The samples used in this study were collected from health facilities within the Kilifi Health and Demographic Surveillance System (KHDSS) on the coast of Kenya.22 Non‐residents and residents of KHDSS presenting to these facilities were included. First, 5304 nasopharyngeal/oropharyngeal (NP/OP) swabs were taken from childhood admissions under the age of 5 years with syndromic severe or very severe pneumonia23 collected as part of continuous viral pneumonia surveillance at the Kilifi County Referral Hospital (KCH) from January 2009 through December 2016.23 Second, 6254 NP swab samples were taken from outpatients of all ages presenting with acute respiratory illness to selected nine outpatient health facilities spread throughout the KHDSS between December 2015 and March 2017. A comprehensive description of the study area and respiratory disease surveillance at KCH and surrounding outpatient health facilities is available from previous reports.22, 23, 24, 25 Samples were stored in viral transport medium (VTM) at −80°C prior to molecular screening and subsequent processing.24, 26, 27

Samples were screened for a range of respiratory viruses, including IAV, using a multiplex (MPX) reverse transcription (RT)‐PCR assay employing Qiagen QuantiFast multiplex RT‐PCR kit (Qiagen),26 with epidemiological surveillance results for the period 2007‐12 previously reported.24 A real‐time PCR cycle threshold (Ct) of <35.0 was used to define virus‐positive samples.26

Viral nucleic acid extraction from IAV positive samples (Ct < 35.0) was performed using the QIAamp Viral RNA Mini Kit (Qiagen). Ribonucleic acid (RNA) was reverse transcribed, and the entire genome of influenza was amplified in a single multi‐segment real‐time polymerase chain reaction (M‐RTPCR) using the Uni/Inf primer set.28 The amplification was performed in 25 μL reactions containing 8 μL nuclease‐free water, 12.5 μL 2X RT‐PCR buffer, 0.2 μL Uni12/Inf1 (10 μmol/L), 0.3 μL Uni12/Inf3 (10 μmol/L), 0.5 μL Uni13/Inf1 (10 μmol/L), 0.5 μL SuperScript III One‐Step RT‐PCR with Platinum Taq High Fidelity (Invitrogen), and 3 μL extracted RNA. Thermocycling conditions were as follows: 42°C for 50 minutes, 50°C for 10 minutes, 94°C for 2 minutes; four cycles (94°C for 30 seconds, 43°C for 30 seconds and 68°C for 3 minutes and 50 seconds) followed by 30 cycles of 94°C for 30 seconds, 57°C for 30 seconds, and 68°C for 3 minutes and 30 seconds (with the 3 minutes and 30 seconds for the 68°C extension step increased by 10 seconds per subsequent cycle after cycle 1), and a final extension step at 68°C for 10 minutes. Successful amplification was evaluated by running the products on a 2% Agarose gel and visualized on a UV transilluminator after staining with RedSafe Nucleic Acid Staining solution (iNtRON Biotechnology Inc).

Following PCR, the amplicons were purified with 1X AMPure XP beads (Beckman Coulter Inc), quantified with Quant‐iT dsDNA High Sensitivity Assay (Invitrogen), and normalized to 0.2 ng/μL. Indexed paired‐end libraries were generated from 2.5 μL of 0.2 ng/μL amplicon pool using Nextera XT Sample Preparation Kit (Illumina) following the manufacturer's protocol. Amplified libraries were purified using 0.8X AMPure XP beads, quantitated using Quant‐iT dsDNA High Sensitivity Assay (Invitrogen), and evaluated for fragment size in the Agilent 2100 BioAnalyzer System using the Agilent High Sensitivity DNA Kit (Agilent Technologies). Libraries were then diluted to 2 nmol/L in preparation for pooling and denaturation for running on the Illumina MiSeq (Illumina). Pooled libraries were NaOH denatured, diluted to 12.5 pmol/L and sequenced on the Illumina MiSeq using 2 × 250 bp paired‐end reads with the MiSeq v2 500 cycle kit (Illumina). Five percent Phi‐X (Illumina) spike‐in was added to the libraries to increase library diversity by creating a more diverse set of library clusters. Low diversity libraries are common for amplicon pools and occur when a significant number of reads have similar sequences, for example, in amplicon pools.

