The predicting roles of carcinoembryonic antigen and its underlying mechanism in the progression of coronavirus disease 2019

This study was approved by the Ethics Committee of Jinyintan Hospital (KY-2020-58.01), followed the Standards for Reporting of Diagnostic Accuracy Studies Statement and Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) [13, 14]. A total of 300 hospitalized adult COVID-19 patients diagnosed by reverse transcription polymerase chain reaction (RT-PCR) in Wuhan Jinyintan Hospital from January 1, 2020, to April 30, 2020, were included in the retrospective study. The exclusion criteria were: (1) patients younger than 18; (2) non-hospitalized patients; (3) patients with follow-up period less than 60 days; (4) patients admitted for another reason than COVID-19-related respiratory failure (as patients with specific malignancies could have increased CEA levels without any correlation with COVID-19, all patients with primary malignancy were excluded from the study); (5) patients whose survival time, endpoint (overall survival), demographic information or treatment data was unknown; and (6) patients whose admission carcinoembryonic antigen (CEA) was unknown.

The clinical data in the study were retrieved from the electronic medical record system of Wuhan Jinyintan Hospital on initial admission, including variables of demographic information (age at diagnosis and gender), symptom (fever, cough, expectoration, shortness of breath and diarrhea), comorbidity (diabetes mellitus, hypertension, cardiovascular disease and cerebral infarction hypertension) and therapeutic information (use of glucocorticoid, imaging score, nasal catheter, high flow oxygen intake, ventilation). Additionally, laboratory indexes were also collected including CEA (ng/ml), albumin (g/l), hemoglobin (g/l), neutrophils (× 10⁹/l), lymphocytes (× 10⁹/l), C-reactive protein (CRP, mg/l), hypokalemia, hypocalcemia, hyponatremia, hyperkalemia and hypernatremia. The cutoff values of laboratory and imaging indexes were determined according to the normal values stipulated by the laboratory and imaging department of Wuhan Jinyintan Hospital.

As the endpoint, the survival time and overall survival status of each patient were retrieved. The endpoint of the present study was the overall death of COVID-19 patient, which presented the outcome and prognosis of patients in this study. Patients who were diagnosed after April 30, 2020, were excluded from the study.

The retrospective study started with descriptive statistic: Discontinuous variables were presented as percentages while continuous variables in normal distribution were described as mean ± standard deviation (SD) or else reported as median (range). Two statistical methods were applied to explore potential significant predictors. As initial parameter or nonparametric tests, the Chi-square test was used to compare the outcomes between discontinuous variables, and variance homogeneous and normal distributed continuous variables were compared by the Student t-test; otherwise, the Mann–Whitney U-test or Kruskal–Wallis H-test was used. Besides, the Kaplan–Meier survival analysis was used to determine the prognostic value of each variable. Furthermore, predictors with statistical significance in both parameter or nonparametric tests and Kaplan–Meier survival analysis were selected to construct the multivariate Cox proportional hazard model. The nomogram was established based on the multivariate model to predict the prognosis of COVID-19 patients. The significant prognostic factors in multivariate Cox model were marked with asterisks (*) in the nomogram (*: P < 0.05; **: P < 0.01). Receiver operating characteristic (ROC) curve and calibration curve were drawn to evaluate the discrimination and calibration of the nomogram.

Single-cell RNA-sequence (scRNA-seq) data of COVID-19 patients’ and healthy volunteer’s bronchoalveolar lavage fluid (BALF, accession no. GSE145926) [15] and peripheral blood mononuclear cells (PBMCs, accession no. GSE150728) [16] were download from the Gene Expression Omnibus (GEO). All BALF and PBMC samples were taken at the initial admission.

The preliminary data processing of single-cell RNA-seq data started from the Cell Ranger Single Cell Software Suite 3.3.1 (http://​10xgenomics.​com/​). The pair-ended reads fastq files were trimmed to remove template switch oligo (TSO) sequence and poly-A tail sequence. Then, command of “cellranger count” was used to quantify the clean reads, aligned to the hg38 human genome. The Seurat method was applied to integrated data analysis [17].

