Ebola survivor cohort
Ebola virus disease survivors (n = 115), previously described42 with certificates (issued by Ebola treatment centres on discharge) were recruited as potential donors through 34 Military Hospital, Freetown, and the Sierra Leone Association of Ebola Survivors as participants in the study ‘Convalescent plasma (CP) for early Ebola virus disease in Sierra Leone’. The study (ISRCTN13990511 & ACTR201602001355272) was approved by the Scientific Review Committee and Sierra Leone Ethics, authorized by the Pharmacy Board of Sierra Leone (PBSL/CTAN/MOHSCST001) and sponsored by the University of Liverpool. All participants provided written consent for data collected in this study.
Volunteers were considered suitable to donate plasma if they tested negative for blood-borne infections (hepatitis B, hepatitis C, HIV, malaria and syphilis), had had two documented negative EBOV PCR tests 72 h apart, had no acute febrile illness and had no comorbidity, such as heart failure, to suggest that they might be at increased risk of adverse events during apheresis. Volunteers were not excluded if they exhibited indications of post-Ebola syndrome (PES; for example, musculoskeletal pain, headache or ocular problems), although such complaints were noted and subsequently contributed to the characterization of PES41,42,44. The majority of the participants were male (n = 82), their age ranging between 18 and 52 with a median of 27 years old. The female (n = 33) age range was between 18 and 42 with a median of 27 years old (Supplementary Table 1a).
For transfusion safety reasons, donor identity numbers were not confidential to donors during the conduct of the study; for the avoidance of doubt, donor identity numbers have since been dissociated. All participants (n = 115) were tested using DABA, blocking EIA and IgG capture immunoassays21. PPV antibody neutralization assays were performed with a subset of participants not selected on any criteria other than sample availability (n = 52). The compartmental population pharmacodynamics model was developed on the more replete DABA dataset using those participants with longitudinal data (n = 51) (Supplementary Table 1f).
HEK293T (ATCC CRL-3216) and TZM-bl45,46,47,48,49 (acquired from NHI AIDS Reagent Program) cells are adherent cell lines cultivated in Dulbecco’s modified Eagle’s medium (Invitrogen: 12491-023), supplemented with 10% heat-treated fetal bovine serum (FBS) (Sigma: F7524), 2 mM/ml l-glutamine (Invitrogen: 25030024), 100 U/ml penicillin (Invitrogen: 15140148) and 100 mg/ml streptomycin (Invitrogen: 15140148), referred to as complete DMEM (Thermo Fisher: 12491023). Cells were grown in a humidified atmosphere at 37 °C and 5% CO2. Vero E6 cells (ECACC: 85020206) were grown in VP-SFM (Thermo Fisher: 11681-020). All cell lines were tested monthly for mycoplasma contamination.
EBOV PPV construct design
Three viral strain glycoprotein genes were cloned into pCDNA3.1 produced by GeneArt using gene synthesis: a 2014 isolate (KP096421)22, a variant carrying the A82V, T230A, I371V, P375T, and T544I mutations (Fig. 1b) identified by analysis of sequenced EBOV strains between March and August 201439 and the AY354458 1995 Kikwit isolate50. The latter was used in ring vaccinations during the 2014 epidemic.
EBOV PPV production
We chose to utilize the HIV-1 SG3 ΔEnv and EBOV-GP expression plasmids, co-transfected into HEK293T cells, to generate infectious PPV stocks47,51,52. The EBOV-GP-pseudotyped lentiviral system generates single-cycle infectious viral particles. HEK293T cells were plated at a density of 1.2 × 106 in a 10-cm diameter tissue culture dish (Corning: 430167) in 8 ml complete DMEM and incubated overnight. The cells were transfected with 2 μg pSG3Δenv along with 0.285 μg of a plasmid expressing EBOV-GP using a cationic polymer transfection reagent (Polyethylenimine, Polysciences: 23966-2), in the presence of OptiMEM (Invitrogen: 31985-070). OptiMEM was replaced 6 h after transfection with 8 ml complete DMEM. Seventy-two hours after transfection, supernatant containing the generated stock of single-cycle infectious EBOV-GP pseudotyped virus particles was harvested, passed through a 0.45-μM filter and stored in aliquots at −80 °C. EBOV-GP plasmid (285 ng per 10-cm culture dish) was used to produce a large virus stock that was tested for infectivity (Fig. 1a) then pooled, aliquoted and stored at −80 °C.
EBOV PPV infection
EBOV infectivity was determined through infection of TZM-bl cell lines where luciferase activity (expressed from LTR promoter) is under the control of Tat expressed from the HIV-1 backbone. We used 100 μl EBOV-GP virus to infect 1.5 × 104 TZM-bl cells per well for 6 h in a white 96-well plate (Corning: CLS3595). Following infection, 150 μl per well DMEM complete was added to the cells. Forty-eight hours after infection, medium was discarded from the wells, cells were washed with phosphate-buffered saline (PBS, ThermoFisher: 12899712) and lysed with 30 μl cell lysis buffer (Promega: E1531), and luciferase activity was determined by luciferase assay (Promega: E1501) using a BMGLabtech FluoroStar Omega luminometer. Negative controls included pseudotyped virus bearing no glycoproteins and TZM-bl cells alone, which routinely resulted in luminescence of 3,000–7,000 relative light units (RLU).
