(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Modeling and predicting individual variation in COVID-19 vaccine-elicited antibody response in the general population [1]

['Naotoshi Nakamura', 'Interdisciplinary Biology Laboratory', 'Iblab', 'Division Of Natural Science', 'Graduate School Of Science', 'Nagoya University', 'Nagoya', 'Yurie Kobashi', 'Department Of Radiation Health Management', 'Fukushima Medical University School Of Medicine']

Date: 2024-07

Abstract As we learned during the COVID-19 pandemic, vaccines are one of the most important tools in infectious disease control. To date, an unprecedentedly large volume of high-quality data on COVID-19 vaccinations have been accumulated. For preparedness in future pandemics beyond COVID-19, these valuable datasets should be analyzed to best shape an effective vaccination strategy. We are collecting longitudinal data from a community-based cohort in Fukushima, Japan, that consists of 2,407 individuals who underwent serum sampling two or three times after a two-dose vaccination with either BNT162b2 or mRNA-1273. Using the individually reconstructed time courses of the vaccine-elicited antibody response based on mathematical modeling, we first identified basic demographic and health information that contributed to the main features of the antibody dynamics, i.e., the peak, the duration, and the area under the curve. We showed that these three features of antibody dynamics were partially explained by underlying medical conditions, adverse reactions to vaccinations, and medications, consistent with the findings of previous studies. We then applied to these factors a recently proposed computational method to optimally fit an “antibody score”, which resulted in an integer-based score that can be used as a basis for identifying individuals with higher or lower antibody titers from basic demographic and health information. The score can be easily calculated by individuals themselves or by medical practitioners. Although the sensitivity of this score is currently not very high, in the future, as more data become available, it has the potential to identify vulnerable populations and encourage them to get booster vaccinations. Our mathematical model can be extended to any kind of vaccination and therefore can form a basis for policy decisions regarding the distribution of booster vaccines to strengthen immunity in future pandemics.

Author summary This study investigates the dynamics of antibody responses following COVID-19 vaccination, with the aim of elucidating individual-level variability in immune responses. Using a mathematical model and longitudinal antibody measurements from a vaccination cohort in Fukushima, Japan, we reconstructed the time course of antibody dynamics after vaccination. The study showed that not everyone’s immune system responds equally to the vaccine. Some people may have lower antibody levels after vaccination, which could make them more susceptible to disease. We identified key factors that influence antibody responses, such as age, adverse reactions, comorbidities and medication use, and developed a personalized antibody score to predict individual antibody levels. These factors play a critical role in shaping the immune landscape following vaccination, highlighting the need for tailored approaches to assessing vaccine efficacy and recommending booster doses for vulnerable populations. The development of a personalized antibody score provides a practical tool for healthcare professionals to stratify individuals based on their predicted antibody levels, facilitating targeted interventions to enhance immune protection. The study underscores the importance of considering personal characteristics when assessing vaccine efficacy and suggests potential applications in guiding booster vaccination strategies.

