T cell and antibody functional correlates of severe COVID-19

Comorbid medical illnesses, such as obesity and diabetes, are associated with more severe COVID-19, hospitalization, and death. However, the role of the immune system in mediating these clinical outcomes has not been determined. We used multi-parameter flow cytometry and systems serology to comprehensively profile the functions of T cells and antibodies targeting spike, nucleocapsid, and envelope proteins in a convalescent cohort of COVID-19 subjects who were either hospitalized (n=20) or not hospitalized (n=40). To avoid confounding, subjects were matched by age, sex, ethnicity, and date of symptom onset. Surprisingly, we found that the magnitude and functional breadth of virus-specific CD4 T cell and antibody responses were consistently higher among hospitalized subjects, particularly those with medical comorbidities. However, an integrated analysis identified more coordination between polyfunctional CD4 T-cells and antibodies targeting the S1 domain of spike among subjects that were not hospitalized. These data reveal a functionally diverse and coordinated response between T cells and antibodies targeting SARS-CoV-2 which is reduced in the presence of comorbid illnesses that are known risk factors for severe COVID-19. Our data suggest that isolated measurements of the magnitudes of spike-specific immune responses are likely insufficient to anticipate vaccine efficacy in high-risk populations.


INTRODUCTION
4 compare frequencies and phenotypes of conventional α β T cells as well as donor-unrestricted T cells 98 (DURTs)(35). Finally, we compared the functional profiles of antigen-specific T cells targeting S1, S2, 99 N, and envelope (E) proteins using intracellular cytokine staining (ICS). In nearly all the parameters 100 tested, we consistently observed both higher magnitudes and increased functional breadth among 101 hospitalized subjects, particularly those with medical comorbidities. However, T cell and antibody 102 responses showed less correlation among hospitalized subjects. Our balanced analysis reveals a 103 qualitative shift in the adaptive immune response to SARS-CoV-2, which may be directly related to the 104 presence of comorbid illnesses that are known risk factors for severe disease. 105 106 107 108 109 110 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint 6 analysis of non-redundant aspects of the SARS-CoV-2 specific antibody response. Relative to non-155 hospitalized subjects, hospitalized subjects demonstrated lower correlation among antibody titers, Fc-156 specificities, and Fc-effector functions ( Figure 2D). This difference was robust to sub-sampling in order 157 to account for the unequal sample sizes in each group (Supplementary Figure 7). Finally, we 158 calculated a polyfunctionality score for each individual for S, RBD and N over the six antibody 159 functionality readouts against three SARS-CoV-2 antigens. Subjects with comorbidities were able to 160 activate a robust polyfunctional antibody response against S, RBD, and N in comparison to subjects 161 without comorbidities ( Figure 2E). Taken together, these results reveal qualitative and quantitative 162 increases in several aspects of the SARS-CoV-2 specific antibody response among hospitalized 163 subjects with comorbidities, many of which are likely the result of differences in innate immune system 164 activation and T cell help. 165 166 Activated CD8 and γ δ T cells are associated with hospitalization after  To investigate the role of T cells, we used multi-parameter flow cytometry to quantify the 168 frequencies and phenotypes of conventional and donor-unrestricted T cell populations, such as 169 invariant NK T (iNKT) cells, mucosal-associated invariant T (MAIT) cells, and γ δ T cells(35). In our 170 matched cross-sectional analysis, we noted that the frequency of CD3+, CD4+, and CD8+ T cells did 171 not vary significantly over time since symptom onset or between hospitalized and non-hospitalized 172 subjects ( Figure 1D, 1E, 1F and Supplementary Figure 3). We also found no difference in the 173 frequency of γ δ

T cells, invariant NKT cells, or mucosal associated invariant T cells as well as B cells, 174
