Prolonged SARS-CoV-2 Infection in Patients with Lymphoid Malignancies

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection can lead to chronic manifestations in healthy and immunocompromised individuals with different clinical presentations and unique risks. In otherwise healthy people hospitalized with coronavirus disease 2019 (COVID-19), up to one third continue to experience physical and cognitive symptoms, a syndrome commonly referred to as “long COVID” (1). In severely immunocompromised hosts, protracted respiratory symptoms with frequent hospital readmissions can occur. The underlying pathogenic mechanisms for the two entities are not well understood. However, for severely immunocompromised patients, isolated cases of persistent COVID-19 clinical manifestations with coincident persistent respiratory SARS-CoV-2 RNA detection and the absence of seroconversion have been reported (2, 3).

Certain immunocompromising conditions, such as hematologic malignancy, are associated with a higher risk for severe COVID-19 and poor outcomes after acute infection (4–6). The unusual and protracted disease course marked by intermittent flares is less defined but a frequently encountered COVID-19 complication in this patient population. Persons with lymphoid malignancy, especially those with diminished T-cell responses and after B cell–depleting therapies, are particularly susceptible (7). A few case reports describe the clinical and viral genomic evolution in chronically infected patients with lymphoid malignancies and highlight the immunologic correlation with impaired CD4 and CD8 virus–specific T-cell responses (3, 8).

In addition to the substantial direct morbidity from COVID-19 and interruption in cancer care, chronic persistent SARS-CoV-2 infection poses the additional threat of extended transmissibility and selection pressure that could generate immune-evading mutations (8, 9). Intrahost evolution in chronic SARS-CoV-2 infection has led to the detection of mutations found in variants of concern, raising intense speculation that immunocompromised patients may serve as reservoirs for the emergence of critical mutations in the receptor binding domain region of the virus (8).

Therefore, a better understanding of chronic infection in severely immunocompromised patients is significant from both a therapeutic and a public health perspective. In this study, we identify predictors of persistent COVID-19 among patients with hematologic malignancies and characterize the clinical course and viral genomic evolution in patients with lymphoid malignancy, including those receiving B cell–depleting therapies.

From March 10, 2020, until February 28, 2021, 382 patients with hematologic malignancy were diagnosed with COVID-19, including 214 patients with lymphoid malignancies (Fig. 1). Persistent PCR positivity, defined as SARS-CoV-2 RNA detection ≥30 days after initial positivity and regardless of symptomatology, was detected in 51 (13.9%) of 368 patients alive at 30 days from COVID-19 diagnosis. In the absence of established criteria, the 30-day time point was selected based on prior studies showing that most patients become PCR negative within 28 days of infection (10) and a median of 31 days of viral shedding from illness onset in patients with severe COVID-19 (10, 11). The demographic and clinical features of the 368 patients as well as univariate and multivariate predictors of persistent PCR positivity are shown in Supplementary Table S1 and Table 1, respectively. Lymphopenia, treatment with anti-CD20 antibodies within 1 year, and cellular therapy including hematopoietic stem cell transplantation (HSCT) within 1 year were independent predictors of prolonged SARS-CoV-2 RNA detection, whereas parenteral cytotoxic chemotherapy within 6 months was not. These time periods reflect estimated duration for immune recovery (12, 13). Age and underlying lymphoid malignancy did not independently influence prolonged viral detection among patients with hematologic cancers.

Given that the strongest predictors of viral persistence are standard-of-care therapies for patients with lymphoid malignancies, we performed further analysis of the 214 patients with lymphoid malignancy and COVID-19 infection with a data cutoff date of April 1, 2021 (Supplementary Table S2). Most patients had B-cell non-Hodgkin lymphoma (77%), and approximately half of the patients had received recent systemic therapy, including anti-CD20 mAbs (32%), Bruton tyrosine kinase (BTK) inhibitors (10%), autologous or allogeneic stem cell transplantation (7%), or CD19-directed chimeric antigen receptor (CAR) T-cell therapy (3%). Seventy-nine patients (37%) were undergoing active workup or treatment at the time of COVID-19 diagnosis.