Citation: Nakamura N, Kobashi Y, Kim KS, Park H, Tani Y, Shimazu Y, et al. (2024) Modeling and predicting individual variation in COVID-19 vaccine-elicited antibody response in the general population. PLOS Digit Health 3(5): e0000497. https://doi.org/10.1371/journal.pdig.0000497 Editor: Dukyong Yoon, Yonsei University College of Medicine, REPUBLIC OF KOREA Received: February 2, 2023; Accepted: February 14, 2024; Published: May 3, 2024 Copyright: © 2024 Nakamura et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The research ethics review application form submitted to and approved by the Ethics Committee of Hirata Central Hospital (number 2021-0611-1) and Fukushima Medical University School of Medicine (number 2021-116) states that the anonymized clinical dataset can only be accessed by the research groups of this study, as it contains potentially identifying and sensitive patient information. In accordance with the PLOS Data Policy, which provides the option of de-identifying human research participant data and adding noise to protect participant privacy, we provide a synthetic patient dataset created using CTGAN (https://github.com/sdv-dev/CTGAN) based on the original dataset (S1 Data). This synthetic patient dataset, while not containing personally identifiable information, retains the same statistical properties as the original and can be analyzed in the same manner as the original. Funding: This study was supported by Medical & Biological Laboratories Co., Ltd., Shenzhen YHLO Biotech Co., Ltd., the distributor and manufacturer of the antibody measurement system (iFlash 3000), Research Center for Advanced Science and Technology in the University of Tokyo, and in part by Scientific Research (KAKENHI) B 18H01139 (to S.I.), 16H04845 (to S.I.); Scientific Research in Innovative Areas 20H05042 (to S.I.); AMED Strategic International Brain Science Research Promotion Program 22wm0425011s0302 (to K.A.); AMED JP22dm0307009 (to K.A.); AMED CREST 19gm1310002 (to S.I.); AMED Development of Vaccines for the Novel Coronavirus Disease, 21nf0101638h0001 (to M.T.), 21nf0101638s0201 (to S.I.); AMED Japan Program for Infectious Diseases Research and Infrastructure, 22wm0325007h0001 (to S.I.), 22wm0325004s0201 (to S.I.), 22wm0325012s0301 (to S.I.), 22wm0325015s0301 (to S.I.); AMED Research Program on Emerging and Re-emerging Infectious Diseases 22fk0108140s0802 (to S.I.); AMED Research Program on HIV/AIDS 22fk0410052s0401 (to S.I.); AMED Program for Basic and Clinical Research on Hepatitis 22fk0210094 (to S.I.); AMED Program on the Innovative Development and the Application of New Drugs for Hepatitis B 22fk0310504h0501 (to S.I.); AMED Research Program on Hepatitis 19fk0210036h0502 (to S.I.), 19fk0310114h0103 (to S.I.); JST MIRAI JPMJMI22G1 (to S.I.); Moonshot R&D JPMJMS2021 (to K.A. and S.I.) and JPMJMS2025 (to S.I.); The National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (2022R1C1C2003637) (to K.S.K.); The Shinnihon Foundation of Advanced Medical Treatment Research (to S.I.); SECOM Science and Technology Foundation (to S.I.); The Japan Prize Foundation (to S.I.); and Kowa Company, Ltd. (to M.T.). URL of funder websites: Medical & Biological Laboratories Co., Ltd. https://www.mblbio.com/e/ Shenzhen YHLO Biotech Co., Ltd. https://en.szyhlo.com/ Research Center for Advanced Science and Technology in the University of Tokyo https://www.rcast.u-tokyo.ac.jp/en/ KAKENHI https://www.jsps.go.jp/english/e-grants/ AMED https://www.amed.go.jp/en/index.html Moonshot R&D https://www.jst.go.jp/moonshot/en/ The National Research Foundation of Korea https://www.nrf.re.kr/eng/main The Shinnihon Foundation of Advanced Medical Treatment Research http://www.shinnihon-zaidan.jp/ The Japan Prize Foundation https://www.japanprize.jp/en/index.html Kowa Company, Ltd. https://www.kowa.co.jp/ Competing interests: YKaneko is employed by Medical & Biological Laboratories, Co. (MBL, Tokyo, Japan). MBL imported the testing material used in this research. YKaneko participated in the testing process; however, he did not engage in the research design and analysis. YKobashi and MT received a research grant from Pfizer Health Research Foundation for research not associated with this work.