In terms of quality control (QC), genes with average read count greater than one and being expressed in at least three single cells were considered for further analysis. Cells with either fewer than 100,000 transcripts or fewer than 1,500 genes were filtered out.

In data processing, first, variance stabilizing transformation (VST) method was used to identify variable genes. Variable genes were input as initial features for principal component analysis (PCA) [17]. Then, the principal components (PCs) with P values < 0.05 were filtered by the jackstraw analysis and were incorporated into further UMAP (uniform manifold approximation and projection) and t-distributed stochastic neighbor embedding (t-SNE) to identify cell subclusters (resolution = 0.50) [18]. Only the genes with |log2 fold change (FC)|> 0.5 and false discovery rate (FDR) value < 0.05 were identified as differentially expressed genes (DEGs) among cell subclusters. Feature plots and violin plots were utilized to illustrate the distribution and expression of DEGs, respectively. Additionally, scMatch [19], singleR [20] and CellMarker [21] were used as references to define each cluster. Cell trajectory and pseudo-time analysis was performed by monocle2 [22]. Furthermore, 50 hallmark gene sets were retrieved from the Molecular Signatures Database (MSigDB) version 7.1 (https://​www.​gsea-msigdb.​org/​gsea/​msigdb/​index.​jsp) and gene set variation analysis (GSVA) algorithm was performed to absolutely quantify the activity of signaling pathways in each single cell [23, 24]. Furthermore, the CellphoneDB algorithm was used to identify the cellular communication between pneumonocyte and immune cells [25].

First of all, the distribution and expression of CEA-related genes (CRGs) including CEACAM1, CEACAM3, CEACAM4, CEACAM5, CEACAM6, CEACAM7, CEACAM8, CEACAM16, CEACAM18, CEACAM19, CEACAM20, CEACAM21, CEACAMP1, CEACAMP2, CEACAMP3, CEACAMP4, CEACAMP5, CEACAMP6, CEACAMP7, CEACAMP8, CEACAMP9, CEACAMP10, CEACAMP11 and CEACAM22P were visualized by feature plot and violin plot in BALF and PBMC scRNA-seq data. Then, co-expression (correlation) analysis was performed among CRGs and 50 hallmark of gene sets to identify the potential downstream pathways. The CellphoneDB algorithm was used to illuminate the cellular communication between cells with high CRG expression and other cells. Besides, two data including scRNA-seq data of acute lung injury (ALI) mouse lung (GSE134383) and idiopathic pulmonary fibrosis (IPF) mouse lung (E-HCAD-14) were downloaded to evaluate the distribution and expression of CRGs, key receptor–ligand pair of cellular communication and potential downstream pathways [26‐30].

Only p value of two-sided statistical testing lower than 0.05 was considered statistically significant. All statistical analysis processes were performed with R version 3.6.1 software (Institute for Statistics and Mathematics, Vienna, Austria; www.​r-project.​org).

A total of 300 hospitalized adult COVID-19 patients diagnosed by RT-PCR in Wuhan Jinyintan Hospital from January 1, 2020, to April 30, 2020, were included in this retrospective study.

The baseline characteristics of 300 COVID-19 patients are described in Fig. 1A and Table 1. The cohort comprised 170 males and 130 females, with a median age of 63.0 (range 21.0–90.0) years. After removing 17 of 28 laboratory indicators with missing values more than 20% of the sample size, the results of initial Kaplan–Meier survival analysis (Fig. 1C–D) and parameter or nonparametric tests (Fig. 1E) revealed that only five indicators (serum CEA, lymphocytes, neutrophils, CRP and albumin) were significantly associated with both the imaging score and prognosis of COVID-19 patients (Fig. 1B).

