Evolution of immunity to SARS-CoV-2

The durability of infection-induced SARS-CoV-2 immunity has major implications for public health mitigation and vaccine development. Animal studies and the scarcity of confirmed re-infection suggests immune protection is likely, although the durability of this protection is debated. Lasting immunity following acute viral infection requires maintenance of both serum antibody and antigen-specific memory B and T lymphocytes and is notoriously pathogen specific, ranging from life-long for smallpox or measles4, to highly transient for common cold coronaviruses (CCC). Neutralising antibody responses are a likely correlate of protective immunity and exclusively recognise the viral spike (S) protein, predominantly targeting the receptor binding domain (RBD) within the S1 sub-domain. Multiple reports describe waning of S-specific antibodies in the first 2-3 months following infection. However, extrapolation of early linear trends in decay might be overly pessimistic, with several groups reporting that serum neutralisation is stable over time in a proportion of convalescent subjects. While SARS-CoV-2 specific B and T cell responses are readily induced by infection, the longitudinal dynamics of these key memory populations remains poorly resolved. Here we comprehensively profiled antibody, B and T cell dynamics over time in a cohort recovered from mild-moderate COVID-19. We find that binding and neutralising antibody responses, together with individual serum clonotypes, decay over the first 4 months post-infection, as expected, with a similar decline in S-specific CD4+ and circulating T follicular helper (cTFH) frequencies. In contrast, S-specific IgG+ memory B cells (MBC) consistently accumulate over time, eventually comprising a significant fraction of circulating MBC. Modelling of the concomitant immune kinetics predicts maintenance of serological neutralising activity above a titre of 1:40 in 50% of convalescent subjects to 74 days, with probable additive protection from B and T cells. Overall, our study suggests SARS-CoV-2 immunity after infection is likely t 66 o be transiently protective at a population level. SARS-CoV-2 vaccines may require greater immunogenicity and durability than natural infection to drive long-term protection.