Introduction The global distribution of vaccines for coronavirus disease 2019 (COVID-19) and the high vaccine potency and coverage will bring the pandemic caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) under control. As we learned from the COVID-19 pandemic, vaccination is an important part of a multi-faceted public health response to future pandemic illnesses. An unprecedentedly large volume of high-quality data have been accumulated during the COVID-19 pandemic, and we should use these data to prepare for future pandemics. In particular, given the limited global supply of vaccines during the early phase of a pandemic, determining vaccination priority is important to effectively distribute doses and achieve early disease control. In addition, waning of vaccine efficacy becomes a major concern during vaccination campaigns, as we observed for COVID-19 vaccinations [1,2]. Because a rapid decline in vaccine-elicited antibodies may result in breakthrough infections [3,4], additional, appropriately timed booster vaccinations will be required to maintain a high-level antibody response at both individual and population levels. Thus, to shape an effective vaccination strategy, it is important to quantify the individual-level time course of antibody dynamics and to identify “vulnerable” populations with sustained low antibody titers. Thus, by using a large volume of high-quality data on COVID-19 vaccination, our purpose in this study was to establish an approach to quantifying vaccine-elicited time-course antibody dynamics and predicting individual-level antibody responses with a simple and noninvasive method. To date, the findings of studies evaluating vulnerable populations for COVID-19 have been biased by the selection of study population. Rather than studying a general population, studies to date have focused on specific subpopulations stratified by variables like age, sex, lifestyle habits, comorbidities, adverse reactions, or medication use [2,3,5–8]. However, variation in vaccine-elicited antibody responses among healthy individuals per se has also been suggested to correlate with breakthrough infection risk [2,9–11]. In fact, although the heterogeneity in antibody responses over time is important for identifying the characteristics of vulnerable populations (in addition to standard risk factors such as age and comorbidities) [2,3,5–8], individual-level variation remains poorly understood. This is also true for viruses beyond SARS-CoV-2, because the dynamics of antibody responses in humans has not been described in detail for any vaccination. Here we used a mathematical model to describe the process of differentiation from naïve B cells to plasma cells to accurately reconstruct individual vaccine-elicited antibody dynamics. To fit the model, we used longitudinal antibody measurements from non-sequential and sequential blood sampling in the Fukushima vaccination cohort (a community-based cohort in Fukushima, Japan). Of note, because our vaccination cohort consisted of participants from a primarily rural area where the prevalence of COVID-19 was relatively low, the samples from this cohort are ideal for modeling vaccine-elicited antibody dynamics. Thanks to the reconstructed time-course of antibody dynamics, we were able to compare the antibody response at the same time points at an individual level. This overcomes another limitation of current vaccination studies, which can only directly compare antibody titers at a population level rather than at an individual level at specific time points (i.e., date on blood sampling) [2,3,5]. The model parameters describe highly variable individual-level antibody responses, allowing us to partially predict variation in vaccine response on the basis of personal information including age, adverse reactions, comorbidities, and medication use. Furthermore, we devised a personalized antibody score that aims to identify individuals with higher or lower antibody titers from their personal information. Although this score may not be able to detect all individuals, it has the potential to be used by medical practitioners to encourage individuals with low predicted antibody levels to get booster vaccinations. We stress that our approach will be easily applied to reconstruct antibody responses even after the third, fourth and fifth booster doses.

Discussion In this study, we created a personalized antibody score that can be used as a basis for identifying populations with low sustained antibody titers. To derive an optimal score, we used a mathematical model of antibody production in response to two-dose mRNA vaccinations and reconstructed the vaccine-elicited antibody dynamics of 2,407 participants from the Fukushima vaccination cohort. In particular, we highlighted that the reconstructed best-fit antibody titer curves perfectly predicted the additional timepoint data sampled from 120 of the 2,407 participants. Our mechanism-based mathematical modeling, in contrast to the statistical modeling used in recent reports [30,31], enables a biologically accurate description and precise comparison of antibody dynamics. The parameters of the estimated dynamics showed a large variation spanning two orders of magnitude. This variation was partially explained by individual characteristics like age, sex, the interval between the two vaccine doses, adverse reactions, comorbidities, and medications taken. This result is consistent with previous studies reporting age, sex, vaccine interval, and comorbidities as factors affecting antibody titers [5,32]. Our antibody scores can be easily calculated from individual demographic and health information, yet identify participants with high and low antibody titers (AUC of the IgG(S) titers) to a certain extent. Given the pleiotropic aspect of humoral immunity, it is surprising that a score consisting of several questionnaire items can provide predictions. The score showed that COVID-19 infection history and adverse reactions positively affect the AUC, whereas age and immunologically compromised conditions negatively affect the AUC. These positive and negative factors are consistent with previous studies [5,19,33–35]. In