Comprehensive analyses of immunoglobulin proteome and clinical variables identify biomarkers to predict mortality in patients with influenza-associated lower respiratory tract infection

Background: Influenza-associated lower respiratory tract infection (I-LRTI) brings a heavy clinical burden, and clinicians lack an effective prognostic evaluation system to control disease progression. Methods: This was a prospective, observational study, and the endpoint was 28-day mortality. Plasma microarrays were used for quantitative analysis of immunoglobulin (Ig) and its subclasses. Prognostic factors from Ig and clinical variables in the patients with I-LRTI were identified to create a prediction model. Results: To address this issue, we prospectively and observationally studied the difference of immunoglobulin proteome and clinical variables between survivors and non-survivals in 107 patients with influenza-associated lower respiratory tract infection (I-LRTI) selected from four hospitals affiliated to Capital Medical University. The results identified 17 variables with significant or marginally statistical differences by univariate analysis, including lymphocyte count (LY), monocytes count (MO), CD3 + CD4 + T-cell count, CD3 + CD8 + T-cell count, IgA, IgA1, IgG2, IgG4, CRP, PCT, D-dimer, oxygenation index, glycosylated hemoglobin, lactic acid (LAC), base excess of blood, lactic dehydrogenase, and α-hydroxybutyrate dehydrogenase. Furthermore, we analyzed the correlations of all the variables by hierarchical clustering analysis in which different functional modules were formed between survival and non-survival groups that are associated with the immunity and severe infection. At last, we built a prediction model with nine variables (D-dimer, days from onset to ED, IgA, IgG2, LAC, LY, MO, Staphylococcus aureus co-infection and age), with which the AUC value of 0.810 (95% CI 0.755-0.839) was achieved with the evaluation of LOO cross validation. The predictive model was further validated by disease severity evaluation. Conclusion: was with regression-based packages

cellular immunity, oxygenation index, HbA1C and age. The combined prediction model with D-dimer, Days from onset to ED, IgA, IgG2, LAC, LY, MO, S.aureus co-infection and age demonstrate the predictive mortality powerfully in patients with I-LRTI.

Background
Seasonal influenza epidemic accounted for an estimated 145 000 (95% uncertainty interval [UI] 99 000-200 000) deaths among all ages worldwide annually. [1] Although influenza vaccine is an available countermeasure to mitigate the considerable influenzarelated mortality burden, [2] vaccine effectiveness is low compared with other viral vaccines, and the induced immune response is narrow and short-lived. [3] Current influenza vaccines are imperfect and the expected benefits of vaccination programs might be overstated, especially in elderly people. [4,5] For people who are not effectively prevented and are infected with influenza, neuraminidase inhibitors (NAIs) are recommended as the most effective antiviral agents. [6] However, in clinical practice, antiinfluenza treatment for hospitalized patients with suspected influenza is not always initiated at admission and may be delayed while patients await results of diagnostic tests.
[7] Even more, not all observed studies have confirmed the benefits of NAIs treatment in hospitalized influenza patients. [8] Therefore, vaccines or antiviral agents alone tend not to reduce the burden of deaths caused by influenza, comprehensive evaluation based on clinical indicators and individual immune status remains critical.
Severe influenza originated from a respiratory infection but involved multiple organ systems, so it should also be caused by highly interacting networks of parameters in body instead of the isolated ones. [9][10][11][12] Particularly notable is the importance of omics analysis and the interrelationship between these multiple indicators, which can reflect both physical and non-physical associations between them. [13] There is no widely accepted clinical system for evaluating the prognosis of influenza patients. Although commonly used clinical evaluation systems, such as pneumonia severity index (PSI) [14] and Sepsisrelated Organ Failure Assessment (SOFA) score [15] were multi-index evaluation systems, they were based on clinical data of patients with community-acquired pneumonia or sepsis, and lacked an assessment of the immune system and co-regulation network analysis. Therefore, the efficacy of these systems are limited in evaluating the prognosis of influenza patients. There have been no new reports of physicians exploring evaluation systems in recent years to assess the prognosis of patients with severe influenza.
Protein microarray has the advantages of high sensitivity and accuracy, which makes it possible to detect low abundant proteins in clinic. [16] In this work, we developed and applied a high-throughput and ultra-micro plasma microarray platform to measure the expression of nine immunoglobulin isotypes (IgG, IgG1-4, IgA, IgA1-2, IgM) in hundreds of clinical plasma samples, which proved to be of high reproducibility and stability. The purpose of this study was to perform comprehensive analysis of the changes of immunological proteome and clinical variables, to explore novel pathological mechanisms through the immunoglobulin proteome-clinical variable co-regulation networks and to establish a prediction model for evaluating the prognosis of influenza patients.

