Influenza viruses remain a severe threat to human health causing up to 650,000 deaths
annually.
1,2
Seasonal influenza virus vaccines can prevent infection, but are rendered ineffective
by antigenic drift. To provide improved protection from infection, novel influenza
virus vaccines that target conserved epitopes of influenza viruses, specifically those
in the hemagglutinin (HA) stalk and neuraminidase (NA), are currently being developed.
3
Antibodies against the HA stalk confer protection in animal studies.
4–6
However, no data exist on natural infections in humans and these antibodies do not
show activity in the hemagglutination inhibition (HI) assay, the HI titer being the
current correlate of protection against influenza virus infection.
7–9
While previous studies have investigated the protective effect of cellular immune
responses and NA-inhibiting antibodies, additional serological correlates of protection
from infection could aid the development of broadly protective or universal influenza
virus vaccines.
10–13
To address this gap, we performed a household transmission study to identify alternative
correlates of protection from infection and disease in naturally exposed individuals.
Using this study we determined 50% protective titers and levels for HI, full-length
HA, NA, and HA stalk-specific antibodies. Further, we found that HA stalk antibodies
independently correlated with protection from influenza virus infection.
We followed 300 household members in a Nicaraguan family cohort who lived with one
of 88 influenza-positive index cases for 3–5 weeks to test for infection and seroconversion
(Figures 1A, Extended Data Fig. 1). The majority of households were recruited during
the 2015 season (n=65) as pandemic H1N1 influenza virus activity was lower in 2013
(n=23). Only 10 household members were vaccinated for the concurrent influenza season
(Table S1) which did not allow for detailed comparisons to unvaccinated individuals.
Individuals who reported prior influenza virus vaccination were distributed evenly
across antibody levels and two had PCR-confirmed influenza virus infection. Overall,
84 (28%) household members had a PCR-confirmed infection and approximately two-thirds
(n=53) of the PCR-positive individuals developed symptomatic influenza.
To identify antibody levels associated with protection from infection and disease,
we tested baseline (collected upon confirmed infection in a household) and follow-up
(3–5 weeks post-enrollment) blood samples using the classical HI assay as well as
in enzyme-linked immunosorbent assays (ELISAs) that measured antibodies against the
full-length HA, the HA stalk domain or the NA.
As expected, we found that individuals with higher pre-exposure HI titers were less
likely to become infected (Figure 1B). The 50% protective HI titer (i.e. the antibody
level at which the risk of contracting influenza is reduced by 50% compared to individuals
without detectable antibodies) was between 1:20 and 1:40, which is consistent with
previous studies (Figure 1B, Table S2).
14
High baseline ELISA levels measuring full-length HA antibodies in individuals who
tested negative in the HI-assay indicate a strong prevalence of non-HI-active antibodies
in the study participants. This can be explained by the presence of antibodies that
bind the HA protein, but do not sterically interfere with receptor binding, which
is the activity measured in the HI assay. ELISAs measuring full-length HA and NA antibodies
correlated with protection from infection with narrow 95% confidence intervals (C.I.).
The estimated 50% protective levels for both assays ranged between areas under the
curve (AUCs) of 20 and 40 (Figure 1B, Table S2). The confidence intervals were wider
for HA stalk antibodies, but similar 50% protective levels were between AUCs of 40
and 80.
We additionally estimated crude (i.e. not adjusted for age or other variables) 50%
protection antibody levels against PCR-confirmed symptomatic influenza and found a
similar good correlation with protection for all measured antibody levels. The antibody
level associated with 50% protection was approximately 1:40 for HI antibodies and
between AUCs of 20 and 40 for full-length HA, HA stalk and NA binding antibodies (Figure
1C, Table S3 and Extended Data Fig. 2, 3).