monocytes, or NK cells ( Figure 1G and Supplementary Figure 4). However, the frequency of activated 175 CD8+ T cells was significantly higher among hospitalized subjects, which is consistent with prior reports 176 ( Figure 1I) (21, 38, 39). The frequency of naive CD8+ T cells was also lower among hospitalized 177 subjects, suggesting differentiation to an effector phenotype after SARS-CoV-2 infection ( Figure 1J). 178 Among total γ δ T cells, the frequency of activated γ δ T cells was higher among hospitalized subjects 179 independent of expression of the Vδ2 gene segment ( Figure 1K). The frequency of activated CD4, 180 CD8, and γ δ T cells was broadly steady over time since symptom onset, which is in contrast to some 181 reports ( Figure 1H, 1I, and Supplementary Figure 4C)(38). These data confirm and extend published 182 studies by revealing the durability of differences in activated CD8 and γ δ T cell but not CD4 T cell 183 populations in a matched cross-sectional analysis stratified by hospitalization status. 184 185 IFN-γ independent CD4 T-cell responses to SARS-CoV-2 structural antigens 186 We next investigated the functional profiles of SARS-CoV-2 specific T cells. PBMCs were 187 stimulated with overlapping peptide pools targeting the S1 or S2 domain of spike, nucleocapsid (N), or 188 envelope small membrane protein (E). We used intracellular cytokine staining (ICS) to identify antigen-189 specific T cells expressing interleukin 2 (IL-2), IL-4/5/13, IL-17a, IFN-γ, tumor necrosis factor (TNF), 190 CD107a, and CD40L (Supplementary Figure 3). To ensure the detection of polyfunctional T cell 191 subsets that may be present at low frequencies, we employed COMbinatorial Polyfunctionality analysis 192 of Antigen-Specific T cell Subsets (COMPASS)(40). Among 128 possible functional profiles, we 193 detected 21 antigen-specific CD4 T cell subsets across all four peptide pool stimulations ( Figure 3A). 194 Notably, the probability of detecting a particular response varied according to the antigen. For 195 example, several profiles containing three or four functions were readily detected after stimulation with 196 S1, S2, or N but not E. However, the two profiles containing five functions (IFN-γ, IL-14/5/13, TNF,  . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint 2, and CD40L) were only detected after stimulation with S1. Stimulation with E resulted in a CD107a 198 monofunctional profile that was also observed after stimulation with S2 ( Figure 3A). 199 200 Notably, 11 (52%) of the 21 CD4 T cell functional profiles identified by COMPASS did not 201 contain IFN-γ ( Figure 3A). Because COMPASS only reports the probability of detecting a particular 202 response, we next examined the magnitude of T cell responses stratified by the presence of IFN-γ. We 203 found nearly equivalent numbers of IFN-γ + and IFN-γ -T cells after stimulation with S1 or N. However, 204 more T cells expressed IFN-γ-independent functions after stimulation with S2 and E ( Figure 3B and 205 3C). These data suggest that a substantial fraction of the SARS-CoV-2-specific T cell response could 206 be missed by conventional assays, such as IFN-γ ELISPOT (41). We used uniform manifold 207 approximation and projection (UMAP) to examine qualitative associations between hospitalization 208 status, stimulation, and T cell functional profiles. Hospitalization appeared to be associated with 209 responses to S1, S2, and N, though there was overlap with non-hospitalized subjects ( Figure 3D). The 210 degree of polyfunctionality appeared to be associated with hospitalization, which was also suggested 211 by COMPASS ( Figure 3A and 3D). Among the 21 functional profiles identified by COMPASS, CD4 T 212 cells simultaneously expressing CD40L, IL-2, and TNF were detected at the greatest magnitudes, 213 regardless of the presence of IFN-γ, and were highest after stimulation with S1 or S2 ( Figure 3E and 214 Supplementary Figure 5). By contrast, ~1% of CD4 T cells expressed CD107a independent of IFN-γ 215 after stimulation with E ( Figure 3F). Finally, CD4 T cells with a detectable cytokine response 216 predominantly expressed a CCR7+CD45RA-central memory phenotype, but very few demonstrated 217 co-expression of the activation markers HLA-DR and CD38 ( Figure 3D). Taken together, these data 218 demonstrate the functional diversity of CD4 T cell responses to SARS-CoV-2 structural antigens driven 219 in large part by IFN-γ-independent profiles that are not typically the focus of vaccine immunogenicity or 220 epitope mapping studies (26, 42, 43). 