Study Design and Participants
We conducted a prospective observational study, and collected patients with suspected influenza-associated lower respiratory tract infection (I-LRTI) in emergency departments

Plasma Sample Preparation
At the same time, 2 mL of peripheral blood was collected in an EDTA anticoagulant tube and thoroughly mixed. At 4 ℃, blood samples were centrifuged at 1300×g for 10 minutes.
After centrifugation, supernatant (plasma) was taken and loaded into a 1.5 mL centrifuge tube. Mark the sample number, date on the centrifugal tube and record the sample information. After processing the sample to save to -80 ℃ refrigerator, avoid freeze and thaw.

Plasma Microarrays
Properly diluted plasma samples (from 10 to 500-fold) and serially diluted standard immunoglobulins and then printed onto a 3D modified slide surface (Capital Biochip Corp, Beijing, China) in two replicates using an Arrayjet microarrayer (Roslin, UK). Phosphatebuffered saline (PBS) and bovine serum albumin (BSA, 1 mg/mL) (Sigma-Aldrich, MO, USA) were used as negative controls. After printing, the plasma microarrays were stored at -20 ℃ until use.
After equilibration to room temperature, the plasma microarrays were assembled into a microarray incubation tray and blocked with 500 μL 1% BSA in each well for 1 hour at room temperature. After removing the BSA, the arrays were incubated with the corresponding fluorescently-labeled antibody combinations for 1 hour at room temperature. The resulting slides were washed three times with PBS containing 0.05% (w/v) Tween 20 (PBST), 5 min/each, and H 2 O twice, 5 min/each. The slide was scanned using a Genepix 4000A microarray scanner (Molecular Devices, CA, USA). The fluorescent images were analyzed and the signal intensity was extracted using a GenePix Pro image analysis software (Molecular Devices, CA, USA). For each Ig subclass, the 4-paramter logistic standard curve between the absolute concentration and the signal intensity was fitted with properly diluted plasma samples by least-squares method and thus the absolute concentration of Ig subclass in plasma samples was imputed by Newton-GC method according to the fitted standard curve.

Statistical Analysis
Mann Whiney U test was used to compare the difference of the means of two groups as the univariate values do not conform to the normal distribution. The exact Fisher's test was used to perform the independence statistical test. If p-value is under 0.05, then the result is considered statistically significant. The Spearman correlation coefficient was calculated and the hierarchical correlation clustering heatmap was plotted with the R package corrplot. Logistic regression-based model was adopted for the discrimination of the survival/non-survival groups of the influenza patients with the recursive feature selection in the LOO cross-validation procedure. ROC curve was plotted and AUC was used to assess the discriminative efficacy of the trained classifier model. The statistical analyses were performed with R. The logistic regression-based prediction and evaluation was performed with the Python packages SciKit-Learn and SciPy.

Patient Characteristics
In the winter of 2018-2019, 280 patients with suspected influenza and accompanied LRTI from the four hospitals were screened. Of these patients, 173 were excluded: 166 had no laboratory positive test evidence for influenza, and 7 had recently been treated with hormones or immunosuppressants for autoimmune disease or solid organ transplantation.

Differential Regulation of Immunoglobulin Proteome and Clinical Indicators between
Survival/Non-survival Groups Each patient in the study cohort was investigated for demographic characteristics (age, sex, and body mass index (BMI)), days from onset to ED, co-morbidities (active cancer, chronic respiratory disease, coronary artery disease, chronic heart failure, chronic hepatopathy, chronic kidney disease, diabetes mellitus), vital signs (heart rate (HR), respiratory rate (RR), mean arterial pressure (MAP), and Glasgow Coma Scale (GCS)), clinical laboratory tests (blood routine examination, blood biochemistry, D-dimer, arterial blood gas analysis, Glycosylated hemoglobin (HbA1C), C-Reactive Protein (CRP), procalcitonin (PCT), and T-cell subset counts), microbiological detections, antiviral administrations, organ supports, and immunoglobulin quantification (IgG, IgG1-4, IgA, IgA1-2, IgM). All of the above indicators were compared between the two groups (Table 1)

Co-regulatory Network Analysis of Clinical Indicators and Immunoglobulin Proteome
Quantification in the Survival/Non-survival Groups Co-regulatory network was used to analyze and compare the correlation characteristics across clinical indicators and immunoglobulin quantification for the survivors and the nonsurvivors respectively, and was showed by hierarchical correlation clustering heatmap To further evaluate the selected optimal indicators for the prediction model, the unsupervised participant-feature hierarchical correlation clustering was also performed and the participants were separated into two clusters. The majority of the actual nonsurvivors were discriminated into the same cluster, which indicated the consistency between the clustered results and the actual grouping (Fig. 4C). This proved the feature combination's discriminative efficacy of the survivors/non-survivors.

