Genome-to-genome analysis highlights the impact of the human innate and adaptive immune systems on the hepatitis C virus

Abstract

Outcomes of hepatitis C virus (HCV) infection and treatment depend on viral and host genetic factors. We use human genome-wide genotyping arrays and new whole-genome HCV viral sequencing technologies to perform a systematic genome-to-genome study of 542 individuals chronically infected with HCV, predominately genotype 3. We show that both HLA alleles and interferon lambda innate immune system genes drive viral genome polymorphism, and that IFNL4 genotypes determine HCV viral load through a mechanism that is dependent on a specific polymorphism in the HCV polyprotein. We highlight the interplay between innate immune responses and the viral genome in HCV control.

Introduction

Hepatitis C virus (HCV) infection presents a major health burden, infecting more than 185 million people worldwide¹ and leading to liver failure and hepatocellular cancer. Both host and virus genetic variations are associated with important clinical outcomes. Host genetic polymorphisms, most notably in the interferon lambda 3 and 4 locus, are associated with spontaneous clearance of the virus, response to treatment, viral load and progression of liver disease^2–6. Viral genotypes and distinct viral genetic motifs have been associated with the response to interferon based therapies^7,8 whilst resistance-associated substitutions (RASs) have been identified for most of the new oral direct-acting antiviral (DAA) drugs^9–12. HCV can be divided into seven major genotypes, and most of the genetic data acquired to date focuses on HCV genotype 1 with a lack of data addressing other genotypes. HCV genotype 3 is of particular interest as this genotype is known to infect 53 million people globally¹³ and is associated with a higher failure rate to DAA therapies^14,15.

Previous work has shown that within-host virus diversity evolves in response to the adaptive immune system, including candidate genes studies of the association between the human leukocyte antigen (HLA) type I proteins with the HCV genome^16,17. HLA molecules are expressed on most cell types and present viral peptides (epitopes) to cytotoxic T lymphocytes (CTLs) that kill infected cells. CTL-mediated killing of virus-infected cells drives the selection of viral polymorphisms (“escape” mutants) that abrogate T cell recognition¹⁸. Understanding how host HLA molecules impact on viral selection has important implications for the development of HCV T-cell vaccines that aim to prevent infection^19,20. A comprehensive host genome to viral genome analysis at scale will assess the relative contribution of host HLA molecules in driving HCV genetic change and this analysis may also reveal other host genes that play a key role in shaping the HCV genome.

We generated data from a cohort of 601 HCV infected patients (from the BOSON²¹ clinical trial) to systematically look for associations between host and virus genomes, exploiting the fact that while host genetics remain fixed, the virus mutates, allowing it to evolve during infection. For this we developed targeted viral enrichment^22,23 to obtain whole HCV genomes, and used high-throughput genotyping arrays in combination with statistical imputation to obtain data on HLA alleles^24,25 and nucleotide polymorphisms across the human genome. We provide evidence that polymorphisms relevant to the innate (IFNL4) and adaptive immune systems (HLA genes) are associated with HCV sequence polymorphisms. We show that an interaction between host IFNL4 genotypes and a viral site in the NS5A protein determines HCV viral load. By assessing viral evolution in individuals with different IFNL4 genotypes, we highlight systematic differences in the innate immune response and discuss how these might relate to previous associations with spontaneous clearance and clinical treatment. We demonstrate the potential for a joint analysis of host and viral genomic data to provide information on underlying molecular interactions and their importance in treating and preventing HCV, and other viral infections, in the era of genomic analysis.

Results

Sample description and genetic structure

DNA samples from 567 patients (out of 601 patients) were genotyped using the Affymetrix UK Biobank array. This array genotypes over 800,000 single nucleotide polymorphisms (SNPs) across the human genome, including a set of markers specifically chosen to capture common HLA alleles. Pre-treatment plasma samples from 583 patients in the same study were analysed to obtain HCV whole genome consensus sequences using a high-throughput HCV targeted sequence capture approach coupled with Illumina sequencing²².

Both full-length HCV genome sequences and human genome-wide SNP data were obtained on a total of 542 patients of mainly White and Asian self-reported ancestry infected with HCV genotypes 2 or 3 (Supplementary Table 1). After quality control and filtering of the human genotype data, approximately 330,000 common SNPs with minor allele frequency greater than 5% were available for analysis along with inferred alleles at both class I and II HLA genes. The full-length HCV genome is approximately 9.5 kb, corresponding to over 3000 encoded amino acids. In our dataset, 1226 sites of the HCV proteome were defined to be variable (where at least 10 isolates have an amino acid which differs from the consensus amino acid) to have minimal statistical power for analysis.

We characterised human genetic diversity in the cohort via principal component analysis (PCA). The first two principal components (PCs) corresponded to the sample’s 83% White and 14% Asian self-reported ethnicity, (Supplementary Figure 1), which differed significantly in some of the inferred HLA alleles frequencies (Supplementary Table 2) consistent with previous observations²⁶. The third PC separates individuals with Black self-reported ethnicity from the rest of the cohort. We summarised virus diversity by constructing a maximum-likelihood tree of the consensus sequences from each patient (Supplementary Figure 2). Major clades in the tree separated HCV genotypes 2 (8% of the sample) and 3 (which in turn comprised clades representing subtype 3a and non-subtype 3a samples (90% and 2% of the total respectively)). We observed that patients of specific ancestries, measured either by genetic ancestry (PCs), or self-reported ethnicity, clustered together on the tree of viral diversity (Supplementary Figure 2). A PCA on virus nucleotides sequence data reflected the structure of the tree, specifically at the level of virus subtypes (Supplementary Figure 3).

