Targeted MRM-analysis of plasma proteins in frozen whole blood samples from patients with COVID-19: a retrospective study

Study population

Participants were recruited in the department for treatment of the novel coronavirus infection at the Federal Research Clinical Center (FRCC) under Federal Medical and Biological Agency of Russia from 25.06.2021 till 19.07.2021. The samples were collected (on the 1st–5th day after admission to the clinic) from 99 patients with COVID-19 infection confirmed by RT-qPCR. All patients (n=99) have been hospitalized; all of them had viral pneumonia confirmed by high resolution computer tomography of the chest. The group of healthy controls (n=32) was recruited among the FRCC staff. Informed consent was obtained from all participants. The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the local Ethics Committee of the FRCC (clinical protocol No. 5, 11 May 2021). Patients were divided into severity groups according to the respiratory support they needed. Mild (n=41) – did not need oxygen supply, moderate (n=39) needed low flow oxygen support, and severe (n=19) needed high flow nasal oxygen, non invasive or invasive mechanical ventilation (11 of them died). The demographic and clinical characteristics of the studied groups are presented in Table 1.

Whole blood sample collection and preparation for MS

Venous blood was collected using vacuum tubes with K₂₊-EDTA, aliquoted and stored at −80 °C. The study was performed using isotope-labeled standard (SIS) “heavy” peptides, which were added to each sample and acted as internal standards for normalization, and unlabeled “light” peptides (NAT), which were used to create quantitative calibration curves. Synthesis and characterization of SIS and NAT peptides was carried out in the Omics lab at Skoltech using standard procedures, which were previously described in detail [23, 24, 31]. The list of peptides and proteins is provided in Supplementary Table S1.

Sample preparation was carried out using 10 μL of whole blood, similar to the protocols applied for plasma [23, 24, 32]. After thawing, the blood samples were carefully mixed and centrifuged at 4,000 g for 10 min. Before trypsinolysis, the samples were denaturated and reduced by incubation with 6 M urea, 13 mM dithiothreitol and 200 mM Tris × HCl (pH 8.0, +37 °C, 30 min). Next, the proteins were alkylated by a 30-min incubation in the dark with 40 mM iodoacetamide. For trypsinolysis, the samples were diluted with 100 mM Tris × HCl (pH 8.0) until <1 M urea; L-(tosylamido-2-phenyl) ethyl chloromethyl ketone (TPCK)-treated trypsin (Worthington) was added at a 20:1 (protein:enzyme, w/w) ratio; and the samples were incubated for 18 h at 37 °C. The reaction was quenched by acidifying the reaction mixture with formic acid (FA) to a final concentration of 1.0 % (pH≤2), and an estimated peptide concentration of ∼1 mg/mL [31]. Then each sample (40 μL) was spiked with 10 μL of the SIS peptide mixture, prepared by solubilization in 30 % ACN/0.1 % FA and dilution to 10× LLOQ per μL with 0.1 % FA. All samples were then concentrated by solid-phase extraction (SPE) using an Oasis HLB μElution plate: (1) the plate was conditioned with MeOH (600 μL), equilibrated with 0.1 % FA (600 μL), and loaded with samples; (2) the wells were washed with H₂O (600 μL,×3); (3) bound peptides were eluted with 70 % ACN/0.1 % FA (55 μL) [24]. For each reference standard and quality control sample, 40 μL of a BSA surrogate matrix digest (143 μg/mL) was spiked with 10 μL of the SIS peptide mixture and 10 μL of a level-specific “light” peptide mixture. The standard curve and quality control samples were subjected to the same SPE procedure as the samples. All SPE eluates were evaporated using a speed vacuum concentrator and stored at −80 °C. Prior to LC-MS/MS analysis, the samples were reconstituted in 34 μL of 0.1 % FA.

Plasma samples

Plasma samples from 32 healthy controls were prepared in parallel with their whole blood samples to study the correlation between target protein concentrations. Plasma was obtained by centrifuging the K²⁺-EDTA whole blood samples at 4,000×g for 10 min at room temperature within 1 h after collection. The aliquots were stored at −80 °C. Sample preparation for MRM-MS was performed as described above for whole blood samples.

