Emergence of the Virulence-Associated PB2 E627K Substitution in a Fatal Human Case of Highly Pathogenic Avian Influenza Virus A(H7N7) Infection as Determined by Illumina Ultra-Deep Sequencing

Abstract

Avian influenza viruses are capable of crossing the species barrier and infecting humans. Although evidence of human-to-human transmission of avian influenza viruses to date is limited, evolution of variants toward more-efficient human-to-human transmission could result in a new influenza virus pandemic. In both the avian influenza A(H5N1) and the recently emerging avian influenza A(H7N9) viruses, the polymerase basic 2 protein (PB2) E627K mutation appears to be of key importance for human adaptation. During a large influenza A(H7N7) virus outbreak in the Netherlands in 2003, the A(H7N7) virus isolated from a fatal human case contained the PB2 E627K mutation as well as a hemagglutinin (HA) K416R mutation. In this study, we aimed to investigate whether these mutations occurred in the avian or the human host by Illumina Ultra-Deep sequencing of three previously uninvestigated clinical samples obtained from the fatal case. In addition, we investigated three chicken samples, two of which were obtained from the source farm. Results showed that the PB2 E627K mutation was not present in any of the chicken samples tested. Surprisingly, the avian samples were characterized by the presence of influenza virus defective RNA segments, suggestive for the synthesis of defective interfering viruses during infection in poultry. In the human samples, the PB2 E627K mutation was identified with increasing frequency during infection. Our results strongly suggest that human adaptation marker PB2 E627K has emerged during virus infection of a single human host, emphasizing the importance of reducing human exposure to avian influenza viruses to reduce the likelihood of viral adaptation to humans.

INTRODUCTION

Influenza viruses have been isolated from a wide range of host species, and they occasionally cross the species barrier to infect humans and vice versa (1). The persistence of avian influenza viruses in their natural wild waterbird reservoir can trigger avian influenza epidemics in poultry, in which these viruses generally cause mild or subclinical infection (2). For this reason, they are referred to as low-pathogenic avian influenza (LPAI) viruses. LPAI viruses of subtypes H5 and H7 have the ability to mutate to viruses with increased virulence for poultry, referred to as high-pathogenic avian influenza (HPAI) viruses, which is why they are targeted through mandatory veterinary surveillance programs. The transmission of LPAI virus from birds to mammals has resulted in the establishment of adapted variants of equine, canine, and multiple swine influenza virus lineages (3). At present, avian influenza viruses of subtypes H5 and H7 have not adapted to mammalian species. However, the emergence and dissemination over a wide geographic region of a lineage of HPAI virus A(H5N1) in wild birds and poultry, with reported transmissions to >600 humans including >300 fatalities, is of concern (4). Similarly, the recent emergence of human cases of influenza A(H7N9) virus infections in China in 2013 poses a public health risk (5, 6). Although evidence of avian influenza virus transmission between humans has to date been limited, possible evolution toward more-transmissible virus variants carries the specter of a potential pandemic. Fortunately, avian influenza viruses have not yet acquired mutations that facilitate efficient aerosol or respiratory droplet transmission between humans. Assessment of HPAI A(H5N1) surveillance data for mutations required for airborne transmission of A(H5N1) viruses between ferrets identified viruses in clade 2.3.2.1 that are only four nucleotide substitutions away from becoming transmissible by the airborne route (7–9). Corresponding A(H5N1) viruses have been sampled in Nepal, Mongolia, Japan, and Korea and lack a glutamate-to-lysine change at position 627 of the viral polymerase basic 2 protein (PB2 E627K), one of the four prerequisite substitutions for airborne transmissibility, which has also been associated with increased viral replication and pathogenicity in mammals (10–15). In addition, viruses obtained from human fatal cases detected during the recent outbreak of A(H7N9) in China are characterized by PB2 E627K, while A(H7N9) viruses obtained from poultry are not (5, 16, 17). This mutation allows increased virus replication at the body temperature of mammals, thereby increasing virulence and the potential for transmission (18). An important question for the assessment of pandemic risk is whether this change occurs in the avian host or emerges during human infection.