Systematic host genome to virus genome analysis

We used the genotyped autosomal SNPs in the host genome to undertake genome-wide association studies, where the traits of interest were the presence or absence of each amino acid at the variable sites of the virus proteome, resulting in nearly one billion association tests. We performed logistic regression assuming an additive model, and adjusting for sex, and population structure by including the first three PCs of the host and the first ten PCs of the virus as covariates. Failure to control for either covariates leads to a significant inflation in the association test statistics (Supplementary Figure 4), as would be expected given the observed correlation in population structure of the virus and the host (Supplementary Figure 2). Assuming a human genome-wide significance threshold²⁷ of 5x10^-8, and that amino acid variants in the viral genome are approximately uncorrelated once the population structure is accounted for, then a Bonferroni correction²⁸ results in a significance threshold of approximately 2x10^-11.

Across the human genome, the most significant associations were observed between multiple SNPs in the major histocompatibility complex (MHC; chromosome 6) locus and a virus amino acid variant in non-structural protein 3 (NS3) (Figure 1). Three other associations were observed between multiple SNPs in the host MHC and virus amino acids in NS3 and NS4B proteins (P<4x10^-9, Figure 1). Outside the MHC, the strongest association between host and virus was detected between the SNP rs12979860 in the IFNL4 gene (chromosome 19) and HCV amino acids at position 2570 in NS5B protein (P=1.98x10^-9, Figure 1). Observed variability in the density of nominally significant associations (Figure 1) is largely explained by variability in host and virus sequences, for example in the hyper-variable region (HVR) of HCV in E2 protein.

Figure 1. Human to hepatitis C virus genome-wide association study in 542 patients.

The lower arc shows the human autosomes from chromosome 1 to 22, and the upper arc shows the HCV proteome from core (C) to NS5B. The red line represents the most significant association, P<2x10^-11. The four blue lines represent suggestive associations, P<4x10^-9. The thin grey lines represent associations with P<10^-5. The outer mini-panels represent, on the upper arc, the viral diversity as measured by Shannon entropy and, on the lower arc, the density of human SNPs in bins of 1Mb, with higher values away from the centre for both upper and lower arcs.

We observed 182 associations between human SNPs mapping to two loci and five HCV amino acid sites (P<4x10^-9) (Supplementary Table 3). Since these associations represent places where host genetic diversity has impacted on virus sequence diversity we refer to them as “footprints”. We interpret the signals of association in the MHC region to indicate the effect of the adaptive immune system on genetic diversity in the virus genome. Whilst the effect of MHC was anticipated, the strong signal of association of the interferon lambda region with viral diversity indicates additional effects of the innate immune response.

Host HLA alleles to virus genome

SNPs in the MHC which show strong association with viral amino acids are likely to be correlated with alleles at HLA genes due to extensive linkage disequilibrium across the region^29,30. The HLA repertoire of a patient defines which viral peptides will be presented to T-cells as part of the adaptive immune response, and can lead to the selection of viral mutations (“escape mutations”)^16,31–34. Upon transmission to another host, with a different HLA repertoire, reversion to the wild type may occur (“reversion mutations”). To test for footprinting of the host HLA alleles along the viral genome (Figure 2a and Table 1), we inferred the changes on the terminal branches of the virus phylogenetic tree for each amino acid site and assessed association with the hosts’ HLA alleles. We repeated the analysis on patients infected with genotype 3a and self-reporting as White (Table 1 and Supplementary Figure 5).

Figure 2. Hepatitis C virus genome-wide association studies.

Association between HCV amino acids and (a) HLA alleles (Fisher’s exact test) (b) IFNL4 genotypes (Fisher’s exact test) and (c) pre-treatment viral load (log₁₀PTVL, linear regression). Sites in experimentally validated epitopes in HCV genotype 3 are indicated by a plus sign. Viral sites 1444, 2414 and 2570 are coloured as red. Dashed lines represent a 5% false discovery rate. (d) HCV polyprotein.

Table 1.

Associations between HLA alleles and viral amino acids at a 5% false discovery rate.

In the whole cohort (All) or in HCV genotype 3a self-reported “White” patients only (G3a White). Significant associations are in bold. For each combination of HLA allele and viral site, only the most significant associated amino acid is reported. Amino acids are ordered by decreasing frequency in the column “variable amino acids at this site”.