LC-MS/MS analysis and MS data processing

All samples were analyzed in duplicate by HPLC-MS using an ExionLC™ UHPLC system (Thermo Fisher Scientific, Waltham, MA, USA) coupled online to a SCIEX QTRAP 6500+ triple quadrupole mass spectrometer (SCIEX, Toronto, ON, Canada). LC and MRM parameters were adapted and optimized basing on the previous studies done with the BAK125/270 kits [31, 32].

The loaded sample volume was 10 μL per injection. HPLC separation was carried out using an Acquity UPLC Peptide BEH column (C18, 300 Å, 1.7 μm, 2.1 mm × 150 mm, 1/pkg) (Waters, Milford, MA, USA) with gradient elution. Mobile phase A was 0.1 % FA in water; mobile phase B – 0.1 % FA in acetonitrile. LC separation was performed at a flow rate of 0.4 mL/min using a 53-min gradient from 2 to 45 % of mobile phase B. Mass spectrometric measurements were carried out using the MRM acquisition method. The electrospray ionization (ESI) source settings were as follows: ion spray voltage 4000 V, temperature 450 °C and ion source gas 40 L/min. The corresponding transition list for MRM experiments with retention time values and Q1/Q3 masses for each peptide is available in Table S1.

For quantitative analysis of the LC-MS/MS raw data, Skyline Quantitative Analysis software (version 20.2.0.343, University of Washington) was used. To calculate the peptide concentrations in the measured samples (fmol per 1 µL of plasma), calibration curves were generated using the 1/(×2)-weighted linear regression method. Calibration points differing from the standard concentration by >15 % were excluded. Proteins with 5 or less calibration points per batch, lying outside the calibration range, showing signs of interference, low signal-to noise identified in less than 70 % of the samples were discarded from further analysis (Tables S2, S3). All experimental results were uploaded to the PeptideAtlas SRM Experiment Library (PASSEL) and are available via link: http://www.peptideatlas.org/PASS/PASS05842 (accessed on 5 July 2024).

Statistical analysis

Statistical analysis and data visualization were performed using Python (3.7.3) scripts with SciPy [33], Seaborn [34], matplotlib [35], and Pandas [36] packages.

After filtering the data 94 proteins features were used for analysis. The data was Log₂(x+1) transformed. “NaN” – values were replaced with random non-zero values using a Gaussian distribution with a shift down=0.4 and width=0.2 of the mean value for each group of patients (control, mild, moderate, and severe).

The Mann–Whitney test was used to evaluate statistical differences between the groups. p-values were adjusted using the Benjamini–Hochberg method. The Cohen`s size (mean difference divided by the variance for different groups of samples) was also considered.

The ordinary least squares (OLS) method was used to construct the linear regression model in order to compare the data from paired whole blood and plasma samples from control patients.

Machine learning

The Scikit-Learn package was used to obtain all machine learning models [37]. To differentiate between pairwise groups, 4 classifying algorithms were considered: k-Nearest Neighbors (kNN), Logistic Regression (LR), Random Forest (RF), and the Support Vector Classifier (SVC). For the application of all algorithms, except RF, normalized data were used. The best models were determined using k-Fold cross-validation (k=4). Features were selected according to preliminary statistical analysis. A logistic regression algorithm with default hyperparameters was also used to select a panel for the classifier (proteins were ranked by “feature importance”). Specific hyper-parameters for each algorithm were selected using a grid search based on the highest ROC-AUC scores for different combinations of compared groups and protein marker panels. The proposed hyper-parameters for the classifiers discussed in the “results” section are presented in the Supplementary Materials (Table S5).

Significantly dysregulated plasma proteins in whole blood

In total, 203 target plasma proteins were analyzed in whole blood samples using LC-MRM MS with corresponding SIS peptides, resulting in 173 measured proteins, 147 of which showed non-empty quantitative results detected proteins, 94 of which passed the quality filters for statistical analysis (Supplementary Tables S2, S3). The later included 80 previously described COVID-19 markers, with 61 reproducible markers according to ≥2 studies, of which 36 were consistently dysregulated in ≥3 different studies and 11 were shown to be significant in ≥5 studies (SERPINA3, SERPING1, HRG, ORM1, APOA1, ITIH2, AFM, AHSG, LBP, SERPINC1, and ACTA2).

