The Decline in Influenza Antibody Titers and Modifiers of Vaccine Immunity from over Ten Years of Serological Data

Abstract

Annual influenza vaccination is the cornerstone for seasonal protection, yet antibody responses are highly variable across individuals and over time. To systematically assess the determinants of this heterogeneity, we compiled 20,449 hemagglutination inhibition and neutralization titers from 4,540 participants enrolled in 14 new vaccine studies we conducted and 50 prior studies that collectively span 2010–2023. Seasonal effects dominated, with pre- and post-vaccination titers declining steadily from 2017 onwards, outweighing the influence of age, sex, or repeated vaccination. Titers to B Yamagata remained steady throughout all years examined, suggesting unique durability and offering a reason for lineage extinction. Vaccine timing emerged as a strong and previously underappreciated determinant of immunity, with individuals vaccinated later in the season exhibiting larger post-vaccination titers. Not being vaccinated or receiving the live-attenuated FluMist vaccine in one year significantly enhanced the response to inactivated vaccines in 45% or 68% of cohorts, respectively, whereas antigen dose and adjuvants had modest impact. These findings identify vaccine timing and seasonal context as underrecognized drivers of immunogenicity and provide actionable insights for optimizing influenza vaccination strategies.

Introduction

Influenza continues to cause substantial global morbidity and mortality, and vaccination remains the most effective strategy to elicit protection. Among the immune pathways engaged by vaccination, the antibody response serves as a primary line of defense, and hemagglutination inhibition (HAI) titers are widely used as a regulatory correlate of protection.^1–3

HAI titers exhibit striking heterogeneity across individuals and seasons, making it challenging to discern consistent patterns.^4–9 Most prior studies enrolled 10–100 participants within a single season, providing valuable but inherently limited snapshots of immune dynamics. For example, the contributions of repeated vaccination, seasonality, or host factors have not been systematically studied at scale, and the field lacks a clear framework for predicting vaccine responsiveness across the population.

Such questions have become especially timely in the wake of the COVID-19 pandemic, which disrupted influenza circulation, vaccination patterns, and caused the extinction of the B Yamagata lineage.^10–12 Simultaneously examining pre- and post-pandemic seasons provides a unique opportunity to quantify how these events shaped influenza antibody responses.

This work tackles these questions by combining 14 new influenza vaccine studies with 50 existing studies to assess antibody measurements from over 4,500 individuals spanning the past decade. We used this dataset to disentangle the effects of host features (age, sex, prior exposure), vaccine formulation (brand, dose), and season-to-season variability. A subset of longitudinal cohorts let us examine the impact of repeated vaccination as well as the antibody response out to two years post-vaccination to advance our understanding of long-term durability.

At present, the CDC lists multiple options for influenza vaccines but few guidelines to choose among them.¹³ Aside from several age restrictions (e.g., high-dose or adjuvanted vaccines are recommended for ≥65 y.o., recombinant vaccines for ≥9 y.o., live-attenuated vaccines for 2–49 y.o.¹⁴), current CDC recommendations state that most healthy individuals 6 months or older should receive an annual influenza vaccine,¹⁵ but they offer no recommendations among the available vaccines (e.g., Fluzone, Flucelvax, Flublok, Fluarix, FluMist, Afluria) or the timing of vaccine administration, nor do they suggest any changes based on a person’s recent vaccination or infection history. By assessing the vaccine options made across numerous studies, this work aims to uncover actionable practices that elicit the best possible antibody response.

Results

Pre- and post-vaccination antibody titers exhibited a small but steady decline from 2017 to 2023

20,449 HAI or neutralization titers were combined from 4,540 participants enrolled in 64 influenza vaccine studies spanning the 2010 to 2023 seasons (Table S1). In all studies, the antibody response was measured pre-vaccination and 1-month post-vaccination against that season’s vaccine strains (Table S2). Cases where HAI and neutralization were both measured in the same individuals showed minimal differences across assays (Figure S1), and thus studies measuring either HAI or neutralization were combined, with the results referred to as antibody titers henceforth. If HAI and neutralization were available for the same participant, only neutralization titers were kept.

Across all studies, there was a progressive decrease in pre-vaccination antibody titers where the geometric mean titer (GMT) across all vaccine strains changed from 38–61 in 2010–2017 to 17 in 2022–2023 (Figure 1A, variant data in Figure S2). This decline in background immunity became especially pronounced after the decrease of influenza cases during 2020–2021 due to the COVID-19 pandemic.^16,17 Despite this shift in baseline immunity, post-vaccination responses were more stable, exhibiting a smaller decrease from a GMT of 85–170 in 2010–2017 to 55–75 in 2022–2023 (Figure 1B), while fold-change ranged between 2–4x across all seasons (Figure 1C). Given the large number of measurements in each year, there were statistically significant differences between the pre-vaccination titers in every influenza season (p<0.01, two-sided T test), and hence we opt to describe the absolute differences in GMTs and reserve significance testing for the later subgroup analyses.

Figure 1. Antibody responses to influenza vaccination across 11 seasons.

(A) Pre-vaccination, (B) post-vaccination, or (C) fold-change in antibody titer across seasons. Panel A shows the number of studies (S) and number of measurements (n) analyzed. The horizontal dashed line in Panel C indicates a fold-change of 1, corresponding to no change in titer. (D) Breakdown of pre-vaccination titers by vaccine component. Unless otherwise noted, titers are averaged over all host traits (e.g., age, sex, vaccine history). All box plots show the interquartile range, horizontal lines the medians, and black diamonds the geometric mean titers (GMTs). Whiskers extend to 1.5 times the interquartile range.

Among the influenza vaccine strains, H1N1 showed the most pronounced decline in pre-vaccination titers that started in 2019 and reached GMT≤15 by 2021 (Figure 1D; post-vaccination and fold-change in Figure S3). H3N2 and B Victoria followed this trend more gradually, dropping to comparable levels by 2022. Interestingly, B Yamagata followed the same pattern as H3N2 and B Victoria until 2021, but pre-vaccination titers then progressively increased. This abrupt change in behavior coincided with the extinction of the B Yamagata during the COVID-19 lockdown, although the influenza vaccine still included this lineage until 2023 before becoming trivalent in 2024.^11,12

Season-to-season variability outweighed age or consecutive vaccination effects

To disentangle whether this decline in titers was due to repeated vaccination or seasonal factors, we tracked antibody responses in individuals with two consecutive vaccinations across seasons, classifying them based on whether their antibody titers were weak (<40) or strong (≥40) against that season’s vaccine strain (e.g., strong-weak means that post-vaccination titers were ≥40 in one season and <40 the next). Prior studies have found that a titer threshold of 40 roughly correlated with 50% protection,^18–22 so that strong-strong and weak-weak represent consistently robust/frail phenotypes, whereas strong-weak or weak-strong represent a change across consecutive seasons.