HLA allele	HCV amino acid position	Viral protein	Variable amino acids at this site	Associated amino acid	P value All	q value All	P value G3a White	q value G3a White	In a known epitope?
B*07:02	42	C	PL	P	1.55x10-07	5.00x10-03	8.07x10-07	5.81x10-03	Yes
C*07:02	42	C	PL	P	2.27x10-07	6.10x10-03	1.99x10-07	1.33x10-03	Yes
B*41:02	109	C	PQSAL	S	4.29x10-06	8.09x10-02	9.84x10-06	4.93x10-02	No
B*13:02	348	E1	ILVM	I	8.13x10-06	1.04x10-01	1.25x10-06	8.78x10-03	No
C*08:02	372	E1	ATVIFC	I	7.33x10-04	4.12x10-01	9.41x10-06	4.93x10-02	No
A*29:02	444	E2	YHFRVWL	Y	1.85x10-04	2.89x10-01	4.02x10-06	2.39x10-02	No
C*15:02	561	E2	VTLI	L	5.70x10-07	1.50x10-02	6.80x10-04	5.92x10-01	No
DRB1*14:04	905	NS2	ATSC	S	4.45x10-03	5.19x10-01	5.58x10-07	4.20x10-03	No
A*31:01	1270	NS3	RKHS	R	1.95x10-09	4.44x10-04	6.20x10-08	3.64x10-04	Yes
A*33:03	1282	NS3	NVTSA	T	1.73x10-07	5.26x10-03	4.14x10-02	9.13x10-01	Yes
B*15:01	1290	NS3	KPARS	R	1.25x10-07	4.47x10-03	3.64x10-06	2.39x10-02	Yes
A*68:02	1341	NS3	VA	V	2.43x10-09	4.44x10-04	9.93x10-07	6.91x10-03	No
A*68:02	1344	NS3	TVA	T	3.68x10-10	<4.00x10-04	9.22x10-11	<2.50x10-04	No
B*51:01	1380	NS3	ILV	L	3.22x10-08	1.67x10-03	2.10x10-07	1.38x10-03	Yes
A*01:01	1444	NS3	FY	F	9.63x10-33	<4.00x10-04	7.18x10-24	<2.50x10-04	Yes
A*68:02	1452	NS3	IV	I	6.98x10-09	4.44x10-04	3.84x10-07	2.89x10-03	No
A*30:02	1585	NS3	YF	Y	1.19x10-06	3.28x10-02	6.89x10-05	1.90x10-01	No
B*13:02	1635	NS3	ITVLAF	T	7.38x10-08	3.13x10-03	1.65x10-07	1.00x10-03	No
B*08:01	1646	NS3	MTAVSI	T	2.89x10-12	<4.00x10-04	2.01x10-10	<2.50x10-04	No
C*07:01	1646	NS3	MTAVSI	T	3.01x10-07	7.39x10-03	1.40x10-08	<2.50x10-04	No
DQA1*05:01	1646	NS3	MTAVSI	T	7.24x10-07	1.76x10-02	1.18x10-05	5.64x10-02	No
B*57:01	1759	NS4B	AVTGNSDI	A	3.38x10-05	1.83x10-01	4.01x10-06	2.39x10-02	No
A*02:01	1873	NS4B	LFKICVPRMTA	F	4.06x10-10	<4.00x10-04	2.58x10-08	2.50x10-04	No
B*38:01	2034	NS5A	STNLPIAQVDKEMGRFW	P	7.56x10-08	3.13x10-03	6.35x10-08	3.64x10-04	Yes
C*12:03	2034	NS5A	STLPQKVAIRMFW	P	9.31x10-06	1.05x10-01	3.99x10-06	2.39x10-02	Yes
B*18:01	2144	NS5A	ED	E	1.16x10-06	3.28x10-02	2.36x10-06	1.67x10-02	Yes
B*51:01	2148	NS5A	MTVSLKIA	V	2.77x10-06	6.07x10-02	4.60x10-07	3.16x10-03	Yes
C*14:02	2148	NS5A	MTVSLKIA	V	1.91x10-05	1.34x10-01	2.32x10-06	1.67x10-02	Yes
B*40:01	2248	NS5A	TSKAN	T	5.76x10-04	3.91x10-01	9.62x10-06	4.93x10-02	No
DQA1*01:01	2486	NS5B	TIANSV	T	4.71x10-05	1.96x10-01	5.12x10-06	2.90x10-02	Yes
A*31:01	2510	NS5B	AKQSEMGLTV	A	4.96x10-08	2.14x10-03	1.79x10-06	1.35x10-02	No
A*32:01	2537	NS5B	NDHSYEAI	N	1.13x10-08	7.27x10-04	4.77x10-06	2.75x10-02	Yes
A*32:01	2540	NS5B	RSKHNQC	R	7.45x10-09	6.00x10-04	5.96x10-12	<2.50x10-04	Yes
A*02:11	2600	NS5B	QKRSAL	Q	1.01x10-06	3.12x10-02	NA	NA	Yes
A*26:01	2605	NS5B	EAGVK	G	2.61x10-07	6.64x10-03	5.91x10-05	1.73x10-01	Yes
B*51:01	2713	NS5B	ILMV	I	7.70x10-06	1.03x10-01	1.25x10-08	<2.50x10-04	No
A*25:01	2821	NS5B	RCQL	R	8.31x10-06	1.04x10-01	2.51x10-06	1.73x10-02	No

In the whole cohort, at a 5% FDR, 24 combinations of HLA alleles and HCV sites were significant (Table 1) and this increased to 153 associations at a 20% FDR (Supplementary Table 4). Out of 21 viral amino acid positions showing signals of association with one or more HLA alleles, 12 were located in previously reported HCV genotype 3 epitopes²⁰ (Table 1), which represents a strong enrichment (odds ratio, OR=5.2, P=2.8x10^-4). We also observed that NS3 protein was strongly enriched for association signals with HLA alleles (OR=5.5, P=2.2x10^-4). The strongest HLA footprints are found with common alleles at HLA-A and HLA-B genes, although signals are also found in HLA-C and the class II gene HLA-DQA1. (See Supplementary notes and Supplementary Figures 6, 7 and 8 for detailed description of the most significant association between HLA-A and Y1444F). At position 1646 in the HCV genome (in NS3), the footprint is seen with multiple HLA alleles (B*08:01, C*07:01 and DQA1*05:01), although this is potentially a result of linkage disequilibrium between these HLA alleles (r² between B*08:01 and C*07:01 = 0.49, and r² between B*08:01 and DQA1*05:01 = 0.22; Supplementary Figure 9). Using the 27 HLA and viral amino acid associations at a 20% FDR that have sufficient observations to estimate the ORs, we observed a negative correlation between the ORs of escape and reversion (r=-0.65, P<0.01; Supplementary Figure 10). These observations are consistent with HLA alleles driving patterns of both escape and reversion at viral amino acids. Our analysis provides a map of their influence across the HCV genome.

IFNL4 host variants to virus genome

Variants in the interferon lambda region have been associated with multiple HCV outcomes including spontaneous clearance, treatment response, viral load and liver disease progression^2,3,35. In our genome-wide analysis, variants around the interferon lambda region showed the strongest association with HCV amino acids outside the MHC region (the top associated SNP is rs12979860, P=1.98x10^-9). For SNP rs12979860, the CC genotype is associated with higher rates of spontaneous clearance and interferon-based treatment response, putatively due to the fact that the C allele tags a dinucleotide insertion polymorphism (rs368234815-TT) which prevents IFNL4 expression³⁶. In our cohort, only two individuals had discordant genotypes for rs12979860 and rs368234815 (which was imputed), resulting in a strong linkage disequilibrium (r²=0.99) between these two SNPs. Therefore, individuals with non-CC genotypes express IFNL4 and have increased expression of hepatic interferon stimulated genes (ISGs)³⁷.