Between February 2003 and May 2003, a large HPAI A(H7N7) outbreak occurred in the Netherlands. This outbreak struck 255 Dutch poultry farms in a 9-week period, resulting in the culling of nearly 30 million chickens. A total of 89 humans were infected, primarily persons involved in the direct handling of infected poultry, resulting in one fatal case of a veterinarian (19–21). Previously, two amino acid mutations had been identified in the A(H7N7) virus obtained from this fatal human case, relative to an avian virus obtained from the source farm (10). The first was a K416R mutation in the hemagglutinin (HA), which was shown to have no impact on replication kinetics in vitro. The second mutation was the well-characterized human adaptation marker PB2 E627K. Previously, investigating the presence of minority variants in clinical samples required either culturing of the clinical sample (disturbing the viral population composition) or using direct Sanger sequencing methods, which are hampered by a high 20% detection limit. With the rise of next-generation sequencing technologies, it is now possible to characterize the composition of the viral quasispecies in great detail. Therefore, we traced previously unanalyzed clinical samples from the fatal A(H7N7) human case as well as original A(H7N7)-positive poultry samples from the farm at which the veterinarian was infected and a control sample from an unrelated farm. Using Illumina Ultra-Deep sequencing, we were able to investigate the viral quasispecies dynamics associated with a fatal human A(H7N7) virus infection.

MATERIALS AND METHODS

Controls, plasmid DNA, and reverse genetics.

Plasmid DNA was isolated using Maxi prep column purification according to the manufacturer's protocol (Macherey-Nagel). The preparations of eight reverse genetic system plasmids containing the individual full-length segments of influenza A/WSN/33 (H1N1) virus PB2, PB1, PA, HA, NA, NP, MP, and NS in a pHW2000 backbone were obtained from R. G. Webster. Plasmid insert sequences were previously confirmed using standard Sanger sequencing (22). An 80% confluent monolayer of epithelial human embryonic kidney (HEK) 293T/17 cells (ATCC no. CRL-11268) in a T25 tissue culture flask was transfected with a mixture of all eight WSN33 reverse genetic plasmids (1 μg DNA each) and 20 μl Lipofectamine 2000 (Invitrogen) according to the manufacturer's protocol. Medium was replaced 16 h after transfection with Dulbecco's modified eagle medium (DMEM; Invitrogen) containing 1% fetal calf serum. Virus-containing supernatant was harvested 48 h after transfection, and cells and cell debris were removed by centrifugation (4 minutes at 400 relative centrifugation force [RCF]) and filtration through a 0.2-μm filter (FP 030/3; Schleicher & Schuell). For RNA extraction and consecutive full-genome amplification, 200 μl supernatant was added to 400 μl lysis-binding buffer (High Pure RNA isolation kit; Roche). The remaining supernatant was aliquoted and stored at −80°C.

Samples. (i) Chicken tracheal samples.

The source of the fatal human A(H7N7) virus infection was shown to be a chicken egg layer farm that housed around 15,000 chickens in battery cages (10). From 27 March 2003 onwards, the farmer reported an increasing number of dead chickens. On 2 April 2003, the poultry farm was visited by the veterinarian for screening purposes, and 2 days later, on 4 April 2003, all chickens from this farm were culled as screening samples tested positive for influenza A(H7N7) virus. Two pools of five influenza A(H7N7) virus-positive chicken tracheal samples were obtained from the source farm. One pool was obtained on 2 April, and the second pool was obtained on 4 April 2003. In addition, we analyzed a control pool of five A(H7N7) virus-positive chicken tracheal samples from an unrelated chicken egg layer farm, located 50 km away from the source farm, sampled on 31 March 2003. From each pool of five tracheal samples positive for A(H7N7) virus, 200 μl was used for RNA extraction.

(ii) Human samples.

Two days after his visit to the source farm, on 4 April 2003, the veterinarian developed influenza-like illness with high fever and severe headache. Throat and eye swabs were obtained on 10 April 2003 and tested negative for influenza virus by two separate laboratories. He was hospitalized on 11 April 2003, 9 days after his farm visit, with high fever and pneumonia with infiltrates visible on the chest X-ray in the lower right lung lobe. On 13 April 2003, the patient was transferred to the intensive care unit due to respiratory distress requiring mechanical ventilation. A bronchoalveolar lavage (BAL) was performed on the same day with concurrent sputum analysis. On 14 April 2003, kidney function decreased, requiring continuous veno-venous hemofiltration, while blood gas values failed to improve despite mechanical ventilation with 100% oxygen. On 17 April 2003, the patient died as a result of respiratory failure (23). After extensive laboratory testing for bacterial and viral pathogens, influenza A(H7N7) virus was detected in the BAL sample on 19 April. Virus culture of the BAL sample provided virus isolate A/Netherlands/219/03 on 23 April 2003 (21).

As part of this study, we obtained the original uncultured sputum sample as well as the original uncultured BAL sample from 13 April 2003 and used a total of 50 μl and 25 μl, respectively, for RNA extraction. In addition, RNA was extracted from a freshly cut section of formalin-fixed paraffin-embedded (FFPE) lung tissue obtained during autopsy, performed on 18 April 2003.