Most patients (83%) met criteria for severe COVID-19, with 73 of 102 (72%) hospitalized patients requiring supplemental oxygen, 32 of 102 (31%) requiring intensive care unit–level care, and 22 of 102 (22%) requiring mechanical ventilation. The management and outcomes of COVID-19 in inpatients are summarized in Supplementary Tables S3 and S4. Inpatients were treated with hydroxychloroquine (40%) in March and April 2020, convalescent plasma (44%), remdesivir (37%), dexamethasone (24%) and/or other systemic steroids (30%), intravenous immunoglobulin (9%), tocilizumab (7%), anakinra (1%), and N-acetylcysteine (17%) on a clinical trial (ClinicalTrials.gov Identifier: NCT04374461). The case fatality rate during the index hospitalization was 22%. Thirty-two patients (31%) experienced an extended hospitalization over 21 days, and 19 of the remaining 77 patients who were discharged following index hospitalization were readmitted with recrudescent respiratory symptoms after initial improvement. Most of these patients had continuous COVID-19 PCR positivity, suggestive of a high rate of incomplete viral clearance in this population.

Univariate analysis identified clinical and laboratory features associated with death from acute COVID-19 infection, as shown in Fig. 2A and Supplementary Fig. S1. Key predictors of death included cardiovascular illness, relapsed/refractory disease, and active treatment, including anti-CD20 and CAR T-cell therapy in the past year. Multivariate analysis revealed cardiovascular disease, active treatment, and CAR T-cell therapy, but not anti-CD20 therapy, to be independently associated with mortality.

In addition, unbiased clustering of 42 immunologic laboratory measurements (listed in Supplementary Table S5) obtained from patients within 72 hours of hospitalization for acute infection generated three semiglobal clusters of patients with distinct immunologic phenotypes (Fig. 2B; Supplementary Fig. S2A and S2B). Cluster 1 was defined by the most significantly reduced levels of T cells as well as the lowest ratio of effector or activated T cells to naïve T cells but the highest levels of acute inflammatory markers, consistent with a blunted T-cell response. Cluster 2 was defined by intermediate T-cell counts, a relatively high incidence of anti-CD20 therapy, and a relatively high level of T-cell activation as well as moderate levels of acute inflammatory markers, consistent with a partially compensated immune response in the presence of B-cell depletion. Finally, cluster 3 was defined by minimal acute inflammatory markers, normal T-cell counts, and higher levels of T-cell activation consistent with a sufficient adaptive immune response. Correlation of immune clusters with COVID-19–specific clinical outcomes demonstrated the highest mortality in cluster 1 (5/11, 45%), the lowest mortality rate in cluster 3 (3/17, 18%; P = 0.041), and intermediate mortality (3/8, 38%) but a high rate of readmission to the hospital for persistent symptomatic disease (4/8, 50%) in cluster 2.

We therefore asked whether specific inflammatory markers were associated with clinical outcomes in COVID-19–infected patients with lymphoma. Hospitalized patients with lymphoma demonstrated similar immunologic profiles to those reported in the broader population, with elevated levels of all inflammatory markers and lymphopenia in approximately half (55%) of patients. Severe lymphopenia (absolute lymphocyte count <500 cells/μL) was present in 27% of hospitalized patients, and flow cytometry analysis revealed undetectable B-cell counts in 57% (25/44) of patients. Laboratory parameters that correlated with mortality included IL6, C-reactive protein (CRP), ferritin, D-dimer, and lactate dehydrogenase (LDH; Fig. 2C; Supplementary Fig. S2C). Notably, inflammatory markers such as CRP were incrementally associated with disease severity (Supplementary Fig. S2D). Conversely, median levels of CD19⁺ B cells, CD4⁺ T cells, and CD8⁺ T cells did not correlate with mortality from acute infection (Supplementary Fig. S2C). However, patients with either severe CD8⁺ lymphopenia (<50 cells/μL) or CD4⁺ lymphopenia (<100 cells/μL) had markedly poor outcomes (Fig. 2D).