We compared viral amino acid changes between hosts with CC and non-CC genotypes at the IFNL4 SNP rs12979860 using the same Fisher’s exact test as described above. The most significant association was with changes to and away from valine (V) at position 2570 in the viral NS5B protein (P=1.94x10^-11, Figure 2b). We replicated this association in an independent study of 360 patients of European ancestry chronically infected with HCV genotypes 2 or 3^38,39 (one-sided P=0.005). In addition, a candidate gene association study in a genotype 1b single source infection cohort has shown an association between IFNL4 genotypes (rs12979860) and an amino acid in the same region of NS5B (position 2609)³² reinforcing the potential role of the locus in interactions with the host innate immune system.

Overall, using a 5% FDR, 11 significant associations were observed between IFNL4 genotypes and amino acid positions located in core, E2, NS5A and NS5B proteins (Supplementary Table 5). Using a permutation approach, we found that the core protein is nominally (P<0.05) enriched, and the NS5A protein is nominally depleted, in the number of associations with the IFNL4 genotypes. Genotype 3a viral sequences in individuals with the CC genotype did have more mutations away from the population consensus than non-CC individuals (See Supplementary Note and Supplementary Table 6). One of the sites (position 109 in the core protein) is also associated with HLA-B*41:02 (P=4.3x10^-6, Table 1). However, we did not find any consistent evidence for interaction between by HLA alleles and IFNL4 genotypes and their association with viral amino acids (Supplementary Figure 11). Nor did we find strong differences in the mean number of escape mutations in CC and non-CC patients when comparing HLA allele carriers and non-HLA allele carriers.

To further investigate the selective pressures on the virus in patients with IFNL4 CC and non-CC genotypes, we estimated the rates of synonymous (dS) and non-synonymous substitutions (dN) in genotype 3a infected patients (Figure 3). Whilst there was no difference in dS in CC and non-CC patients (P=0.68, Figure 3a), dN was significantly higher in CC patients (P=1.6 x10^-8, Figure 3b). The lower dN/dS ratio in non-CC patients (P=1.3x10^-10) is potentially indicative that the virus is under a stronger purifying selection (Figure 3c). This hypothesis is supported by the observation that for the same rate of synonymous substitutions dS (a surrogate for the time and amount of divergence), the HCV genome is under a larger purifying selection in patients with IFNL4 non-CC genotypes than CC genotype (Figure 3d).

Figure 3. Association between IFNL4 genotypes and substitution rates in genotype 3a infected patients.

Stratification of substitution rates by IFNL4 CC and non-CC genotypes. (a) Rate of synonymous substitutions (dS, P=0.68), (b) Rate of non-synonymous substitutions (dN, P=1.6x10^-8) and (c) dN/dS ratio (P=1.28x10^-10). Each blue dot represents the mean dS, dN or dN/dS ratio per patient. The mean and 95% confidence intervals are shown as black dots and bars. (d) The joint distribution of dS and dN/dS in individuals with the IFNL4 non-CC genotypes (red dots) and with the CC genotype (blue dots).

Estimating dN/dS ratio per viral gene showed that this ratio was significantly higher (P<0.05) in CC patients compared to non-CC patients in E1, E2, NS3 and NS5B (Supplementary Figure 12). A sliding window analysis across the HCV genome showed that in the full cohort E1 and E2 genes had a much higher dN/dS ratio compared to the rest of the genome. These envelope genes include hyper-variable regions (HVRs), and are thought to be the primary targets of the antibody-based immune response (Supplementary Figure 13).

Host and virus genetic determinants of viral load

We performed a genome-wide association study in patients infected with HCV genotype 3a using an additive linear regression model adjusted for sex and the three first host PCs for log₁₀ transformed of pre-treatment viral load (log₁₀PTVL) (Figure 4a). We replicated the known association between IFNL4 variants on chromosome 19 and viral load³ (rs12979860, P=5.9x10^-10) with the non-CC genotypes conferring an approximately 0.45-fold decrease in viral load (mean for non-CC=3.4760x10⁶ IU/mL and for CC=6.3447x10⁶ IU/mL). We also performed a genome-wide association study to detect associations between viral amino acids of genotype 3a virus and viral load (Figure 2c). The only amino acid significantly associated with log₁₀PTVL at a 5% FDR was a change from a serine (S) to an asparagine (N) at position 2414 in NS5A protein (P=9.21x10^-7, Figure 4b). This site is one of the 11 sites significantly associated (5% FDR) with IFNL4 genotypes (Figures 2b and 2c).

Figure 4. Association between viral load and human and virus genetic variants in genotype 3a infected patients.

(a) Association between human SNPs and log₁₀PTVL, (b) Distribution (blue dots), estimated mean and 95% confidence interval (shown as black dots and bars) of log₁₀PTVL stratified by amino acids present at viral position 2414 (P=9.21x10^-7), (c) by IFNL4 (CC, CT and TT) genotypes in patients whose virus carries a serine at position 2414 (P=9.37x10^-9) or (d) in patients whose virus does not carry a serine amino acid at position 2414 (P=0.9).

In patients with a serine at position 2414, the association between IFNL4 genotypes and log₁₀PTVL is significant (P=9.37x10^-9, Figure 4c). However, we observed no association (P=0.9) between IFNL4 genotypes and log₁₀PTVL in patients infected with a virus that has a different amino acid (Figure 4d). In other words, the host’s IFNL4 genotypes determine viral load only if they are infected by a virus with the serine amino acid at position 2414 in NS5A protein (Figures 4c and 4d). The interaction is statistically significant when analysing either the whole cohort (P=0.017) or just patients with genotype 3a infections who self-report as being White (P=0.017). Together the combinations of non-CC genotypes and a serine viral amino acid at position 2414 are inferred to result in a 0.57-fold decrease in viral load compared to all other combinations (mean viral load for non-CC and serine at position 2414 =2.81x10⁶ IU/mL and all other combinations =6.47x10⁶ IU/mL). Introduction of the change from a serine to an asparagine at position 2414 into a modified S52 replicon of genotype 3a⁴⁰ resulted in a approximately 10-fold increase in replication in Huh7.5 cell culture (Supplementary Note and Supplementary Figure 14).