RNA extraction and viral load determination.

RNA extraction was performed using the High Pure RNA isolation kit (Roche) with an on-column DNase treatment according to the manufacturer's protocol. RNA was extracted from the FFPE lower lung tissue using the RNeasy FFPE kit (Qiagen) according to the manufacturer's protocol. Total RNA was eluted in a volume of 50 μl. All A(H7N7)-positive RNA extracts were subjected to a semiquantitative RT-PCR targeting the influenza virus matrix gene segment to prevent possible resequencing due to low-copy-number input in the reverse transcription (RT)-PCR (24).

Universal influenza virus full-genome amplification.

Viral RNA was RT-PCR amplified using a universal 8-segment PCR method as described previously (25). In short, two separate RT-PCRs were performed for each sample, using primers common-uni12R (5′-GCCGGAGCTCTGCAGATATCAGCRAAAGCAGG-3′), common-uni12G (5′-GCCGGAGCTCTGCAGATATCAGCGAAAGCAGG-3′), and common-uni13 (5′-CAGGAAACAGCTATGACAGTAGAAACAAGG-3′). The first RT-PCR mixture contained the primers common-uni12R and common-uni13. The second RT-PCR mixture contained the primers common-uni12G and common-uni13, which greatly improved the amplification of the PB2, PB1, and PA segments. Reactions were performed using the One-Step RT-PCR kit High Fidelity (Invitrogen) in a volume of 50 μl containing 5.0 μl eluted RNA with final concentrations of 1× SuperScript III One-Step RT-PCR buffer, 0.2 μM each primer, and 1.0 μl SuperScript III RT/Platinum Taq High Fidelity Enzyme Mix (Invitrogen). Thermal cycling conditions were as follows: reverse transcription at 42°C for 15 min, 55°C for 15 min, and 60°C for 5 min; initial denaturation/enzyme activation of 94°C for 2 min; 5 cycles of 94°C for 30 s, 45°C for 30 s, slow ramp (0.5°C/s) to 68°C, and 68°C for 3 min; 30 cycles of 94°C for 30 s, 57°C for 30 s, and 68°C for 3 min; and a final extension of 68°C for 5 min. After the PCR, equal volumes of the two reaction mixtures were combined to produce a well-distributed mixture of all 8 influenza virus segments. All RT-PCRs were performed in duplicate.

Influenza A(H7N7) virus HA and PB2 fragment amplification.

Although the influenza virus matrix RT-PCR demonstrated that the viral load of the human samples was higher than that of the poultry samples, the universal influenza virus full-genome amplification of human samples was unsuccessful. The largest RT-PCR products detectable after the full-genome amplification of the human samples had a length of ∼1 kb, illustrating that the amplification of larger influenza virus gene segments had failed due to RNA fragmentation. As mutation PB2 E627K is located on the largest of eight influenza virus gene segments with a length of >2,000 nucleotides, an amplification strategy targeting smaller fragments was needed to allow successful characterization of minority variants. Therefore, specific amplification of a 410-nucleotide PB2 fragment covering codon 627 and a 362-nucleotide HA fragment covering codon 416 was applied in duplicate. A two-step protocol with SuperScript III and 1.0 μM specific primers followed by amplification with HotStarTaq Master Mix (Qiagen) and 0.5 μM each primer was performed (26). PB2 primers H7-PB2F (5′-GAGCCCTTTCAATCCTTGGT-3′) and H7-PB2R (5′-CTTGCCCTATCAATACATTAGCCT-3′) and HA primers H7-HAF (5′-ATGTCCGAGATATGTTAA-3′) and H7-HAR (5′-GATTCTCCATTGCTACTA-3′) were used. Thermal cycling conditions were as follows: reverse transcription at 50°C for 1 h and denaturation at 95°C for 5 min; enzyme activation of 95°C for 15 min; 45 cycles of 94°C for 30 s, 50°C for 30 s, and 72°C for 1 min; and final extension of 72°C for 10 min.

Illumina sequencing. (i) Genome Analyzer IIx (GAIIx).

Products from the full-genome influenza virus RT-PCR of the pooled chicken samples were diluted to a DNA concentration of 50 ng/μl and sheared to a length of 200 to 400 bp using a Covaris AFA (Covaris, Woburn, MA). The sheared fragments were end repaired, A-tailed, and ligated to Illumina sequencing adaptors containing molecular identifier (MID) tags using the NEBNext DNA Library Prep reagent set (New England BioLabs, United Kingdom) to allow for multiplex sequencing per lane. Cluster generation was done using the TruSeq PE cluster generation kit version 4 (Illumina) according to the manufacturer's instructions. The three samples were sequenced in the same run on an Illumina GAII sequencer with a paired-end 54-bp run using the Sequencing by Synthesis kit version 5 according to the manufacturer's instructions (Illumina). All sample processing from shearing to sequencing was performed by the Wellcome Trust Sanger Institute (Hinxton, Cambridge, United Kingdom) as part of the European FP7 framework Emperie.