Given the high rate of persistent PCR positivity, protracted hospital stay, and rehospitalization in this population, we next asked whether distinct clinical or immunologic factors were associated with persistent COVID-19 infection. To uncouple factors associated with prolonged disease from factors associated with severe initial disease, we focused on the outcomes of patients whose COVID-19 was sufficiently severe to require hospitalization (n = 102) to determine whether patients who died during their primary hospitalization (n = 26), patients who were discharged and not readmitted to the hospital (n = 57), and patients who were readmitted to the hospital for COVID-19–related symptomatology (n = 19) exhibited distinct clinical and laboratory manifestations (Supplementary Fig. S3A). The clinical description of these 19 cases is presented in Table 2. Notably, 13 of 19 patients were on active treatment and 15 had received anti-CD20 therapy in the preceding year. Five patients were allogenic HSCT or CAR T recipients and were in remission at the time of COVID-19 diagnosis. The initial presentation at the time of diagnosis was characterized as asymptomatic in 3 patients, mild in 13, moderate in 2, and severe in 1. In between hospitalizations, all 19 cases had recurrent, persistent, or progressive respiratory symptoms accompanied by new fever in 53% of cases. CT demonstrated progressive ground-glass pulmonary infiltrates in all cases without a clear alternate etiology (Supplementary Fig. S3B). The median duration of PCR positivity was 59 days (range, 26–344 days). During a median follow-up of 277 days, 13 (68%) individuals with persistent COVID-19 manifestations required discontinuation, interruption, or modification of lymphoma-directed therapy (Table 2). Five patients died during the follow-up period; four deaths were attributed to progressive respiratory decline due to symptomatic COVID-19, and one death was attributed to both progressive symptomatic COVID-19 and concurrent pulmonary coccidiomycosis.

Relapsed or refractory lymphoma and recent systemic therapy were independent predictors of hospital readmission for symptomatic COVID-19, the strongest individual treatments of which were anti-CD20 therapy [HR, 6.57; 95% confidence interval (CI), 2.37–18.26] and bendamustine (HR, 4.72; 95% CI, 1.37–16.23); notably, BTK inhibitors were not associated with readmission (Supplementary Fig. S3C). On multivariate analysis, only anti-CD20 therapy was independently associated with hospital readmission (HR, 4.23; 95% CI, 1.38–13.02; Fig. 3A and B). Among the laboratory parameters, a decrease in CD4⁺ but not CD8⁺ T cells correlated with risk of readmission (Fig. 3C). Notably, there was no significant difference in initial inflammatory markers between patients who were or were not rehospitalized for COVID-19 (Supplementary Fig. S3D). Taken together, these data indicate that immunosuppression, particularly loss of B-cell and CD4 T-cell immunity driven at least in part by B cell–directed therapies, is a risk factor for persistent symptomatic COVID-19 infection and suggests that CD4⁺ T-cell and/or B-cell function may contribute to complete viral eradication. Consistent with this hypothesis, patients who were ultimately rehospitalized for COVID-19 exhibited an increased proportion of activated CD4⁺ and CD8⁺ T cells, suggesting ongoing active infection in these patients (Fig. 3C).

Viral evolution, reinfection, and immune escape have been described in immunocompromised patients with COVID-19 infection (8). To determine whether there was any evidence of viral evolution, coinfection with different strains, or reinfection in patients with lymphoma with persistent RNA detection and clinical symptoms, we analyzed 71 longitudinal samples from 18 of 19 persistently infected patients. Overall, there was >90% coverage for 64 analyzed samples. The time span of the analyzed samples relative to the time of diagnosis is depicted in Fig. 4A. The frequency of accumulation of the variants compared with the reference genome (NC_045512), and sustained versus sporadic minor variant detection is demonstrated in Fig. 4A and Supplementary Fig. S4A. Variants and their allelic fractions are available in Supplementary Tables S6 and S7. Overall, single-nucleotide variant (SNV) analysis of consensus sequences (color-coded for the same patient) showed evidence of intrahost evolution suggesting ongoing infection. Conversely, there was no indication of reinfection or coinfection in any of the analyzed samples.