We also observed that non-consensus amino acids, which are increased in frequency in individuals with a CC genotype, tend to also associate with increased viral load (r=0.42, P=0.005, Supplementary Figure 15). The same positive relationship was observed when estimating the effect on viral load in individuals with a CC genotype only (r=0.35, P=0.02) or non-CC genotypes only (r=0.4 and P=0.007). A nominally significant trend was observed when the analysis was done for all variant positions in the viral genome (r=0.099 and P=0.04) (Supplementary Figure 16).

Discussion

Here we report the first systematic analysis of associations between variation in human and HCV genomes in a large patient cohort. Advances in DNA and RNA sequencing technology and new bioinformatics tools have allowed full-length viral consensus sequences to be obtained in large number of patients for reasonable cost (approximately £100 per sample), as well as host genetic data at millions of directly assayed and imputed polymorphisms (approximately £75 per sample). We apply a fast and simple approach to test for association between host and pathogen variants, using a logistic regression analysis corrected for both human and viral population structures by using the principal components of the genome-wide data as covariates (60 hours for approximately 2500 association studies of 330,000 SNPs). We also applied a contingency table analysis based on the inferred viral amino acid changes since infection. We anticipate that with the reduction in the cost of sequencing and genotyping, and the increasing interest in studying large patient cohorts, analyses of this kind will become a powerful approach to understanding infectious diseases. Confirmation of the specific associations reported here, and the extent to which they are specific to viral genotype, will require replication analysis in independent cohorts.

We found strong evidence for the adaptive immune system exerting selective pressures on the HCV genome, presumably by preferentially selecting viral mutations that avoid antigenic presentation by the host’s HLA proteins. Some of the observed associations are located in experimentally validated viral epitopes²⁰, however others have not been described experimentally and most likely represent sites of novel T-cell escape mutations. Assuming that our analysis removes biases associated with population structure and incorrect ancestral inference, 5% of the viral amino acids (153/3021) are associated with HLA alleles (at 20% FDR). These data highlight the importance of the adaptive immune system in driving viral evolution, and serves as a map of the targets of T-cell based immunity along the HCV genome which can aid vaccine design and development²⁰.

In addition to the HLA, we now show that IFNL4 activity may significantly shape the HCV viral genome. Previous studies have shown that the “favourable” CC IFNL4 genotype increases the chances of spontaneous resolution and interferon-based treatment success^2–6. The IFNL4 TT/TT genotype (rs368234815), which is strongly linked to the “favourable” CC genotype (rs12979860), abolishes the expression of IFNL4⁴⁶, whereas in individuals with the IFNL4 ΔG/TT or ΔG/ΔG genotypes (linked to“unfavourable” non-CC genotypes), IFNL4 is expressed, leading to the downstream up-regulation of hepatic ISGs expression via the JAK-STAT pathway³⁷. The expression of ISGs has been shown to render the host less susceptible to exogenous IFNα/γ stimulation and is associated with more infected cells in the liver⁴². To date it has been presumed that this fully explains why patients with specific IFNL4 genotypes have differential outcomes during primary infection or with drug therapy^42,43.

Our analysis adds to this hypothesis, with the observation that IFNL4 variants also impact on the HCV genome at multiple amino acid sites. Indeed, these were the strongest footprinting signals in our systematic analysis outside HLA region. The most significant association was at position 2570 in the viral NS5B protein; an association that we replicated in an independent cohort infected with HCV genotypes 2 and 3. The additional signals in NS5B protein associated with HLA alleles or IFNL4 genotypes at a 5% FDR did not replicate (P > 0.05) potentially due to a lack of power resulting from the smaller sample size of the replication cohort and/or the fact that the phylogenetically corrected Fisher’s tests could not be performed in an equivalent way. An association between IFNL4 and amino-acid variability in the same region of NS5B protein, has previously been reported in a candidate gene study³² in a single source infection cohort. IFNL4 has also been associated with a viral mutation associated with DAA resistance in HCV genotype 1 infection⁴⁴ although this association has not been replicated in our HCV genotype 3 cohort⁴⁵. The additional signals in NS5B protein associated with IFNL4 genotypes (or HLA alleles) at a 5% FDR did not replicate (P > 0.05) potentially due to a lack of power resulting from the smaller sample size of the replication cohort and/or the fact that the phylogenetically corrected Fisher’s exact tests could not be performed in an equivalent way (only NS5B sequences were available in the replication cohort). However, the broader impact of interferon lambda host genes that are associated with, and potentially select for, specific viral variants has not been previously recognised.

HCV viral load is an important and clinically relevant parameter since patients with higher HCV viral loads have lower response rates to IFN and DAA based therapy⁴⁶ (independent of IFNL4 status). Paradoxically, the “favourable” IFNL4 variants have also been associated with both an increase in disease progression⁴⁷ and high viral load^3,48. We report an association between a virus amino acid site (serine vs. non-serine at position 2414 in NS5A) and HCV viral load. This site is one of the 11 sites that are putatively associated with IFNL4 genotypes in our dataset. Further, our data shows that a decrease in viral load was only observed in those patients with non-CC genotypes whose virus carried the serine amino acid at site 2414 in NS5A protein. Since the 2414 S variant is found in 85% of non-CC patients (compared to 67% of CC patients), this interplay between host and viral genes helps explain the previous observation that non-CC patients have a lower HCV viral load (Figure 5). Our in vitro data from a genotype 3 replicon assay shows that a change from a serine to asparagine at site 2414 is associated with an increase in RNA replication and perhaps hyper-phosphorylation⁴⁹, which is a negative regulator of virus replication.

