Comparison Between QPCR And RNA‑seq Reveals Challenges Of Quantifying HLA Expression Part 2

Normalization

We used expression estimates in transcripts per million (TPM), which is the standard normalization produced by Salmon and corresponds to the relative amount of a given transcript in a sample. For any given gene, the estimate is simply the sum of TPMs for its transcripts. In some cases when we show standard normal transformed estimates, we performed a rank normal transformation of the RNAseq data using the GenABEL R package (Aulchenko et al. 2007), which is usually applied, for example, in linear models of eQTL mapping (Delaneau et al. 2017).

QTL mapping is a method to study the regulatory mechanism of gene expression by comparing the association between gene expression and gene polymorphism in a population. The linear model is a commonly used eQTL mapping method, which can use a linear regression model to estimate the correlation between gene expression and gene polymorphism.

The immune system is a complex biological system that protects the body from infections and diseases such as cancer. Regulation of gene expression plays an important role in the immune system and can affect the development, differentiation, and function of immune cells.

Therefore, linear models of eQTL mapping can be used to analyze the relationship between gene expression regulatory mechanisms and immunity. For example, studying the correlation between a certain genotype and the expression of a key immune gene in a population can reveal the influence of this genotype in regulating immune gene expression. Such research can provide important enlightenment for the treatment and prevention of immune diseases. This shows the importance of immunity, so we need to improve our immunity every day. Minced meat also has anti-virus, anti-cancer, and other effects, which can strengthen immunity The system's ability to resist and improve the body's immunity.

Click health benefits of cistanche

Read alignment to the reference genome

For the analysis of read coverage at HLA genes reported in Fig. S7, we aligned reads to the reference genome GRCh38 with STAR v2.7.3a (Dobin et al. 2013), using Gencode v37 gene annotations. To control for mapping bias at HLA genes, we further processed the BAM files with hlamapper v4.3 (Castelli et al. 2018).

Simulation

Ground‑truth data

To generate simulated data, we first ran Salmon v1.3.0 (Patro et al. 2017) on the real sample #66K00003 to learn the expression levels of Gencode v37 transcripts. Then we used the Polyester package (v1.26.0) to generate 50 synthetic samples with identical transcriptome-wide expression levels, except for HLA-A, -B, and -C. The expression levels for these genes were based on 50 randomly chosen individuals from our real data (for which we have HLA allele data available). For each HLA gene, we selected the isoforms that accounted for at least 90% of the total protein-coding transcript expression in a Salmon run on the real dataset (which resulted in only 1 transcript per gene) and personalized the transcript sequences according to the HLA alleles carried by each individual.

This procedure allowed us to synthetically generate 50 individuals with identical background expression levels, but with variable HLA expression and with HLA polymorphism built into the simulated reads.

To mirror our real data, thirty million 126 bp paired-end reads with a mean fragment size of 261 bp were simulated for each individual, using the defaults for other polyester parameters (e.g., the standard deviation of the fragment length=25 bp, error rate=0.005, uniform distribution of reads, and no bias). Polyester outputs FASTA files, from which we produced FASTQ files with a constant quality score (corresponding symbol “F”).

Metrics for accuracy

TPMs were computed on simulated counts given the transcript lengths and average fragment size of 261 bp. The “Estimated TPM/True TPM” ratio is used to assess performance in recovering simulated expression levels and allows us to observe down or overestimation.

Graphics

We prepared all the plots in this article using the ggplot2 package v3.3.2 (Wickham 2016) in R.

Results

Accuracy of RNA‑seq HLA quantification

Given the absence of a method that can be considered the experimental gold standard for HLA expression quantification from RNA-seq data, we initially assessed the accuracy of RNA-seq quantification methods for HLA using simulated data where true expression levels are known since they are generated in a computer to emulate real experiments. This was done to choose the best computational approach among RNA-seq-based methods, allowing a subsequent contrast with non-RNA-seq approaches.