(ii) HiSeq 2000.

Products from the HA and PB2 RT-PCR and plasmid pool RT-PCR (PCR control) as well as a plasmid pool sample (no-PCR control) were diluted to a DNA concentration of 50 ng/μl and sheared by nebulization. Samples were subjected to end repair, A-overhang, and adaptor ligation with MID tags using the Illumina TruSeq DNA sample preparation kit. The libraries were multiplexed, clustered, and sequenced on an Illumina HiSeq 2000 (TruSeq v3 chemistry) with a paired-end 100-cycle sequencing protocol with indexing. All sample processing from shearing to sequencing was performed by BaseClear BV (Leiden, The Netherlands).

Data analysis. (i) Demultiplexing.

The raw fastq files containing the sequences with corresponding quality scores were split into separate sample-specific files based on the corresponding MID sequence using the readset parser function of the Quality Assessment of Short Reads (QUASR) package version 7.0.1 (25).

(ii) Quality control.

To correct for the inaccurate calling of nucleotides by the sequencer, an error correction step was performed on all MID-split data files using the paired-end quality control function of the QUASR pipeline version 7.0.1. In short, the median phred quality score of each read was calculated, and if this value was below a (user-defined) threshold, the read was trimmed from the 3′ end until either the median quality cutoff requirement was met or the read length reached a user-defined minimal length cutoff and was discarded. In paired-end mode, the quality control step requires both mates to pass quality control. In this study, for the three chicken samples that were sequenced on an Illumina GAIIx, a median read phred quality score cutoff of 30 (corresponding to a 99.9% base call accuracy) was used in the quality control process and a length cutoff of 50. For all HiSeq 2000 data sets, a median read phred quality score cutoff of 33 was used in the quality control process and a length cutoff of 90. This difference in phred cutoff scores was determined empirically and was necessary due to inherent differences in the overall quality produced by the two different types of Illumina sequencers. Full details concerning error correction of deep sequencing data will be published elsewhere (M. R. A. Welkers, M. Jonges, R. E. Jeeninga, M. Koopmans, M. D. de Jong, unpublished data).

Primer removal.

The fastq files containing the reads that passed quality control were parsed to remove any sequences with exact matches to any of the primers used in the amplification or sequencing process in both the 5′and 3′ orientation using in-house-generated PHP scripts.

Mapping strategy.

All MID-split files containing quality-controlled primer-removed reads of the plasmid controls were mapped to the WSN reference sequence (NCBI taxonomy ID 382835) using the Burrows-Wheeler Aligner (BWA) version 0.5.9-r16 with the paired-end mapping option (27). All data files containing quality-controlled reads from the human clinical samples as well as the pooled chicken samples were mapped to the consensus sequence of virus isolate A/Nederland/219/03, which was isolated previously from the BAL specimen of the fatal case (NCBI taxonomy ID 680693). A quality-controlled consensus sequence was generated for each of the chicken samples using only nucleotides with a phred score above 30. This consensus sequence was subsequently used to remap all reads. Due to the internal concatenation of reference sequences by BWA, the resulting sequence alignment map (SAM) files were checked for mapping inconsistencies by removing reads with a mismatch rate of >10% for further analysis. In addition, mapped reads with either hard or soft clipping were removed for further analysis. As we mapped to a near-homologous consensus sequence, we removed reads from the SAM file with excessive mutations that might overestimate the viral diversity and limited the per-read mutation frequency to 2% by analyzing the NM tag (genetic distance from consensus sequence) from the SAM file (28). Reads remaining after the removal of high mutation reads formed the cleaned (RHM) data sets. To characterize the type of influenza virus-specific sequence reads that were removed from the RAW data set to obtain the RHM data set, all quality-controlled but unmapped and clipped reads were split in half, and both parts were remapped against their sample-specific consensus sequence. Defective influenza virus RNA was defined when the difference in starting mapping positions of both read parts was greater than the length of the read (29).

Generation of coverage overview and quality-based consensus sequence.

Both the RAW and RHM SAM files of each sample were analyzed to generate a coverage overview per reference position disregarding nucleotides with a phred score below 30. Indel correction was done by analyzing the read CIGAR string (28), removing insertions and adding gaps to sites of deletions.

Determining SSEs.