addition, autoimmune diseases have been reported to be risk factors for breakthrough infections [36,37]. Considering that individual antibody titers are partially predictive of the likelihood of breakthrough infections [11,38], this suggests that our antibody score has potential to be used as a risk score for breakthrough infections. On the other hand, the antibody score also has some similarities with COVID-19 severity scores [39–43]. In fact, both include age as a factor likely to lead to low antibody titers and critical illness, both of which may be related to a defective immune system, as observable in cytokine signatures [44,45] or immunoglobulin responses [46,47]. However, whereas COVID-19 severity scores use the results of laboratory tests and clinical symptoms to assess the patient’s condition in the hospital, our antibody score can be easily calculated on the basis of questionnaire responses provided by individuals (i.e., not limited to patients) themselves. So far, 13.01 billion doses of COVID-19 vaccines have been administered globally, and 2.51 million are now administered each day [48]. Whereas 68.5% of the world population has received at least one dose of a COVID-19 vaccine, 24.6% of people in low-income countries have received at least one dose as the result of disparities in global access to vaccines [49,50]. For example, only a small percentage of persons have received one vaccine dose in African countries [51]. Although the third and fourth doses of COVID-19 vaccines have already begun to be offered, vaccine uptake after the primary two doses is challenged by (booster) vaccine hesitancy, waning of vaccine coverage due to painful adverse reactions [49], and a drop-off in interest in COVID-19 [52]. Even during the unprecedented but successful worldwide vaccination programs for COVID-19, we experienced difficulty in the global distribution of vaccine doses. One of the lessons we learned was the importance of identifying vulnerable populations to achieve early control of infectious disease. In future pandemics, the score that we have developed may play an important role in identifying vulnerable populations. There are limitations to this study that we can improve for a better understanding of the role of vaccination. First, we measured IgG(S) and neutralization activity against the Wuhan strain as the humoral immune status, and analyzed IgG(S) because of the high correlation between these measurements. Although it is reported that antibodies induced by infection with the ancestral SARS-CoV-2 strain and/or the first of the primary two-dose vaccinations show cross-neutralization of variants from Alpha to Omicron BA.1 [53,54], further analysis of antibody responses (in particular, for neutralization) against variants of concern may be important to extend our scoring to prediction of vaccine efficacy for corresponding strains. Although our current score is limited in its detection capacity, the creation of such a personalized antibody score in the future by using individual-level demographic and health information will be an important tool for designing optimal vaccination strategies in future pandemics (not only COVID-19 but also other infectious diseases). Second, although one of the advantages of the Fukushima vaccination cohort is the extremely low number of natural infections, and that we can evaluate antibody responses minimizing the effect of infection, large portions of the population have been infected with COVID-19 in many parts of the world. In addition, the third and fourth doses of COVID-19 vaccines have already begun to be administered. Our primary purpose in this study was to establish an integrated framework for scoring, but an extended score considering breakthrough infection and booster vaccinations is worth evaluating. Another limitation is that the model fitting was based on antibody measurements and the antibody score was solely based on information available from the questionnaire. This is reflected in the rather low R2 values (around 0.2–0.3) of the regression models of the antibody features. Therefore, further refinement of the score using additional information will be worthwhile. It is worth mentioning that the immune system is affected by multiple factors, including genetics, the environment (such as cohabitation), and markers of metabolic health [55–58], all of which likely influence individual antibody status but were not considered here. The biological determinants of antibody variation will be further revealed in future studies addressing not only B cell subsets but the whole immune system encompassing adaptive as well as innate immunity. In conclusion, quantifying the variability in antibody dynamics can be a basis for policy decisions regarding the distribution of booster vaccines to strengthen immunity [59] or the use of oral antiviral drugs for the treatment of breakthrough infections [60]. As we learned from the COVID-19 pandemic, determining the priority for (booster) vaccines is an important public health concern given a limited global supply of vaccines at the beginning of a pandemic. For example, since we have data on what percentage of people with certain top/bottom AUC scores have low antibody levels, we can estimate how many people should be prioritized for booster vaccination based on these scores. Thus, the score developed here can be used to estimate and allocate necessary vaccination resources to “priority groups” for vaccination [61,62] in addition to the standard prioritized population [1,2,5,6]. Moreover, the score may be used by medical practitioners to encourage individuals with low predicted antibody levels to get (booster) vaccinations. However, we caution that our current score does not identify all individuals who need booster vaccinations and should not be the sole resource for guiding individual decisions. Taking advantage of our Fukushima vaccination cohort, we will further evaluate the impact of booster vaccinations and post-vaccination infections on personalized antibody scores. It is important that we make use of the lessons learned from COVID-19 vaccination to combat future infectious diseases.