Figure 5. Overview of the observations relating to the interplay between innate immune response and the viral genome in hepatitis C virus control.

(a) Infected individuals with IFNL4 CC genotype (37% of the HCV genotype 3a infected patients) show high rates of spontaneous and treatment-induced clearance of HCV. IFNL4 is not expressed, which in turn induces a weaker and possibly differential interferon-stimulated genes (ISG) expression. The host environment is associated with weaker purifying selection and allows viral mutations associated with a better replicative fitness to accumulate, leading to higher viral load. In this group of patients, 67% are infected by a virus with a serine at position 2414 and 33% with a different amino-acid. (b) Infected individuals with IFNL4 non-CC genotypes (63% of the HCV genotype 3a infected patients), have lower rates of spontaneous and treatment-induced clearance of HCV. IFNL4 is expressed and induces ISGs that collectively establish an antiviral state hostile to viral replication. This hostile environment induces a high selective pressure and fewer viral mutations can accumulate. Although the serine at position 2414 (compared to non serine) is associated with lower viral load, it is highly prevalent in this group of patients (85% are infected by a virus with a serine at position 2414 and 15% with a different amino-acid).

Our interpretation of the data (Figure 5) is that the expression of IFNL4 by ΔG/TT or ΔG/ΔG genotypes (tagged in our cohort by the “unfavourable” non-CC genotypes) leads to the activation of additional components of the immune response, likely driven by ISGs, which interact directly with specific amino acids in the viral genome (most notably amino acid 2414 in NS5A which has a significant impact on viral load (Figures 2b, 4 and Supplementary Figure 15). Our data suggest that this also leads to an overall increase in the strength of purifying selection (decrease in dN/dS, Figure 3), and together this leads to lower viral load. However, viruses that establish chronic infections in non-CC patients have evolved to survive in a more hostile environment (for example mutating the serine at amino acid 2414 of NS5A), which makes them less likely to respond to interferon-based therapy. In contrast, our analysis suggests that the “favourable” CC genotype and the inactivation of IFNL4 gene (by the IFNL4 TT genotype) disables components of the immune response (therefore removing the effect of amino acid 2414 in NS5A on viral load), which leads to a reduced level of purifying selection (Figure 3). It is possible that this then permits a range of mutations which confer higher replicative fitness and therefore higher viral load (Supplementary Figure 15), but these viruses are more susceptible to interferon-based treatments. At the population level, we would expect a balance in the relative contribution of these mechanisms as viruses move between individuals with CC and non-CC genotypes. Our results make the prediction that the outcome of a new infection will be dependent on both the HLA alleles and the IFNL4 genotype of the patient who is the source of the new infection. Further analysis is required to fully understand the impact of “favourable” CC and “unfavourable” non-CC genotypes on the different components of the immune system and to establish their clinical relevance before, during and after infection.

In conclusion, we provide a comprehensive genome-to-genome analysis in chronic HCV infection. Using this genome-wide, hypothesis-free, approach we show that the host’s HLA alleles leave multiple footprints in the HCV genome and that the host’s innate immune environment also influences the amino acid polymorphisms in the virus, both at specific loci, and genome-wide. We observe a common viral amino-acid residue that is associated with HCV viral load only in patients with the “unfavourable” non-CC IFNL4 genotypes. These observations suggest that the innate and adaptive immune system jointly impact on HCV genome evolution and likely together determine the establishment of infection and its control over time. The new insights into the biological mechanisms that drive HCV evolution in vivo, and the identification of specific interactions between viral and host polymorphisms are relevant for future approaches to treatment stratification and vaccine development.

Online Methods

Patients and sample

Plasma and DNA samples came from patients enrolled in the Boson study. Boson study is a phase 3 randomized open-label trial to determine the efficacy and safety of sofosbuvir with and without pegylated-interferon-alfa, in treatment-experienced patients with cirrhosis and hepatitis C virus (HCV) genotype 2 infection and treatment-naïve or -experienced patients with HCV genotype 3 infection²¹. All patients provided written informed consent before undertaking any study-related procedures. The study protocol was approved by each institution’s review board or ethics committee before study initiation. The study was conducted in accordance with the International Conference on Harmonization Good Clinical Practice Guidelines and the Declaration of Helsinki. The study reported here is not a clinical trial, but is based on the analysis of patients from a clinical trial (registration number: NCT01962441). Sample sizes were determined by the available data. All samples for which both viral sequencing and host genetics were available were included in the final analysis unless otherwise specified.

Host genotyping and imputation

Informed consent for host genetic analysis was obtained from 567 patients. Genotyping was performed using Affymetrix UK Biobank arrays. We imputed the MHC class I loci HLA-A, HLA-B, HLA-C and class II loci HLA-DQA1, HLA-DQB1, HLA-DPB1, HLA-DRB1, HLA-DRB3, HLA-DRB4, HLA-DRB5 using HLA*IMP:02²⁴ accessed 22 March 2015. HLA amino acids were also imputed by SNP2HLA⁵⁰ using the T1DGC as the reference panel, which contains 5225 unrelated individuals (10,450 haplotypes). Logistic regression using posterior genotype probabilities (allele dosages) for each HLA allele from SNP2HLA were carried out using PLINK2⁵¹ (https://www.cog-genomics.org/plink2).

Virus sequencing

Sample collection and preparation

RNA was isolated from 500µl plasma using the NucliSENS magnetic extraction system (bioMerieux) and collected in 30µl of kit elution buffer for storage in aliquots at -80°C.

Sequencing library construction, enrichment and sequencing

Libraries were prepared for Illumina sequencing using the NEBNext® Ultra™ Directional RNA Library Prep Kit for Illumina® (New England Biolabs) with 5µl sample (maximum 10ng total RNA) and previously published modifications of the manufacturer’s guidelines (v2.0)²², briefly: fragmentation for 5 minutes at 94°C, omission of Actinomycin D at first-strand reverse transcription, library amplification for 18 PCR cycles using custom indexed primers⁵² and post-PCR clean-up with 0.85× volume Ampure XP (Beckman Coulter).