CEA is the only laboratory indicator with significant difference in all the univariate and multivariate analysis. To identify the optimal cutoff point of CEA, the cyclic log-rank test was conducted. The results revealed that CEA = 7.3 ng/ml was the optimal cutoff point with the most significant P value in the log-rank test (Fig. 2A, B). The variable of nasal catheter is integrated into a variable named “Mode of ventilation,” which is a variable with five levels (no oxygen; oxygen with nasal canula (OWNC); oxygen through high flow nasal canula (HFNC); noninvasive ventilation (NIV); and invasive ventilation (IV)). Since mode of ventilation has the property of the ordered categorical variable, even if a patient only takes HFNC on admission, then the condition deteriorates and he or she receives invasive ventilation, the variable "Mode of ventilation" will be marked as IV but not HFNC. Then, 11 potential indicators showing prognostic values in Kaplan–Meier analysis were incorporated into the initial Cox proportional hazard models, along with two demographic information (age and gender). The final multivariate models were constructed to confirm the effects of significant covariates in the initial models of the overall survival (OS) of COVID-19 patients (Fig. 2C and Table 2). Patients with lower CEA had better OS (HR 0.57; 95% CI 0.354 to 0.920; P = 0.020) in the multivariate model, suggesting CEA as a prognostic indicator for COVID-19 patients independently.

The prognostic nomogram was constructed based on the multivariate Cox model including CEA to predict the 3-week and 5-week overall survival probability of COVID-19 patients (Fig. 3A). The calibration curve and the ROC curve (AUC = 0.783) suggested acceptable calibration and discrimination of the nomogram, respectively (Fig. 3B; Additional file 5: Figure S1A-B). Besides, the risk score (RS) was calculated by the formula generated by the multivariate Cox model. The scatter plot (Additional file 5: Figure S1C) and risk curve (Additional file 5: Figure S1D) of the model demonstrated the RS distribution based on risk score of each patient. The Kaplan–Meier curve suggested the prognostic value of the RS (Fig. 3C, P < 0.001). Additionally, the residual distribution of the multivariate model was accessed by the residual plot (Additional file 5: Figure S1E). Eventually, the RS was shown to be an independent prognostic indicator for COVID-19 patients in both univariate (HR = 4.105, 95% CI (2.140 − 7.874), P < 0.001, Fig. 3D) and multivariate (HR = 1.053, 95% CI (1.026 − 1.082), P < 0.001, Fig. 3E) Cox regression model corrected by demographics.

Additionally, since many laboratory values were not analyzed as there were more than 20% missing data for 17 out of 28 variables, raising some concern on the fact some prognosis factors could remain unidentified, the results of Kaplan–Meier analysis of 17 laboratory values with more than 20% missing data were illustrated by survival curves (Additiona file 1: Table S1; Additional file 5: Figure S1F). Some laboratory indicators, such as ALT (alanine aminotransferase), AST (aspartate aminotransferase), PLT (platelet), PCT (procalcitonin), HBDH (α-hydroxybutyrate dehydrogenase), LDH (lactate dehydrogenase), ferritin, IL-6 (interleukin 6), D-dimer, myoglobin and HsTNT (high-sensitivity troponin) did show prognostic values in univariate analyses.

In order to further prove whether these laboratory indicators with significance in univariate analysis can be used as independent prognostic factors, we, respectively, incorporated these variables into the multivariate Cox regression model and conducted multivariate Cox regression analysis for 22 times (keep or remove missing values). In the cohort where missing values were removed, only the regression models, respectively, including AST, ALT, D-dimer and PLT were converged (due to the uneven distribution of some variables level and events number when removing missing values), suggesting that CEA was an independent prognostic factor in all multivariate models and both normal PLT (HR = 0.635, 95% CI (0.408 to 0.990), P = 0.045) and normal ferritin (HR = 0.094, 95% CI (0.010 to 0.860), P = 0.037) were also independent favorable factors (Additiona file 2: Supplementary material 1) compared abnormal levels. In the cohort keeping missing values, a total of 11 regression models were converged and CEA was an independent prognostic factor in all multivariate models. Additionally, patients with normal PLT (HR 0.624; 95% CI 0.406 to 0.960; P = 0.031), ferritin (HR 0.089; 95% CI 0.010 to 0.750; P = 0.026), IL − 6 (HR 0.494; 95% CI 0.264 to 0.930; P = 0.028) and myoglobin (HR 0.520; 95% CI 0.303 to 0.890; P = 0.017) had better OS than patients with abnormal levels of these laboratory indicators in the multivariate models (Additiona file 3: Supplementary material 2). However, most of these laboratory indicators did not meet the requirements for inclusion in multivariate analysis [31, 32]. Inclusion of more covariates does not necessarily lead to higher accuracy, but instead to overfitting, and should be avoided. Thus, only five indicators (serum CEA, lymphocytes, neutrophils, CRP and albumin) were incorporated into the multivariate Cox model. And all these results suggested CEA as a prognostic indicator for COVID-19 patients independently.