Methods Ethics statement The study was approved by the ethics committees of Hirata Central Hospital (number 2021-0611-1) and Fukushima Medical University School of Medicine (number 2021–116). Written informed consent was obtained from all participants individually before the survey. Participant recruitment and sample collection The candidates were mainly recruited from Hirata village, Soma city, and Minamisoma city in rural Fukushima prefecture. We conducted non-sequential blood sampling and sequential blood sampling. A total of 2526 individuals participated in non-sequential blood sampling (Fig 1E). Health care workers, frontline workers, and administrative officers from each municipality were intentionally recruited to keep the cohort size large and the dropout rate low. Although most of the health care workers, frontline workers, and administrative officers were under the age of 65, relatively healthy community-dwelling older adults living in the community and in long-term care facilities were also recruited to maintain a wide age range for the cohort. Blood sampling was performed once during each period in June, September, and November 2021, respectively. The first vaccine dose was administered between March 10 and August 20, 2021, and the second dose between March 31 and September 14. The median (interquartile range) interval for the two-dose vaccination was 21 days. A total of 226 health care workers participated in the first blood sampling between May 31 and June 6, 2021. A total of 2526 individuals participated in the second blood sampling between September 8 and October 8, 2021. A total of 2443 individuals participated in the third blood sampling between November 21 and December 25, 2021. In contrast, 12 health care workers participated in the sequential blood sampling, and their vaccination and blood sampling schedule are shown in Fig 1E. Note that the 12 health care workers (described in S5 Fig A) with sequential blood sampling were not included in the non-sequential population of 2526 participants. In conclusion, of the total 2526 participants, those eligible for analysis were 2459 participants who completed the second vaccination and at least two blood samplings. Out of these 2459 participants, 43 individuals had higher IgG(S) titers on the second measurement than on the first, although their IgG(N) titers were negative (meaning no infection history). We speculated that this was due to measurement error and did not attempt model fitting. Furthermore, 9 individuals who had an interval of longer than 40 days between the first and the second vaccination were also removed. Because most of our participants had an interval of 21 or 28 days and there were very few cases of longer intervals, we decided that there were not enough data to reliably perform predictions among the participants with longer intervals. This was a conservative decision to ensure the high quality of our model fitting. As a result, we performed parameter estimation on the remaining 2407 participants (Fig 1B). Information on sex, age, daily medication, medical history, date of vaccination, adverse reaction after vaccination, type of vaccination, blood type, bacillus Calmette–Guérin (BCG) vaccine history, smoking habits, and drinking habits was retrieved from the paper-based questionnaire (summarized in Table 1). All blood sampling was performed at the medical facilities with 8 mL, and serum samples were sent to The University of Tokyo. SARS-CoV-2-specific antibody measurement All serological assays were conducted at The University of Tokyo. Specific IgG (i.e., IgG(S)) and neutralizing activity against the Wuhan strain were measured as the humoral immune status after COVID-19 vaccination. Specific IgG antibody titers (IgG(N)) were used to determine past COVID-19 infection status. Chemiluminescent immunoassay with iFlash 3000 (YHLO Biotech, Shenzhen, China) and iFlash-2019-nCoV series (YHLO Biotech, Shenzhen, China) reagents were used in the present study. The measurement range was 2–3500 AU/mL for IgG(S) and 4–800 AU/mL for neutralizing activity. For neutralizing activity, AU/mL×2.4 was used to convert to International Units (IU/mL); for IgG(S), AU/mL×1.0 was used to convert to binding antibody units (BAU/mL). The testing process was as per the official guideline. Quality checks were conducted every day before starting the measurement. Modeling vaccine-elicited antibody dynamics We developed a mathematical model describing COVID-19 vaccine-elicited antibody dynamics to evaluate the impact of primary two-dose COVID-19 vaccination on rapid immunity at the individual level and reconstructed the best-fit antibody titer curves of 2,407 participants in the Fukushima vaccination cohort. Here we explain the derivation and formulation of the mathematical