The fraction of participants in each group revealed a steady decline among strong-strong responders that have been replaced by weak-weak responders (Figure 2A). This shift started in the 2017–2018 seasons where 18% of participants went from a strong to a weak response, while only 1% improved from weak to strong. At this inflection point, the previously dominant strong-strong group comprising ~60% of participants began to decline, and weak-weak responses became the dominant phenotype in 2021–2022. Note that these trends started before the COVID-19 pandemic, and they have persisted into 2022–2023.

Figure 2. Changes in post-vaccination antibody titers for consecutively vaccinated individuals.

(A) Antibody titers were categorized as weak (<40) or strong (≥40) for participants vaccinated in two consecutive seasons (x axis, e.g., 2017–2018 implies vaccination in 2017 and 2018, irrespective of vaccines in other seasons). Percent of participants in each category (y axis) was computed across all vaccine strains and vaccine histories. (B) Post-vaccination titers or fold-change stratified by age. (C) GMT of participants consecutively vaccinated in ≥3 consecutive seasons, subset by immunological age (young: 10–34 y.o., adults: 35–96 y.o.) and by the first season of their consecutive vaccines (Methods). Lines show GMTs and 95% confidence intervals. The number of measurements (n) is shown in each panel.

To determine if this effect was caused by differences in age groups (e.g., more elderly participants were measured in later seasons), we first determined how antibody titers vary with age and then calculated the repeated vaccine responses in each age group. Surprisingly, GMTs of the post-vaccination antibody titers showed minimal differences across every age group, neither showing a dip for elderly individuals ≥65 y.o. nor significantly stronger responses in children (Figure 2B). Instead, post-vaccination titers showed a mild decrease of 2x between ages 25–35 but otherwise remained stable from ages 10–85. As a result, we denoted participants as immunologically young (<35 y.o.) or immunologically adult (≥35 y.o.). Repeating the consecutive vaccination analysis in both groups showed that in adults, weak-weak responses became dominant even earlier in 2019–2020, while strong-strong responses have remained dominant in the young but are also trending towards a dominant weak-weak phenotype (Figure S4).

To further investigate whether the decline in antibody titers can be attributed to repeated influenza vaccination, we analyzed individuals that received a vaccine in three or more consecutive seasons. Antibody responses were stratified by immunological age group and vaccine component. Participants were further subset based on the first year they participated in these consecutive vaccine studies to quantify seasonal variation (Figure 2C, Methods).

Individuals showed highly comparable dynamics across all age groups and number of consecutive vaccinations, suggesting that variability in antibody titer was primarily due to the influenza season and the vaccine strain considered. For example, H1N1 GMTs for adults in 2019 ranged between 20–36 regardless of whether this was their 1st (teal), 3rd (purple), or 4th (dark blue) vaccination (Figure 2C).

A similar pattern was observed across all age groups and vaccine components, with no clear patterns to the deviations. For example, while H1N1 titers in the young during 2017 or 2018 appeared to be higher in the year they started their consecutive vaccinations, this pattern did not hold up for 2019 or 2021, nor for other vaccine strains (Figure 2C). Altogether, these data indicate that while repeated annual vaccination is associated with broad trends of declining titers at the population level (Figure 2A), seasonal factors, rather than the cumulative number of doses, appeared to be the primary driver of variability (Figure 2C).

Long-term antibody titers for all vaccine strains tend to remain within 25% of baseline

Despite the pressing need to elicit durable immunity, few studies directly measure the long-term antibody response. However, a number of studies included repeat participants that were vaccinated over multiple seasons whose long-term decay could be assessed. We examined two such groups of individuals: 1) participants vaccinated in two consecutive seasons and 2) participants who skipped vaccination for one season but were vaccinated in the preceding and following seasons. In the first group, pre-vaccination titers in the second season represent day 365 titers from the prior season. In the second group, the pre-vaccination titers after the skipped season represent the day 730 post-vaccination titers from the previous vaccine. In both cases, day 0 represents the pre-vaccination titer from the prior season, and later time points are always measured against the day 0 vaccine strain even if the vaccine is updated in later seasons (Figure 3A).

Figure 3. Substantial effects from vaccine strain, vaccine formulation, and vaccination history on the magnitude and durability of antibody titers.

(A) Long-term titers of individuals vaccinated in two consecutive years (left panel) and individuals that skipped a year in between vaccines (right panel). Antibody titers were measured at days 0, 28, 365, and 730 against the same day 0 variant, with GMTs shown across all participants and seasons. Vertical bars denote the 95% confidence intervals, and x coordinates are jittered for clarity. (B) Boxplots comparing post-vaccination antibody titers and fold-change from baseline for individuals with different vaccination histories over the preceding three years (0=not vaccinated, 1=vaccinated, tuples represent vaccine history [3 years ago, 2 years ago, 1 year ago]). GMTs are averaged across all seasons. (C) Density plots of the log₂ post-vaccination titers (titer=5 → 0, titer=10 → 1…) of immunologically young (<35 y.o.), and middle-aged adults (35–64 y.o.). Titers are shown for participants who received Fluzone or Flucelvax [IIV] in consecutive years (blue), no vaccination [Skip] followed by IIV (teal), or FluMist [LAIV] followed by IIV (red). The symbols indicate a statistically significant difference compared with the IIV-IIV group (†), the Skip group (*), or the LAIV group (^) using a Mann-Whitney U test. The number of measurements (n) is shown in each panel.

At day 365, antibody titers relative to day 28 markedly differed across vaccine strains, with a slower decrease for B Yamagata and B Victoria that reached 58% and 48% of their day 28 titers compared to 36% and 39% for H1N1 and H3N2. Relative to day 0, all four vaccine components remained within 15% of baseline values, with mild increases seen for both B viruses and H3N2. Thus, antibody responses show little-to-no durability at the population-level.

At day 730, titer levels relative to day 28 moderately decreased (B Yamagata = 52%, B Victoria = 32%, H1N1 = 21%, and H3N2 = 27%). Relative to day 0, the H1N1, H3N2, and B Victoria titers remained within 25% of their baseline values, although B Yamagata showed a 68% elevation in titers at day 730 for this smaller set of participants. Hence, B Yamagata may exhibit a distinct durability compared to the other vaccine strains, although this was not seen in the far larger sampling of day 365 responses.

