Publicly available deep sequencing datasets associated with influenza vaccine trials and heavy chain sequences of healthy donors were downloaded using accession numbers listed in supplemental Tables S1A, B . The Stand‐alone IgBLAST (57) was used for V(D)J germline gene assignments, and IgBLAST output was parsed and analyzed by in‐house developed python script. For 454 sequencing, non‐immunoglobin reads and non‐productive reads were removed. For Illumina sequencing, non‐Ig reads, non‐duplicate and non‐productive reads were filtered out. The remaining reads were sieved by the HV6-1 class signatures previously published (9) or developed in this paper. The CDR-H3 of HV6-1 class signature positive reads were clustered with 97 percent sequence identity using CD-HIT (58). The centroid sequence was selected as representative sequence of each cluster for neutralization assay. The germline amino acids were used to repair missing N- and C- terminal residues.

24 h prior to DNA-transient transfection, 100 μl per well of log-phase growing HEK 293T cells were seeded into a 96-well microplate at a density of 2.5x10⁵ cells/ml in optimized expression medium (RealFect-Medium, ABI Scientific, VA), and incubated at 37°C, 5% CO₂ for 24 hours. Prior to transfection, 40 μl per well of spent medium was removed. For transient transfection, 0.15 ug of each heavy chain variant plasmid DNA was paired with 0.15 ug of light chain plasmid DNA of 56.a.09, 54.f.01, or MEDI8852, respectively, in 10 μl of Opti-MEM medium (Invitrogen, CA) per well in a 96-well plate, and then mixed with 0.9 μl per well of TrueFect-Max transfection reagent (United BioSystems, VA) in 10 μl of Opti-MEM medium, followed by an incubation for 15 min at room temperature (RT). The DNA-TrueFect-Max complex was mixed with growing cells in the 96-well microplate and incubated at 37°C, 5% CO₂. In day one post transfection, 30 μl per well of enriched expression medium (CelBooster Cell Growth Enhancer Medium for Adherent Cell, ABI Scientific, VA) was fed. After five days post transfection, the antigenic analysis of paired antibodies was characterized by 96-well-formatted ELISA. Briefly, 100 ul per well of Flu hemagglutinin of A/California/04/2009 (CA2009, H1 subtype) or H3 A/Hong Kong/1/1968 (HK1968, H3 subtype), respectively, at a concentration of 4 ug/ml in phosphate buffered saline (PBS) was captured in Ni-coated 96-well ELISA plates (Thermo, IL)) and incubated for two hours at room temperature, followed by the removal of the Flu HA antigen solution and incubation of 200 μl per well of CelBooster Cell Growth Enhancer Medium for Adherent Cell for one hour at RT. After washing with PBS + 0.05% Tween 20, 30 μl per well of the expressed supernatant antibody mixed with 70 μl of PBS was incubated for one hour at RT. After washing, 100 μl per well of Horseradish peroxidase (HRP)-conjugated goat anti-human IgG antibody (Jackson ImmunoResearch Laboratories Inc., PA), diluted at 1:10,000 in CelBooster Cell Growth Enhancer Medium for Adherent Cell with 0.02% tween 20, was incubated for 30 min at RT. After washing, the reaction signal was developed using 100 μl of BioFX-TMB (SurModics, MN) at RT for 10 min, and then stopped with 100 μl of 0.5 N H₂SO₄. The readout was measured at a wavelength of 450 nm, and OD₄₅₀ values were normalized and analyzed. All samples were performed in duplicate.

H1 CA09, H3 HK68, and H3 VIC11 HA constructs C-terminally fused to a thrombin cleavage sequence, T4 fibritin trimerization motif followed by hexahistidine affinity tag were synthesized (Genscript) and subsequently cloned into a pVRC8400 expression plasmid, as previously described (12). HA proteins were expressed by transfection in 293F cells (Thermo Fisher) using Turbo293 transfection reagent (SPEED BioSystem) according to the manufacturer’s protocol. Transfected cells were incubated in shaker incubators at 120 rpm, 37°C, 9% CO₂ overnight. On the second day, one tenth culture volume of CellBooster medium (ABI scientific) was added to each flask of transfected cells. Cell cultures were then incubated at 120 rpm, 37°C, 9% CO₂ for an additional 5 days. 6 days post-transfection, cell culture supernatants were harvested, clarified by centrifugation at 2,000 × g and filtered. The supernatants were loaded on Complete His-Tag Resin (Roche) by gravity flow. The resin was washed with three column volumes of PBS with 50 mM imidazole (Roche) and the target protein was subsequently eluted in three column volumes of PBS with 300 mM imidazole. The eluted protein was concentrated and further purified on a Superdex 200 16/60 size exclusion column (GE Healthcare) in PBS.