We simulated an RNA-seq experiment for 50 individuals using the Polyester package (Frazee et al. 2015). These synthetic individuals have the same expression levels for all genes in the genome, except for HLA-A, -B, and -C, for which we varied the expression levels. We also personalized the annotated HLA transcript sequences from Gencode v37 to introduce real genetic variation observed in randomly chosen individuals from a dataset of 96 individuals (which were HLA genotyped by Sanger sequencing as described below). The resulting personalized transcripts had a median sequence identity with a reference greater than 95% for all HLA loci.

We compared estimates of HLA expression obtained by two bioinformatic methods: (1) “Ref transcriptome,” which uses Salmon (Patro et al. 2017) to align reads to the standard reference transcriptome, quantifying transcript abundance and (2) “Personalized,” which also uses Salmon, but maps reads to personalized HLA transcripts, reflecting the individual’s HLA genotype (Fig. 1). The “Personalized” approach extends our previous strategy (Aguiar et al. 2019) by using a personalized transcript, rather than a single canonical coding sequence for each allele carried by the individual.

The “Ref transcriptome” method underestimated expression levels, in particular for alleles with a greater proportion of sequence differences concerning the reference genome (Fig. 1). This is expected since a higher mismatch rate between reads and the reference negatively impacts alignment (Brandt et al. 2015). This approach also overestimated HLA-C expression for some individuals, a consequence of reads from HLA-B being mapped to the HLA-C reference transcripts (Fig. S1). The “Personalized” approach, on the other hand, controls the mapping bias and achieves optimal accuracy.

Although our simulation provides encouraging results regarding the quantification of HLA expression using RNAseq, we must consider some caveats. We modified the sequence of the annotated isoforms according to the individuals’ HLA alleles using a single set of isoforms for all alleles at a given HLA gene. Those sequences were used both in the simulation of reads as well as in the quantification of expression; thus, we expect optimal accuracy. In a real scenario, different HLA alleles might be associated with different isoforms. Later in this paper, we discuss a specific example that we observed for HLA-A, consistent with the hypothesis that certain isoforms are exclusive to specific alleles. Nonetheless, given that we are mainly interested in the gene-level and HLA allele-level expression estimates, we expect that personalized sequences represent an improvement over a single reference transcriptome by reducing mapping bias.

Estimating HLA expression from real RNA‑seq data

We performed expression estimation on whole-transcriptome RNA-seq data for 96 individuals, for which qPCR for HLAA, -B, and -C, and HLA-C surface expression levels were previously estimated (Kulkarni et al. 2013; Ramsuran et al. 2015, 2017), and could be used to compare with the RNAseq results (see Fig. S2 for QC analyses on RNA-seq data).

Given the higher accuracy of the personalized approach in the simulation, we contrast this method of RNA-seqbased expression estimates to that of other non-RNA-seq approaches, but provide the results for the reference transcriptome-based approaches in the “Supplementary information.” We personalized the transcript sequences given the individual HLA genotypes obtained by Sanger sequencing. We ran HLApers (Aguiar et al. 2019) and Kourami (Lee and Kingsford 2018) to infer alleles directly from the RNA-seq data and confirm the Sanger calls (see “Materials and methods”).

The gene-level expression estimates show that HLA-B has the highest expression among HLA loci in our dataset, followed by HLA-C and HLA-A (Fig. 2A). This ordering is consistent with the GTEx whole blood dataset (GTEx Consortium 2020) and with a previous HLA-capture RNAseq method applied to PBMCs (Yamamoto et al. 2020). However, this pattern differs from that seen by Boegel et al (2018), who observed similar levels across genes using a different strategy for dealing with reads mapping to multiple loci, which may contribute to the lack of distinction across loci in terms of expression level). Future studies will have to tease apart the contribution of differences in methodologies or cell-type composition to these differences.