Individuals vaccinated in the prior season elicit 2x smaller post-vaccination titers

Although repeated vaccinations have been shown to attenuate immune responses,^9,23–26 it is unclear how many years of prior vaccinations are informative nor whether more recent years have a larger impact. Across all datasets, vaccination history was often available for the past three years, which we encoded as a binary string showing either no vaccination ‘0’ or vaccination ‘1’ (e.g., ‘011’ means no vaccine 3 years ago followed by vaccination 2 years ago and also 1 year ago).

Individuals with no vaccinations in the preceding three years (‘000’) had 2.4x larger post-vaccination titers and 3.0x larger fold-change than individuals vaccinated consecutively for all three years (‘111’) (p<0.05, Mann-Whitney U test) (Figure 3B). By considering all vaccination paths, the most recent season had the strongest effect, with individuals not vaccinated in the prior season (‘100’, ‘010’, and ‘110’) displaying post-vaccination titers statistically comparable to the vaccination-naïve (‘000’) group. In contrast, participants with only one prior vaccine from the prior season (‘001’), showed a significant decrease in titers compared to the naïve group (p<0.05, Mann-Whitney U test). These results suggest that the prior year’s vaccination status is the primary determinant of this effect, while the total number of vaccinations in the past three years has a smaller impact. This phenomenon was highly robust and persisted after stratifying by age, season, pre-vaccination titer, and vaccine strain (Figure S5), and is therefore unlikely to be a statistical artifact.

Two types of “skipped” vaccination improved the subsequent vaccine response

Skipping vaccination in the prior season and then receiving an inactivated vaccine (Skip-IIV) led to significantly larger post-vaccination titers in 45%=8/18 of seasons and immunological age groups, compared to individuals vaccinated in both seasons (IIV-IIV) (Figure S6). Prior work suggested that antibodies may mediate this effect by blocking the new vaccine’s antigen or by focusing on the response towards previously-seen epitopes.^27–30

If this is an antibody effect, then we hypothesized that another way to “skip” vaccination is to receive FluMist, a live-attenuated vaccine (LAIV) administered as a nasal spray that typically does not elicit a systemic antibody response in adults with prior influenza exposure.^31–33 Thus, we expected individuals receiving FluMist in 2018 and IIV in 2019 (LAIV-IIV) would behave identically to the Skip-IIV group. Interestingly, while immunologically young participants (<35 y.o.) showed a statistically similar response from LAIV-IIV and IIV-IIV, middle-aged adults (35–64 y.o.) had significantly higher titers from LAIV-IIV than from IIV-IIV or Skip-IIV, suggesting that FluMist enhanced the subsequent IIV response (Figure 3C), consistent with prior observations.^33,34 Children in the subsequent 2019–2020 seasons showed significantly larger titers from LAIV-IIV than IIV-IIV or Skip-IIV (Figure 3C, LAIV-IIV not given to adults in 2019–2020). While preliminary, these results showed enhanced responses in 67%=2/3 of the seasons and age groups examined.

Viral antigen concentration minimally impacted antibody titers across age groups

Different influenza vaccine formulations have been designed to elicit stronger antibody responses and greater protection by varying the antigen dose or the vaccine platform.^35,36 For example, compared to standard dose Fluzone (15μg/antigen, hereafter called “Fluzone”) or the adjuvanted vaccine Fluad (15μg/antigen), the Flublok vaccine has 3x antigen content (45μg/antigen) while Fluzone High-Dose has 4x antigen content (60μg/antigen). Fluad also includes the adjuvant MF59 to enhance the immune response. We assessed 15,000 antibody titer measurements from participants receiving one of these vaccines (Figure 4).

Figure 4. Antibody titers as a function of antigen dose.

Post-vaccination titers for (A) all seasons or only the (B) 2018 or (C) 2019 season. Lines show GMTs with 95% confidence intervals to standard dose Fluzone (15μg/antigen, teal), Fluad (15μg/antigen + adjuvant MF59, purple), Flublok (45μg/antigen, blue), or Fluzone High-Dose (60μg/antigen, grey), with participants dynamically binned by rounding their age down to the nearest decade so that each point represents at least 10 participants (Methods). The number of measurements (n) is color-matched for each age group and vaccine dose, with significance marked by stars with the corresponding brand color.

When combining data from all seasons, the most striking feature of these responses was their similarity. Across most age groups, there was a small but in some cases significant difference in antibody titers between the three antigen concentrations. With the notable exception of 20–29 y.o. (n=42), GMTs in each age group were always ≤2x different between any two vaccines (Figure 4A). These differences were also impacted by seasonality, with more recent vaccines eliciting slightly smaller antibody titers (Figure 4B,C, Figure S7).

Higher vaccine doses had mixed effects. In two seasons (2015, 2018), Fluzone High-Dose titers were significantly higher (absolute effect: ≥1.5x larger GMTs) than Fluzone in all age groups ≥60 years old. In six other seasons (2014, 2016, 2017, 2020, 2022, 2023), high dose was either comparable or only slightly better than standard dose in these age groups (<1.4x larger GMTs). In two seasons (2019, 2021), high dose elicited slightly but significantly smaller titers than standard dose in the 60–69 age range (0.65x smaller GMTs), with poorer performance also seen in the 70–79 age range in 2019.

While Fluzone High-Dose is only approved for individuals ≥65 y.o., our studies included two clinical trials in 2017 and 2019 that assessed Fluzone High-Dose responses in younger age groups.^33,37 In both seasons, standard- and high-dose vaccines often elicited comparable titers across all age groups. A notable exception was that participants 20–29 y.o. in 2017 (but not 2019) showed significantly higher titers with Fluzone High-Dose, suggesting that young adults may benefit from higher antigen content.

Flublok also had mixed performance across seasons, eliciting smaller titers in 2016, 2017, and 2021 in at least one age group while eliciting slightly but consistently higher titers in 2018–2019 and dramatically higher titers in 2023 for 20–30 y.o. (Figure S7). In most seasons and age groups, Flublok titers were not significantly different from Fluzone. Fluad titers were only available in 2018 and 2023, yet in both seasons its titers were comparable to Fluzone High-Dose in all age groups.

Multiple influenza vaccines elicit similar antibody titers for influenza A and B

We next expanded this analysis to compare the titers from multiple influenza vaccine brands to identify which formulations elicited the strongest antibody responses (Figure 5). Most vaccine brands elicited comparable titers, with ~2x fold-change. However, a few notable exceptions were observed.