221 222 Functional diversity of CD4 T cell responses to SARS-CoV-2 are associated with hospitalization 223 Since UMAP revealed a qualitative association between T cell functional profile and 224 hospitalization, we wanted to next explore that relationship quantitatively. To accomplish this, we used 225 COMPASS to calculate a 'functionality score' (FS), which summarizes the functional breadth for each 226 subject and stimulation into a continuous variable that can be incorporated into standard statistical 227 models (40). Among CD4+ T cells, we found the highest functionality scores after stimulation with N, 228 followed by S1, S2, then E ( Figure 4A). However, the correlation between stimulations was modest, 229 even between S1 and S2, confirming the importance of examining each antigen and functional domain 230 independently ( Figure 4B). CD4 functionality scores were not associated with age or sex for any of the 231 antigens tested ( Figure 4C and 4D). Notably, the functional breadth of CD4 T cell responses was 232 stable over time ( Figure 4E). Finally, we investigated whether functionality scores were associated with 233 clinical risk factors and outcomes. We found higher functionality scores to S1, S2, and N but not E 234 among hospitalized subjects and in the presence of medical comorbidities ( Figure 4F and 4G). We 235 examined this association using magnitudes of polyfunctional (CD40L+IL-2+TNF+) CD4 T cells and 236 found the same to be true independent of the production of IFN-γ ( Figure 4H). Thus, our data reveal 237 that increased functional breadth of CD4+ T cell responses to spike and nucleocapsid are associated 238 with known risk factors for severe COVID-19 independent of the production of IFN-γ. 239 240 CD8 T cell responses to SARS-CoV-2 structural antigens are not associated with hospitalization 241 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint We next explored the functional breadth of CD8 T cell responses and its association with  242  hospitalization. In contrast to the CD4 T cell response, COMPASS analysis identified seven T cell  243 subsets, of which only two lacked IFN-γ ( Figure 5A). IFN-γ independent T cell responses were 244 dominant after stimulation with S2 and E ( Figure 5A and 5B) and were characterized by expression of 245 CD107a ( Figure 5A and 5C). Both UMAP and COMPASS revealed polyfunctional profiles consisting of 246 IFN-γ, IL-2, and TNF that were largely detected after stimulation with N in both hospitalized and not 247 hospitalized subjects ( Figure 5A and 5D). Similar to CD4 T cells, CD107a monofunctional CD8 T cells 248 were mostly detected after stimulation with S2 and E ( Figure 5E). Cytokine producing CD8 T cells were 249 distributed across effector memory, central memory, and TEMRA phenotypes and did not co-express 250 activation markers HLA-DR and CD38 ( Figure 5D). Analysis of CD8 functionality scores revealed the 251 greatest breadth after stimulation with N and very little correlation between antigens ( Figure 5F and 252 5G). Again, we noted a surprisingly poor correlation between S1 and S2 that was driven by the 253 dominance of polyfunctional responses to S1 and CD107a monofunctional responses to S2 ( Figure  254 5A). Only S2 functionality scores were negatively correlated with age ( Figure 5H). Finally, none of the 255 stimulations were associated with sex, days post symptom onset, or hospitalization ( Figure 5I, 5J, and 256 5K). Together, these data reveal that a thorough assessment of CD8 functional responses requires 257 assays that examine more than IFN-γ, and that IFN-γ production and cytotoxic function are poorly 258 correlated, even between the S1 and S2 domains of spike glycoprotein. 259 260 Antigen-specific T cell and antibody responses are less coordinated among hospitalized subjects 261 Our results indicated a consistently higher magnitudes and increased functional breadth of 262 several antibody and T cell features among hospitalized subjects. Thus, we next sought to identify the 263 minimum set of features that could differentiate between hospitalized and non-hospitalized subjects. 