Libraries were quantified using Quant-iT™ PicoGreen® dsDNA Assay Kit (Invitrogen) and analysed using Agilent TapeStation with D1K High Sensitivity kit (Agilent) for equimolar pooling, then re-normalized by qPCR using the KAPA SYBR® FAST qPCR Kit (Kapa Biosystems) for sequencing. A 500ng aliquot of the pooled library was enriched using the xGen® Lockdown® protocol from IDT (Rapid Protocol for DNA Probe Hybridization and Target Capture Using an Illumina TruSeq® or Ion Torrent® Library (v1.0), Integrated DNA Technologies) with equimolar-pooled 120nt DNA oligonucleotide probes (IDT) followed by a 12-cycle, modified, on-bead, post-enrichment PCR re-amplification. The cleaned post-enrichment ve-Seq library was normalized with the aid of qPCR and sequenced with 151b paired-end reads on a single run of the Illumina MiSeq using v2 chemistry.

Sequence data analysis

De-multiplexed sequence read-pairs were trimmed of low-quality bases using QUASR v7.0120⁵³ and adapter sequences with CutAdapt version 1.7.1⁵⁴ and subsequently discarded if either read had less than 50b remaining sequence or if both reads matched the human reference sequence using Bowtie version 2.2.4⁵⁵. The remaining read pool was screened against a BLASTn database containing 165 HCV genomes⁵⁶ covering its diversity both to choose an appropriate reference and to select those reads which formed a majority population for de novo assembly with Vicuna v1.3⁵⁷ and finishing with V-FAT v1.0 (http://www.broadinstitute.org/scientific-community/science/projects/viral-genomics/v-fat). Population consensus sequence at each site is defined as the most common variant at that site among all the patients.

Phylogenetic and ancestral sequence reconstruction

Whole genome viral consensus sequences for each patient were aligned using MAFFT⁵⁸ with default settings. This alignment was used to create a maximum likelihood tree using RAxML⁵⁹, assuming a general time reversible model of nucleotide substitution under the gamma model of rate heterogeneity. The resulting tree was rooted at midpoint. Maximum likelihood ancestral sequence reconstruction was performed using RAxML⁵⁹ with the maximum likelihood tree and HCV polyprotein sequences as input.

Association analysis

To test for association between human SNPs and HCV amino acids at genome-to-genome level, we performed logistic regression using PLINK2⁵¹ (https://www.cog-genomics.org/plink2) adjusted for the human population structure (three first PCs assessed using EIGENSOFT v3.0⁶⁰) and the virus population structure (10 first PCs). For the viral data PCA was performed on the nucleotide data. Tri and quad-allelic sites were converted to binary variables and the amino acid frequencies were standardised to have mean zero and unit variance. MATLAB (Release 2015a, The MathWorks) was used to perform the PCA using the singular value decomposition function.

To test for association between imputed HLA alleles and HCV amino acids, we used Fisher’s exact test, correcting for the virus population structure as described in Bhatacharaya et al.⁶¹. We used the inferred ancestral amino acid to estimate changes for each site on the terminal branches of the virus phylogenetic tree. Inferring the changes along the terminal branches of the tree aims to control for the confounding between host and virus population structures⁶¹ by looking at viral mutations after infection. We construct a 2x4 contingency table where rows denote presence or absence of a host HLA allele, and the columns denote changes to and away from a specific amino acid in viruses with and without the amino acid inferred to be ancestral. To test for association between IFNL4 SNP rs12979860 genotypes and HCV amino acids, we used the same Fisher’s exact test with a dominant model for IFNL4 rs12979860 by encoding genotypes as CC and non-CC.

Permutation was used to estimate the FDR for association tests that used Fisher’s exact test. The rows of matrix M (where rows correspond to study participants and columns to HLA alleles or the IFNL4 genotypes) were randomly permuted 500 times and in each case the p-values of the associations were calculated. For each threshold t, the expected number of false significant associations was estimated by the mean number of false positives across permuted null data sets. The FDR for threshold t was then estimated as the mean number of false positives divided by the observed number of significant associations in the actual data at threshold t.

To test for enrichment of association signals in epitope regions, viral proteins or with a specific HLA allele we used Fisher’s exact test. Each site is either within a target region (epitope regions or a specific viral protein) or not and the most associated test with it, is significant or not with an FDR of 5%. The resulting contingency table was tested using Fisher’s exact test to assess enrichment or depletion of signals of association.

To assess the relationship between rates of escape and reversion in HLA presentation, we estimated the odds ratio for each 2x2 sub-table used in the Fisher’s exact test. This was done only for viral sites associated with HLA alleles at 20% FDR and which there were sufficient observations in both tables to estimate the odds ratio where confidence interval did not go to infinity. Pearson’s correlation coefficient was used to assess the relationship between the log₁₀(OR) of escape and reversion.

To test for enrichment of viral amino acid associations with host IFNL4 genotypes in viral proteins, the null distribution of number of association in each protein was estimated using 10,000 permutations of IFNL4 labels and performing the same tests. The estimated null distribution of number of associations for each viral protein was compared to the observed number of associations in the data to test for enrichment or depletion of number of associations. To test if HVR1, HVR2, HVR3, ISDR and PKR-BD regions show differences in the number of changes away from the population consensus in CC and non-CC hosts, we used a Poisson regression. In each individual and each locus we determined the number of differences to the population consensus. We then estimated the effect of IFNL4 genotypes on the mean number of differences to the population consensus using Poisson regression. The same procedure was used to test if the total number of differences across the whole poly-protein relative to the population consensus was influenced by IFNL4 genotypes.