Figure 5. Pre- and post-vaccination titers against each vaccine component and vaccine brand.

Box plots comparing (A) pre-vaccination titer, (B) post-vaccination titer, and (C) fold-change across all seasons. The numbers in Panel A show the number of measurements (n) and number of studies (S) analyzed. The horizontal dashed line in Panel C indicates a fold-change of 1, corresponding to no change in titer. Box plots show the interquartile range, horizontal lines the medians, black diamonds the GMTs, and whiskers extend to 1.5 times the interquartile range.

First, FluMist showed no measurable increase in serum antibody titers with GMT fold-change consistently around 1 (Figure 5C), even though these participants had very low pre-vaccination titers, so any other vaccine would have likely elicited an excellent response. Second, Vaxigrip elicited an unusually high fold-change (~8x GMT), although these responses were from a single study of adults in Vietnam that had never received a prior influenza vaccine. Third, Fluad showed elevated fold-changes for H1N1 and B Victoria, which appeared to be driven by unusually low pre-vaccination titers. The Fluad H3N2 pre-vaccination titers, which were more comparable to the other vaccine brands, resulted in the same ~2x fold-change as other brands. Finally, the exceptionally high Fluad pre-vaccination titers against the B Yamagata lineage likely decreased its fold-change due to the antibody ceiling effect where larger baseline immunity leads to diminished fold-change.^38,39

In summary, vaccine responses consistently showed 2–3x fold-change across all vaccine brands, with the key exceptions arising from: 1) FluMist that elicited little-to-no vaccine response, 2) large vaccine responses for adults receiving their first influenza vaccine, and 3) cases where pre-vaccination titers were substantially smaller/larger than normal, leading to larger/smaller fold-change.

Minor sex differences were only seen against influenza A viruses in the elderly

Overall, sex was not a major determinant of post-vaccination antibody titers, with minimal differences between males and females across vaccine strains and age groups that were consistently <2x (Figure 6A,B). However, following prior work on frequently vaccinated individuals,⁴⁰ a modest sex difference emerged in elderly individuals (age ≥65) who received multiple consecutive vaccinations. In this subgroup, females exhibited slightly higher antibody titers than males against influenza A viruses (Figure 6C). This difference only reached statistical significance during the third year of consecutive vaccination against H3N2 (p=0.02, Mann-Whitney U test), yet in absolute magnitude women only had 1.5-fold larger titers. No significant difference was observed after 4 or 5 consecutive vaccinations, nor for H1N1 or either influenza B virus, suggesting that sex-based differences in antibody titers are very minimal, although age and repeated vaccination may reveal subtle, strain-specific effects.

Figure 6. Minimal instances differing post vaccination antibody titers across sex.

Post vaccination titers of males and females for each (A) vaccine component and (B) age. Box plots show the interquartile range, horizontal lines the medians, black diamonds the GMTs, and whiskers extend to 1.5 times the interquartile range. The numbers at the top show the amount of participants (n) and studies (S) included in this analysis. (C) Lines show GMT of post vaccination titers for groups of elderly (age ≥65) females and males that received 1 to 5 consecutive vaccinations. Colored numbers above each column denote the number of participants receiving their k^th vaccine, with stars indicating when the sex difference is significant (Mann-Whitney U rank test).

Vaccinating later in the season elicited stronger antibody responses

To complement the known determinants of immunity described above, we also used an AI-guided discovery engine⁴¹ that aimed to find groups of features associated with the strongest and weakest post-vaccination titers or fold-change. Extreme subgroups were identified based on four patterns. The first two patterns involved age and pre-vaccination titers, which were described above and mildly affected antibody fold-change (Figure S8A,B). Interestingly, the combination of age and pre-vaccination titers led to larger effects, where participants ≤35 y.o. with low pre-vaccination titers achieved fold-change≈8x while older adults with low pre-vaccination titers only had 2–4x fold change (Figure S9).

The third AI-guided pattern focused on BMI, which in prior influenza and SARS-CoV-2 studies led to faster antibody waning and higher risks of hospitalization or death^42–45 In our cohorts, BMI did not show a noticeable effect on the antibody response, suggesting that the AI method overfit on this parameter (Figure S8C).

The final pattern revealed a broad and robust pattern linked to vaccine timing, where the ~10% of participants vaccinated between early January and mid-February exhibited higher post-vaccination antibody titers than those vaccinated earlier in the season. Extending to the full range of vaccine timing (September-February) showed a consistent increase in fold-change, with a ~2x increase in geometric mean fold-change for those vaccinated in February relative to September (Figure 7A).

Figure 7. Antibody titers through the season.

(A) Fold-change and (B) pre-/post-vaccination titers of individuals vaccinated on different days of the year for northern hemisphere studies. The pre-vaccination, post-vaccination, and fold-change points are all shown at the x coordinate corresponding to the day of vaccination. Solid lines show the GMT of individuals at each day of the year with reported measurements. Shades represent the 95% confidence intervals. The number of measurements (n) and studies (S) are shown for each month.

Importantly, pre-vaccination titers remained largely stable, while post-vaccination titers progressively increased with later vaccination dates, suggesting that this effect is not caused by influenza infections priming the immune response (Figure 7B). These results held when stratified by age, season, or the time since last vaccination (Figure S10A–C). Increasing pre-vaccination titer led to a smooth and consistent decrease in fold-change (Figure S10D), as expected from the antibody ceiling effect that was clearly seen across studies (Figure S11). Thus, subsetting to participants with pre-vaccination titers≤10 revealed an even larger ~5x increase in fold-change when vaccinated in February relative to September. Overall, these results suggest that vaccine timing can substantially affect immunogenicity.

Discussion

By integrating influenza serology data from more than 4,500 individuals across multiple seasons and study designs, this work provides an unusually broad view of how host, vaccine, and seasonal factors shape antibody responses. The following sections highlight the insights learned from analyzing influenza vaccine immunogenicity at scale.