264 We used least absolute shrinkage and selection operator (LASSO) and identified eight features that 265 consistently distinguished the two clinical groups via partial least squares discriminant analysis (PLS-266 DA) ( Figure 6A and Supplementary Figure 6). With the exception of the induction of CD107a 267 expression on NK cells by anti-RBD antibodies, all features were consistently enriched among 268 hospitalized subjects ( Figure 6B). When we examined the correlation between the selected features 269 and all measured features, we noted that ADNP and FcR2A targeting spike were highly correlated with 270 other features of humoral immunity ( Figure 6C). Further, the four CD4 polyfunctional T cell features did  271  not correlate with each other or with the humoral features, indicating a non-redundant contribution of T  272  cell functions to the classification. Finally, we examined how T cell and antibody features correlated  273 with each other in the two groups. Among non-hospitalized subjects, we noted more significant positive 274 correlations between T cell and antibody features as compared to subjects who were hospitalized, even 275 when the two groups were downsampled to account for the different sample sizes ( Figure 6D and 276 Supplementary Figure 7). These data suggest that non-hospitalized subjects are able to better 277 coordinate antigen-specific T cells and antibody responses to SARS-CoV-2 despite having reduced 278 functional breadth compared to subjects that were hospitalized. 279 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint

281
In summary, we performed a cross-sectional study comprehensively examining the functional 282 profiles of T cells and antibodies targeting SARS-CoV-2 spike, nucleocapsid, and envelope proteins in 283 convalescent subjects who were either hospitalized or not hospitalized. We consistently found the 284 magnitude and functional breadth of measured responses to be higher among hospitalized subjects 285 and in the presence of medical comorbidities. However, these responses were more poorly correlated 286 with each other when compared to non-hospitalized subjects. Since the presence of medical 287 comorbidities are a known risk factor for severe disease and were over-represented among 288 hospitalized subjects, these data support the possibility that virus-specific responses may contribute to 289 immunopathology and severe COVID-19. 290 291 In contrast to most studies in which T cells or antibodies are studied in isolation, we 292 comprehensively profiled both and analyzed them together in the context of detailed clinical 293 information. In almost every respect, we find that they track together and show high levels of 294 coordination among non-hospitalized subjects. The lack of coordination observed among hospitalized 295 subjects may reflect a failure to control the virus at early stages, resulting in increased inflammation and 296 virus load. Comorbid diseases were over-represented among hospitalized subjects, suggesting that 297 they may be related to the increased functional breadth among T cells and antibodies that we describe 298 here. Supporting this hypothesis are studies examining the effect of diabetes on adaptive immunity to 299 M. tuberculosis (44). These studies have shown increased production of antigen-specific Th1 and 300 Th17 cytokines in the presence of chronic hyperglycemia which is associated with an increased 301 inflammatory state (45, 46). Whether SARS-CoV-2 specific T cell and antibody responses with 302 increased functional breadth are the cause of poor clinical outcomes is not addressed by the cross-303 sectional design of our study and more definitively assessed in longitudinal studies or animal models. 304 305 Notably, we did not observe an association between neutralizing antibody titers and 306 hospitalization in our study, which contrasts with emerging data examining patients much earlier in their 307 disease course (47, 48). However, we did find that several functional attributes of spike-specific 308 antibodies, including Ig subclass titers, were poorly correlated with neutralization yet associated with 309 hospitalization. We have also previously shown that the ratio of spike:nucleocapsid antibodies is more 310 predictive of death among hospitalized subjects than neutralization titers (37). These data add to a 311 growing body of literature showing that several attributes of virus-specific antibodies are associated 312 with clinical outcomes, including hospitalization (49). In general, we found that Ig subclass titers and 313 Fc-specificity, and Fc-effector functions were lower among non-hospitalized subjects yet were more 314 highly correlated with each other compared to hospitalized subjects. These findings may be the result 315 of differences in innate immune activation, which may contribute to increased viral clearance and lower 316 antigen loads. Innate immunity is known to be impaired in older subjects and in the presence of co-317 morbidities like diabetes (50). 