model in detail. (i) Vaccination-elicited antibody dynamics after the first dose. After the first vaccination, naïve B cells encounter antigens and differentiate into short-lived antibody-secreting cells (ASCs), plasmablasts, germinal center (GC) B cells, or GC-independent memory B cells depending on BCR affinity for their cognate antigen [63]. Then, the GC B cells undergo rapid proliferation with somatic immunoglobulin hypermutation and subsequently differentiate into GC-dependent memory B cells or long-lived antibody-secreting cells, which are plasma cells with immunoglobulin class switching. To describe this antigen-specific B cell expansion and the induction of antibody-secreting cells and memory B cells after the first vaccination (S4 Fig A), we developed a simple but quantitative mathematical model as follows: (1) (2) (3) where the variables M(t) = M 1 (t), B(t), and A(t) are the amount of mRNA inoculated by the vaccination, the number of antibody-secreting cells, and the antibody titers at time t, respectively. The parameters τ 1 and d represent the timing of the vaccination and the decay rate of mRNA. We here considered D 1 to be the inoculated dose of mRNA by the vaccination, that is, M 1 (τ 1 ) = D 1 . Because the data we used here were limited (i.e., only time-course vaccine-elicited IgG(S) titers), one compartment of B cells including heterogeneous cell populations that produce antibodies (i.e., short-lived and long-lived antibody-secreting cells) was assumed. Therefore, we modelled the average B cell population dynamics in Eq (2), where the product of P 1 (t) and M(t)m/(M(t)m + Km) represents the average de novo induction of the antibody-secreting cells. P 1 (t) is a step function defined as P 1 (t) = P 1 for τ 1 + η 1 ≤ t, where η 1 is the delay of induction of antibody-secreting cells after vaccination: otherwise P 1 (t) = 0. The parameters m, K, and μ correspond to the steepness at which the induction increases with increasing amount of mRNA (i.e., the Hill coefficient), the amount of mRNA satisfying P 1 /2, and the average decay rate of the antibody-secreting cell compartment, respectively. The other parameters, p and c, represent the antibody production rate and the clearance rate of antibodies, respectively. (ii) Vaccination-elicited antibody dynamics after the second dose. After the second vaccination, memory B cells are reactivated by re-exposure to the antigen. Some differentiate into short-lived antibody-secreting cells (plasmablasts) or memory B cells outside the GC. Others enter the GC to become secondary GC B cells. Subsequently, these secondary GC B cells differentiate into GC-dependent memory B cells or long-lived antibody-secreting cells (plasma cells). To describe these recall B cell responses and their secretion of antibody after the second vaccination (S4 Fig B), we modified the above mathematical model, Eqs (1–3), as follows: (4) (5) (6) where M 2 (t) is the amount of mRNA by the vaccination inoculated at τ 2 satisfying M 2 (τ 2 ) = D 2 . In addition, P 2 (t) = P 1 and M(t) = M 1 (t) for τ 2 ≤t<τ 2+ η 2 , P 2 (t) = P 2 and M(t) = M 1 (t) + M 2 (t) for τ 2 + η 2 ≤ t, where η 2 is the delay of induction of antibody-secreting cells after vaccination; otherwise P 2 (t) = 0. Here we ignored the GC-independent memory B cells induced by the first vaccination because of their minor effect. In general, once reactivated, memory B cells can re-enter the GC more rapidly than naïve B cells, and therefore the secondary antibody responses are much faster and larger (i.e., η 1 >η 2 and P 1 < P 2 , respectively). In the main recall immunity, the quantity and quality of memory B cells established by the first vaccination is included in P 2 . (iii) Mathematical model for data fitting. Since the clearance rate of antibody is much larger than the decay of antibody-secreting cells (i.e., c ≫ μ), we made a quasi-steady state assumption, dA(t)/dt = 0 dV(t)/dt = 0, and replaced Eqs (3) and (6) with A(t) = pB(t)/c. Moreover, since Eqs (1) and (4) are linear differential equations, M 1 (t) = D 1 e-dt for t ≥ τ 1 and M 2 (t) = D 2 e-dt for t ≥ τ 2 : otherwise M 1 (t) = M 2 (t) = 0, respectively. Thus, the above Eqs (1–6) are further simplified assuming τ 1 = 0 and D 1 > 0, and we obtained the following single ordinary differential equation, which we used to analyze the antibody responses (i.e., IgG(S) titers (AU/mL)) in this study: (7) Where H(t) = H i = pP i /c for τ i + η i ≤ t<τ i+1 + η i+1 (i = 1 or 2), η 3 = ∞ and D 2 > 0 for τ 2 + η 2 ≤ t: otherwise D 2 = 0. This simple model can quantify the vaccine-elicited time-course antibody dynamics as described in Fig 2A under an arbitrary threshold of antibody titers A TH (see below). Quantifying vaccine-elicited time-course antibody dynamics In addition to the