When pre-existing antibody levels of individuals who either were PCR-positive, or
negative for influenza virus were plotted side-by-side, a clear trend could be observed
(Figures 2, Extended Data Fig. 4 and Tables S4–S6). Participants who became infected
had very low HA and NA antibody levels. In addition, HA stalk antibody levels measured
by ELISA were lower in individuals who developed PCR-confirmed symptomatic influenza
compared to PCR-confirmed asymptomatic individuals (p<0.01) suggesting that these
antibodies correlated with protection from infection and disease.
These crude analyses do not account for the effect of age and other antibodies in
protected individuals. Thus, we adjusted for potential correlations by comparing the
calculated protective effects associated with a 4-fold increase in antibody levels
(seroconversion) in a single-assay model, to a multi-assay model that adjusts for
correlation with other assays and age.
When adjusted for the effects of other measured antibody levels and age (multi-assay
model), we found that HI, full-length HA, and HA stalk antibody levels remained independent
predictors of protection against both PCR-confirmed infection and symptomatic influenza
(Figure 3). Antibodies measured against NA showed a similar trend and were associated
with protection against symptomatic influenza when adjusted for HI antibodies. However,
they were not independently associated with protection when adjusted for both HI and
HA stalk antibodies, indicating that NA antibodies in these individuals correlated
with antibodies that were induced against HA. Using the multi-assay model, we found
a four-fold increase in HA stalk antibodies to be associated with a 42% (C.I.,15%,60%)
reduction in risk of being infected, which was slightly lower than the effect observed
for a four-fold increase of HI antibodies (57%;C.I.,35%,72%). A similar reduction
in infection risk was observed for symptomatic influenza. Age-stratified results are
shown in extended data figures 5 and 6. These findings provide important support that
non-HI active antibodies can be independently predictive of protection from influenza
virus infection.
To compare the usefulness of ELISA-based readouts to HI for assessing seroconversion,
we calculated fold-inductions of antibody levels post-infection for PCR-positive and
PCR-negative cases (Figure 4). Consistent with previous studies, we found that 22%
of individuals did not respond to infection as measured by HI (Figure 4, light-blue
peak at 1-fold).
15
Interestingly, we did not detect any apparent non-responders using ELISAs measuring
full-length HA and HA stalk-specific antibodies. Many infected children (64%) did
not show an increase in anti-NA antibody levels. Infection is generally thought to
boost NA antibody levels,
16
but measured responses against NA may be generally low as we found previously.
17
A proportion of PCR-negative individuals seroconverted in all assays which might be
attributable to these individuals not shedding enough virus for detection via PCR
while still being infected. Additional sensitivity and specificity analyses were performed
and indicate that ELISAs are useful to assess seroconversion in addition to HI assays
(Extended Data Fig. 7, 8).
Novel universal influenza virus vaccines that elicit broadly-reactive antibodies against
conserved epitopes in the HA stalk domain are currently in clinical development. However,
HA stalk antibodies have not been shown to correlate with protection against natural
influenza virus infection in humans. In this study, we used samples from a household
transmission study to examine HI, full-length HA, HA stalk and NA antibodies as potential
correlates of protection from influenza infection and disease. Importantly, using
multiple statistical approaches we showed that HA stalk antibodies (which cannot be
detected in HI assays
10
) were associated with protection against pandemic H1N1 influenza virus infection
and disease.
Consistent with previous studies, a baseline HI titer between 1:20 and 1:40 was predictive
of a 50% reduction in PCR-confirmed H1N1 influenza virus infection.
14
Interestingly, only few individuals had baseline HI titers ranging from 1:10 to 1:40.
Instead, titers were either undetectable or higher than 1:40 (Figure 1). A possible
explanation for this could be that the antibodies measured in this largely unvaccinated
population were elicited by recent infections, because the virus only circulated for
4–6 years prior to the study and it has been shown that that HI antibodies elicited
by infection are maintained at titers >1:40 for many years.