Seasonality and the impact from COVID-19

The largest effect on the influenza antibody response was from season-to-season variation. While we expected that virus evolution and updates to the vaccine strains would yield highly variable antibody responses across seasons, we instead found a smooth and progressive decline in post-vaccination antibody titers. From 2017–2023, pre-vaccination titers decreased by ~4x while post-vaccination titers decreased by ~3x, with this process beginning before the COVID-19 pandemic and becoming more pronounced from 2020–2023. As a result, unvaccinated individuals or those who vaccinate late in the season are expected to have reduced protection in recent years. This may have contributed to the large number and severity of infections in 2024, where the rate of hospitalizations was the highest observed since 2010.⁴⁶

These low pre-vaccination titers were most pronounced for H1N1, followed by H3N2 and B Victoria. This hierarchy of decline was unexpected, as it does not mirror known antigenic and genetic evolutionary rates: H3N2 undergoes the fastest antigenic drift and highest HA substitution rate,^47,48 yet H1N1 titers declined most rapidly, reaching especially low levels (GMT ≤15) as early as 2020. This suggests that background immunity is not solely determined by viral antigenic drift.

We hypothesize that low levels of influenza infections during COVID-19 (2020–2021) drove the decrease in titers from 2021–2023, whereas the small decrease from 2017–2020 represents typical seasonal variation due to noise or vaccine strain selection (e.g., pre-vaccination titers in 2019–2020 were comparable to pre-vaccination titers in 2010 or 2014–2016). Future studies will reveal whether this downward pattern continues, although we expect that antibody responses will rebound given the resurgence of influenza after 2021.

Distinct response to B Yamagata

B Yamagata exhibited a strikingly different behavior than the other vaccine components, where its pre- and post-vaccination titers remained unexpectedly stable and even increased in 2023 after it went extinct. This raises the question: Why did B Yamagata pre-vaccination titers not decrease from 2021–2023 when there were few influenza cases?

Corroborating evidence of B Yamagata durability is seen from year two of the DRIVE study, where the placebo group that received no vaccination in 2020 or 2021 nevertheless had stable HAI GMT≈40 for B Yamagata pre- and post-vaccination in both years (compared to GMT≲10 for H1N1 and H3N2 and GMT≈20 for B Victoria).⁴⁹ Given that: 1) there were few influenza infections in Hong Kong during 2020–2021, 2) the inclusion criteria for this study required no vaccination during 2018 or 2019, and 3) vaccine coverage for adults in Hong Kong is generally low, these elevated B Yamagata titers have persisted for multiple years.

One explanation is that some B Victoria antibodies can cross-react and target B Yamagata,^50,51 although we did not find a strong correlation between these two vaccine components. A second explanation is that vaccination elicits especially durable B Yamagata antibody titers. Despite its strategic importance for a universal influenza vaccine, vaccine-elicited durability remains understudied.⁵² The few prior works suggest that after reaching the peak antibody response 1-month post-vaccination, H1N1 and H3N2 HAI titers will drop by 2–4x by 12 months^9,53 while H1N1 binding titers drop by 2–4x by 18 months for HA stem-targeting antibodies.⁵⁴

Our work extended these efforts, suggesting that relative to 1-month post-vaccination, B Yamagata titers only decay by ~2x and maintain those levels for at least 24 months, whereas all other vaccine strains decayed by ≥3x at 24 months. When compared to pre-vaccination titers, B Yamagata appeared to show durable fold-change=1.7x (95% CI: 0.9–3.3x) at 24 months (n=20), yet we found a smaller and tighter fold-change=1.2x (95% CI: 1.1–1.2x) at 12 months with a far larger set of individuals (n=1500). Taken together, this suggests that current seasonal vaccines do not elicit a durable antibody response at the population-level, and hence does not explain the steady B Yamagata titers in 2021–2023.

We instead speculate that infections elicit antibody titers that last multiple years, and that periodic reinfections help maintain the pre-vaccination titers ≈40 observed before 2020. With the absence of influenza infections in 2020 and 2021, this long-term immunity started to decline, revealing the baseline antibody levels. B Yamagata titers remained elevated because it has evolved more slowly both genetically⁴⁷ and antigenically (Figure S2) than the other three vaccine components, with all B Yamagata variants since lineage introduction in 1988 until its 2021 extinction effectively representing the same strain. Thus, each B Yamagata exposure matched the imprinted strain first encountered throughout life and elicited an especially strong and durable response. In contrast, H1N1, H3N2, and B Victoria have all antigenically evolved, and hence each exposure elicits a less durable response (Box 1).

Box 1. Rationale for Decreasing Influenza Antibody Titers from 2020–2023 and B Yamagata Stability.

1. Pre-vaccination titers represent long-term immunity primarily conferred by infections.

   • In most vaccine studies, titers return to baseline at 1-year post-vaccination.

   • The two exceptions were adults given an influenza vaccine for the first time in their life,^49,53 where fold-change was ≥8x at 1-month and ≥4x at 1-year post-vaccination in most subjects. These adults likely had prior influenza infections, suggesting that prior infections do not hinder this first durable vaccine response.

2. The lack of infections in 2020–2021 led to waning long-term immunity, revealing the baseline antibody titers against the current vaccine strains. Titers should therefore rise in the coming years.

3. B Yamagata titers have stayed high because it mutates slower than H1N1, H3N2, and B Victoria.

   • We hypothesize that all B Yamagata variants are antigenically related, and that receiving any B Yamagata infection or vaccination yields an especially strong recall response.

   • In contrast, H1N1, H3N2, and B Victoria have all antigenically drifted, with the variants encountered in early life different from those in the 2020–2023 vaccines (that were used to assess pre-vaccination titers).

Indeed, experiments in mice have shown that exposure to similar H1N1 influenza strains preferentially boosts antibody titers to the prior strain, while exposure to antigenically different strains predominantly elicits titers against the new variant.⁵⁵ Another intriguing point of comparison is the H2N2 subtype that circulated for 12 years (1957–1968), slowly evolving until it was replaced by H3N2. Individuals born during or before the 1950s still had inhibiting H2N2 A/Singapore/1/1957 antibodies (HAI GMTs≳40) more than 40 years after its extinction, while people born after 1968 had GTM=5.^56–58 These results further support that the first encounter to a slowly-evolving subtype leads to highly durable immunity, and this statement may hold for other pathogens beyond influenza.

Looking ahead, the elimination of B Yamagata provides a rare natural experiment to examine immune memory in the absence of both infection (since its extinction in 2021) and vaccination (since 2024 when vaccines became trivalent). Future studies should examine how long B Yamagata titers remain elevated in children versus adults, as well as from infection- or vaccination-primed individuals. Based on the H2N2 data, we predict that the B Yamagata titers will remain elevated for multiple decades.^57,58

Age effects

Given the extensive literature on the effects of age on the antibody response,^{4,6–8,59–61} it was surprising to find that age had a mild influence on the vaccine response. Elderly individuals had ~2x lower post-vaccination titers than children, with the transition happening between ages 25–35. No further drop in antibody titers was seen for adults 35–64 y.o. or elderly individuals ≥65 y.o., suggesting that the higher mortality in this later group is not driven by lower antibody titers. Although age effects were mild, age significantly affected the capacity to elicit large post-vaccination titers when pre-vaccination titers are low.