participants in the cohort, we included 12 health care workers whose serum was sequentially sampled for 40 days (on average 25 samples per individual) for validation and parameterization of a mathematical model for vaccine-elicited antibody dynamics. A nonlinear mixed effects model was used to fit the antibody dynamics model, given by Eq (7), to the longitudinal antibody titers of IgG(S) obtained from the 12 health care workers. The mathematical model included both a fixed effect and a random effect in each parameter. That is, the parameters for individual k, θ k (= θ×eπ k ) are represented as a product of θ (a fixed effect) and eπ k (a random effect). π k follows a normal distribution with mean 0 and standard deviation Ω. As we described above, since η 1 > η 2 and P 1 < P 2 , we assumed η 1 = η and η 2 = f delay η, H 2 = H and H 1 = f degree H, and estimated η, H, 0 < f delay < 1 and 0 < f degree < 1 for conducting biologically reasonable estimations. We here assumed that the parameters η, H, f delay , f degree and m varied across individuals, whereas we did not consider interindividual variability in other parameters to ensure parameter identifiability. Note that the half-life of mRNA (i.e., log 2/d) and dose of mRNA (i.e., D i ) are assumed to be 1 day [64] and 100 (μg/0.5 mL) [65], respectively. Fixed effect and random effect were estimated by using the stochastic approximation expectation-approximation algorithm and empirical Bayes’ method, respectively. Fitting was performed using MONOLIX 2019R2 (www.lixoft.com) [66]. The estimated (fixed and individual) parameters are listed in S1 Table. With the estimated parameters for each individual, the dynamics of IgG(S) titers, A(t), and the average de novo antibody response elicited by the first and second vaccinations, , were calculated in S5 Fig AB, respectively. Interestingly, we observed that the variations induced by the second vaccination were much larger than those induced by the first vaccination. We found that most of the best-fitted estimated parameters in the mathematical model (i.e., D 1 , D 2 , d, μ, K, η, f delay , f degree ) were the same or similar across the 12 individuals compared with those of parameters of m and H (see S1 Table). We note that m and H, which showed wide variation of estimated values, contributed mainly to the vaccine-elicited antibody dynamics after the second vaccine dose, whereas the other parameters contributed to that after the first dose. In fact, we are interested in the large variation after the second dose (but not the negligible variation after the first dose). Therefore, we hereafter fixed the parameters in our mathematical model to be the estimated population parameters listed in S1 Table, except m and H. These assumptions enabled us to accurately reconstruct the large variations in antibody dynamics after the second dose, and these two parameters were independently estimated from each IgG(S) by a nonlinear least-squares method, even if 2,407 participants had only 2 or 3 measurements of antibody titers at different time points. Using the estimated parameters for each individual, we fully reconstructed the dynamics of IgG(S) titers of all participants after the first vaccination in S5 Fig C. We summarized the distribution of parameter values m and H for 2,407 participants in S5 Fig D, and the best-fit antibody titer curves of 200 randomly selected individuals are plotted along with the observed data for visualization in S2 Fig. Mathematical model validation Outside the study period described in Fig 1E, to validate our mathematical model (i.e., Eqs (4–5)), we first prepared Validation dataset A: the additional sequentially performed serum sampling of 12 health care workers, that is, data for 3 to 4 timepoints before the booster dose (on average 4 additional samples per individual for 165 days after the first vaccine dose). Employing our estimated parameters, which were estimated from the original dataset (i.e., the black circles in S6 Fig), listed in S1 Table, we found our mathematical model predicted the values for the additional 3 to 4 timepoints before booster vaccination with high accuracy. Next, to validate the reconstructed best-fit antibody titer curves described in S5 Fig C, we also prepared Validation dataset B: an additional third or fourth antibody measurement from 110 of the 2407 participants of our Fukushima vaccination cohort around 3 to 4 months after the second or third timepoint data but before the booster vaccination. We compared the reconstructed antibody titer curves and the additional data and confirmed that the additional data points obtained from participants who were not infected with COVID-19 (i.e., IgG(N)-negative participants) during that period perfectly matched the prediction of our mathematical model (S7 