Immunoglobulin heavy chain or light chain sequences were constructed by gene synthesis and then cloned into human IgG1, lambda, or kappa expression plasmids as previously described (12, 59). Heavy and light chain expression plasmid DNA was transfected into Expi293F cells (Thermo Fisher) in 1:1 (v/v) ratio using Turbo293 transfection reagent (60). Monoclonal antibodies from the culture supernatants were purified using recombinant Protein-A Sepharose (GE Healthcare) as per the manufacturer’s instructions.

The purified human IgG proteins were cleaved by LysC enzyme (1:4000 w/w) (Roche) at 37°C overnight to yield Fabs. On the next day, the enzymatic digestion reaction was terminated by addition of protease inhibitor (Roche). The cleavage mixture was then passed through a protein A column to separate the Fc fragments from the Fab. The Fab collected in the flow-through was loaded onto a Superdex 200 16/60 column for further purification to be used for structure determination.

The H3N2 hemagglutinin (A/Victoria/361/2011) HA trimer was incubated with a molar excess of Fab SRR2899884.46167H+MEDI8852L and 2.3 μl of the complex at 1 mg/ml concentration was deposited on a C-flat 1.2/1.3 carbon grid (protochip.com). The grid was then vitrified using an FEI Vitrobot Mark IV with a wait time of 30 seconds, blot time of 3 seconds, blot force of 1 and humidity of 100%. Data collection on a Titan Krios was performed through Leginon (61) equipped with a Gatan K2 Summit direct detection device. Exposures were collected in movie mode for a 10 s with the total dose of 71.06 e–/Ų fractionated over 50 raw frames. Pre-processing was performed through Appion (62, 63); frames were aligned and dose-weighted with MotionCor2 (64). CTFFind4 (65, 66) was used to estimate the CTF and DoG Picker (62, 63) was used for particle picking, RELION (67) was then used for particle extraction. CryoSPARC 2.15 (68) was subsequently used for the remaining processing of 2D classifications, ab initio 3D reconstruction, homogeneous refinement, and nonuniform 3D refinement. Initial 3D reconstruction was performed using C1 symmetry, confirming 3 Fab molecules per trimer, whereupon C3 symmetry was applied for the final reconstruction and refinement. Model building through coot was followed by simulated annealing and real space refinement in Phenix (69) and then iteratively improved with manual fitting of the coordinates in Coot (70). Geometry and map fitting were evaluated using Molprobity (71) and EMRinger (72). PyMOL (www.pymol.org) and chimera (73) were used to generate figures.

Influenza HA-NA pseudotyped lentiviruses that harbor a luciferase reporter gene were produced as described previously (74, 75). Pseudovirus was produced by transfection of 293T cells of HA and corresponding NA along with the lentiviral packaging and reporter plasmids. For all pseudoviruses except H5N1, a human type II transmembrane serine protease TMPRSS2 gene was also contransfected for proteolytic activation of HA to HA1/HA2. Cells were transfected overnight by use of Fugene6 (Promega, Madison, WI) and then replaced with fresh medium. Forty-eight hours after transfection, supernatants were harvested, filtered through a 0.45- μm syringe filter and frozen at -80°C before use.

Neutralization assays were carried out as follows: pseudovirus was mixed with serial dilutions of monoclonal antibodies for 45 minutes followed by addition to 293A cells (Thermo Fisher Scientific) in 96-well plate white/black isoplates (PerkinElmer, Waltham, MA) in triplicate. Three days after infection, cells were lysed in 20μl of cell culture lysis buffer (Promega, Madison, WI) and 50 μl of luciferase assay reagent (Promega) was added. Luciferase activity was measured according to relative light unit (RLU) by MicroBeta luminescence counter (PerkinElmer). IC50 were generated using Prism 8 (GraphPad, San Diego, CA).