Sequence-specific errors (SSEs) are associated with a discrepancy in the mismatch frequency of mutations observed in mapped reads in forward and reverse orientation. To investigate the impact of SSEs, we split the SAM files in reads mapped in the forward orientation as well as the reverse orientation based on the bitwise flag value (28). From the two resulting SAM files, a coverage overview was generated, and the mismatch frequency per position (i.e., the frequency of nucleotides not being the consensus nucleotide) was determined. To determine whether a nucleotide variation was an SSE, the difference in mismatch variation between the two orientations had to exceed the mean mismatch variation.

Mapping position of mutants in sequence reads.

In case a particular mutation is due to a technical error, the mutation would be primarily located on a single specific location on all sequenced reads, as it originates from a single source read. A true variant is more evenly distributed over the entire length of the sequence reads, as it originates from more than 1 source read. To characterize the distribution of a particular mutation on the read, SAM files were parsed to extract all reads with a mutation on a predetermined position within the viral genome. The location of the mutation on the read was then determined for all individual reads. Mutational position graphs were generated and visually inspected for presence of a “hairy caterpillar”-like shape that indicates efficient shearing, unbiased amplification, and elimination of variation associated with the terminus of PCR products. When a mutational position graph was characterized by either one peak of >50% or more peaks of >20%, the corresponding minority variant was marked as an artifact and discarded. For the plasmid data set, the 200 nucleotide positions with the highest MMF (range, 0.078 to 0.38%) were analyzed, while all positions with a MMF of >0.5% were analyzed for the RT-PCR and A(H7N7) data sets.

Confirmation of influenza A(H7N7) virus defective RNA.

To confirm the presence of defective influenza virus RNA in A(H7N7)-positive material obtained from the 2003 outbreak in the Netherlands, 32 influenza A(H7N7) virus-positive poultry samples and 6 human samples were analyzed (26, 30). Using a two-step protocol with SuperScriptTM III followed by amplification with HotStarTaq Master Mix and 0.5 μM H7-PB2ncrF (5′-AGCAAAAGCAGGTCAAATATATTC-3′) and H7-PB2ncrR (5′-GGTCGTTTTTAAACAATTC-3′), the presence of PB2-defective RNA was examined in both human and poultry samples and Sanger sequenced with an ABI 3730 sequencer (see Fig. S1 in the supplemental material). In addition, this RT-PCR was performed on the BAL fluid, sputum, and lung tissue samples from the fatal case.

RESULTS

The entire process from influenza A(H7N7) virus-positive samples to deep sequencing data consists of many steps in which a possible bias in the detection and quantification of minority variants can be introduced. Therefore, we first determined the technical error rates induced by the sequencing platform by directly sequencing a plasmid pool, as well as after the required RT-PCR, and optimized the data analysis in favor of sensitive detection of true virus minority variants.

Quantification of technical errors.

A pool of 8 WSN reverse genetics plasmid DNA preparations were sequenced directly to determine the Illumina HiSeq2000-associated technical error. The average coverage per nucleotide position in the WSN genome was 66.100 and displayed a mean mismatch frequency (MMF) of 0.079% per reference position. The duplicate RT-PCR data sets (the average WSN genome coverage was 43.300) displayed a 3-fold increase of the mean MMF compared with the directly sequenced WSN plasmids and were characterized by the presence of nucleotide positions that reached 20% MMF. Although random induction of RT-PCR errors was expected, correlating (R² = 0.89) nucleotide positions in both RT-PCR data sets were characterized by a high MMF (Fig. 1A).

FIG 1.

Steps of the data analysis pipeline and their effect on the influenza virus data set. (A) From the three control data sets, the mismatch frequency is expressed for each influenza virus nucleotide. This shows that the enhanced mapping algorithm was insufficient in removing the variation observed at specific nucleotides of the duplicate RT-PCR sets. (B) Focusing on 1% of the nucleotides that displayed the highest mismatch frequency, the sequence-specific errors are expressed by error bars. While plasmid nucleotides with the highest mismatch frequency were associated with sequence-specific errors, RT-PCR nucleotides with the highest mismatch frequency were not. (C) During the final step of the data analysis pipeline, all nucleotide positions displaying variation were assessed individually by mutational position graphs that demonstrate the location within Illumina sequence reads that code for variation. A mutational position graph of variation associated with the 5′-terminus of a PCR product is presented (top) together with a “hairy caterpillar”-shaped plot showing variation regardless of the position in the sequence fragment, thus considered a true variant (bottom). (D) Effect of each type of data filtering on the mismatch frequency expressed as mean (+ range) for the plasmid, RT-PCR 1, and RT-PCR 2 data sets. While the enhanced mapping algorithm (RHM) and removal of sequence-specific errors (SSE) successfully reduced the mismatch frequency of nucleotides in the plasmid data set, mutational position graphs (MPG) were required to reduce the maximum mismatch frequency to less than 2% of the RT-PCR data sets.

