Computational profiling of predicted neoantigen landscape reveals differential immunogenicity at AML diagnosis and relapse

Abstract Introduction Although AML carries a relatively low mutational burden and was once considered poorly immunogenic, immune escape has now been recognized as a major driver of relapse. Mechanisms implicated include HLA down-regulation, up-regulated immune checkpoint signaling, and an expansion of immunosuppressive cells within the tumor microenvironment (Vago & Gojo, 2020). However, little is known about how the neoantigen repertoire in AML evolves over time. In this study, we computationally predicted neoantigen landscapes at diagnosis and relapse to further characterize how AML adapts to host immunity. Methods Our study included 491 pediatric patients with AML at diagnosis and 91 patients at relapse including 18 paired samples. We applied the published Landscape of Effective Neoantigens Software (LENS) developed in our lab to predict neoantigens in each patient using paired tumor-germline whole-exome sequencing and tumor RNAseq. LENS infers HLA type, calls expressed variants, translates coding alterations (SNVs, indels, fusions, splice events, CTAs, ERVs) into 8–11-mer peptides. We annotated peptides with predicted HLA binding affinity using NetMHCpan-4.1, and those with a binding affinity in the top 2 percentile were retained to filter for peptides most likely to be immunogenic. Statistical analyses were performed in R 4.3.1. Across all subjects, we compared 1) neoantigen count and 2) neoantigen expression between diagnosis and relapse, both total and stratified by antigen class, using two-sided Wilcoxon rank sum tests. For each peptide-HLA pair present at both diagnosis and relapse, we compared neoantigen expression across both time points using a two-sided Wilcoxon rank-sum test, correcting for multiple testing within each HLA allele using FDR adjustment. To perform differential gene expression analysis, RNAseq reads were aligned to GRCh38 with STAR, and gene-level counts were generated with htseq-count. The resulting count matrix was analyzed with DESeq2. Results Normalized expression of neoantigens was significantly lower at relapse compared to diagnosis (diagnosis: mean standardized count 231, SE 0.89; relapse: mean 182, SE 0.5; p<2.2e-16). When stratified by antigen type, expression of neoantigens derived from CTAs, ERVs, indels, SNVs, and splice variants were all significantly lower at relapse compared to diagnosis (p<0.05). The number of predicted neoantigens per patient was not significantly different at diagnosis and relapse (diagnosis: mean 2274, SE 79; relapse: mean 2035, SE 95; p=0.22). When comparing counts stratified by antigen type, patients at relapse had significantly more neoantigens derived from SNVs compared to patients at diagnosis (p<0.01). There were no significant differences in neoantigen count for other antigen sources. Among patients with HLA-C*08:02, the cancer testis antigen SPAG9-derived peptide FTDPLGVQI had significantly lower expression at relapse compared to diagnosis (p<0.01). We performed differential gene expression analysis comparing diagnosis and relapse to query whether any immune-related genes were differentially expressed across time points. From an unbiased analysis, we found statistically significant downregulation at relapse of genes involved in antigen processing and presentation including IFI30, CALR, CTSB, HLA-DQB1, HLA-DQA1, PDIA3, and a cluster of heat-shock proteins (padj<0.05). Conclusions Our analyses implicate downregulation of neoantigen presentation as a key driver of AML relapse. At relapse, tumors showed silencing of neoantigen expression and down-regulation of antigen processing/presentation genes. One CTA-derived peptide from gene SPAG9 was significantly downregulated at relapse and represents a potentially targetable immunogenic neoantigen.