318 319 We observed that T cell responses to envelope protein were qualitatively different from spike 320 and nucleocapsid. Among both CD4 and CD8 T cells, a uniform functional profile of CD107a 321 expression emerged, which was not seen with the other antigens. One potential explanation for this is 322 that while E is abundantly expressed, very few molecules are incorporated into virions. Rather, E 323 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint titers as well as IFN-γ production by S-specific T cells as evidence of immunogenicity (42, 43). 335 However, we show that neutralizing antibody titers are poorly correlated with several important 336 functional qualities of S-specific antibodies. We also show that a significant fraction of the CD4 T cell 337 response to S does not include IFN-γ and depends on which domain is being examined. For example, 338 CD4 and CD8 T cell responses to S2 were notable for having a cytotoxic phenotype compared to S1. 339 In the integrated analysis, eight T cell and antibody features primarily focused on S1 were sufficient to 340 classify hospitalized subjects with near perfect accuracy. These data raise the possibility that some 341 vaccine-induced immune responses to spike glycoprotein might be harmful. Phase I studies that report 342 safety are typically tested on young, healthy volunteers that are not representative of the target 343 populations for candidate COVID vaccine, likely older and with medical comorbidities (55). This is a 344 particularly important concern as several of the platforms being used, such as mRNA and adenoviral 345 vectors, have limited experience in large clinical efficacy studies. An expanded analysis of the 346 functions of vaccine-specific T cells and antibodies beyond what is required for regulatory approval will 347 be required to understand the full benefits or risks of each approach. 348 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020.

350
Study population 351 Whole blood samples were collected from individuals with laboratory-confirmed SARS- Both plasma and PBMC were frozen within six hours of collection time. 391 392 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint

Antibody neutralization 393
The SARS-CoV-2 pseudoviruses expressing a luciferase reporter gene were generated in an 394 approach similar to as described previously (9, 10, 21). Briefly, the packaging plasmid psPAX2 (AIDS 395 Resource and Reagent Program, Germantown, MD), luciferase reporter plasmid pLenti-CMV Puro-Luc 396 (Addgene, Watertown, MA), and spike protein expressing pcDNA3.1-SARS CoV-2 SΔCT were co-397 transfected into HEK293T cells by lipofectamine 2,000 (Thermo Fisher Scientific, Waltham, MA). The 398 supernatants containing the pseudotype viruses were collected 48 hours post-transfection, which were 399 purified by centrifugation and filtration with a 0.45 µm filter. To determine the neutralization activity of 400 the serum or plasma samples from cohorts, HEK293T-hACE2 cells were seeded in 96-well tissue 401 culture plates at a density of 1.75 x 10 4 cells/well overnight. iQue with a PAA robot arm and analyzed using Forecyt software. The readout was mean fluorescence 430 intensity (MFI) of PE. All experiments were performed in duplicate while operators were blinded to 431 study group assignment, and all cases and controls were run at the same time to avoid batch effects. 432 433 Functional antibody measurements 434 Bead-based assays were used to quantify antibody-dependent cellular phagocytosis (ADCP), 435 antibody-dependent neutrophil phagocytosis (ADNP), and antibody-dependent complement deposition 436 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint (ADCD), as previously described (58-60). Fluorescent neutravidin beads (red for ADCD, yellow for 437 ADNP, and ADCP) (Thermo Fisher Scientific, Waltham, MA) were coupled to biotinylated SARS- CoV-2  438 antigens RBD, S, and N and incubated with diluted plasma (ADCP and ADNP 1:100, ADCD 1:10 439 dilution) for 2 hours at 37°C. For measuring monocyte phagocytosis, 2.5x10 4 THP-1 cells (ATCC,  440 Manassas, VA) were added per well and incubated for 16 hours at 37°C. For ADNP, Ammonium-441 Chloride-Potassium ACK lysis was performed on whole blood from healthy blood donors (MGH blood  442 donor center), and 5x10 4 cells were added per well and incubated for 1 hour at 37°C. 10 minutes and rested overnight at a density of 2 million cells/mL. The following day, the cells were 477 enumerated using the Guava easyCyte and analyzed using two multiparameter flow cytometry assays. 478 479 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint San Jose, CA). After lysis, the cells were washed with FACS buffer twice then permeabilized by 529 incubating for 10 minutes at room temperature with 1x FACS Perm II (BD Biosciences, San Jose, CA). 