Finally, we computed the viral entropy, as defined in the Methods, across variant allele frequencies of all mutations that differed from the reference sequence, with a higher entropy indicating greater heterogeneity in the viral population. We observed higher viral entropy in patients with impaired CD8⁺ T-cell counts but not in patients who had received anti-CD20 therapy in the past year (Fig. 4B), suggesting that viral diversity may be constrained by an active CD8⁺ T-cell immune population and that patients who harbor higher viral diversity during chronic infection have weaker adaptive responses. To investigate this further, we computed the ratio of divergence at nonsynonymous and synonymous sites (dN/dS) across all SARS-CoV-2 proteins (ORF1ab, ORF1a, S, ORF3a, E, M, ORF6, ORF7a, ORF7b, ORF8, N, ORF10) and found that cluster 3 had a distribution of dN/dS that was lower than that in cluster 1 (median 0 vs. 0.28; Welch t test, P = 0.28), consistent with constrained evolution (Supplementary Fig. S4B). Furthermore, patients in cluster 1 had nonsynonymous mutations in Spike, whereas patients in clusters 2 and 3 did not. Overall, there was an overrepresentation of nonsynonymous mutations in Spike compared with synonymous mutations (Fisher exact test, P = 0.086).

The unique contribution of this report is the description of persistent symptomatic COVID-19 in patients with lymphoid malignancy, a condition predisposed by previous anti-CD20 therapy and observed in 21% of anti-CD20–treated patients in our cohort. The syndrome manifests as a chronic protracted illness marked by slowly progressive or relapsing respiratory illness with progressive lower airway changes on CT imaging and persistent viral RNA detection. This complication affected 19 of 76 hospitalized patients with initial improvement and resulted in delay of cancer-directed therapy for 13 of 19 patients; 4 of 19 eventually died of the direct long-term sequelae of chronic COVID-19 infection. At a virologic level, we did not find any evidence of reinfection or coinfection with reactivation. Intrahost evolution of the virus suggested ongoing replication. Our analysis points to immune dysregulation characterized by combined loss of B-cell function and impaired CD4⁺ T-cell counts (CD4 <100 cells/μL) and compensatory innate immune and CD8⁺ T-cell activation as the basis for this chronic syndrome.

Multiple studies thus far have established a high mortality rate in patients with cancer who have acute COVID-19 infection (4, 6, 14, 15); the mechanistic explanations for this increased mortality rate have not been fully elucidated but may involve immune dysregulation and decreased functional status. The toxicity from COVID-19 infection clearly extends beyond the period of acute viral infection; the so-called long COVID syndrome is still an area of active exploration but is thought to be largely due to sequelae from prior infection rather than ongoing active infection (16, 17). We had previously reported the presence of persistent PCR positivity with associated culturable virus in HSCT recipients lasting up to 3 months following initial COVID-19 infection (18–22).

We now describe chronic persistent COVID-19 infection as a separate clinical entity in a significant proportion of patients with hematologic malignancy and identify unique clinical and virologic characteristics associated with this syndrome. The immunologic factors associated with death during acute severe COVID-19 infection among patients with hematologic cancer include increased acute-phase reactants, decreased CD4⁺ and CD8⁺ lymphopenia, and limited diversity in T-cell responses. These deficits are distinct from those observed in chronic SARS-CoV-2 infection, which is characterized by CD4⁺ lymphopenia and B-cell aplasia. Furthermore, the immune profile risk is also different from prolonged infection in immunocompromised patients with other severe respiratory virus illnesses such as influenza and respiratory syncytial virus, in which the extent of cell-mediated immunosuppression drives viral persistence and is almost exclusively observed after HSCT (23, 24). Overall, our findings suggest that distinct arms of the immune response play dominant roles in controlling acute infection due to a novel virus and influence permanent SARS-CoV-2 viral clearance (25–28). B-cell depletion due to anti-CD20 antibody therapy has been associated with viral persistence and reactivation, especially in chronically infected patients with hepatitis B following chemoimmunotherapy, and an elevated risk of progressive multifocal leukoencephalopathy, herpes simplex virus, and varicella-zoster virus has also been reported, but an association with prolonged viral shedding from respiratory viruses had yet to be described (29, 30). The role of CD4⁺ T-cell depletion in viral persistence is less clear. CD4⁺ T-cell help is critical for both B-cell maturation and optimal CD8⁺ T-cell survival and function (31); conversely, B-cell depletion may impair naïve CD4⁺ T-cell homeostasis (32). Interestingly, greater intrahost viral diversity was observed mainly in patients with impaired CD8⁺ T-cell immunity. This may reflect greater evolution in the setting of partial immunity, as has been reported previously (20). Whether B-cell depletion alters the COVID-19–reactive CD8⁺ T-cell repertoire is an area of ongoing investigation.