Characterization and elimination of artifacts.

To identify the source of the high MMF in the RT-PCR data sets, we investigated the impact of unspecific primer binding and observed strong increases of the MMF at WSN genome locations that had up to 85% homology with the uni13-primer, indicative for nonspecific binding. To improve the overall sequence quality, clipped sequence reads, reads with indels, and reads that deviated more than 2% from the sample-specific consensus sequence were removed (RHM data set; Fig. 1A). Next, we examined the occurrence of sequence-specific errors for all nucleotide positions that displayed variation (Fig. 1B). For the RHM plasmid data set, all nucleotide positions with an MMF of >0.16% were characterized by SSE. For the RHM RT-PCR data sets, however, removal of SSE proved insufficient in eliminating the RT-PCR-induced variation. As primer sequences are associated with the 5′ and 3′ termini of PCR products, mutational position graphs were lastly applied to assess the distribution of sequence variation in the sequence reads (Fig. 1C). The optimized data analysis pipeline that allowed detection of minority variants present at ≥2% in RT-PCR products was subsequently applied on the deep sequence data sets of A(H7N7)-positive specimens (Fig. 1D). The full details of the analysis pipeline are described elsewhere (Welkers et al., unpublished).

Characterizing the avian source of infection.

In a previous study, the consensus sequence of A/Netherlands/219/03 obtained from the fatal case contained two additional amino acid substitutions, PB2 E627K and HA K416R, compared with the consensus sequence of A/chicken/Netherlands/03010132/03 (H7N7) obtained from the source farm of the fatal human infection (10). Deep sequence analysis of A(H7N7)-positive chicken specimens obtained from the source farm at the day of the visit of the fatal case and the day of culling, with acceptable threshold cycle (C_T) values of 30.5 and 28.3 in the matrix RT-PCR, failed to detect the presence of human adaptation marker PB2 E627K as well as the HA K416R change (Fig. 2).

FIG 2.

Timeline integrating the data on poultry and the fatal case and the corresponding virological data. It shows the increase of PB2 E627K during human infection in the absence of PB2 E627K detected at the avian source of infection. A quarter of the influenza A(H7N7) virus population in the lower respiratory tract of the veterinarian contained avian PB2 627E virus 11 days postexposure, while the postmortem lower lung tissue contained fully adapted (PB2 E627K) virus 16 days postexposure, suggesting that the PB2 E627K mutation emerged during infection of the human host.

For each of the three poultry deep sequence data sets, variation from the consensus sequence was observed (Table 1). While the full-genome A(H7N7) consensus sequences of the two samples obtained from the source farm were identical except for one synonymous mutation at codon 72 in the Matrix gene segment, the A(H7N7) deep sequence data displayed variation at 20 nucleotide positions in the April 2 sample and 5 positions in the April 4 sample. Of the observed variation, only the nucleotide variation within codon 72 in the Matrix gene was present in A(H7N7) viruses obtained from both samples. The A(H7N7) viruses obtained from the control farm, with an acceptable C_T value of 28.2 in the matrix RT-PCR, displayed 7 positions with variation. Human adaptation markers including PB2 E627K as well as HA K416R substitutions were absent in the A(H7N7) viruses obtained from both the source and control farms.

TABLE 1.

Summary of within-farm influenza virus sequence diversity

Farm and collection date	No. of synonymous substitutions	No. of nonsynonymous substitutions	List and percentage of nonsynonymous substitutions
Source farm, 2 April 2003	8	12	3% HA I335V, 3% HA M429I, 2% HA V441A, 2% HA I481T, 2% HA Y498H, 2% NA T137I, 3% NP R446I, 2% NP E454G, 2% PA D294Y, 3% PB1 P698S, 2% PB1 Q210P, 5% PB2 M315V
Source farm, 4 April 2003	4	1	34% HA S102N
Control farm, 31 March 2003	3	4	10% HA I211T, 9% NP I116V, 3% PA E181D, 27% PB2 I90 M

In addition to the screening for human adaptation markers, sequence diversity can be used to characterize transmission chains. Conceivably, the observed deep sequence variation might harbor information that could enhance the resolution of transmission markers obtained by Sanger sequence data. The previously reported transmission network of the A(H7N7) outbreak demonstrated that our source farm was part of a transmission chain, characterized by the accumulation of mutations in HA, NA, and PB2, that ultimately led to introduction of the A(H7N7) virus in Belgium (26, 30, 31). None of the mutations that characterized the A(H7N7)-positive poultry farms in the transmission chain before and after our source farm could be detected as minority variants using deep sequencing. In contrast, our control farm demonstrated minority variant PB2 I90M (27%), which characterized the transmission chain prior to infection of our control farm, providing a direct virological relation between viruses sampled at the control farm and its suggested source of infection (Table 1) (26).