Validation of the Model by Disease Severity Evaluation
The risk of progression to critical conditions, including mechanical ventilation requirements or vasoactive agent requirements, can also be evaluated using the predictive model and for mechanical ventilation or vasoactive agent apparently there exists statistical difference between survivors and non-survivors in practice. For the requirement of mechanical ventilation, 15 (15/37) predicted non-survivors and 12 (12/70) predicted survivors should receive mechanical ventilation respectively, exhibiting the statistical difference between two predicted groups (Fig. 5A), which is in consistency with the actual facts (In the study cohort, 18 (18/27) non-survivors and 9 (9/80) survivors actually received mechanical ventilation (Fig. 5B)).For the requirement of vasoactive agent, 10 (10/37) predictive non-survivors and 8 (8/70) predicted survivors should receive vasoactive agents respectively, exhibiting the marginally statistical difference between two predicted groups (Fig. 5C), which is in accordance with the actual facts (In the study cohort, 10 (10/27) non-survivors and 8 (8/80) survivors received vasoactive agents, respectively (Fig. 5D)). The results of both mechanical ventilation and vasoactive agents demands proved the potential of the predictive model. Another study argued that outpatient data on patients with relatively mild illnesses should not form the basis for policies on the management of more severe disease.

Discussion
[27] Therefore, comprehensive clinical evaluation and multi-system support are still crucial and indispensable.
Clinicians tend to evaluate the prognosis of I-LRTI patients with multiple factors rather than single factor, but there is currently no efficacious evaluation system, which could reflect disturbances in the physiological system. Host immunity, especially humoral immunity, plays an important role in the defense and elimination of influenza viruses.
Influenza virus infection, as an initiating factor, can lead to multiple organ dysfunction and usually co-infect with pathogenic bacteria. This process is associated with higher mortality rate than individual viral infection.
[28] The complex interaction of initial influenza virus, co-pathogen invasion and host immune response complicates the pathophysiological process, and its molecular biological mechanism needs to be explored urgently.
[29] In order to facilitate clinical application, we explored the changing pattern of simple and easily available biomarkers to reveal the underlying pathophysiological signals in I-LRTI status and their changes prior to death. In this study, co-regulation networks and a multi-factor model were established by combining immune cell count, immunoglobulin quantification and clinical indicators routinely measured in clinic.
We developed a large-scale, ultra-micro plasma immunoglobulin quantification chip, which has high reproducibility (Pearson's r > 0.9) and high sensitivity (~ attomolar Second, a decrease in immune cell count predicts a poor prognosis. Lymphocytopenia was considered as a risk factor for mortality in patients with pneumonia caused by influenza virus. [38] Decreased monocytes was found to have a marginal difference between the survivors and the non-survivors, which may also be a prognostic factor. So far, few studies have been reported, and further large sample validation is required. The decrease of CD3 + CD4 + T-cell and CD3 + CD8 + T-cell had been proved to be closely associated with increased mortality and deteriorated primary and subsequent infections in influenza A virus infected animals. [39,40] Similar results were found in our cohort study, where decreased CD3 + CD4 + T-cell and CD3 + CD8 + T-cell counts were associated with mortality and bacterial co-infection (especially S. aureus co-infection). Third, some biomarkers such as D-dimer, LDH, HBDH, HbA1C, LAC, HCO 3 − , BEB, CRP and PCT were also strong prognostic factors. A surge of LDH suggested metabolic reprogramming, and was associated with increased mortality in patients with sepsis. [41] Currently, we found a similar phenomenon in the influenza cohort. Our previous study confirmed that elevated PCT levels in influenza patients reflected bacterial co-infection, which was an independent risk factor for mortality. [42] We further comprehensively analyzed the immunoglobulin profile and laboratory parameters in patients with I-LRTI between survivors and non-survivors and established corresponding co-regulatory network models for different groups. In the cluster heatmap, including deranged T-cell counts. [44] In our previous studies, patients with influenza was found to be susceptible to community-acquired bacterial co-infections, and was characterized by a synergistic lethal effect of influenza and S. aureus.
[20] In view of the low sensitivity and lagging results of bacteriological investigation, our another study have quantified bacterial co-infection in influenza patients using PCT and identified an independent association between PCT and 28-day mortality. [42] However, the correlation between bacterial co-infection in hospitalized patients with influenza and their immune indicators has not been revealed. In the current study, we further elucidated the correlation between this phenomenon and T cell mediated immunity. This finding is of great clinical significance. One is that T cell counts should be instantly detected to assess the risk stratification when a severe influenza patient visits ED. The other is to predict the risk of bacterial co-infection in patients with I-LRTI based on the T cell counts, and then to manage antibiotics and regulate immune function as early as possible.
In this study, we have created an assessment model to predict mortality due to I-LRTI by

Consent for publication
Not applicable.

Availability of data and materials
The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request.

Competing interests
The authors declare that they have no competing interests.         The efficacy of the model in predicting critical condition (mechanical ventilation and vasoactive agents). Fisher's accurate test was used to calculate the p value, and different statistical significance levels were noted (p <0.001 ***, p <0.01 **, p <0.05 *, p < 0.1).