530 The PBMC were again washed twice with FACS buffer then stained with the remaining markers for 30 531 minutes at 4°C and then washed with FACS buffer: anti-CD3 ECD (clone UCHT1) (Beckman Coulter, 532 Brea groups were evenly distributed in each batch and operators were not blinded to study group 542 assignments. 543 544 Statistics 545 Flow cytometry data analysis 546 Initial compensation, gating, and quality assessment of flow cytometry data was performed 547 using FlowJo version 9.9.6 (FlowJo, TreeStar Inc, Ashland OR) for T cell data or Forecyt software 548 (Intellicyt, Albuquerque, NM) for the antibody data. Representative gating trees for the surface marker 549 panel and ICS data are shown in Supplementary Figure 3. The surface marker and ICS flow cytometry 550 data were then processed using the OpenCyto framework in the R programming environment (63). 551 Samples with poor viability defined on the basis of low CD3 counts (<10,000 cells) or low CD4 counts 552 (<3,000 cells) were excluded from analysis. For the ICS panel, data from 20 convalescent hospitalized 553 and 37 convalescent non-hospitalized subjects were ultimately analyzed. For the surface marker 554 panel, data from 15 convalescent hospitalized and 36 convalescent non-hospitalized subjects were 555 analyzed. 556 557 To achieve a comprehensive and unbiased analysis of the functional profiles of antigen-specific 558 T cells, we used COMPASS(40). COMPASS uses a Bayesian hierarchical framework to model all 559 observed cell subsets and select those most likely to have antigen-specific responses. Notably, 560 COMPASS reports only the probability of detecting a particular T cell functional profile, rather than the 561 absolute magnitude, which we calculated separately. For a given subject, COMPASS was also used to 562 compute a functionality score that summarizes the entire functionality profile into a single continuous 563 variable that can be used for standard statistical modeling (e.g. regression). For the data presented 564 here, COMPASS was applied to each of the antigen stimulations separately for CD4+ and CD8+ T 565 cells. Each one of the analyses was unbiased and considered all of the 128 possible boolean 566 combinations of cytokine functions. Subjects with a high probability of response across many subsets 567 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint were accordingly assigned a high functionality score. Magnitudes of T cell responses were calculated 568 independent of COMPASS as the proportion of gated events in the stimulated condition minus the 569 proportion of gated events in the unstimulated condition. Statistics were performed using background 570 subtracted magnitudes, although data are plotted as the maximum of zero or this value. The R package 571 ComplexHeatmap (64)  Integrated analysis of T cell and antibody functional profiles 586 Classification models were trained to discriminate subjects between hospitalized and non-587 hospitalized subjects using all the measured humoral and T-cell responses. Models were built with an 588 approach similar to what we have previously published, using a combination of the least absolute 589 shrinkage and selection operator (LASSO) for feature selection and then classification using partial validation run, subjects were randomly stratified into five subsets ensuring that both groups were 596 represented in each subset, with four subsets serving as the training set and the fifth as the test set. 597 Each subset served as the test set once; therefore, each individual was in the test fold exactly once for 598 each cross-validation run. For each test fold, LASSO-based feature selection was performed on 599 logistic regression using the four subsets designated as the training set for that fold. Fold specific 600 LASSO was repeated ten times and features, which are selected nine times out of ten, were identified 601 as selected features. Using these selected features, a fold-specific PLS-DA was trained on training data 602 for that fold. A set of predicted group labels were recorded for each subset. Significance of model performance was evaluated using "negative control" models of permuted 610 data and randomly selected size-matched features. The repetitions of five-fold cross-validation 611 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint generated a distribution of model classification accuracies.