Vaccine dose and adjuvants

Fluzone High-Dose did not always elicit a substantially stronger antibody response in the elderly (≥65 y.o.) than standard dose Fluzone. Sanofi’s 2006 clinical trial found an average of 1.6x (range 1.3–1.8x) higher HAI titers post-vaccination from Fluzone High-Dose than Fluzone across all vaccine components.⁶² A 2012–2013 follow up trial by Sanofi found an average of 1.7x (range 1.4–2.0x) higher HAI titers from high dose.³⁶ In contrast, other studies showed more modest increases of <1.5x in HAI titers from high dose vs standard dose vaccines.^63,64 Dose-independence was also found in a recent phase 1/2 clinical trial on influenza mRNA vaccines where the standard dose (10μg/antigen) and higher dose (12μg/antigen) mRNA-1012 constructs elicited comparable fold-change, with the higher dose exhibiting smaller but not statistically significant responses.⁶⁵ Compared to standard dose, we found that high dose elicited 1.38x larger antibody titers for ≥60 y.o. (95% CI: 1.11–1.66x), and 1.70x for all ages (95% CI: 1.15–2.25x), across all seasons. Similarly, Flublok elicited 1.76x larger antibodies titers (95% CI: 1.05–2.47) and Fluad elicited 1.60x larger titers (95% CI: 0.23–2.96x) than standard dose across all ages and seasons.

Ultimately, these antibody effects must not only be statistically significant, but clinically relevant, which is why we quantified the absolute size of each effect. Although our analysis was limited to antibody titers, the prior 2012–2013 clinical study demonstrated that high dose vaccines reduced laboratory-confirmed influenza to 1.4%, providing slightly better protection compared with the 1.9% infection rate seen with standard dose.³⁶

Recently, these results have been extended to hospitalizations and all-cause deaths in elderly individuals, where results have been mixed. Among 812,000 participants, the adjusted relative vaccine effectiveness for influenza-related hospitalizations from high dose vs standard dose was 27% (95% CI: −12–61%) in 2022 and 7% (95% CI: −36–42%) in 2023.⁶⁶ In terms of hospitalizations for influenza or pneumonia, one recent study assessed 134,000 elderly individuals and found a smaller hospitalization risk for high dose (0.26%) than the standard dose influenza vaccine (0.34%), yet there were too few cases for the results to be statistically significant.⁶⁷ A second study of 332,000 elderly participants also found a smaller but not significant hospitalization risk for high dose (0.68%) than standard dose (0.73%), and comparable risks of all-cause death (0.67% vs 0.66%).⁶⁸ Despite these two latter studies not reaching significance in their primary endpoints for influenza and pneumonia hospitalizations, the absolute hospitalization risk from influenza alone substantially decreased from 0.14%→0.09%⁶⁷ and 0.11%→0.06%⁶⁸ from standard dose to high dose recipients.

Most influenza vaccine brands behave similarly, except FluMist

The vaccine brands used in these studies spanned several distinct technologies.^2,69 Fluzone is an inactivated vaccine (containing both external and internal viral proteins), Flublok is recombinant (containing only hemagglutinin), and Fluad contains an adjuvant along with influenza surface proteins (hemagglutinin and neuraminidase, although only hemagglutinin content is standardized). Flucelvax uses cell-grown virus to avoid egg-adapted mutations. Each vaccine is administered intramuscularly, and only the hemagglutinin content is regulated. FluMist is a live attenuated nasal spray vaccine using hemagglutinin and neuraminidase from circulating strains coupled with internal proteins from H2N2 A/Ann Arbor/6/1960 and B/Ann Arbor/1/66 strains. In theory, these vaccines could trigger different immune pathways and lead to distinct antibody responses.

In practice, most vaccine brands elicited comparable antibody titers. One striking exception was FluMist, which aims to elicit mucosal responses (not examined in this work), but that consistently produced no measurable systemic antibody response across all cohorts (mean fold-change=1.2x). Most adult studies similarly report that FluMist elicits no systemic antibody response.³¹ Indeed, FluMist has no known correlate of protection, and it was licensed based on vaccine efficacy and not immunogenicity.³¹ Studies in young children have been mixed, with some reporting fold-changes ranging from 1–8x.^70–72 Recent work in 2–5 y.o. receiving their first influenza vaccine showed that only one third of children elicited a strong IgG response.⁷³ Our cohorts consistently showed no antibody response, even in children as young as 4 years old.

Aside from FluMist, especially large fold-change was observed for Vaxigrip and Fluad. The former vaccine (equivalent to Fluzone but produced by Sanofi for non-US markets) was given to a vaccine-naïve cohort, which likely led to its large fold-change. The latter vaccine was given to participants with unusually low pre-vaccination titers, resulting in large fold-change but comparable post-vaccination titers to other vaccines. Nevertheless, both brands are worth assessing in additional cohorts.

Taken together, these results suggest that vaccine brand has a subdominant effect compared to seasonality. As a rule of thumb, FluMist elicits no antibody response while every other vaccine elicits a geomean fold-change≈3x. From this vantage, the enhancing effect of a FluMist skip discussed below is even more surprising.

Vaccinations across consecutive seasons

Prior work has suggested that more frequently vaccinated individuals have worse antibody responses and smaller vaccine effectiveness.^9,26,74,75 In those studies, the biggest decline in antibody titers was seen between groups that received 0 vs 1 vaccinations in the past 5 years, with more gradual decreases seen between groups vaccinated 1–5 times in that same period.^9,26 This suggests a “start of study” effect, where participants entering an influenza vaccine study with no recent vaccination history will have a transient boost to antibody titers that will rapidly fade in subsequent years of the study.

Our cohorts showed that seasonality was a far stronger effect than vaccination history. When controlling for seasons, we only observed a start of study effect in the young (<35 y.o.) in 2018–2019 for H1N1 and only in 2018 for H3N2, but not in the other seasons (2019–2023), other vaccine components (B Victoria or B Yamagata), nor any adult responses.