. Kinetics of decay were broadly consistent between IgG1 and IgG2 91 subclasses, with IgG3 displaying a more rapid, two phase decline (Extended Data Fig.  92 2). Consistent with a previous report 26 , we find N-specific IgG decays significantly 93 more rapidly than S-specific IgG (t1/2 = 71 and 229 days respectively, p < 0.00001, 94 Fig 1D). In contrast to IgG, S-specific IgM and IgA1 fit a two-phase decay, with a 95 more rapid early decay (t1/2 = 55 and 42 days respectively) followed by a slower 96 decay in late convalescence (t1/2 = 118 and >1000 days respectively; Fig 1D). A 97 comparison of decay rates between neutralising activity and antibody binding 98 demonstrated that early neutralisation decay occurs at a similar rate to the early 99 decline in S, RBD and S1-specific IgM (Fig 1E, Extended Data Fig 2). Neutralisation 100 titre at both early and late convalescence was well correlated with serum inhibition of 101 RBD-ACE2 binding and S, S1 and RBD specific IgG, IgM (and to a lesser extent 102 IgA1) responses, as well as with S2 and N specific IgG responses (Extended Data Fig  103   3). Neutralising activity during early convalescence was the best correlate of long-104 term maintenance of neutralisation responses (Spearman rho=0.88, p<0.00001; 105 Extended Data Fig 3). Serum inhibition of RBD-ACE2 binding and S1-specific IgG 106 responses in early infection were also well correlated with neutralisation titre in late 107 convalescence (Spearman rho=0.79, 0.81, respectively; Extended Data Fig 3). 108 However, in a multiple regression model, once early neutralisation activity was 109 included no other significant predictors were identified (p>0.15 for all other 110 variables). 111 based quantitative proteomics workflow developed for serum autoantibody 116 profiling 27,28 to track unique CDR-H3 peptides matching recovered S-specific 117 immunoglobulins sequences from convalescent subjects 19 (n=4; Extended Data Fig  118   4A, B). Consistent with the decay of polyclonal S-specific antibody in the blood, we 119 find a decline in the relative abundance over time for each unique clonotype 120 (Extended Data Fig. 4C), although absolute rates of decay did vary, suggesting the 121 kinetics might to some degree be clonotype-, epitope-or subject-specific. for respiratory infection such as influenza 33 . S-specific cTFH and conventional CD4+ 152 and CD8+ memory T cells (Tmem), were quantified using activation induced marker 153 (AIM) assays 19,22 (Methods) following stimulation with overlapping S (split into S1 154 or S2) peptide pools (Fig. 3A, 3B). Frequencies of S-specific memory T cells were 155 dynamic over time and varied between individuals, with evidence of either rapid 156 decline or stable maintenance (Fig. 3C). Pairwise comparison of cTFH or CD4+ 157 Tmem frequencies at the final visit relative to the first available sampling 158 demonstrated a significant reduction in S-specific responses over time (p=0.0031 for 159 cTFH, p=0.0224 for CD4+ Tmem; Fig. 3D). In contrast, frequencies of S-specific 160 CD8+ Tmem were stable at a population level (p=0.3247), although individual 161 responses were varied (Fig. 3D) Fig 9A,B). Analogous patterns were observed for the CD4+ Tmem 176 cells (Extended Data Fig 9C,D). Consequently, S2-specific cTFH and CD4+ Tmem 177 populations predominated over S1-directed responses (p=0.0147 and p=0.0021 178 respectively) in late convalescence (Extended Data Fig 9B,D). 179 180 Polyclonal T cell responses to S comprise an array of immunodominant and 181 subdominant epitopes; we therefore additionally tracked single CD4+ T cell epitopes 182 in a subset of 9 donors (Extended Data Fig 10A). Strikingly, we observed substantial 183 inter-and intra-individual variability in longitudinal epitope-specific responses 184 (Extended Data Fig 10B,C); in some subjects, all epitope-specific responses tracked 185 similarly while in others distinct epitope-specific responses would vary independently 186 over time. In most, but not all, cases, peptide responses tracked similarly between the 187 cTFH and Tmem populations (Extended Data Fig 10C). Overall, some degree of T 188 cell immunity remains readily detectable in most subjects 4 months after infection, 189 although longitudinal epitope-specific frequencies were markedly less predictable. 190 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint Deconvoluting the protective potential of the suite of concomitant immune responses 192 elicited by SARS-CoV-2 infection is challenging. The general decline of serological 193 immunity over time (Fig 4A) was similarly observed for most memory immune cell 194 subsets except for IgG+ and IgA+ MBC populations ( Fig 4B). Importantly, rates of 195 immune decay are likely to stabilise over time to levels of homeostatic maintenance 37 , 196 although this set point is not yet clear for SARS-CoV-2. Neutralising antibody is the 197 most widely accepted protective correlate against a range of human respiratory 198 viruses 38 . However, any relationship between in vitro neutralisation titres and in vivo 199 protection for SARS-CoV-2 is unclear at present. We therefore developed a 200 simulation model (see Methods) employing the estimated initial distributions of 201 neutralisation titres and decay rates across subjects, to predict the time for titres to 202 drop below a nominated cut-off of 1:40, selected based on the 1:40 hemagglutination 203 inhibition titre (a surrogate for neutralisation activity) widely used as the 50% 204 protective titre for influenza 39 . Notably, 43% of our cohort were already below this 205 threshold in early convalescence, with 64% of subjects dropping below this threshold 206 in late convalescence. Simulating a population of 1000 individuals, and running the 207 model 1000 times, we find the median time for 50% of the population to drop below a 208 titre of 1:40 was 74 days (Fig 4C; 95% confidence interval 46 to >1000 days). 209 Assuming early neutralisation titres predicted titres into late convalescence, our 210 simulation also allows us to estimate how higher initial levels of neutralisation may 211 affect the proportion of individuals maintaining titres above 1:40. We found that if 212 aiming for a median of 50% of individuals with a titre above 1:40 at one year, initial 213 neutralisation titres at about day 30 would need to be in the order of 2.1-fold higher 214 than that observed in our convalescent cohort (95% CI = no increase to 16.9-fold 215 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint increase required). It is important to emphasise that at present the in vitro 216 neutralisation titre required and the additive contribution of other immune responses 217 to protective immunity are unknown. In addition, our analysis assumes that immunity 218 to vaccination decays at a similar rate to infection, and that the decay of neutralisation 219 titre from day 70 to around 140 predicts immune decay over the first year. Despite the 220 limitations inherent in these assumptions, this analysis provides an approach to 221 estimating the target level of immune response necessary for effective vaccination. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint cells incubated in serum-free media containing 1 µg/ml TPCK trypsin at 37°C/5% 266 CO2; viral cytopathic effect was read on day 5. The neutralising antibody titre is 267 calculated using the Reed/Muench method as previously described 43,44 . All samples 268 were assessed in two independent microneutralisation assays. MabTech) was added at 1.3µg/ml, 25µl per well followed by streptavidin-PE (SA-PE; 291 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint Thermo Fisher) at 1µg/ml. Plates were acquired by a FLEXMAP 3D (Luminex). 292 Median fluorescence intensity (MFI) for each isotype/subclass detector was assessed. 293 Background subtraction was conducted, removing background of blank (buffer only) 294 wells. Multiplex assays were repeated twice as two independent experiments. 295 296 RBD-ACE2 binding inhibition multiplex bead-based assay 297 RBD protein was coupled to bioplex beads (Biorad) as described above. 20µl of RBD 298 multiplex bead suspension containing 500 beads per well, 20µl of biotinylated 299 Avitag-ACE2 (kindly provided by Dale Godfrey and Nicholas Gherardin), final 300 concentration of 12.5µg/ml per well, along with 1:100 dilution of each subject's 301 plasma were added to 384 well plates. Plates were covered and incubated at room 302 temperature (RT) whilst shaking for 2 hours, and then washed twice with PBS 303 containing 0.05% Tween20 (PBST). Biotinylated Avitag-ACE2 was detected using 304 40µl per well of SA-PE at 4µg/ml, incubated with shaking for 1 hour at RT. 10µl of 305 PE-Biotin amplifier (Thermo Fisher) at 10µg/ml was added and incubated for 1 hour 306 with shaking at RT. Plates were washed and acquired on a FLEXMAP 3D (Luminex). 307 Anti-SARS-CoV-2 RBD neutralising human IgG1 antibody (ACROBiosystems, 308 USA) was included as a positive control, in addition to COVID-19 negative plasma 309 and buffer only negative controls. The MFI of bound ACE2 was measured after 310 background subtraction of no ACE2 controls. Maximal ACE2 binding MFI was 311 determined by buffer only controls. % ACE2 binding inhibition was calculated as 312 100% -(% ACE2 binding MFI per sample/ Maximal ACE2 binding). RBD-ACE2 313 binding inhibition multiplex assays were repeated independently twice. 314 315 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint

Estimating the decay rates 367
We sought to predict the response variable (yij for patient i at timepoint j) as a 368 function of days post symptom onset, assay replicate (as a binary categorical variable) 369 and a random effect for each individual (both in in intercept and slope). The 370 dependency of the response variables on days post symptom onset can be modelled by 371 using one or two decay slopes. The model can be written as below: 372 !" = # + #! + $ !" + % !" + %! !" -for a model with a single slope; and 373

376
The parameter # is a constant (intercept), and #! is a patient-specific adjustment to 377 the overall intercept. The slope parameter % is a fixed effect to capture the decay 378 slope before # ; which also has a subject-specific random effect %! . To fit a model 379 with two different decay rates, an extra parameter & (with a subject-specific random 380 effect &! ) was added to represent the difference between the two slopes. Assay 381 variability between replicates was modelled as a single fixed effect $ , in which we 382 coded the replicate as a binary categorical variable !" . 383 The response variables obtained were highly variable, containing zeros where the 384 value was below the limit of detection and contrasted with samples where very high 385 levels were observed. Thus, we performed log transformations of the non-zero data to 386 help normalize variability and censored every value less than 40 for the 387 microneutralisation data; every value less than 0.01 for the T cell and B cell data; and 388 every negative value for the multiplex data. More specifically, a mixed-effect 389 regression method that allows for censoring at the limit of detection was used to 390 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. Rank Sum test in GraphPad Prism 8. All statistical tests used were two sided. 430