As smokers or patients with specific malignancies could have increased CEA levels without any correlation with COVID-19, all patients with primary malignancy were excluded from the study. Furthermore, in order to identify the association between CEA levels and smoking, two subgroup Cox proportional hazard regression models including smoking status (keep or remove missing values) were constructed, suggesting that the CEA level (HR 0.547; 95% CI 0.318 to 0.940; P = 0.037) (remove missing values) (HR 0.620; 95% CI 0.384 to 0.990; P = 0.048) (keep missing values) was still an independent prognostic indicator for COVID-19 patients (SAdditiona file 4: Supplementary material 3). Moreover, CRG expression levels were retrieved from the RNA-seq data of lung squamous cell carcinoma (LUSC) available from The Cancer Genome Atlas (TCGA). The results of rank-sum tests showed that CEA levels in both the serum of COVID-19 patients (Additional file 6: Figure S2A) and the tissues of lung cancer (Additional file 6: Figure S2B-C) were significantly higher in smokers than in non-smokers.

Additionally, to further evaluate the prognostic value of the mode of ventilation, the Kaplan–Meier analysis was performed. The results suggested that the mode of ventilation was significantly associated with the prognosis of COVID-19 patients (P < 0.001) (Additional file 7: Figure S3).

The scRNA-seq data of bronchoalveolar lavage fluid (BALF) from three patients with moderate COVID-19 (C141, C142 and C144), six patients with severe or critical infection (C143, C145, C146, C148, C149 and C152) and three healthy controls (C51, C52 and C100) [15] were download from the GEO database. (This part was a secondary analysis of published data.) A UAMP analysis was performed in 63,010 cells in BALF and clearly identified 20 clusters and 11 cell types including B cell, CD4 + T cell, CD8 + T cell, dendritic cell, macrophage, monocyte, natural killer cell, neutrophil, T cell: gamma–delta, type I pneumocyte, type II pneumocyte (Fig. 4A, B, Additional file 8: Figure S4A-B). The expression levels and expression percentages of the marker genes in each cell type were displayed in Additional file 8: Figure S4C and S4D, respectively. Except for macrophages and type I and type II pneumocytes, all other immune cells (B cell, CD4 + T cell, CD8 + T cell, dendritic cell, monocyte, natural killer cell, neutrophil and T cell: gamma–delta) were dominantly differentiated and chemotactic in the BALF of COVID-19 patients compared to healthy volunteer (Fig. 4C). Furthermore, in terms of the expression and distribution of CRGs, CEACAM1, CEACAM3, CEACAM5, CEACAM6, CEACAM7, CEACAM8 and CEACAM21 were differentially expressed among moderate, severe/critical COVID-19 patients and healthy controls while CEACAM5 and CEACAM6 were significantly localized in the type II pneumocytes of COVID-19 patients (Fig. 4D–E). Additionally, the cell cycle analysis suggested that COVID-19 patients were more likely to have cells in the G2M and S stages (Additional file 9: Figure S5A-B). Moreover, cellphoneDB analysis illustrated that pneumocytes of COVID-19 patients communicated extensively with other immune cells through CRGs. In particular, the number of type II pneumocyte was found to significantly increase in COVID-19 and have cross talk with neutrophils via CEACAM8-CEACAM6 (Fig. 4F, Additional file 9: Figure S5C). Figure 4G summarizes the absolute quantification of 50 hallmark gene sets calculated the GSVA in type I and type II pneumocytes, suggesting that the interferon response and cell proliferation signaling pathways were significantly activated in type II pneumocytes highly expressing CRGs of COVID-19 patients.