Skipping vaccination or the FluMist-skip

We also found a small and sporadic boost to antibody titers when participants did not vaccinate in the prior season. Between seven sets of consecutive seasons (2016–2017 to 2022–2023) and three age groups (children, adults, elderly), there were 8/18 instances where skipped vaccination (Skip-IIV) led to significantly higher post-vaccination titers than two consecutive vaccination (IIV-IIV), although the absolute increase in GMT was only 1.41x in these 8 seasons and 1.35x across all 18 cases. Year two data from the DRIVE study corroborated these results, where groups receiving the Flublok (FB) vaccine or no vaccine (Skip) showed (post-vaccination GMT of FB-FB)/(post-vaccination GMT of Skip-FB) to H1N1 (0.73x), H3N2 (1.42x), B Vic (0.68x), or B Yam (0.84x) were all ≈1x.⁴⁹ Future years of this study will investigate whether having more skipped years will yield a larger effect.

A distinct form of “skipped” vaccination involved administering FluMist followed by an inactivated influenza vaccine. Because FluMist does not elicit a systemic antibody response, this strategy could offer a similar immunological reset while still providing protection. To our knowledge, this vaccination sequence has not been formally tested, yet a subset of individuals in our cohorts followed this sequence. Notably, in 2018–2019, adults who received FluMist followed by Fluzone exhibited significantly higher post-vaccination titers compared with those vaccinated with Fluzone alone (GMT: 166 vs 58). A similar pattern was observed in young individuals vaccinated with FluMist in 2019 and Flucelvax in 2020 (GMT: 340 vs 107), and further subsetting this group into younger (2–18 y.o., n=44) and older (18–34 y.o., n=40) age groups continued to show enhanced responses in each case with GMTs of 387 and 296, respectively. These results are also consistent with prior studies that also reported enhanced IIV responses following FluMist vaccination.^34,76 It is important to note that only 20 individuals received the FluMist→Fluzone sequence and 84 received FluMist→Flucelvax, and that differences were not significant in young individuals in 2018–2019. Thus, these findings are suggestive but not conclusive.

Sex differences

Across all cohorts, sex had little impact on antibody titers. Some prior studies showed modest sex differences where females had 1–2x larger titers than males,^40,77 while other work found no difference in antibody titers but saw sex differences in ELISPOT or NAI responses.^78,79 The only reproducible signal in our cohorts was a modest female advantage in elderly individuals repeatedly vaccinated against influenza A viruses, which reached statistical significance for H3N2 after exactly three consecutive vaccinations. The magnitude of this effect was modest (1.5-fold), and no similar pattern was seen for influenza B viruses. These subtle differences may reflect interactions between sex hormones, immune senescence, and repeated antigen exposure, but they do not appear to be a dominant determinant of influenza vaccine responsiveness.

Time of Year Vaccinated

A pattern from the discovery engine⁴¹ suggested that the day-of-year someone is vaccinated affects the magnitude of their vaccine response. Such a feature has been underreported and underexplored, and indeed, among our 64 datasets it was only available in the 8 UGA studies. Yet across those 4,700 vaccine responses, delayed vaccination in February led to ~2x larger fold-change than early vaccination in September (or ~5x larger fold-change for subjects with weak baseline titers). Since pre-vaccination titers remained flat across all months, this effect was not caused by infections inflating baseline immunity. Previous studies noted that getting vaccinated too early in the season can increase the risk of infection due to within-season antibody waning,^80–82 yet to our knowledge, increased titers from vaccinating later in the season have not been reported.

Mechanistically, we hypothesize that these changes in the immune response could be related to annual fluctuations in immune cells such as lymphocytes and neutrophils that prior reports have associated with daylength.⁸³ Although lymphocyte counts remain relatively stable from September through February, neutrophil counts show a marked increase, which may contribute to enhanced vaccine responses through their role in coordinating adaptive immune responses.⁸⁴ Future work should investigate such mechanisms, and also explore this effect in southern hemisphere cohorts.

Limitations

Our analyses were performed on pooled data from multiple studies. While this approach increases sample size and provides greater flexibility to examine subgroups, it also introduces heterogeneity arising from differences in study design. In addition, we focused exclusively on antibody responses that constitute a widely used correlate of protection, yet antibody titers do not fully capture clinical outcomes, hospitalizations, or mortality. Finally, the majority of data analyzed were derived from cohorts in the United States, which may limit the generalizability of our findings to populations with different demographic or epidemiological characteristics.

Translational applications

The progressive decline in baseline and post-vaccination titers underscores the need to understand why antibody levels fall across the population. The remarkable durability of B Yamagata responses demonstrates that such strong and lasting influenza immunity is possible, offering a benchmark for what effective influenza immunity could look like. Similarly, the observation that post-vaccination titers do not decline after age 65 suggests that the challenges of immunizing older adults may lie in immune mechanisms beyond antibody generation.

Vaccine timing emerged as a powerful yet underutilized lever to partially improve post-vaccination titers, suggesting that prospective trials are needed to determine whether timing recommendations could be optimized to align with anticipated influenza circulation. The timing of influenza seasons can be inferred from prior seasons, near-real-time surveillance of the current season, or forecasts of hospitalizations in the coming weeks.^85–87 Importantly, vaccinating early risks both smaller post-vaccination titers and within-season waning.^82,88

In addition, while skipped vaccinations slightly enhanced vaccine responses the following year in some age groups and some seasons, this effect was inconsistent and often mild, and it leaves the population unprotected during the skipped season. In contrast, 2 out of 3 cases where individuals were primed with FluMist in one season and given an inactivated vaccine (Fluzone or Flucelvax) the following season led to noticeable increases in the antibody response, suggesting that alternating across vaccine brands can yield better results. If some feature of FluMist durably enhances inactivated vaccines, that feature could potentially be incorporated into future vaccines.

Current and next-generation vaccine design often relies on introducing novel components (e.g., adjuvants, neuraminidase) to test their efficacy. An alternative approach is to identify unifying principles of the most potent or durable immune responses observed in large cohorts, and then leverage these principles to improve the vaccine. For example, we found far more variability within than across age groups, with similar results for participants given different vaccine doses or with different frequencies of vaccination. Future efforts may benefit from extreme-phenotype analysis that intensively characterize individuals with the strongest and weakest responses to uncover additional actionable immunological mechanisms that govern vaccine immunogenicity.

Methods

Datasets Analyzed

Prior vaccine studies in Table S1 have been previously described.^{33,37,38,53,89–95} The new vaccine studies introduced in this work were the 2019–2023 WHO-CBER studies run by the Center for Biologics Evaluation and Research within the FDA as part of the WHO influenza surveillance efforts using human sera from individuals vaccinated in the prior season.