18
Using ELISAs that measured antibodies against full-length HA, NA or specifically the
HA stalk, we were also able to identify crude estimates of protection. We found that
these results were consistent between two influenza seasons (Extended Data Fig. 2,
3). Importantly, ELISAs against HA can measure antibody levels irrespective of the
ability of the virus to agglutinate red blood cells, which is required for HI assays
and has posed a problem for serology against recent H3N2 virus strains.
19
These assays can furthermore detect non-neutralizing (but potentially protective)
antibodies, which is of importance for anti-stalk antibodies that confer the majority
of their protective effect through Fc-mediated functions.
20,21
This study demonstrates that levels of HA stalk antibodies are a correlate of protection
against natural pandemic H1N1 influenza virus infection. While previously published
findings from a human challenge model did not find HA stalk antibodies to be predictive
of protection from infection, they found an association with a reduction in viral
shedding.
22
A possible explanation for this difference is that human volunteers were intranasally
inoculated with high doses of infectious particles (approximately 10^7), whereas natural
infection is likely caused by much lower particle numbers.
23
This difference is further highlighted by the fact that even individuals with high
HI titers (> 1:1000) were not protected from infection in the challenge setting, which
is not consistent with the findings of the majority of vaccine efficacy studies.
24
Of note, the confidence intervals in this study for the predicted protective effect
of HA stalk antibodies for PCR-confirmed infection were wider for adults compared
to children (Figure 1A). Multiple factors may have contributed to this observation.
There were few adult individuals who had baseline HA stalk antibody levels of <10,
which may have contributed to the lower than expected number of cases. An important
observation is the higher than expected number of infections at AUC levels from 160–640.
This could be an indication that high titers of HA stalk antibodies are required for
complete virus neutralization, which is consistent with previous observations that
HA stalk antibodies have lower neutralizing activities compared to HI-active antibodies.
25
Importantly, the correlation of HA stalk antibodies with protection from symptomatic
influenza was consistent for both children and adults, which may indicate that HA
stalk antibodies can reduce symptoms at sub-neutralizing levels.
While NA antibody levels also correlated with protection, the majority of our subjects
had low baseline NA antibody levels which limited the power of NA antibodies as an
independent correlate of protection, after adjusting for age and HI titers. Furthermore,
the results indicated that antibodies against NA correlated with HA antibodies in
these individuals, potentially because the antibodies in this largely unvaccinated
population were mainly elicited by infections, which would elicit antibodies against
both HA and NA. NA antibodies correlated more with antibodies against the HA stalk
(Pearson’s r = 0.35) compared to HI active antibodies (Pearson’s r = 0.25; Extended
Data Fig. 9) Previous studies have shown that NA inhibition assays (NI) could be a
useful correlate of protection from infection
12,26
but HA stalk antibodies can contribute to NA inhibition measured in the traditional
enzyme-linked lectin assay (ELLA).
17,27–29
Based on these findings, it is possible that the correlation with protection reported
for NI is partially conferred by HA stalk antibodies. Unfortunately, we did not have
sufficient serum specimens to perform NI in this study. We also did not perform microneutralization
assays, which have been previously shown to correlate with protection from infection,
but may not fully capture the specific effects of HA stalk antibodies. The protective
effect of these non-neutralizing antibodies should be investigated in future studies
using assays that measure Fc-mediated functions of antibodies to dissect the mechanisms
of antibody-mediated protection. Similarly, cell-mediated immunity could not be assessed
here and will need to be further investigated. Importantly, these additional immune
mechanisms could explain why a subset of adults did not have PCR-confirmed infection,
despite low antibody levels.
ELISAs are used as standard assays for a number of other pathogens and are comparatively
easy to standardize.
30
While antibody binding as measured in these assays may not directly translate into
functionality, our findings indicate that in a human cohort study setting with individuals
who have acquired immunity primarily through virus exposure, results from binding
assays could accurately predict protection from infection. We have also previously
shown that ELISA antibody levels after vaccination can predict the protection of mice
in a human serum transfer experiment.