There are several limitations in our current study. Laboratory markers, including flow cytometry data and viral sampling, were ordered at the discretion of the treating physician and were therefore not uniformly available across patients. In addition, viral viability was not assessed in patients with persistent COVID-19 PCR positivity. We also detected an association of viral entropy with viral sequencing depth (Supplementary Fig. S4C), likely due to the increased ability to detect low-frequency variants at greater sequencing depth; whether the presence of intact CD8⁺ T-cell immunity altered viral sequencing depth requires further investigation. Finally, the management of these patients was heterogeneous and physician-dependent with rapidly emerging evidence. Despite this, the high volume of patients with hematologic malignancy and immune-depleting therapies at our institution enabled the characterization of this distinct complication of COVID-19.

In summary, the results of this work have significant implications for devising effective preventive and treatment strategies for COVID-19–infected patients with lymphoid malignancies. The high risk for adverse outcomes with acute COVID-19 in patients with lymphoid cancers is well recognized in multiple previous reports (7, 33). We corroborate this finding among our large study cohort. Most important, we demonstrate the substantial morbidity and mortality associated with the long-term effects of chronic COVID-19 infection in patients with lymphoma unable to attain adequate viral clearance and potentially posing a transmission risk to others. Delaying B cell–depleting therapy should be considered in select cases, such as noncurative intent treatment, with individualized decision-making between the patient and the physician. Future studies should elucidate the immunologic defects associated with this clinical phenomenon beyond the recent anti-CD20 treatment risk identified in our report. When taken in conjunction with emerging evidence on suboptimal COVID-19 vaccine response in anti-CD20–treated patients (34–37), our findings urge for prioritization of passive immunoprophylactic approaches in high-risk patients with lymphoid malignancy.

Memorial Sloan Kettering Cancer Center (MSKCC; New York, NY) is 514-bed tertiary care cancer with 22,417 annual admissions and 160,298 inpatient days in 2020. In 2019, MSKCC providers saw 18,800 patients with a new cancer diagnosis, including 946 patients with lymphoma and 335 and 235 cases of leukemia and myeloma, respectively. From March 10, 2020, until February 28, 2021, this study included all consecutive adults with laboratory-confirmed SARS-CoV-2 infection and underlying hematologic malignancy. Identification of case patients and their medical background and clinical course during COVID-19 illness were abstracted from electronic medical records. Longitudinal respiratory samples from 18 patients with lymphoid malignancies and persistent COVID-19 were retrieved for whole-genome sequencing.

SARS-CoV-2 RNA was detected in nasopharyngeal swabs (NPS) or saliva samples using a laboratory-developed test as previously described (4, 38). Testing was also performed using several commercial assays, including the TaqPath COVID-19 Combo Kit (Thermo Fisher Scientific) targeting the N, S, and ORF genes; the Cobas SARS-CoV-2 test (Roche Molecular Diagnostics) targeting the ORF1 a/b and E gene; and the Xpert Xpress SARS-CoV-2 test (Cepheid) targeting the N and E genes. Samples were reported as positive per manufacturers' instructions. The cycle threshold (C_t) value, a semiquantitative estimate of the viral SARS-CoV-2 RNA load, was retrieved for all gene targets from each instrument record.