Fig). On the other hand, interestingly, for those who got infected with COVID-19, the additional data were significantly larger than the reconstructed curves (S8 Fig). Additionally, we assessed the accuracy of the reconstructed best-fit antibody titers by calculating the Pearson’s correlation coefficient between the reconstructed antibody titers and the observed (or additionally observed) antibody titers (S9 Fig). The Pearson’s correlation coefficients between the observed antibody titers, which were used in parameter estimation, and the reconstructed antibody titers, were 0.99 for both of the 12 health care workers and the 2407 participants in the Fukushima vaccination cohort. Furthermore, the Pearson’s correlation coefficients between the additionally observed antibody titers, which were not used in parameter estimation, and the reconstructed antibody tiers, were 0.94 and 0.92 for the 12 health care workers and the 110 participants in the Fukushima vaccination cohort, respectively. Taken together, our additional data (i.e., Validation dataset A and B) and quantitative analysis demonstrated that our mathematical model and the reconstructed best-fit antibody titer curves using the same are validated for the purposes of predicting antibody dynamics in a generalized population. Building optimized antibody scores A Python implementation of Ustun et al.’s [29] algorithm (risk-slim, https://github.com/ustunb/risk-slim) was used to build optimized AUC scores. Briefly, the algorithm searches for the best linear combination of features with integer coefficients that minimizes the sum of the logistic loss and the l 0 -norm of the coefficients. The range of coefficients was set to -5 to +5; the l 0 -penalty parameter C 0 was set to 1×10−3. Evaluation of antibody scores When an individual’s top (or bottom) AUC score is above a threshold, that individual is predicted to be in the top 1/3 (or the bottom 1/3) of the population. The accuracy, precision, and recall of the score are defined as (TP + TN)/(TP + TN + FP + FN), TP/(TP + FP), and TP/(TP + FN), respectively, where TP, TN, FP, and FN denote true positives (i.e., the number of individuals predicted to be positive that were actually positive), true negatives, false positives, and false negatives, respectively. Statistical analysis The answers from the 2,407 participants who completed the paper-based questionnaire were converted into a set of categorical and numerical variables. Numerical variables included age, BMI, and the interval between the two doses. These variables were then used in a multiple regression analysis to explain the three antibody dynamics features. Six participants who did not fully answer the questionnaire were excluded from the analysis. The variables used here belonged to any of the five categories: (i) basic demographic information and lifestyle habits, (ii) information on vaccinations, (iii) underlying medical conditions, (iv) adverse reactions, and (v) medications being taken. When necessary, the same variables were compared among different generations or different groups using Pearson’s chi-square test (for categorical variables), analysis of variance (ANOVA, for more than two numerical variables), or Welch T-test (for two numerical variables). A Bonferroni correction was applied for multiple comparisons. All statistical analyses were performed using R (version 4.1.2) or JMP Pro 16. Map of Japan The basemap shapefile of Japan and Fukushima Prefecture was downloaded from National Land Numerical Information (https://nlftp.mlit.go.jp/ksj/gml/datalist/KsjTmplt-N03-v2_4.html) and edited using the R package sf.

Acknowledgments We would like to thank all the staff from Fukushima Medical University, Seireikai health care group, Hirata Village office, Soma City office, Soma Central Hospital, Soma General Hospital, Minamisoma City office, Minamisoma City Medical Association, Minamisoma Municipal General Hospital, Shindo Clinic and Medical Governance Institute, who contributed significantly to the accomplishment of this research, especially, Ms. Yuka Harada, Ms. Serina Noji, Ms. Naomi Ito, Dr. Makoto Kosaka, Mr. Anju Murayama, Mr. Sota Sugiura, Mr. Manato Tanaka, Ms. Yuna Uchi, Mr. Yudai Kaneda, Mr. Masahiko Nihei, Mr. Hideo Sato, Ms. Rie Yanai, Ms. Yasuko Suzuki, Ms. Keiko Abe, Dr. Hidekiyo Tachiya, Mr. Kouki Nakatsuka, Dr. Ryuzaburo Shineha, Ms. Miki Sato, Dr. Masahiko Sato, Mr. Naoharu Tadano, Mr. Kazuo Momma, Mr. Shu-ichi Mori, Ms. Saori Yoshisato, Ms. Katsuko Onoda, Mr. Satoshi Kowata, Mr. Masatsugu Tanaki, Dr. Tomoyoshi Oikawa, Dr. Joji Shindo, Ms. Xujin Zhu, Ms. Asaka Saito, Ms. Yuumi Kondo, and Ms. Tomoyo Nishimura.

[END]
---
[1] Url: https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000497

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.