we also examined the effect of tumor purity on ASHLI/LEAHLA performance using synthetic cell line mixtures of tumor cell lines and their corresponding matched normal. We show that it could estimate HLA LOH even when the sample is at 20–50% tumor purity.

Conclusions We have developed a computational workflow that has been shown to be performant in detecting HLA-I LOH in both cancer cell lines and primary samples, and when compared to a recently published tool.