EBOV PPV neutralization
Plasma samples (n = 52) from Ebola convalescent plasma and healthy blood donors (n = 6) were heat treated at 56 °C for 30 min and centrifuged for 15 min at 13,000 RPM. Aliquots were then stored at −80 °C. Plasma samples were serially diluted 50% with complete DMEM; 13 μl plasma dilution was incubated with 200 μl EBOV-GP PPV for 1 h at room temperature. We used 100 μl of virus/plasma dilution to infect TZM-bl cells as described above. Luciferase activity readings of neutralized virus were analysed (i) by considering 0% inhibition as the infection value of the virus in the absence of convalescent plasma included in each experiment, (ii) by considering 0% inhibition as the infection value of two consecutive high dilutions that did not inhibit virus entry. Both methods produced highly correlated results (Extended Data Fig. 2d) and the latter was used. The neutralization potential of a CP was represented as the plasma dilution that reduced viral infectivity by 50% (IC50) or by 70% (IC70).
Enzyme immune assays
Samples were diluted in 0.1% Empigen (Sigma: 30326) in TBS before the ELISA assay (Fisher: 10167481). The p24 assays were conducted using the Aalto Bio Reagents Ltd protocol and recombinant p24 standard, p24 coating antibody (polyclonal sheep anti-HIV-1-p24 gag, Aalto Bio Reagents Ltd: D7320), secondary conjugate (alkaline phosphatase conjugate of mouse monoclonal anti-HIV-1-p24, Boehringer Mannheim: 1089-161) and ELISA light assay buffer. Plates were incubated for 30 min at room temperature before we measured luminescence with a FLUOStar Omega luminometer (BMG LabTech).
Double antigen bridging assay (DABA)
We measured EBOV GP targeting antibody present in Ebola survivor CP samples. EBOV GP antigen, Mayinga Zaire EBOV strain (IBT Bioservices: 0501‐016) was pre-coated onto the ‘solid phase’, while a second antigen conjugated to horseradish peroxidase (HRP) acted as the detector, binding to EBOV antibodies captured on the solid-phase antigen in the first incubation step. Antibody reactivity was expressed as arbitrary units per ml (AU/ml) as compared to a standard comprising five reactive donor samples that were pooled and set as 1,000 AU/ml21.
Antibody levels in CP to EBOV GP (glycoprotein), VP40 and NP (nucleoprotein) were determined by blocking of the binding of specific rabbit EBOV anti-peptide (GP, VP40, NP) antibodies (IBT Bioservices) to EBOV Makona virion-coated microplates. Microplate wells were coated with a 10,000-fold dilution of concentrated Ebola virions. EBOV patient CP and negative control CP dilutions (1/100) were reacted on virion-coated microplates for 4–6 h. CP dilutions were removed and plates were then reacted with EBOV anti-peptide antibodies. Bound rabbit antibodies were detected by species-specific horseradish peroxidase conjugate (DAKO: P03991-2). Evidence of EBOV protein-specific human antibodies in CP was determined by blocking of the binding of the antipeptide antibody compared to the blocking of binding by the CP negative control. Results were expressed as a percentage of blocking of the CP negative control reactivity.
IgG capture assay
IgG antibodies present in CP were captured onto a solid phase coated with rabbit hyperimmune anti-human γ-Fc and interrogated in a second incubation with HRP-conjugated EBOV GP as above. Reactivity was expressed as binding ratios derived as sample OD/cut-off OD21.
Plaque reduction neutralization test
The wild-type strain used for assays was EBOV Makona (GenBank accession number KJ660347)21, isolated from a female Guinean patient in March 2014 (virus provided to PHE Porton by S. Günther, Bernhard-Nocht-Institute for Tropical Medicine, Hamburg, Germany). The virus was propagated in Vero E6 cells and culture supernatant virions were concentrated by ultracentrifugation through a 20% glycerol cushion; pellets were resuspended in sterile PBS at a titre of 109 focus-forming units (FFU) per ml.
The wild-type virus neutralizing antibody titre in CP was determined by reacting serial dilutions of CP with 100 FFU of EBOV virions for 1 h at room temperature to allow antibody binding. The EBOV virion CP mixture was adsorbed to Vero E6 monolayers for 1 h and then overlaid with cell growth medium containing 1% (v/v) Avicel (Sigma-Aldrich). After 80–90 h, EBOV foci were visualized by immunostaining with anti-VLP (Zaire EBOV) antibodies (IBT Bioservices). All work was undertaken under ACDP containment level 4 conditions.