4
Furthermore, this study shows that these assays can be useful in combination with
HI assays to assess seroconversion after influenza virus infection.
HA stalk antibodies were measured using a chimeric HA antigen, which has an exotic
H6 head domain to which humans are generally naïve. However, some rare cross-reactive
head antibodies have been previously described that could recognize conserved epitopes
on this antigen.
31,32
It can therefore not be excluded that part of the measured response is provided by
non-HI active cross-reactive head antibodies. Since these antibodies have been rarely
isolated from humans, the majority of the measured responses are likely HA stalk antibodies.
Accordingly, antibody levels measured using a chimeric HA have been previously shown
to correlate well with antibodies measured using headless HA probes in ferrets that
were vaccinated with multiple heterologous HA head domains.
5
A particular strength of our study was the intensive follow-up which allowed us to
capture both symptomatic and asymptomatic individuals, which translated to a high
number of observed infections and provided statistical power for our detailed analyses.
The study was performed in Nicaragua where influenza virus vaccination was introduced
recently and is not widely used; therefore, the majority of the pre-existing antibody
response was likely induced by repeated natural infections. This differs from the
situation in some countries, where vaccination rates are high. Similar studies that
test highly vaccinated individuals will be required to detect potential differences
in protective antibody levels elicited by vaccination vs. infection.
In summary, we found that HA stalk antibodies are an independent correlate of protection
from pandemic H1N1 infection and disease in a natural transmission setting. Further,
antibodies measured by ELISA can be used as a powerful correlate of protection and
to assess seroconversion, which will be important for novel universal influenza virus
vaccine development.
10,33,34
Additional resources should be allocated to standardize these assays to enable their
use in both research and clinical settings. Further studies are required to examine
the role of these antibodies as potential correlates of protection against influenza
A(H3N2) and influenza B in natural transmission settings.
Methods
Participants and Study Procedures
As a part of an observational household transmission study in Nicaragua, members who
lived with an influenza index case in their household were monitored for influenza
virus infection. Daily symptoms were assessed, nasal and oropharyngeal swabs were
taken every 2–3 days for 10–14 days and blood samples were collected at enrolment
as well as 3–5 weeks later to determine infection outcomes and antibody responses.
Eligible households included those that 1) had an index case that had a positive QuickVue
Influenza A+B rapid test result and with acute respiratory infection (ARI) symptom
onset within the previous 48 hours; and, 2) had at least one person living with the
index case. Details of the study design are published.
35,36
Participants were excluded from this analysis if sufficient blood samples were not
available. The principles of the Declaration of Helsinki were strictly followed. Ethical
approval was obtained from institutional review boards of the University of Michigan
(HUM 00091392) and the Ministry of Health, Nicaragua (CIRE 06/07/10–025). Written
informed consent was obtained from all adult participants and proxy written informed
consent was obtained for all children. Assent was obtained from children aged 6 and
older.
Laboratory Methods
Respiratory samples were tested in the Nicaraguan National Virology Laboratory by
real-time reverse-transcription polymerase chain reaction (PCR) using standard protocols.
37
HI assay
38
was performed to determine HI titers while ELISA was performed to measure binding
antibodies to full-length HA, HA stalk and NA. ELISAs were performed as described
elsewhere.
4
The HA full-length constructs corresponded to the vaccine strains during the respective
seasons (H1 A/California/4/09 – 2013 season; H1 A/Michigan/45/15 – 2015 season). A
chimeric HA expressing the head domain from an H6N1 virus (to which humans are naïve)
and the stalk domain of pandemic H1N1 influenza virus A/California/4/09 was used to
measure HA stalk antibodies. The HAs were expressed as soluble proteins with a trimerization
domain to maintain correct protein folding and conformational epitopes as previously
described.
39
The NA of A/California/4/09 was used to measure NA-specific antibodies. The NA was
expressed as soluble antigen with a tetramerization domain to maintain correct folding
and enzymatic activity (as measured in NA star assays).