Influenza neutralization assay using replication-restricted reporter(R³) influenza viruses was described previously (76). Briefly, R³ ΔPB1 influenza viruses, which have PB1 coding sequence replaced with TdKatushka2S fluorescent reporter, were rescued by reverse genetics and propagated in PB1-expressing MDCK-SIAT1 cells. For neutralization assay, 4-fold serial antibody dilutions with a starting concentration of 25 ug/ml were made in OptiMEM (Thermo Fisher) supplemented with TPCK-trypsin and incubated with pre-titrated viruses for 1h at 37°C. After incubation, antibody-virus mixtures were transferred in quadruplicates to 384-well plates (Greiner) and 3 ×10³ MDCK-SIAT1-PB1 cells were added to each well. Plates were incubated for 18-20h at 37°C in a humidified 5% CO₂ atmosphere. The number of fluorescent cells in each well was obtained using a Celigo image cytometer (Nexcelom Biosciences) with customized red channel to enhance detection of mKate2/TdKatushka2 reporter (EX 540/80 nm, DIC 593 nm and EM 593/LP nm). The percent neutralization was calculated by constraining the VC control as 0% and the CC control as 100% and plotted against antibody concentration. A curve fit was generated by a four-parameter nonlinear fit model in Prism (GraphPad). The 50% (IC₅₀) inhibitory concentrations were obtained from the curve fit for each antibody.

Publicly available paired heavy- and light- chain datasets were downloaded using accession numbers from references (77–79). Deep sequencing reads were filtered and sieved by the HV6-1 class signatures. The CDR-H3 sequences were clustered as described in section 2.1, and the centroid antibody of each cluster was selected for modeling of HA/antibody complex structures. Germline amino acids were used to repair the missing residues at N-termini of heavy and light chains. Repaired antibody sequences were threaded on 56.a.09/HA complex structure (PDB:5K9K) (9) and MEDI8852/HA complex structure (PDB:5JW4) (56) structures respectively using software NEST (80). Binding energy between antibody and influenza A virus HA was predicated by Rosetta package, Interface Analyzer, using default parameter (81). The average of the binding energies calculated from the two homology models were used to compare antibodies that satisfied the signatures versus antibodies that satisfied the version 1 signature but not the version 2 signature; two-tailed Mann-Whitney test were used to compare the statistical significance.

The hexahistidine affinity tagged CA09 or PR34 HA proteins (30 μg/ml) diluted in PBS was captured on Ni-NTA sensor tips to a level of approximately 1.0 – 1.2nm. The duration of protein capture was 300 s. The sensor tips were then washed with PBS for 60 s to allow baseline adjustment. After that, the sensor tips were dipped into the Fabs with eight different concentrations (400 nM, 200 nM, 100 nM, 50 nM, 25 nM, 12.5 nM, 6.25 nM, 0 nM) for an association time of 300 s. Subsequent to Fab binding, the sensors were placed back into PBS for dissociation for 300 s. Antibody binding experiment was performed on a ForteBio OcteRed 96 machine. Subsequent data analysis was performed using the ForteBio Data Analysis 12.0 software.

We developed a workflow to delineate antibody sequence signature in general and applied it to delineate the sequence signature of HV6-1 class influenza antibodies ( Figure 1 ). Briefly, we start with a broader signature that are able to accommodate all known HV6-1 class influenza antibodies, use this sequence signature to identify antibody sequences from sequence databases, and then characterize their binding to hemagglutinin. As we only defined the signature for heavy chain, we pair the identified heavy chain sequences with three known HV6-1 antibodies: MEDI8852, 56.a.09, and 54.f.01. If all or majority of the sequences are functional, the sequence signature would be made broader and the sequence search and binding analyses would be repeated again. If substantial number of the sequences are not functional, this would suggest the sequence signature is too broad, and the sequence signature would be made more restrictive by adding amino acid requirements based on the informatics analyses from the binding data. Select number of antibodies would also be selected to assess for their neutralization against group 1 and group 2 viruses.