Comparing RNA‑seq and qPCR on real data

We next compared RNA-seq expression estimates to those obtained with qPCR (Fig. 2B). Although the correlation between RNA-seq and qPCR expression was statistically significant for all genes (p=0.024, 0.002, 0.000000016, for HLA-A, -B, and -C, respectively; Spearman’s test for positive association), the magnitude of the correlations were modest for HLA-A and -B, and higher for -C. The use of a personalized reference for RNA-seq modestly increased the correlation with qPCR compared to a standard reference (Fig. S3). This agrees with our previous observation that gene-level expression estimates are not substantially different between reference-genome-based or personalized approaches for HLA class I genes (Aguiar et al. 2019), with the main benefit of personalized approaches being the estimates at the HLA allele level, which we explore below. The use of bias correction in Salmon (GC bias, sequence-specific bias, and position-specific bias) improves the correlation with qPCR, with the highest impact for HLA-B (compare Figs. 2B and S4, for corrected and uncorrected data, respectively).

Comparing mRNA levels with surface expression

Because RNA expression is informative about the initial steps of cell signaling and response to stimuli, analyzing its relationship to downstream molecular phenotypes (such as protein expression on the cell surface) can help us understand the role of post-transcriptional and post-translational regulation on HLA expression. Differences between RNA and protein abundances are expected since they are subject to distinct modes of regulation. Technical effects can also introduce differences since RNA and protein techniques differ and are affected by uncorrelated types of error (Li and Biggin 2015; Kaur et al. 2017; Carey et al. 2019). Furthermore, in the case of our study, gene expression was measured on total PBMCs, whereas protein expression was measured on sorted CD3+ cells. With this difference in mind, we measured the degree to which HLA protein on the cell surface can be predicted by mRNA expression. This analysis was performed exclusively for HLA-C since it is the only locus for which an antibody that can bind all alleles with equal affinities is available. Interestingly, there was a high correlation between mRNA and protein expression for HLAC, with a slightly higher correlation for RNA-seq (Fig. 2C).

HLA allele‑level expression

HLA genes harbor regulatory elements associated with constitutive transcription and dynamically activated transcription (René et al. 2016). As a result, HLA expression varies across tissues and can be modulated by regulatory networks triggered by different stimuli (Anderson 2018; Carey et al. 2019). There is increasing interest in understanding whether distinct HLA alleles are associated with different basal expression levels and regulatory programs (Aguiar et al. 2019; Gutierrez-Arcelus et al. 2020), and whether this variation contributes to disease phenotypes or transplantation outcomes (Petersdorf et al. 2014, 2015; René et al. 2016; Bettens et al. 2022; Johansson et al. 2022). Therefore, HLA allele-level expression estimates for qPCR and RNA-seq were compared. Because individual alleles are often quite rare in the dataset, we grouped them by allelic lineages (i.e., groups of alleles that are phylogenetically defined by the relationship of exons) (Elsner et al. 2002).

We ranked lineages according to their expression levels based on both RNA-seq and qPCR data and assessed the concordance of rankings across methods (Fig. 3). Our personalized RNA-seq approach directly provides allele-level estimates, since HLA allele sequences are used to index the alignments, so we ordered allelic lineages according to their median expression levels. Because our qPCR expression estimates are at the gene level and do not directly provide allele-level estimates, we ordered allelic lineages according to their effects in a linear model of expression levels explained by the HLA genotype (see Ramsuran et al. 2015). In Fig. 3, the expression values are plotted twice for each level for each allele of the individual, and for qPCR, it is simply the gene-level expression plotted twice, reflecting the presence of two alleles.

The ordering of expression estimates is more similar between RNA-seq and qPCR for HLA-C than it is for -A and -B (average absolute order difference, where order difference refers to the observed difference in positions within a ranked order of expression values, between RNA-seq and qPCR quantification, of 2.3 for HLA-C, 3.1 for -A, and 3.9 for -B), following a similar pattern of an agreement to that of gene-level expression, for which we found the highest correlation between RNA-seq and qPCR for HLA-C.