Whole-genome sequencing was performed on all samples with high enough viral load (i.e., C_t value <30). Total viral nucleic acids were isolated from 200 μL NPS or saliva samples on the KingFisher Flex Magnetic Particle Processor using the MagMAX Viral/Pathogen Nucleic Acid Isolation Kit (Thermo Fisher Scientific). Amplicon sequencing was performed following the Artic protocol with version 3 primers (Integrated DNA Technologies). Following cDNA synthesis and multiplexed PCR, libraries were prepared for the two amplicons pools using the Nextera XT DNA Kit followed by sequencing on an Illumina Miseq platform (Illumina) as a paired end (2 × 150 base pair reads). We mapped the reads using STAR (version 2.7.7a) to the combined human (hg38 p.13, gencode release 36) and reference SARS-CoV-2 (NC_045512) genome (39). Duplicates were removed using picard (version 2.24.2; http://broadinstitute.github.io/picard/). Reads were subset to those mapping to the viral genome and subsequently postprocessed as follows: first, ends of the reads were trimmed until the 10 bases closest to the end mapped to the reference with no mismatches. After that, the trimmed reads with overall identity to the reference less than 95% and reads shorter than 40 base pairs were discarded. In addition to that, if there were two mismatches to the reference within 5 bases of each other, the quality score for the corresponding bases was set to 2. After that, the reads were written to FASTQ format and mapped to the SARS-CoV-2 genome using bwa (version 0.7.17-r1188; refs. 40, 41). SNVs were called using exactSNP program from the Subread package (v 2.0.0; ref. 42). When we encountered major variants with respect to the reference genome, we applied those variants and reran SNV calling with the updated reference. Settings that we modified from the default for SNV calling are -Q 7 -n 2 -s 20 -T 5. We increased the default P value cutoff (changed the -Q parameter from the default value of 12 to 7) to account for the fact that the size of the SARS-CoV-2 genome is 10⁵ times smaller than the size at which the human genome defaults are targeted. When calling minor variants, we required the depth of coverage to be at least 50.

Consensus sequences of the viral genomes were assembled using bcftools (43, 44). SARS-CoV-2 lineages were assigned using pangolin https://pangolin.cog-uk.io/ (45).

The Shannon entropy was computed for each variant site with the major and minor variant allele fractions:

where p_a is the frequency of the minor allele at the site a. The total “viral entropy” per sequencing sample is the sum across the sites within that sample. This ignores the possible correlation between the polymorphic sites (i.e., minor variants observed at sites A and B can be present in the same clone or in two different clones), which generally cannot be determined using the short reads. On a per-patient basis, we used the sample viral entropy corresponding to the latest time point. Categorical tests were conducted using SciPy's ttest_ind function (46). Linear and rank correlations to flow cytometry data were determined using SciPy's pearsonr and spearmanr functions (46).

Peripheral blood was collected into BD Horizon Dri tubes (catalog no. 625642; BD Biosciences). Immunophenotyping of peripheral blood mononuclear cells via flow cytometry was performed in the MSKCC clinical laboratory. The lymphocyte panel included CD45 FITC (BD 340664, clone 2D1), CD56+16 PE (BD 340705, clone B73.1; BD 340724, clone NCAM 16.2), CD4 PerCP-Cy5.5 (BD 341653, clone SK3), CD45RA PC7 (BD 649457, clone L48), CD19 APC (BD 340722, clone SJ25C1), CD8 APC-H7 (BD 641409, clone SK1), and CD3 BV 421 (BD 562426, clone UCHT1). The naïve/effector T panel included CD45 FITC (BD 340664, clone 2D1), CCR7 PE (BD 560765, clone 150503), CD4 PerCP-Cy5.5 (BD 341653, clone SK3), CD38 APC (BioLegend, 303510, clone HIT2), HLA-DR V500 (BD 561224, clone G46–6), CD45RA PC7 (BD 649457, clone L48), CD8 APC-H7 (BD 641409, clone SK1), and CD3 BV 421 (BD 562426, clone UCHT1). The immune phenotypes were based on NIH vaccine consensus panels and the Human Immunology Project. Samples were acquired on a BD FACSCanto using FACSDiva software. Eight-color peripheral blood cytometry for annotation of T-cell, B-cell, and natural killer (NK)–cell subsets was performed in 54.7% of hospitalized patients, and 12-color peripheral blood flow cytometry for T-cell subset analysis was performed in 38.4%.