Contiguous nucleotide sequence (contigs) assembly was carried out using the FLU module of the Iterative Refinement Meta‐Assembler (IRMA),29 which performs iterative segment‐level read sorting based on Lineage Assignment By Extended Learning (LABEL),30 and iteratively refines the references to optimize the final assembly based on the Striped Smith‐Waterman (SSW) algorithm.31 IRMA quality control, variant calling and phasing, and assembly pipelines were all implemented in the EDGE Bioinformatics environment32 using IRMA default settings. All the sequence data were deposited in the Virus Epidemiology and Control (VEC) research group's Data Repository in Harvard Dataverse under the doi https://dataverse.harvard.edu/ and the Global Initiative on Sharing All Influenza Data (GISAID) EpiFlu^™ database (https://www.gisaid.org/ ) under the accessions EPI_ISL_393682, EPI_ISL_393684‐393703, EPI_ISL_393705‐393709, EPI_ISL_393711‐393723, EPI_ISL_393725‐393753, EPI_ISL_393936‐393946, EPI_ISL_393949, EPI_ISL_393951‐393953, EPI_ISL_393955‐393956, EPI_ISL_393960, EPI_ISL_393963, EPI_ISL_393965‐393969, EPI_ISL_394051‐394052, EPI_ISL_394107‐394112.

Consensus nucleotide sequences of HA gene were aligned and translated into amino acids using MUSCLE program implemented in mega v7.0.26.33 The sequences of reference strains of known clades and Northern Hemisphere vaccine strains recommended by WHO included in the phylogenetic analysis were obtained from the GISAID EpiFlu^™ database (https://www.gisaid.org/). Phylogenetic trees were constructed with maximum likelihood (ML) and bootstrap analysis of 1000 replicates using the GTR + G model implemented in mega 7.0.26. The full‐length HA sequences were used to characterize A(H3N2) strains into genetic clades and subclades according to ECDC guidelines.34, 35

Potential N‐linked glycosylation sites on the HA were determined using the NetNGlyc 1.0 server (http://www.cbs.dtu.dk/services/NetNGlyc/ ), in which a threshold of >0.5 suggests glycosylation.

Ethical clearance for the study was granted by the KEMRI‐Scientific and Ethical Review Unit (SERU# 3103) and the University of Warwick Biomedical and Scientific Research Ethics Committee (BSREC# REGO‐2015‐6102). Informed consent was sought and received from the study participants for the study.

The specimen processing flow for IAV NGS is shown in Figure 1. A total of 124/5304 (2.3%) and 168/6254 (2.7%) IAV positive specimens were identified from the inpatient and outpatient facilities, respectively. Of the 292 IAV positive specimens identified for this study, 186 (63.7%) were retrieved for this study; the remainder were limited in quantity for RNA extraction, that is, <140 μL. Of these 186 specimens, 152 (81.7%) that passed pre‐sequencing quality control checks were loaded onto the Illumina MiSeq: 80 (52.6%) from the inpatient and 72 (47.4%) from the outpatient studies, respectively, with corresponding success in generating whole‐genome sequences (WGS) for 71 (88.8%) and 71 (98.6%) specimens. A total of 101 (71.1%) A(H3N2) and 41 (28.9) A(H1N1)pdm09 WGS were generated. For this report, 101 determined A(H3N2) virus full‐length HA sequences were used: 35 from the KCH inpatient and 66 from the KHDSS outpatient studies, respectively. The sociodemographic characteristics of these patients are shown in Table 1.