To estimate the rate of synonymous and non-synonymous mutations, we used dndsml function from MATLAB (Release 2015a, The MathWorks) that uses Goldman and Yang’s method. It estimates (using maximum likelihood) an explicit model for codon substitution that takes into account transition/transversion rate bias and base/codon frequency bias. Then it uses the model to correct synonymous and non-synonymous counts to account for multiple substitutions at the same site. To estimate dN and dS, each sequence was compared to the population consensus which indicates the most common nucleotide observed in our data set at each position along the genome.

To determine whether IFNL4 genotypes impact on HLA alleles presentation of epitopes, logistic regression with interaction term was used. The outcome for individuals was whether on the terminal branches of the tree a specific had changed amino or not. We tested for interaction between presence and absence of the associated HLA allele and IFNL4 genotypes for all combinations of HLA alleles and viral sites associated at a 20% FDR. In addition, we tested for an overall effect of IFNL4 genotypes on HLA alleles’ presentation of epitopes. We used our 2x4 contingency tables and the odds ratios estimated from the 2x2 sub-tables to infer the antigenic amino acids. If the odds ratio indicates that in individuals with HLA allele present, the •X ancestral amino acid -> any other amino acid• element is enriched relative to individuals without the HLA allele then we assume the X amino acid is the antigenic amino acid and escape occurs away from X to any other amino acid. For these cases, we count how many of •X ancestral amino acid -> any other amino acid• occur in CC and non-CC individuals across all combinations of HLA alleles and viral sites associated at 20% FDR. If IFNL4 genotypes have no impact on the HLA presentation (null hypothesis), then the mean number of escape mutations in CC and non-CC hosts should be proportional to the frequency of hosts with CC and non-CC genotypes (null distribution is binomial with parameters n equal to the total number of observed escape mutations and p equal to the proportion of CC hosts).

We used linear regression in PLINK2⁵¹ (https://www.cog-genomics.org/plink2) to test for association between human SNPs and log₁₀ transformed pre-treatment viral load (PTVL) including sex and first three PCs of the host genome as covariates. We used linear regression to test for association between HCV amino acids (with a minimal count of 10 at each site) and viral load. We used linear regression in R (version 3.2.4 (2016-03-10))⁶² to analyse the interaction between IFNL4 genotypes and amino acids at viral site 2414 and to quantify their impact on viral load.

For all viral sites associated with IFNL4 genotypes at 20% FDR, we assessed if there is a relationship between effect size of non-consensus amino acids on viral load and effect size of IFNL4 genotypes and changes to non-consensus amino acids. We estimated the odds ratio of enrichment of changes away from the consensus amino acid on the terminal branches of the virus tree in CC and non-CC genotypes (2x2 contingency table, only using data in which the consensus is inferred to be ancestral). We also estimated the effect size of non-consensus amino acids on the log₁₀ of viral load using a linear regression with IFNL4 genotype as a covariate. Additionally, we estimated the effect size of non-consensus amino acid on the log₁₀ viral load in CC and non-CC hosts. We used Pearson’s correlation coefficient to measure the strength of relationship between the effect size of non-consensus amino acids on viral load and log of odds ratio of enrichment of non-consensus amino acid changes in CC hosts.

Code Availability

R and MATLAB code used to generate the results and figures from the primary analyses described above are available from the authors on request.

Replication study

To replicate the IFNL4 SNP rs12979860 results, we ran the association analysis on an independent HCV infected population that was recruited to the FISSION, FUSION and POSITRON phase 3 clinical studies^38,39. The Material Transfer Agreements under which the data were shared limited analysis to the NS5B protein. Paired human genome-wide genotyping and HCV sanger sequencing data for NS5B amplicon were obtained from DNA and plasma samples collected from 360 Caucasian patients chronically infected with HCV genotype 2 (N=153) or 3 (N=208). We searched for association between the IFNL4 SNP rs12979860 and viral position 2570 using logistic regression with the outcome indicated by the presence or absence of amino acid V at position 2570. To help prevent spurious associations due to host and viral stratification, we included human principal components and viral genotype as covariates.

Replicon assay

Cell Culture

Huh7.5-Sec14L2 cells, previously reported⁴⁰ were grown in Dulbecco’s modified Eagle’s medium (DMEM, Life Technologies) supplemented with 10% fetal calf serum, 2 mM L-glutamine, 100 U/ml penicillin, 100 0 enicillin, 100 100 lmM HEPES and 0.1M nonessential amino acids as described⁶³. Huh7.5 cells are heterozygous CT for the IFNL4 rs12979860 SNP⁶⁴.

HCV Mutant Replicons

The enhanced version of the subgenomic replicon of genotype 3a strain S52, lacking of neomycin resistance gene, has been previously described⁴⁰. Site-specific mutations were introduced using QuikChange II XL Site-Directed Mutagenesis Kit (Agilent Technologies) following manufacturer instructions and confirmed by direct sequencing. HCV mutant plasmids were linearized by XbaI digestion (New England Biolabs; NEB), mung-bean treated (NEB) and purified. Linearized DNA was then used as template for in vitro RNA transcription (IVT) (Megascript T7, Life Technologies) according to manufacturer protocol. Finally, IVT RNA has been DNAse treated, purified and stored at -80°C.

Electroporation and Luciferase Detection

For electroporation, cells were counted and then washed twice in ice-cold PBS. Typically, for each mutant 4x10⁶ cells were mixed with 1 µg of replicon RNA in a 4mm cuvette and electroporated in the Gene Pulser Xcell (Bio-Rad) at 250 V, 950 µF using exponential decay setting. Cells were immediately recovered in pre-warmed complete DMEM, seeded in a 24-well plate and incubated at 37°C. After 5, 24, 48 or 72 hours, medium was removed and cells lysed with Glo Lysis Buffer (Promega). Cell lysates were then transferred in a white 96-well plate (Corning) and the luciferase expression was quantitated in a luminometer (GloMax 96 Microplate Luminometer, Promega) using Bright-Glo assay system (Promega).