Characterizing the human A(H7N7) virus infection.

Three A(H7N7) virus-positive specimens were obtained from the fatal case: BAL fluid and sputum obtained 1 day after hospitalization (11 days postexposure) and postmortem lung tissue (16 days postexposure). The samples demonstrated acceptable C_T values of 29.8, 25.1, and 27.5 in the matrix RT-PCR for the BAL fluid, sputum, and lung tissue, respectively. Since the influenza virus RNA in these specimens was too degraded for successful amplification of the larger influenza virus gene segments, specific RT-PCRs targeting small PB2 and HA fragments were performed in duplicate to allow characterization by deep sequence analysis. After quality control and data filtering, the only subpopulation that was detectable in the duplicate PB2 and HA fragments was located within PB2 codon 627, verified by the “hairy caterpillar”-shaped mutational position graph (Fig. 1C, bottom). The presence of the avian PB2 627E was detected at 30.0% and 20.5% in the duplicate BAL samples and at 7.3% and 6.3% in the duplicate sputum samples obtained in parallel. In the postmortem lung tissue obtained 5 days later, PB2 627E was not detectable. This demonstrated that the A(H7N7) virus infection was characterized by a mean 75% and 93% PB2 E627K mutation in the BAL and sputum samples obtained 11 days after exposure, followed by a dominant (100%) PB2 E627K mutation in the sample obtained during autopsy 5 days later (Fig. 2). The sequence variation observed at HA codon 416 was marked as artificial in the data analysis pipeline, based on the absence of a “hairy caterpillar”-shaped mutational position graph. These plots illustrated that >50% of the observed variation at HA codon 416 was associated with position 11 in sequence reads (Fig. 1C, top). Overrepresentation of one specific DNA template followed by resequencing was found to be the cause of this phenomenon. After correcting for the resequencing bias, we were able to demonstrate that the HA mutation K416R was the dominant population (∼80%) at both sample days.

Identification of A(H7N7) virus defective RNA.

In the data analysis pipeline, on average, 9% of the deep sequence data were removed from the RAW data sets to obtain the RHM data sets. To investigate the type of influenza virus-specific information that was lost by the removal of high mutation reads, these sequence reads were split and remapped against their sample-specific consensus sequence. Influenza A(H7N7) virus-specific sequence reads that demonstrated the presence of internal deletions for all eight influenza gene segments were encountered in the three A(H7N7) chicken samples, consistent with the detection of defective RNA (Fig. 3). The majority of the defective RNAs (80.0%) were derived from the gene segments that form the polymerase complex. Although internal deletions were observed with various sizes, the PB1 and PB2 subgenomic RNA (sgRNA) fragments were characterized by an ∼1,900-nucleotide internal deletion and PA by an ∼1,800-nucleotide deletion. In addition, 6% of the detected influenza virus defective RNAs comprised two different influenza virus gene segments. To independently confirm the presence of defective RNA in A(H7N7)-positive samples, additional A(H7N7)-positive samples obtained from 32 poultry farms and six human conjunctivitis cases during the 2003 outbreak in the Netherlands were analyzed by conventional RT-PCR and Sanger sequencing. This illustrated the presence of influenza virus PB2-defective RNA in 17/32 (53%) A(H7N7)-positive poultry samples and in 3/6 (n = 50%) A(H7N7)-positive human conjunctivitis samples (see Fig. S1 in the supplemental material). No PB2-defective RNA was detected in the samples obtained from the fatal case.

FIG 3.

Detection of influenza virus defective RNA by deep sequencing. Characterization of the start and endpoint of internal deletions in A(H7N7) virus gene segments by deep sequence analysis identified a multitude of influenza virus defective RNAs in chicken samples obtained from the source and control farm. Of the total reads that contained an internal deletion, the pie charts illustrate the contribution for each of the influenza A(H7N7) virus gene segments.