All sequences were uploaded to GISAID (please see the accession numbers in the Supplementary Data file). The MSKCC Institutional Review Board granted a Health Insurance Portability and Accountability Act waiver of authorization to conduct this study. Clinical data were abstracted from the electronic medical record into standardized case report forms. Clinical laboratory data were abstracted from the date closest to research blood collection. In addition to standard laboratory analyses performed on patients during acute COVID-19 infection, inflammatory markers including CRP, IL6, D-dimer, ferritin, fibrinogen, and LDH were measured in most inpatients. In addition, 8-color peripheral blood cytometry for annotation of T-, B-, and NK-cell subsets was performed in 54.7% of hospitalized patients, and 12-color peripheral blood flow cytometry for T-cell subset analysis was performed in 38.4%.

Descriptive statistics were used to summarize the data. Continuous variables were presented as means and SDs or as medians and ranges, as appropriate, and categorical variables were presented as counts and percentages. Overall survival (OS) was measured from the date of laboratory-confirmed COVID-19 to last follow-up or death. Data were censored on April 3, 2021. The probability of OS was estimated using the Kaplan–Meier method and differences compared using the log-rank test. Cox proportional hazard regression models were used to identify predictors of OS. Multivariate analysis using the Cox proportional hazards model was used to identify the potential independent effects of those factors that attained a P value of 0.1 in the univariate analysis. Risk factors for persistent SARS-CoV-2 positivity were analyzed using multivariate logistic regression; the final model was constructed based on a univariate P value of 0.2. Differences in distributions of characteristics of patients were analyzed with the use of Pearson's χ² test for nominal or categorical variables and the Mann–Whitney test for ordinal or continuous variables. Analyses were performed with the use of SPSS software, version 26 (SPSS, Inc.) and R version 4.0 (R Development Core Team). Exact methods were used to calculate confidence intervals for the rate ratios. P values were calculated with the use of two-sided exact tests; a P value of less than 0.05 was considered to indicate statistical significance.

To determine whether there were distinct immunologic responses to COVID-19 in our patient population, we identified 36 patients in whom at least 90% of laboratory analyses were performed and imputed missing data in these patients with the median of each feature (Fig. 2B). We projected the multivariate laboratory data by conducting a Uniform Manifold Approximation and Project (UMAP) in two dimensions (Supplementary Fig. S2A; ref. 47). Next, we computed the Earth Mover's Distance of each point in UMAP space to each other point (48). We clustered these distances and identified three clusters based on immunologic laboratory markers alone.

N.E. Babady reports personal fees from Roche Molecular and grants from GenMark DX outside the submitted work. B. Greenbaum reports grants and personal fees from Merck and Bristol Myers Squibb; personal fees from ROME Therapeutics, Tx, Chugai Pharmaceutical Co., and Darwin Health, and personal fees from PMV Pharma outside the submitted work. S.A. Vardhana reports personal fees from Immunai and personal fees from ADC Therapeutics outside the submitted work. No disclosures were reported by the other authors.

C.Y. Lee: Conceptualization, data curation, formal analysis, investigation, writing–original draft, writing–review and editing. M.K. Shah: Data curation, supervision, investigation, writing–review and editing. D. Hoyos: Data curation, formal analysis. A. Solovyov: Data curation, formal analysis. M. Douglas: Data curation. Y. Taur: Data curation, formal analysis, writing–review and editing. P. Maslak: Data curation, investigation. N.E. Babady: Data curation, formal analysis, investigation, writing–review and editing. B. Greenbaum: Data curation, supervision, writing–review and editing. M. Kamboj: Data curation, supervision, writing–review and editing. S.A. Vardhana: Conceptualization, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing.

This work was supported by P30 CA 008748 from NIH/NCI Cancer Center Support Grant; U01 CA228963 and V Foundation (to B. Greenbaum and A. Solovyov); R01 CA240924, Stand Up To Cancer, and Lustgarten Foundation (to B. Greenbaum and D. Hoyos); Lymphoma Research Foundation (to C.Y. Lee); philanthropic support from Jack and Dorothy Byrne Foundation (to M. Kamboj and N.E. Babady); Pershing Square Sohn Cancer Research Foundation (to S.A. Vardhana and B. Greenbaum); and Conrad Hilton Foundation and Parker Institute for Cancer Immunotherapy (to S.A. Vardhana).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.