As opposed to the HV6-1 class influenza antibody signature initially described in Joyce et al., 2016 (termed HV6-1 class signature version 0 in this paper), which is not compatible with other known HV6-1 antibodies such as MEDI8852, we started with a broader sequence signature in this paper (HV6-1 class signature version 1, Figure 2A ). Specifically, we specified the CDR-H3 length to be 16-18 amino acids long as opposed to limiting the CDR-H3 length to only 16 amino acids long as in the version 0 signature, as 56.a.09 had a CDR-H3 length of 16 amino acids while MEDI8852 had a CDR-H3 length of 18 amino acids. In addition, version 1 signature allows full amino acid flexibility at residue 98, and allowing amino acid flexibility limited to similar residue types for the ⁹⁹IFGI motif in version 0 signature, as residue 98 did not contact hemagglutinin for both 56.a.09 and MEDI8852 ( Figure S1 ). In addition, we removed the requirement of HD3-3 germline gene as D gene determination can be sometimes ambiguous. We used the version 1 signature to search for heavy chain sequences from three NGS datasets from three influenza vaccination three studies ( Table S1A ), and identified a total of 62 cluster representatives ( Figure S2A , Table S1 , and Dataset S1 ). We co-expressed the heavy chains with three different light chains from known HV6-1 class antibodies respectively, including MEDI8852, 56.a.09, and 54.f.01, and assessed the binding to hemagglutinin of A/California/04/2009 (CA2009, H1 subtype) and H3 A/Hong Kong/1/1968 (HK1968, H3 subtype) using a 96-well transient expression ELISA assay (see Methods). We observed that 24 of the 62 heavy chain sequences, when paired with one of the three HV6-1 class antibody light chains, have high binding towards CA2009 HA, and 19 of the 62 heavy chains, when paired with one of the three HV6-1 class antibody light chains, have high binding toward HK1968 HA (see Methods) ( Figure 2B , Dataset S2 ). Overall, 25 of the 62 heavy chains (40.3%), when paired with one of the three HV6-1 class antibody light chains, have high binding toward CA2009 HA or HK1968 HA. Specifically, 24 of these 25 heavy chains are functional when paired with MEDI8852 light chain (KV1-39 derived), while 16 of these 25 heavy chains are functional when paired with 56.a.09 or 54.f.01 light chain (KV3-20 derived) ( Figure 2C ).

As majority of the version 1 sequence signature-identified heavy chains were not functional, we examined the association between amino acid types at each residue position and functional outcome. We observed that 23 of the 25 functional heavy chains had a phenylalanine at residue 100 (P=0.0011) and 22 of the 25 functional heavy chains had a glycine at residue 100a (P=0.0311) ( Figure 2D ). Examination the structure of antibody 56.a.09 in complex with hemagglutinin revealed that Phe100 binds to a cavity lined by hydrophobic residues such as Trp350, Ile372, and Il377 (PDB:5K9K) ( Figure 2E ). We also noticed that among the 62 antibodies, larger fraction of antibodies with CDR-H3 length of 17 were functional compared to antibodies with CDR-H3 length of 16 or 18 ( Figure S3 ), suggesting that CDR-H3 length of 17 is compatible with HV6-1 class antibody signature.

To evaluate if the signature-identified HV6-1 class antibodies can neutralize influenza viruses in addition to binding to hemagglutinin, we evaluated entry-inhibition for 10 of the 22 heavy chains that, when paired with one of the three known HV6-1 class antibody light chains, showed the highest binding to hemagglutinin, using a pseudovirus assay with a panel of six influenza strain ( Dataset S3 ). All ten tested antibodies inhibited the entry of all six viruses. Next, we selected the top three antibodies based on the results of the six strain panel, and evaluated their entry inhibition against a panel of 31 influenza strains, including 16 group 1 strains and 15 group 2 strains ( Figure 2F , Dataset S4 ), and observed that all three signature-identified heavy chains, when paired with the light chain from MEDI8852, showed comparable neutralization potency and breadth to MED8852. We also assessed the neutralization of these three signature-identified antibodies with a high-throughput influenza neutralization assay using engineered reporter viruses (see methods), and showed that two of these three antibodies (SRR2899884.46167 and 39218, paired with MEDI8852 light chain respectively) showed comparable neutralization potency and breadth to MEDI8852 ( Figure 2G and Dataset S5 ).