For each of those WHO-CBER studies, the quadrivalent influenza vaccine from that season was administered (e.g., 2019 WHO-CBER administered the 2019–20 influenza vaccine). H1N1 and influenza B were always measured by HAI, while H3N2 was measured using a neutralization assay. Metadata and vaccine dose are supplied in the supplementary information file, where repeatedly vaccinated individuals have the same Base ID. (Note that WHO-CBER participants were always unique across years, and hence they have no Base ID values). Post-vaccination samples were targeted for 28 days post-vaccination. As is standard procedure, infants under 13 months old were given two full vaccine doses 28 days apart, with their post-vaccination sample collected 28 days after the second dose.

For the WHO-CBER studies, vaccine strains were either cell-grown (for Flucelvax) or egg-grown (for all other vaccine brands). In the supplementary information file, cell-grown viruses have “_Cell”, “_MDCK”, or “_SIAT”, while all other variants were egg-grown. The vaccine formulation LAIV corresponds to the FluMist brand, while IIV corresponds to Fluzone, Vaxigrip, Afluria, and Fluarix.

Vaccine Study Participants

The five new influenza vaccine studies presented in this work included 224 participants for 2019 WHO-CBER, 208 in 2020, 320 in 2021, 240 in 2022, and 560 in 2023. Each participant’s age, vaccine dose, and vaccine brand is provided in the supplementary information file, and no other metadata was collected. These studies used existing sera from prior vaccine studies, and hence they did not qualify as human subjects research and did not require an IRB.

Analyzing Antibody Titers

Antibody titers from HAI or neutralization were collectively analyzed to quantify how much the antibody repertoire inhibits or neutralizes the influenza virus. Both assays are done using 2-fold dilutions, so antibody titers can equal 10, 20, 40… Titers below the lowest dilution (<10) were denoted by titer=5.

In consecutive vaccination analysis (Figure 2A), each vaccine component is assessed separately, so that one participant can have up to four points in each column for every vaccine strain.

In the age analyses (Figures 4, 6B), points were first binned into decades (10–19, 20–29…80–89 years old), and then dynamically binned so that no bin had <10 points to ensure robust statistics. Any bin with fewer points was moved into its neighboring bin with the least number of points, with this process repeated until all bins had ≥10 points. Few participants were <10 or ≥90 years old, and these were discarded from the age analysis to ensure robust statistics, but these participants are included in the supplementary information file.

For the date-of-vaccination analyses (Figure 7), each of the 365 days of the year constituted one bin, and points were then dynamically binned until no bin had <10 points.

Hemagglutination Inhibition and Neutralization Protocols

Hemagglutination inhibition assays were performed using standard methods. Each volume of serum was treated with 3 volumes of receptor destroying enzyme (Accurate Chemical, Westbury, NJ, USA) at 37 °C overnight, which was then inactivated by heating in a 56°C water bath for 30 min. Six volumes of PBS were added to make the final serum concentration of 1:10. Sera were diluted in a series of 2-fold serial dilutions in 96-well plates (Thermo Fisher). An equal volume of influenza virus, adjusted to 8 hemagglutination units (HAU)/50μL diluted in 1x PBS, was added to each well of the plate. The plates were then mixed by gentle agitation, covered, and allowed to incubate for 30 minutes at room temperature. Freshly diluted (0.5%) turkey red blood cells (Lampire Biologicals) for H1N1 and B antigens, were added to all wells and the plates incubated for an additional hour. After incubation, the last dilution of sera that completely inhibited agglutination was recorded. The antibody titer was determined by taking the reciprocal dilution of the last well that contained non-agglutinated red blood cells.

An enzyme-linked immunosorbent assay–based microneutralization assay was performed to quantify antibody titer toward H3N2 viruses. The same sera from the hemagglutination inhibition assay, 1:10 diluted and treated with receptor destroying enzyme, were used for the neutralization assays. Sera were incubated with virus titrated to one hundred 50% tissue culture infectious dose at 37°C with 5% carbon dioxide for 1 hour, and then were added to 1.5×10⁵ MDCK-SIAT1 cells (Sigma). After overnight incubation, infected cells were detected using influenza A nucleoprotein-specific monoclonal antibodies (Millipore). Neutralization titers represent the reciprocal of the highest serum dilution resulting in ≥50% neutralization.

Vaccine History

For the consecutive season analysis (Figure 2), each trajectory begins in the first season when a participant was enrolled in an influenza vaccine study and continues through subsequent seasons as long as the participant either participated in a vaccine study or self-reported as receiving an influenza vaccine. Self-reported vaccinations were recorded for up to three years prior to study enrollment. When a participant reported vaccination but did not participate in a vaccine study, antibody titers were unavailable and therefore excluded from the GMT calculations. In cases where an individual had two independent blocks of ≥3 consecutive seasons (e.g., three seasons vaccinated, a skip year, and then three more seasons vaccinated), each block was treated as a separate trajectory defined by the first year of the series of consecutive vaccinations.

When linking responses across multiple influenza seasons (Figure 3A,C), antibody titers were always assessed against the same day 0 influenza strain, even if the vaccine strain was updated. For example, if a person is vaccinated in 2017 with H3N2 A/Hong Kong/4801/2014, and then in 2018 with H3N2 A/Singapore/INFIMH-16–0019/2016, we used their pre-vaccination antibody titer in 2018 against H3N2 A/Hong Kong/4801/2014 to quantify their day 365 response in 2017. In other words, durability relied on repeated participation in a vaccine study as well as that study measuring prior influenza variants when the vaccine strain changes.

AI-Guided Discovery Engine

The discovery engine⁴¹ is a data-driven tool that combines machine learning’s pattern-recognition capabilities with state-of-the-art interpretability methods^96,97 to illuminate model decision-making. This allows extraction of human-understandable patterns from complex data at unprecedented speed.

The system automatically preprocesses data and evaluates a range of models, from simple statistical approaches to deep learning. The best-performing models were probed with interpretability methods to elicit patterns. In this study, the goal was to predict day 28 HAI titers using features such as day 0 titers, age, sex, BMI, vaccine dose, and the day-of-year vaccinated. A fully connected neural network achieved the best performance and extracted feature combinations that maximized day 28 HAI titers.

Data and Code Availability

For reviews: All data is provided as a supplementary information file.

When the manuscript is accepted: All data and code will be made available at https://github.com/TalEinav/DeterminantsOfInfluenzaImmunity.