Competing interests 431
The authors declare no competing interests. 432

Data Availability 433
All data are available from the corresponding author upon reasonable request. 434

Code Availability 435
All data analysis code can be made available on request. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. identification and quantification of circulating anti-S1 antibodies. S1-specific IgG was 541 purified from plasma of SARS-CoV-2 convalescent subjects using antigen-coupled 542 magnetic beads and heavy chains subject to LC-MC/MS. Peptide spectra are searched 543 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. Statistics assessed by two-tailed Wilcoxon test. 568 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. The best-fit half-lives are shown for the fitting of the decay of cTFH, CD4+ Tmem 579 and CD8+ Tmem specific to total S (S1+S2 responses combined), S1 or S2 peptide 580 pools (n=31 subjects). In all cases decay was fit with a single-phase decay model with 581 the half-lives shown. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020. . https://doi.org/10.1101/2020.09.09.20191205 doi: medRxiv preprint (A) Representative staining of AIM markers following S1 or S2 peptide pool or 593 individual peptide stimulation among the CD4+ Tmem population. (B,C) 594 Longitudinal peptide-specific frequencies in individual subjects (n=9; solid line, 595 CD4+ Tmem; dashed line, cTFH) for whom (B) multiple or (C) single epitopes were 596 identified. 597 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for this this version posted September 10, 2020.        The best-fit model and half-lives are shown for the fitting of the decay of antibody binding to different SARS-CoV-2 antigens (n=64 subjects). Two-phase decay is indicated by red (before day 70) and blue (after day 70) shaded areas. No shading indicates where single-phase decay provided the best fit.

Extended Data Figure 3: Correlation of antibody binding and ACE2 inhibition with neutralisation.
A heat-map of Spearman correlations between neutralisation titre and the serological measurements of antibody binding (by isotype and antigen). Correlations were assessed in early (≤50 days, left column n=54 subjects) and late (≥100 days, right middle column, n=47 subjects) convalescence in all subjects were data was available. The association between early antibody binding and late neutralisation is also shown (right column, n=47 subjects). All correlations are Spearman correlations. *P≤0.05, **P≤0.01, ***P≤0.001.  Extended Data Figure 4: MS-based quantification of immunoprecipitated S1-specific clonotypic antibodies.
(A) Combined B cell receptor sequencing and proteomics platform enables identification and quantification of circulating anti-S1 antibodies. S1-specific IgG was purified from plasma of SARS-CoV-2 convalescent subjects using antigen-coupled magnetic beads and heavy chains subject to LC-MC/MS. Peptide spectra are searched against B-cell receptor sequencers recovered from single sorted S-specific memory B cells from matched individuals to identify clonotypes based upon CDR-H3 amino acid sequence. Clonotype specific peptides are then used as barcodes for relative quantitative parallel reaction monitoring (PRM) for tracking in longitudinal plasma samples. Targeted peptides are monitored during elution from HPLC and individual peptides quantified based on abundance chromatography curves. (B) Clonotypes identified based on matched CDR-H3 sequences from S1-specific plasma IgG and B cell receptor sequences from SARS-CoV-2 convalescent subjects (n=4). (C) Longitudinal changes in the relative plasma abundance of anti-S1 clonotypes within four convalescent subjects over time. The quantity of each reference peptide is expressed as area under the curve (AUC) derived from extracted ion chromatography.     (A, C) Representative staining of AIM markers following S1 and S2 peptide pool stimulation among (A) cTFH (CD3+CD4+CD8-CD45RA-CXCR5+) or (C) CD4+ Tmem cells and longitudinal cohort analysis (n=31). (B, D) Comparison of S1 or S2-specific (B) cTFH or (D) CD4+ Tmem responses at the earliest and latest visit for each participant, as well as paired frequency of S1 versus S2 responses at the initial or final visit (n=31). Statistics assessed by two-tailed Wilcoxon test.