EBOV antibody decay and restimulation modelling
Compartmental population analysis was performed to model the stimulation and decay of antibody levels. All modelling and simulations were performed using Pmetrics version 1.453 within R version 3.2.254. Antibody levels of EBOV survivors were sampled at different number of instances, at varying intervals post convalescence due to limitations of follow-up adherence in the field. Different parts of decay–stimulation profiles were therefore captured, with only a few instances of contiguous decay–stimulation or stimulation–decay profiles being captured. Stimulation and decay data were therefore modelled separately to most efficiently use the data. Antibody stimulation–decay trends with 2 or more data points were included in population analysis as this methodology has been proven to maximally use sparse clinical data for drug development55,56. All points were plotted and visualized. An ‘ascend’ or a ‘descend’ was defined according to the prevailing trend. A 20% alteration in direction was tolerated as part of the prevailing ascend or descend as appropriate.
Structural model selection was performed for the most replete DABA data set. Model fitting and selection were performed using previously published protocols for fitting clinical data sets as described below57,58. In brief, linear regression (intercept close to 0, slope close to 1) was used to assess the goodness-of-fit of the observed–predicted values, the coefficient of determination of the linear regression and minimization of log-likelihood, AIC and BIC values were used for model selection.
A change in BIC drop of more than 2 is generally considered to be significant; with 2–6 indicating positive-to-strong evidence, 6–10 indicating strong evidence and >10 indicating very strong evidence59.
Further details of this analysis leading to the choice of models and analysis of the fit of models to data can be found in Supplementary Tables 3, 4 and Extended Data Figs. 8, 9.
All chosen structural models showed strong-to-very-strong evidence of describing the data the best out of the compared models. Two structural models were tested for antibody stimulation, a one-compartmental stimulation model and a one-compartmental model with saturable stimulation, based on the logistic growth model. The logistic growth model framework allows for plateauing antibody levels, as observed for a subset of stimulation profiles. For antibody decay, four structural models were tested; a one-compartment decay model with first order elimination, a two-compartment decay model with first order elimination from the central compartment, and the above two structural models with saturable recycling offsetting the endogenous elimination rate.
Antibody stimulation was best modelled using the one-compartmental model with saturable stimulation as described by equation (1):
where X1, kgrowth and Kmax denote antibody level in the compartment, the first order rate constant for endogenous antibody stimulation and the maximal antibody level at which stimulation plateaus, respectively.
For antibody decay, the two-compartment decay model with saturable FcRn-dependent recycling (equations (2–4)) as used to model antibody decay in multiple laboratory studies37 was found to best describe the data.
where X1 and X2 are the antibody levels in the central and peripheral compartments. The rate constants kdecay, kcp and kpc denote the empirically observed antibody level dependent rate constant and the first order rate constants to and from the peripheral compartment, respectively. kdecay is in turn dependent on the endogenous decay rate kend, which is offset by an antibody-dependent saturable recycling rate described by a Michaelis–Menten term with parameters Vmax and Km denoting the maximal recycling rate and antibody level at which half the maximal recycling rate occurs, respectively. The optimal structural models above were then used to model the more sparse nAb assay datasets, allowing for comparability between DABA and nAb model parameters. Generally, individual predicted versus observed value correlations were excellent (R2 > 0.8) and population predictions versus observed values were good (R2 > 0.6).
Monte Carlo simulations were performed using Pmetrics as previously described57,58. In brief, 1,000 individuals were randomly sampled from parameter distributions defined in the population models of antibody stimulation and decay. The interquartile range of modelled antibody levels was then plotted longitudinally for average starting antibody levels for decay and stimulation profiles (Fig. 3c–h).
With regard to the choice to model the stimulation and decay data separately: in principle, an immune response followed by a gradual return to baseline post-stimulus could be characterized by a single pharmacodynamic model. In the simplest form, the dynamics can be described by a single-compartment model with the stimulus placed on the input rate and first order elimination, although more mechanistic models based on known pharmacology may also be appropriate if the data are of sufficient quality to estimate the unknown model components. In a controlled trial setting, the onset of a stimulus event would be controlled and the subsequent immune response measured relative to this origin with sufficient frequency to capture the dynamics over time. By contrast, this study was observational with plasma samples taken intermittently that captured only part of the changing levels in the nAbs—either the growth or decay phase in most cases, but on occasion both. Given the lack of detectable viral load and the observational nature of the nAb response data, the ability to fit a single, integrated pharmacodynamic model to the data is limited. The most tractable solution in this case was to split the data into two groups and model them separately: the first model quantifying the rate of increase in nAbs and the second model describing the subsequent decay. The antibody decay was based on ref. 37. While this two-stage approach did not allow data from the ‘stimulation’ phase to inform the model fit of the ‘decay’ phase—and vice versa—it did enable accurate and quantitative characterization of both the stimulation and decay dynamics, which have not been characterized for EBOV disease before this study, and which may be used to inform future work in this area and other impactful viral diseases such as COVID-19.
Statistical analyses of data were implemented using GraphPad Prism 6.0 software. Unpaired sample comparisons were conducted for all data; individual figure legends state the corresponding statistical tests performed. These include parametric and non-parametric t-tests (Student’s t-test and Mann–Whitney U-test); parametric and non-parametric ANOVAs (ordinary ANOVA and Kruskal–Wallis test). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.
Further information on research design is available in the Nature Research Reporting Summary linked to this paper.