40
ELISA values are reported as area under the curve (AUC). AUC was chosen, as it considers
both the endpoint and the absolute levels of optical density measured at all tested
serum dilutions. AUC calculation (optical density multiplied by serum dilution over
the entire curve) was performed in GraphPad Prism. All assays were performed by personnel
who were blinded to infection status.
Outcomes
The primary outcome was PCR-confirmed influenza and the secondary outcome was symptomatic
influenza (PCR-confirmed infection with an episode of fever with cough or sore throat).
41
Antibody response was measured by the ratio between post- and pre-exposure level (pre-existing
antibody level).
Statistical Analyses
Antibody level-specific attack rates were calculated by dividing the number of infected
contacts who had a specific baseline antibody titer by the total number of contacts
who had the same level of antibody titer. To infer the crude estimates of the 50%
and 80% protective levels, we used a 3-parameter logistic regression model (nplr R
package) that allowed for a ≤ 1 probability of infection at the lowest detectable
level, and a ≥0 probability of infection at the highest observable level, meaning
that incomplete protection can occur at high levels and participants could have pre-existing
antibodies at levels that were below what was required for complete protection. Two
multivariable logistic regression models were used to study the effect of a 4-fold
antibody level increase on infection outcome, including 1) a single-assay model in
which levels of one serology assay and age are predictors and 2) a multi-assay model
where levels of multiple assays and age are predictors. The level of pandemic H1N1
influenza virus activity differed between study years and was adjusted for in the
analyses. In models 1 and 2, a smoothing spline function was used to model the effect
of age on infection risk (mgcv R package). Antibody levels were log transformed for
all analyses, levels below the lowest detectable limit of 1:10 were imputed as 1:5.
Individual antibody titer data points were visualized and compared between disease
outcome groups using a two-tailed Wilcoxon rank sum test. A finite mixture model was
used to explore underlying non-responder sub-populations based on the observed distribution
of the antibody response (mixtool R package). The model estimates the mean and standard
deviation for each component of the Gaussian mixtures which were visualized to illustrate
results on antibody response. Receiver Operating Characteristic Curve (ROC) analysis
(pROC R package) was used to estimate the sensitivity and specificity of each assay.
Classification and Regression Trees analysis (rpart R package) were performed to identify
the best combination of assays indicated by their positive and negative predictive
values in identifying PCR positive individuals. False discovery rates (FDR) were calculated
in GraphPad Prism using the two-stage step-up method of Benjamini, Krieger and Yekutieli.
42
Extended Data
Extended Data Fig. 1
Participant follow-up timeline
Participant sample collection timeline with number of samples collected from unique
individuals (n=300 individuals). Day ranges are represented as quintiles.
Extended Data Fig. 2
Pre-existing antibodies and corresponding secondary attack rates in 2015 (N=198 individuals)
Results are shown for A) PCR-confirmed infection and B) Symptomatic influenza. Note
that geometric mean baseline HI titer for this year was 1:10. Grey tags indicate a
50% protection level and black tags indicate an 80% protection level. Grey bars present
the proportion of household contacts having a certain level of pre-existing antibody
levels. The bars group individuals between the antibody levels covered by the bars
on the x-axis (e.g. left-most bar includes all individuals with antibody levels <10,
followed by 10 but less than 40, etc.). Red lines fit the antibody level-specific
secondary attack rate based on the observed rates which are indicated as cyan points.
The attack rate was calculated by dividing the number of infected contacts who had
a specific baseline antibody level by the total number of contacts who had the same
level of antibodies. Shaded area represents the 95% confidence intervals.