Among the lineages with the largest difference between RNA-seq and qPCR is A*11. We measured surface expression on a subset of heterozygotes for A*03 or A*11 using an antibody that has equal affinity for both lineages and observed that qPCR correlates more robustly with cell surface expression of these two allotypes than does RNA-seq (Fig. S5). Next, we present a more extensive evaluation of allele orderings through comparisons with previous studies of HLA mRNA expression.

While there is interest in comparing expression differences among HLA alleles, various studies show that variation in expression within an allele or allelic lineage is often quite high, and differences between alleles of different ranks are often small and non-significant. As a consequence, it may be unrealistic to expect maintenance of ranks across multiple alleles, and it may be preferable to compare expression estimates for alleles at the extremes of expression.

For our RNA-seq data, we compare our estimates with those from two previous HLA-tailored RNA-seq approaches on PBMCs. There is an overall good concordance with Yamamoto et al. (2020), where A*24, A*02, C*04, and C*06 are highly expressed, and A*03, C*03, and B*15 are expressed at low levels, although we also see differences such as for B*35, which would agree more with our qPCR data. When we compare our RNA-seq data with Johansson et al. (2021), however, we see many more differences, although they have very small samples for many lineages.

We also contrast our results with those from two previous qPCR studies that applied allele-specific primers. Bettens et al. (2014) used allele-specific primers for some HLAC lineages and saw C*04 and C*06 as highly expressed, whereas C*07 and C*03 were expressed at low levels, in concordance with what we have for both RNA-seq and qPCR. René et al. (2015) applied allele-specific primers for HLA-A, and observed A*02 (high) and A*29 (low) at the extremes of expression, which agrees more with our RNAseq results than our qPCR; however, we see many differences at other allelic lineages

In some cases, we can also evaluate agreement with functional studies. For example, previous analyses of transcription factor binding sites (TFBS) and promoter activity (reviewed in Anderson 2018), and studies on miRNA regulation (Kulkarni et al. 2011), show that C*03 and C*07 are weakly expressed alleles, which agrees with our observations for both RNA-seq and qPCR.

Potential sources for differences

We next investigated whether processing of the samples used for RNA-seq could have contributed to the differences between expression estimates obtained with qPCR and RNA-seq.

One specific concern was the length of time the samples were stored in a –80 °C freezer (approximately 4 years between the qPCR and RNA-seq assays), as well as other steps specific to the RNA-seq experiment, including thawing of samples. To address this, we performed a second RNAseq experiment on fresh blood redrawn from 11 individuals, which are a subset of the 96 analyzed in this study, and we compared the expression estimates between the two-time points. Even though this second assay carries both technical and biological differences concerning the first RNA-seq experiment (Fig. S6A and B), the transcriptome-wide correlation in expression estimates between time points is high (Fig. S6C). 

Although assessing correlation with 11 individuals can be noisy, correlations at HLA genes are among the largest gene-wise correlations between the two samples (Fig. S6D and F). We also computed within-individual allele ratios, which is the ratio of expression between the two HLA alleles of a heterozygous individual, and compared them between time points. The correlation was greater than 0.94 for HLA-A, -B, and -C (Fig. S6E). Therefore, we saw no evidence of a major contribution of RNA degradation to explain the low correlation between RNA-seq and qPCR in our original sample.

Another possible contribution to differences between RNA-seq and qPCR is that specific HLA allele may be more biased in one method or the other, in which case individuals carrying such alleles would contribute to large differences. For example, for individuals carrying A*03, or for homozygotes for C*07, there is a negative relationship between qPCR and RNA-seq (Fig. 4A).