Corresponding model accuracy 612 distributions were measured for two negative control models. The first approach consisted of 613 permutation testing by randomly shuffling the group labels, within the cross-validation framework 614 described above (i.e., a cross-validation framework matched to the actual model) (70). The second 615 approach was to randomly select a set of features the same size as the LASSO-selected feature set. 616 These control processes were repeated 100 times to generate a distribution of model accuracies 617 observed in the context of permuted data and randomly selected, size-matched feature sets. The 618 predicted group label for each subject was compared to the true group label to obtain a classification 619 accuracy. Exact p-values were obtained as the tail probability of the true classification accuracy in the 620 distribution of control model classification accuracies. Because one of the LASSO-selected features 621 (ADNP Spike) was highly correlated with 54% of all features, we further assessed the performance of 622 randomly selected features by selecting only from the remaining 46% features. Further, we additionally 623 built an alternative model by excluding ADNP Spike to examine whether the separation between the 624 groups would be achieved in the absence of this feature and to identify the strongest surrogate of 625 ADNP Spike that can discriminate subjects between the two groups. These analyses were performed 626 using R package "ropls" version 1.20.0 (71) and "glmnet" version 4.0.2 (72). . 627 628 Correlations were performed using Spearman method followed by Benjamini-Hochberg multiple 629 correction (73). The co-correlate network was generated using R package "network" version 1.16.0 630 (74) and the chord diagram was generated using R package circlize version 0.4.10 (75). 631 632 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review) preprint
The copyright holder for this this version posted November 30, 2020. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. COVID-19 study subjects that were previously hospitalized (purple, n=20) or non-hospitalized (green, 865 n=40) were selected based on matching for age, sex, ethnicity, and date of symptom onset. Samples 866 were comprehensively profiled for SARS-CoV-2 specific T cell and antibody phenotypes and functions. 867 Data were analyzed to identify differences between the groups and to build a classifier. DURTs = 868 Donor-unrestricted T cells (B) Antibody neutralization titers were compared between hospitalized and 869 non-hospitalized subjects (left) and graphed according to days since symptom onset (right). is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint functions were summed, resulting in a polyfunctionality score per individual. Polyfunctional scores are 905 displayed as percent positivity of the whole cohort. Antibody phenotypes and effector functions 906 excluding neutralization were compared across cohorts using Mann-Whitney U tests followed by 907 correction for multiple hypothesis testing using the Bonferroni method. Median, 25th, and 75th quartiles 908 are indicated for violin plots. The black line on the scatter plot represents the best fit linear regression 909 line, and the grey-shaded area represents the 95% confidence interval of the predicted mean. If not 910 shown, p-values for Mann-Whitney tests and regression were not significantly different. CD40L+IL-2+TNF+ functional profile in the presence or absence of IFN-γ are compared between 950 groups after stimulation with S1, S2, and N. CD4 functionality scores were compared using Wilcoxon 951 signed-rank tests or Mann-Whitney U tests were used followed by correction for multiple hypothesis 952 testing using the Bonferroni method except for panels D and F. Supplementary Figure 5  cells specific for the S1 and S2 domains of spike, nucleocapsid (N), and envelope small membrane 959 protein (E). Data were analyzed using COMPASS, and results are displayed as a probability heatmap 960 in which the rows represent study subjects and the columns represent CD8 T cell functional subsets. 961 The depth of shading within the heatmap represents the probability of detecting a response above 962 background. Responses are stratified by group to enable comparisons across stimulation conditions. 963 In Partial least squares discriminant analysis (PLS-DA) was used to identify T cell and antibody features 991 that could discriminate between hospitalized (blue) and non-hospitalized (green) subjects. The PLS-DA 992 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint scores plot shows the separation between the groups using the first two latent variables (LVs     IFN-γ + CD40L + TNF + IL-2 + IL-4/5/13 -IL-17a -CD107a -IFN-γ -CD40L + TNF + IL-2 + IL-4/5/13 -IL-17a -CD107a -  Scores on LV1 (45%) RBD: CD107a S1: CD4 + IL2 + IL4/5/13 -IFNγ + TNF -CD40L + IL17a -CD107a -S1: CD4 + IL2 + IL4/5/13 + IFNγ + TNF + CD40L + IL17a -CD107a -N: CD4 + IL2 -IL4/5/13 -IFN-γ -TNF -CD40L + IL17a -CD107a -

Table 1. Summary of demographics of the SARS-CoV-2 Convalescent Cohort
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.25.20235150 doi: medRxiv preprint