We investigated the HA gene sequences to characterize the amino acid variations affecting antigenic epitopes, receptor‐binding sites, and potential glycosylation. Full‐length segment coverage of HA was obtained for all the A(H3N2) strains (35 KCH inpatient and 66 KHDSS outpatient) using NGS and sufficient sequence data was available to characterize them into genetic clades and subclades. Among the virus strains identified in hospital inpatient specimens (Figure 2), all the 35 A(H3N2) strains from 2009‐15 diverged from the vaccine strains A/Perth/16/2009 (H3N2)‐like virus, A/Texas/50/2012 (H3N2)‐like virus, and A/Switzerland/9715293/2013 (H3N2)‐like virus. The A(H3N2) strains from 2009 to 2012 (23/35) diverged from the 2009‐12 vaccine strain A/Perth/16/2009 (H3N2)‐like virus and fell into clades A/Victoria/208/2009‐like clade (3/24) and clades 7 (17/24), 4 (1/24), and 3B (2/24), respectively. A virus from 2014 diverged further and fell into the 3B clade (1/35). One virus from 2014 belonging to the 3C clade diverged from the 2014‐15 vaccine strain A/Texas/50/2012 (H3N2)‐like virus (1/35). A virus from 2014 (1/35) and all the 2015 viruses (9/35) belonged to the 3C.2a clade and diverged from the 2015‐16 vaccine strain A/Switzerland/9715293/2013 (H3N2)‐like virus. Among the virus strains identified in the outpatient specimens, the A(H3N2) viruses belonged to genetic clade 3C.2A(34/66), subclades 3C.2a2 (3/66) and 3C.2a3 (6/66), and subgroup 3C.2a1b (23/66) all of which significantly diverged from the 2016‐17 vaccine strain A/Hong Kong/4801/2014 (H3N2)‐like virus (Figure 3).

The high numbers of A(H3N2) viruses in the outpatient setting presented an additional opportunity to investigate the diversity of the virus in a rural community across multiple epidemic seasons. Viruses identified from the outpatient setting were sampled continuously between December 2015 and March 2017. Comparison of the deduced amino acid sequences of all the viruses identified in outpatient specimens to that of the vaccine strain A/Hong Kong/4801/2014 (H3N2)‐like virus, from which all the 2015‐17 strains significantly diverged (Figure 3), showed that the 3C.2a strains (n = 34) differed from the vaccine strain by amino acid variations D53N + R142K + Q197R (HA1). All these strains had additional variations S96N + P194L (HA1). Some 3C.2a strains also possessed additional variations N31S (HA1; n = 10) and N121D (HA1; n = 4). The 3C.2a1b (n = 23) strains were defined by amino acid variations K92R + N121K + N171K + H311Q and I77V + G155E (HA2). Some strains also possessed an additional amino acid variation S144G (HA1; n = 7). The 3C.2a3 strains (n = 6) were defined by amino acid variations N121K + S144K (HA1), whereas the 3C.2a2 strains were defined by N121K + R261Q (HA1) variations. All the 3C.2a, 3C.2a3, 3C.2a2, and 3C.2a1b amino acid substitutions were part of the overall evolution of the strains.

Importantly, there was gain of an N‐linked glycosylation site in antigenic site B of HA due to a K160T amino acid substitution in all but three of the strains identified in the outpatient specimens (63/66; Figure 4). This substitution, which emerged in the 2013‐14 Northern Hemisphere influenza season among clade 3C.2a viruses, results in the glycosylation of residue 158 in antigenic site B of HA that is near the receptor‐binding site and alters the antigenic properties of HA resulting in reduced vaccine effectiveness. This substitution was also observed in all but one of the strains identified in inpatient specimens (8/9) in 2015.

The receptor‐binding site of A(H3N2) virus is highly conserved at amino acids positions 98, 136, 153, 183, 190, 194, 195, and 228 on HA1. However, we did not observe any substitution in these conserved positions in the Kilifi strains.