An additional source of differences between the methods might arise from the fact that, in our RNA-seq approach, we personalize all Gencode annotated transcripts for every HLA allele; however, true transcript diversity and its association with specific HLA alleles are not well understood. For example, Kulkarni et al. (2017) showed that A*01 and A*11 produce shorter 3′-UTRs. To investigate if we can replicate that funding in our RNA-seq data, we mapped the reads to the reference genome and corrected for mapping bias at HLA genes with an hla-mapper (Castelli et al. 2018). Indeed, for individuals carrying A*01 or A*11, the read coverage at the 3′-UTR of HLA-A shows a sharp drop at ~120 bp before the annotated gene end (Fig. S7).

Because transcript per million (TPM) values are computed taking into account the reference length, using a reference that is longer than the true transcript leads to an underestimation of expression. We attempted to control for the possibility of such shorter transcripts by including a version of each HLA-A transcript with a shortened 3′-UTR in our index for read alignment. However, we found no evidence of expression of the shorter isoform (Fig. S8), possibly because these shorter isoforms are contained within the normal length isoforms, and Salmon’s implementation assigns all reads to the larger isoform. Interestingly, expression at the isoform level reveals an isoform with a longer 5′-UTR exclusive to A*11, which contributes a large proportion of the total expression for this allele (Fig. S8).

We also tested a normalization of our expression estimates, in which we adjusted the read lengths given the read coverage supporting a proximal or distal 3′-UTR terminus (weighted average of transcript lengths using read coverage as weights). Although we observe a gain of up to 20% in the expression levels for individuals carrying A*01 and/or A*11, we see only a small improvement in the correlation with qPCR after this adjustment (from rho=0.20 in Fig. 2 to rho=0.24 in Fig. 4B).

A*01 and A*11 are among the alleles with the largest rank differences between RNA-seq and qPCR (Fig. 3), and an imperfect representation of their associated transcripts in the annotation may introduce bias in our RNA-seq estimates.

Finally, normalization methods used to obtain final expression estimates from the raw qPCR data can also be a source of differences between qPCR and RNA-see estimates. Quantitative PCR assays for HLA Class I genes usually amplify regions within exons 1 to 4, and standardization by the expression of a housekeeping gene such as B2M (β2-microglobulin) is typically carried out (as was the case in the present study). The rationale for this procedure is that if expression levels are standardized by a stably expressed reference, the estimates for different individuals are put on the same scale, thus allowing for comparisons across individuals.

B2M encodes for the light chain in the HLA Class I molecule, and it is plausible that B2M and HLA class I genes have some coordination of expression, since they share similar promoter architectures (Kobayashi and van den Elsen 2012; Vijayan et al. 2019), and can be regulated by shared transcription factors (for example, NLRC5/CITA induces the expression of both HLA class I and B2M in Jurkat cell lines (Meissner et al. 2010). The normalization of HLA gene expression by correlated values may introduce bias in our qPCR estimates, especially for HLA-B, for which we see a high correlation with B2M expression (Fig. 4C). Scaling a variable by a different but correlated variable can introduce perturbation by bringing extreme values to the middle of the distribution and reducing variance; consistent with this hypothesis, the coefficients of variation for the qPCR data are 0.61 and 0.50 for HLA-A and -C, respectively, but drops to 0.17 for HLA-B (as a comparison, CVs for RNA-seq data are 0.20, 0.14, and 0.29 for HLA-A, -B, and -C, respectively). However, using the same qPCR design, Ramsuran et al. (2017) normalized HLA-B expression by either B2M, GAPDH, 18 s, and b-Actin gene, and observed very consistent results, which does not support an impact of B2M normalization to the qPCR estimates.