To examine if the sequence-signature-identified HV6-1 sequences have similar recognition mode as known HV6-1 antibodies, we solved the structure of SRR2899884.46167H+MEDI8852L in complex with an H3N2 hemagglutinin (A/Victoria/361/2011) using cryo-EM at 3.41 angstroms ( Figures 3A, B and S4 ). When the structures of hemagglutinin in complex with MEDI8852 (PDB:5JW4) and 56.a.09 (PDB:5K9K) were aligned on top of this structure based on hemagglutinin ( Figure 3C ), the ⁹⁹VFGV^100b motif of MEDI8852 CDR-H3, the ⁹⁹IFGI^100b motif of 56.a.09 CDR-H3, and ⁹⁹IFGL^100b motif of SRR2899884.46167 CDR-H3 align almost perfectly on top of each other (RMSD<1 Å, Figure 3D ), demonstrating that the signature-search-identified HV6-1 heavy chain SRR2899884.46167 can recognize hemagglutinin in the similar manner as previously identified HV6-1 class antibodies.

Based on the antigenic and bioinformatics analyses of the sequences identified from HV6-1 class signature version 1 ( Figure 2 ), we came up with the version 2 of HV6-1 class signature by allowing only phenylalanine at residue 100 and glycine at residue 100a ( Figure 4A ). We used the HV6-1 class signature version 2 to search for HV6-1 class sequences from another NGS dataset (dbGaP Study Accession: phs000666.v1.p1), which had no sequence overlap to the NGS datasets used in the prior search. Sixteen out of 1,611,992 productive heavy sequences were compatible with the signature version 2. After clustering, we co-expressed nine heavy chains ( Figure S2B , Dataset S6A ) with light chains from MEDI8852, 56.a.09, and 54.f.01, respectively, and assessed the binding to hemagglutinin of A/California/04/2009 (CA2009, H1 subtype) and H3 A/Hong Kong/1/1968 (HK1968, H3 subtype) using a 96-well transient expression ELISA assay (see Methods). We observed that five of the nine heavy chain sequences (56%), when paired with one of the three HV6-1 class antibody light chains, have high binding towards CA2009 HA, and two of these nine heavy chains, when paired with one of the three HV6-1 class antibody light chains, have high binding toward HK1968 HA (see Methods) ( Figure 4B , Dataset S7 ). We assessed entry-inhibition of eight antibodies, including five unique heavy chains, with the highest ELISA signal using three group 1 and three group 2 viral strains. All eight antigenic-positive antibodies showed activity towards at least two different strains ( Figure 4C , Dataset S8 ), confirming the signature version 2-identified antibody sequences to be functional in terms of neutralization. Notably, version 0 signature did not identify any reads, and version 1 signature identified the same set of sequences as identified by version 2 signature ( Table S1A ).

In addition, we used versions 0, 1, and 2 HV6-1 class antibody signatures to search for paired heavy and light chain sequences from three different deep sequencing samples ( Figure 4F ). We found no sequences satisfying version 0 signature and 22 sequence cluster representatives satisfying version 1 signature, among which 13 also satisfying version 2 signature ( Figure 4F ). Nine different light chain germline genes were observed among the 22 sequence cluster representatives from version 1 signature search, including KV1-39 and KV3-20 used by MEDI8852 and 56.a.09, respectively ( Dataset S6B ); version 2-selected antibodies derived from 8 different light chain germline genes and showed an increase in the fraction from KV-gene origin ( Dataset S6B ). We constructed homology model for the 22 antibodies, analyzed the predicted binding energy between Flu HA and antibody, and observed the 13 antibodies satisfying the version 2 signature to have lower binding energy than the 9 antibodies satisfying version 1 but not the version 2 signature ( Figures 4G, H ).

Finally, we used versions 0, 1, and 2 HV6-1 class antibody signatures to search NGS samples of Naïve B cells and cord blood (78, 82), and found that while there are no sequences that satisfied version 0 signature, sequences that satisfied versions 1 and 2 signature were present in these NGS reads ( Table S1B ), suggesting that antibodies with HV6-1 class antibodies characteristics are present before immunization.