Similarly, the scRNA-seq data of 94,448 PBMCs from six patients with moderate COVID-19 and six healthy volunteers were also downloaded (This part was a secondary analysis of published data) [16]. The UAMP analysis identified 18 clusters and 10 cell types including B cell, B cell Naïve, CD4 + T cell, CD8 + T cell, macrophage–monocyte, myelocyte, natural killer cell, neutrophil, plasma cell. platelets (Fig. 5A, B; Additional file 10: Figure S6A-B). All types of immune cell were significantly differentiated and chemotactic in COVID-19 patients’ PBMCs compared to healthy controls (Fig. 5C). CEACAM1, CEACAM4, CEACAM6 and CEACAM8 were differentially expressed between PBMCs of COVID-19 patients and healthy controls, while CEACAM1, CEACAM6 and CEACAM8 were significantly localized in a novel cell subtype annotated as developing neutrophils, which was significantly differentiated and chemotactic (Fig. 5D, E). Additionally, dot plots summarized the results of Gene Ontology (GO) and the Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis. Based on GO analysis, the DEGs were associated with the neutrophil activation and degranulation (Fig. 5F). According to KEGG analysis, the DEGs were related to protein processing in endoplasmic reticulum, phagosome, Epstein–Barr virus infection and tuberculosis (Fig. 5F). Additionally, cell cycle analysis suggested that the developing neutrophils in COVID-19 patients’ PBMCs were all engaged in the G2M and S stages (Additional file 10: Figure S6C-D). And more extensive cellular communication analysis performed by iTALK algorithm (https://​github.​com/​Coolgenome/​iTALK/​) further illustrated mechanisms between the developing neutrophils and the other PBMCs (Additional file 10: Figure S6E-F). Eventually, all neutrophils were extracted separately and re-analyzed for dimensionality reduction. The UAMP analysis identified two cell types including canonical neutrophils and developing neutrophils (Additional file 11: Figure S7A-B). A significant increase in the number of developing neutrophils was found in COVID-19 while CEACAM1, CEACAM6 and CEACAM8 were also significantly co-localized developing neutrophils (Additional file 11: Figure S7C-D).

The scRNA-seq data of ALI mouse lungs (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​acc=​GSE134383) and IPF mouse lungs (https://​www.​ebi.​ac.​uk/​gxa/​sc/​experiments/​E-HCAD-14/​results/​tsne) were also downloaded to evaluate the distribution and expression of CRGs, key receptor–ligand pair of cellular communication and potential downstream pathways [26‐30]. The UAMP analysis identified 18 clusters and 6 cell types in ALI mouse lungs while there were no abnormal expressions of CRGs (Fig. 6A–C). The interferon response and cell proliferation signaling pathways were not significantly activated in type II pneumocytes of ALI mouse lungs (Fig. 6D). Similarly, abnormal expressions of CRGs were also not detected in 31 clusters and 10 cell types of IPF mouse lungs (Fig. 6E–G). Besides, the heatmap of GSVA also showed that the interferon response and cell proliferation signaling pathways were not activated in type II pneumocytes of IPF mouse lungs (Fig. 6H). Therefore, abnormal expressions of CRGs in COVID-19 patients were COVID-19-specific and not related to CEA involvement in ALI and IPF.

String [33] database was utilized to construct the PPI network of CRGs, illustrating that several CRGs had direct interaction with a variety of immune cell surface markers (Fig. 7A–C). Besides, the protein expression levels of CRGs in normal lung samples of The Human Protein Atlas were also checked [34], showing that only CEACAM21 was stained moderately in pneumocytes while the proteins of CEACAM5, CEACAM6 and CEACAM8 were not detected in normal lung samples (Fig. 7D). To sum up, the prognostic value of CEA was identified in COVID-19 patients and the developing neutrophils/neutrophil progenitors (highly expressed CEACAM8, ELANE and LYZ) could have the cross talk with type II pneumocyte (highly expressed CEACAM5 and CEACAM6) via CEACAM8-CEACAM6.