Discussion

Reliable estimates of HLA transcript expression can contribute to diverse research questions, and although disease outcome is frequently explored in the context of HLA coding variation, expression levels are also likely to explain variation in clinical outcomes (reviewed in Dendrou et al. 2018; and in Johansson et al. 2022). Expression levels also have the potential to inform decisions when planning hematopoietic stem cell transplantation; for example, if a perfect match is unavailable in selection for allogeneic donors, it appears beneficial to select those that are mismatched at low expression alleles (Petersdorf et al. 2014, 2015). Reliable estimates of transcript expression may also assist in the identification of eQTLs that underlie the control of HLA expression, which could be integrated into GWAS findings, by querying if known hits in the MHC region coincide with eQTLs for HLA genes (see, e.g., Table S6 in Aguiar et al. 2019). More generally, improved estimates of HLA transcript expression will help us understand the genetic architecture of HLA regulation, identifying the relative contribution of cis-acting variants (i.e., those in the neighborhood of the HLA gene they regulate) and transacting variants (those in distant genomic locations, including on other chromosomes). This will provide information regarding the degree to which variation in HLA expression is an allele-specific property vs. an inter-individual characteristic independent of allelic identity (see Bettens et al. 2022).

Quantitative PCR techniques have enabled us to uncover associations between HLA expression and disease phenotypes. More recently, RNA-seq has become the method of choice to assess gene expression in large whole-transcriptome datasets of different populations. Being able to extract accurate information for HLA expression from such data is an important challenge, and many methods have been proposed to achieve this goal. However, the degree to which the results emerging from RNA-seq analyses agree with those accumulated by the use of qPCR is currently unknown. Although these methods target the same molecular phenotype (RNA abundance), they differ markedly in the experimental techniques used, the forms of analyzing and normalizing the data, the bioinformatic procedures, and the biases they are subject to.

To our knowledge, previous studies comparing HLAtailored RNA-seq approaches with qPCR included small samples. For example, Johansson et al. (2021) validated their HLA-targeted RNA-seq with qPCR on only 5 samples at HLA-C, funding a Pearson correlation coefficient of 0.9, which was not significant (p=0.08).

In the present study, we compared quantitative PCR and RNA-seq expression estimates for the classical HLA class I gene HLA-A, -B, and -C in a matched set of 96 individuals. We found modest but significant correlations in expression over a sample of 96 individuals. Given the lack of a gold standard with which to compare these estimates, estimation errors and biases associated with both methods may contribute to the overall result.

We explored the effects of various factors that may explain the low correlation between RNA-seq and qPCR estimates, such as poor estimation of expression for specific HLA alleles and normalization by a single housekeeping gene in qPCR. Our results cannot be generalized to every qPCR design or RNA-seq pipeline, for which there are a wide variety of different approaches. However, to our knowledge, this is the first direct comparison between qPCR and RNA-seq for the estimation of HLA expression.

Our study suggests areas that require improvement in the determination of HLA transcript expression. Comparisons between RNA-seq and qPCR, for example, should employ uniform processing of samples across methods (e.g., same RNA isolation protocol, storage/thawing time, RNA integrity) to limit artifactual differences associated with these methods. Mapping short reads to single reference genomes or transcriptomes generates biases, and strategies that map reads accounting for HLA polymorphism are necessary. Given that there are several strategies to accomplish this (Boegel et al. 2012; Lee et al. 2018; Aguiar et al. 2019; Gutierrez-Arcelus et al. 2020; Darby et al. 2020), it will be key to compare the relative accuracy of these approaches. 

There is also a need to develop methods that adequately account for isoform variation, not only to provide another layer of information but also more accurate expression estimates, since the normalization of read counts by an incorrect transcript length is a potential source of error. In this context, long-read data, which directly generates full transcript information, can be a powerful tool (Cornaby et al. 2022). Finally, copy number variation, a known feature for certain HLA loci (e.g., DRB), should also be considered when quantifying expression levels.

Acknowledgments

We thank Tatiana Torres (University of São Paulo), Yang Luo (Harvard Medical School), and the members of the Joint Biology Consortium in Boston, for their helpful discussions.