Extended Data Fig. 3
Pre-existing antibodies and corresponding secondary attack rates in 2013 (N=102 individuals)
Results are shown for A) PCR-confirmed infection and B) Symptomatic influenza. Note
that geometric mean baseline HI titer for this year was 1:34. Grey tags indicate a
50% protection level and black tags indicate an 80% protection level. Grey bars present
the proportion of household contacts having a certain level of pre-existing antibody
levels. The bars group individuals between the antibody levels covered by the bars
on the x-axis (e.g. left-most bar includes all individuals with antibody levels <10,
followed by 10 but less than 40, etc.). Red lines represent the sigmoid function fitted
to the observed antibody level-specific secondary attack rates (SAR) which are indicated
as cyan points. The attack rate was calculated by dividing the number of infected
contacts who had a specific baseline antibody level by the total number of contacts
who had the same level of antibodies. Shaded area represents the 95% confidence intervals
for the predicted antibody level-specific SAR.
Extended Data Fig. 4
Influenza outcome specific distribution of pre-existing antibodies
Results are shown for A) 2015 A(H1N1)pdm epidemic and B) 2013 A(H1N1)pdm epidemic.
Antibody levels for each individual, and the median and inter quantile range are shown.
Y-axis indicates antibody levels. Individuals were separated by PCR-positivity status
(blue dots) and by symptomatic influenza (green dots). Individual antibody titer data
points were compared between disease outcome groups using a two-tailed Wilcoxon rank
sum test. Analyses were performed combined (all ages; 2013: n=102 individuals; 2015:
n=198 individuals) as well as separately for children (0–14 years old; 2013: n=38
individuals; 2015: n=64 individuals) and adults (15–85 years old; 2013: n=63 individuals;
2015: n=135 individuals). Please see Table S5 and S6 for false discovery rate analyses.
Age groups and outcomes were pre-specified before analyses.
Extended Data Fig. 5
Protective effects associated with a 4-fold increase in antibody level amongst children
Results are shown for three different sets of assays for A) PCR-confirmed infection
and B) symptomatic influenza (n=101 individuals). Assay set 1 combines HI, HA stalk
and NA ELISAs. Assay set 2 combines full-length HA and NA ELISAs. Assay set 3 combines
HI and NA ELISAs. Adjusted odds ratios for the single-assay model are shown as green
squares and the multi-assay model as orange circles. Black lines denote 95% confidence
intervals. OR: odds ratio.
Extended Data Fig. 6
Protective effects associated with a 4-fold increase in antibody level amongst adults
Results are shown for three different sets of assays for A) PCR-confirmed infection
and B) symptomatic influenza (n=199 individuals). Assay set 1 combines HI, HA stalk
and NA ELISAs. Assay set 2 combines full-length HA and NA ELISAs. Assay set 3 combines
HI and NA ELISAs. Adjusted odds ratios for the single-assay model are shown as green
squares and the multi-assay model as orange circles. Black lines denote 95% confidence
intervals. OR: odds ratio.
Extended Data Fig. 7
Positive and negative predictive values of the best serology testing strategy identified
by decision tree analyses
True positive cases were individuals who had PCR confirmed influenza virus infection.
True negatives were individuals who had neither a positive PCR nor a four-fold rise
in antibody serology tests. The model also suggested optimal cutoff points to use
when defining seroconversion. PPV: positive predictive value; NPV: negative predictive
value.
Extended Data Fig. 8
Sensitivity and Specificity of HI and ELISA in detecting PCR-confirmed Infections
Curves are plotted as solid lines for sensitivity (Sn) in blue and specificity (Sp)
in green. Shaded areas indicate 95% confidence intervals. X-axes show fold induction
for the respective assay. Analyses were performed combined (all ages, n=300) as well
as separately for children (0–14 years old, n=101) and adults (15–85 years old, n=199).
Extended Data Fig. 9
Antibody titer correlations
Correlation analyses for antibody titers were performed A) combined (all ages, n=300
individuals) as well as separately for B) children (0–14 years old, n=101 individuals)
and C) adults (15–85 years old, n=199 individuals). Pearson’s r value is plotted in
each figure.
Supplementary Material
Supplemental Materials