Immune checkpoint blockade (ICB) therapies have shown success in cancer treatment by enabling the immune system to recognize and eliminate tumor cells presenting neoantigens. However, a significant portion of patients do not respond to ICB, highlighting the need for methods to predict response and understand resistance mechanisms. Antigen presentation plays a crucial role in tumor evolution, and human leukocyte antigen (HLA) genes, responsible for presenting antigens to the immune system, are frequently mutated in cancers. Somatic loss of heterozygosity (LOH) in the HLA region can significantly reduce HLA sequence diversity, leading to immune escape and ICB resistance. Existing methods for detecting HLA LOH from sequencing data have limitations in accuracy and sensitivity, particularly in samples with low tumor purity or subclonal deletions. This study aims to develop a more accurate and sensitive algorithm for detecting HLA LOH to improve prediction of immunotherapy response and enhance our understanding of immune evasion in cancer.

Many copy number detection algorithms exist, but they are unreliable for HLA LOH detection due to the polymorphic nature of HLA genes and challenges in accurate genome alignment. Existing HLA LOH detection algorithms often rely on tumor-only sequencing or standard copy number variant (CNV) algorithms, which are not designed for the specific challenges presented by the HLA region. These methods also struggle with low tumor purity and subclonal deletions, leading to inaccuracies in LOH detection. Validation methods also present challenges, with existing approaches failing to validate the specific allele lost or address accuracy in challenging samples. Therefore, the field lacks a robust, sensitive, and specific method for HLA LOH detection.

This study introduces DASH (Deletion of Allele-Specific HLAs), a machine learning algorithm designed to detect HLA LOH with improved accuracy and sensitivity. The algorithm leverages paired tumor-normal sequencing data and incorporates multiple features to address the specific challenges of the HLA region. These features include an adjusted b-allele frequency to account for probe capture variability, allele-specific sequencing depth ratio, consistency of sequencing depth, total sequencing depth ratio to differentiate LOH from amplification, tumor purity, tumor ploidy, and deletion in flanking regions. DASH was trained using data from 279 patients with manually curated HLA LOH labels. The algorithm's performance was evaluated using 10-fold cross-validation. To assess the limit of detection, in silico cell line mixtures were created by combining tumor and normal cell line sequencing data at varying purities and subclonalities. Allele-specific digital PCR (dPCR) was developed as an orthogonal validation method, with patient-specific primers designed to target the predicted lost and retained alleles. Quantitative immunopeptidomics was used to explore the functional impact of HLA LOH on peptide presentation. Finally, DASH was applied to a pan-cancer cohort of 610 patients across 15 tumor types to assess the prevalence of HLA LOH and its correlation with other clinical and genomic features.

DASH outperformed existing algorithms (LOHHLA) in cross-validation, achieving higher sensitivity and specificity, especially in samples with higher tumor purity. In silico cell line mixture analysis showed DASH's high sensitivity and specificity across a range of tumor purities and subclonalities, demonstrating superior ability to detect subclonal HLA LOH events. Allele-specific dPCR validation confirmed DASH's predictions in patient tumor samples. While quantitative immunopeptidomics showed a trend toward increased peptide intensity for peptides predicted to bind to lost alleles, the results were inconsistent across samples, highlighting the influence of other factors on peptide presentation. Pan-cancer analysis of 610 patients revealed HLA LOH in 18% of patients, varying across tumor types (40% in HNSCC to 4% in liver cancer). Patients with HLA LOH frequently lost all three HLA genes. HLA LOH was significantly associated with higher genome-wide LOH rates, increased neoantigen burden, CD274 (PD-L1) expression, and microsatellite instability (MSI) status, suggesting its role as an immune resistance mechanism. More neoantigens were predicted to bind to lost HLA alleles compared to retained alleles.

DASH's superior performance in detecting HLA LOH compared to existing methods highlights the importance of using specialized features tailored to the unique challenges of the HLA region. The development of allele-specific dPCR provides a crucial orthogonal validation method, furthering confidence in DASH's accuracy and demonstrating its potential for clinical translation. The observed association between HLA LOH and immune resistance mechanisms, such as increased neoantigen burden, PD-L1 expression, and MSI, supports its role in immune evasion. The widespread prevalence of HLA LOH across various tumor types underscores its importance as a potential biomarker for immunotherapy response prediction.

DASH is a validated machine learning algorithm for accurate and sensitive detection of HLA LOH, particularly in samples with low tumor purity and subclonal events. The high prevalence of HLA LOH across various tumor types and its association with immune resistance mechanisms highlight its significance in cancer immunotherapy. Future research could focus on integrating HLA LOH into predictive models for immunotherapy response and investigating its mechanistic effects on peptide presentation in more detail.

The manual curation of HLA LOH labels in the training dataset, while providing high-quality labels, is labor-intensive and may not be fully scalable. While immunopeptidomics showed a trend, inconsistent results across samples emphasize the need for further investigation to account for other biological factors influencing peptide presentation. The current study focuses on HLA LOH and doesn't consider other potential allelic imbalance mechanisms, such as single-allele amplification.