Author contribution

Diogo Meyer, Mary Carrington, Richard M. Single, and Vitor R. C. Aguiar contributed to the study's conception and design. Material preparation, data collection, and experiments were performed by Maureen P. Martin, Veron Ramsuran, Smita Kulkarni, Arman Bashirova, Danillo G. Augusto, and Mary Carrington. Data analysis was performed by Vitor R. C. Aguiar, Erick Castelli, Richard M. Single, Maria Gutierrez-Arcelus, and Diogo Meyer. The manuscript was written by Vitor R. C. Aguiar and Diogo Meyer. All authors read, made contributions, and approved the final manuscript.

Funding

The São Paulo Funding Agency (FAPESP, http://www.fapesp. br/en/) provided funding to DM (2012/18010-0 and 2013/22007-7) and to VRCA (2014/12123-2 and 2016/24734-1). The National Institutes of Health, USA provided funding to DM (NIH R01 GM075091), which supported part of the VRCA postdoc. Conselho Nacional de Desenvolvimento Científco e Tecnológico (CNPq) & Ministry of Health, Brazil provided funding for the RNA-seq experiments and research travel visits, as part of a US-Brazil joint proposal awarded to MC and DM (470043/2014-8). NIH/NIAID R01AI157850 supports SK. VR was funded by the South African Medical Research Council (SAMRC) with funds from the Department of Science and Technology (DST); and also supported in part through the Sub-Saharan African Network for TB/HIV Research Excellence (SANTHE), a DELTAS Africa Initiative (Grant # DEL-15-006) by the AAS. 

This project has been funded in whole or in part with federal funds from the Frederick National Laboratory for Cancer Research, under Contract No. HHSN261200800001E. The content of this publication does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organizations imply endorsement by the U.S. Government. This Research was supported in part by the Intramural Research Program of the NIH, Frederick National Lab, Center for Cancer Research.

Data availability

The RNA-seq data presented in the current publication have been deposited in and are available from the dbGaP database under dbGaP accession phs003177.v1.p1.

References

1. Aguiar VRC, César J, Delaneau O, et al (2019) Expression estimation and eQTL mapping for HLA genes with a personalized pipeline. PLoS Genet 15:e1008091. 

2. Alcina A, Abad-Grau MDM, Fedetz M et al (2012) Multiple sclerosis risk variant HLA-DRB1*1501 associates with high expression of DRB1 gene in different human populations. PLoS One 7:e29819. 

3. Anderson SK (2018) Molecular evolution of elements controlling HLA-C expression: Adaptation to a role as a killer-cell immunoglobulin-like receptor ligand regulating natural killer cell function. HLA 92:271– 278.

4. Apps R, Meng Z, Del Prete GQ et al (2015) Relative expression levels of the HLA class-I proteins in normal and HIV-infected cells. J Immunol 194:3594–3600.

5. Apps R, Qi Y, Carlson JM, et al (2013) Influence of HLA-C expression level on HIV control. Science 340:87–91.

6. Arshad N, Laurent-Rolle M, Ahmed WS et al (2023) SARS-CoV-2 accessory proteins ORF7a and ORF3a use distinct mechanisms to down-regulate MHC-I surface expression. Proc Natl Acad Sci USA 120:e2208525120.

7. Aulchenko YS, Ripke S, Isaacs A, van Duijn CM (2007) GenABEL: an R library for genome-wide association analysis. Bioinformatics 23:1294–1296.

8. Bachtel ND, Umviligihozo G, Pickering S, et al (2018) HLA-C downregulation by HIV-1 adapts to host HLA genotype. PLoS Pathog 14:e1007257.

9. Bettens F, Brunet L, Tiercy J-M (2014) High-allelic variability in HLA-C mRNA expression: association with HLA-extended haplotypes. Genes Immun 15:176–181.

10. Bettens F, Ongen H, Rey G et al (2022) Regulation of HLA class I expression by non-coding gene variations. PLoS Genet 18:e1010212

11.Boegel S, Bukur T, Castle JC, Sahin U (2018) In Silico Typing of Classical and Non-classical HLA Alleles from Standard RNA-Seq Reads. Methods Mol Biol 1802:177–191.

For more information:1950477648nn@gmail.com