DISCUSSION

In this study, next-generation sequencing was applied to investigate the emergence of human adaptation marker PB2 E627K and HA K416R during a fatal case of HPAI A(H7N7) virus infection. Two uncultured samples from the avian source of infection and three uncultured samples obtained from the human fatal case were RT-PCR amplified prior to ultradeep sequencing using the Illumina platform. This platform was chosen for its ability to generate deep sequence data with the lowest reported error rates of all currently available platforms, allowing for ultradeep sequencing of our specimens with high resolution. For usage in next-generation sequencing, a high input DNA quantity is needed, requiring RT-PCR amplification of the influenza viral RNA. Our results illustrate the paradox in using RT-PCR for generating sufficient sequencing depth, as RT-PCR was associated with artifacts that hamper minority variant calling. By extended analysis of the data sets obtained from plasmid DNA and RT-PCR products, we developed a universal approach for distinguishing variation of viral RNA from artifacts and sequence bias that was subsequently applied on A(H7N7)-positive human and chicken samples.

The Illumina HiSeq 2000 platform had a maximum error below 0.2% for all nucleotides, which rose to a maximum of 20% after applying the RT-PCR. The “hairy caterpillar” plot analysis reduced the maximum error to less than 2% for all nucleotides, permitting characterization of viral diversity when present above 2%. Subsequent analysis of A(H7N7)-positive poultry samples illustrated the presence of a dominant influenza virus genotype for each farm supplemented with a multitude of low-frequency sequence variants that are directly linked to the dominant variant, in agreement with conventional cloning results (30). Interestingly, one minority variant (PB2 I90M) correlated with mutations from a previously characterized farm-to-farm transmission chain (26), suggesting that ultradeep sequence can be applied for more robust determination of a transmission network than that obtained with consensus sequencing approaches. In addition, the poultry samples were characterized by the presence of defective influenza virus RNA, as reported previously (32). The formation of virus-derived defective RNA requires a high multiplicity of infection and has been reported under laboratory conditions (33, 34). Similarly, HPAI virus infection of poultry is associated with extremely high titers that might facilitate the emergence of defective influenza virus RNA. The presence of defective RNA corresponding with all eight influenza virus gene segments in the three A(H7N7) poultry samples illustrated the variety of defective RNA segments generated during HPAI virus infection of a chicken farm. It has been illustrated that defective influenza virus RNAs compete with full-length viral gene segments at the step of virion assembly and thereby inhibit the production of infectious progeny virus (35, 36). Subsequent virions that carry such an incomplete genome are known as defective interfering (DI) viruses. It has been demonstrated that DI influenza viruses in vivo provide protection against homologous and heterologous influenza virus infection in laboratory animals (37–39). The presence of DIs in HPAI-infected poultry could attenuate the virulence for exposed humans, possibly aided by enhanced interferon (IFN) host response (40, 41). Additional experiments are needed to confirm a potential inhibitory role of DI influenza viruses during human HPAI virus infection.

For the samples obtained from the fatal human A(H7N7) case, mean frequencies of PB2 E627K ranged from 75% in the BAL fluid to 93% in the sputum and 100% in the postmortem sample (Fig. 2). The HA K416R mutation was present with a frequency of ∼80% at both sample days and is in agreement with the previously determined consensus sequence A/Netherlands/219/03 (21). Additional minority variants could be present, but failure of the universal influenza virus genome RT-PCR did not allow a full-genome analysis. The highest percentage of PB2 E627K was found in the sputum sample, containing mucus and spit from the carina of the lungs and up, while the BAL sample is restricted to the lower part of the airways. The difference in PB2 E627K frequency between BAL fluid and sputum might be due to the preferential replication temperature difference between the human-adapted and avian type PB2, allowing increased virus replication of the PB2 E627K variant at lower temperatures (11, 18). Modeling of the emergence and positive selection of human adaptation mutations, including PB2 E627K, during human influenza A(H5N1) virus infection demonstrated the plausibility of such a variant to become dominant during infection of a single human host, consistent with our in vivo data (8). The absence of PB2 E627K variants in poultry samples, combined with the detection of initial mixed PB2 627E and E627K variants during human infection, strongly suggests that the PB2 E627K mutation emerged during infection of the human host and correlated with disease severity.

Herfst et al. demonstrated that PB2 E627K was a prerequisite for the development of airborne A(H5N1) virus in ferrets (7). The accumulation of other human adaptation markers than PB2 E627K observed in avian influenza viruses from poultry and the wild bird population suggests that the introduction of this particular mutation in avian influenza viruses dramatically increases the virus pandemic potential and public health risk (8, 26). Moreover, the fatal human cases observed during the recent LPAI A(H7N9) outbreak in China are characterized by a virus carrying the PB2 E627K substitution, while viruses obtained from poultry are not (5, 16, 17). Here, we illustrate that the conversion of an avian influenza virus toward a more human-tropic and potentially transmissible variant may occur during one infection cycle in humans, emphasizing the importance of reducing human exposure to avian influenza viruses on a global scale.