Pairwise comparison of different groups of patients with each other and with healthy controls revealed a very large number of proteins with statistically significant differences (Supplementary Table S4). A total of 41 proteins passed the 5 % FDR cutoff after the Benjamini–Hochberg multiple testing correction in at least one pairwise comparison: 21 were up-regulated, 20 were down-regulated (Table 2).

Seven of the revealed significantly different proteins (down-regulated APOA1, APOA4, APOA2, AHSG, TF, and up-regulated APOE and NRP2) deserve special attention as they were found to be different between all groups presented in Table 2. Thus, they not only distinguish patients from healthy controls but also distinguish patients by severity and risk of mortality (Figure 1). Nevertheless, some other proteins (including SERPINA3, SERPING1, HRG, CPN2, LUM, APOH, FN1 etc.) demonstrate an even greater difference between groups of patients and healthy controls (Table 2, Figure 1).

Figure 1:

Significantly different proteins that distinguish various groups. *Indicates a significant difference at p<0.05 after the Benjamini–Hochberg correction. Protein levels are shown in determined concentrations (fmol/μL).

Correlation of protein concentrations between whole blood and plasma

To better understand the relevance of comparing results obtained on whole blood with published results for plasma/serum, analysis of the correlations is particularly important. For this analysis of paired whole blood and plasma samples from 32 healthy controls was carried out. The constructed linear regression for the average protein concentrations confirms a linear relationship between measurements in whole blood and plasma (Figure 2). Overall, the data are well described by the proposed linear model, with 93 % of the proteins falling into the 95 % prediction interval with only 4 outliers. Even taking into account the data for all proteins, the coefficient of determination R² turned out to be 0.64 (>0.5, p=2.95 × 10⁻²⁵), what suggests a good fit of the model to the data. After removing the outliers the R² coefficient increased to 0.90, indicating a much better agreement between the data and the constructed linear model.

Figure 2:

Comparison of target protein concentrations between paired frozen whole blood and plasma samples from 32 healthy controls. (A) Linear regression plot of average protein concentrations. Red dots indicate proteins that fall outside the predicted interval (95 %) for the constructed linear regression model. (B) Distribution profiles of average protein concentrations in plasma and whole blood samples.

To assess the coherence of plasma-blood data for each individual protein Pearson correlation coefficients for paired plasma and blood samples were estimated (Supplementary Table S5). According to these calculations, 39 proteins showed a strong positive correlation (R≥0.8, p-value <0.001), 17 proteins were moderately correlated (R≥0.7, p<0.001), and 25 were weakly correlated (R>0.35, p<0.05). The non-correlating proteins (R<0.35, p-value >0.05) included only 3 of the 44 significantly different proteins mentioned in Table 2. Two of the outliers also turned out to be non-correlating: ACTA2, and PRDX2.

Building classifiers distinguishing between patients and healthy controls

Differences in the proteomic profiles among severity groups can be used to identify unique features for COVID-19 stage classification with a machine-learning approach. Four algorithms were considered to build a classifier: logistic regression (LR), random forest (RF), k-nearest neighbors (kNN) and the support vector classifier (SVC) (Supplementary Table S7).

Sorting proteins by their feature importance values using a logistic regression method clarified the most important features for distinguishing between healthy individuals and COVID-19 patients, as well as patients with mild and severe disease (Supplementary Table S6). A combination of just 10 proteins including PZP, THBS1, APOC4, APOE, APOM, AHSG, APOA4 (7 top features in ‘mild vs. severe’ comparison), and LUM, HRG, and APOC3 (3 top features in ‘control vs. severe’ comparison) made it possible to achieve a very good combination of ROC-AUC and accuracy metrics. The LR-10 classifier particularly showed the highest values for ‘mild vs. severe’ comparison (0.98/0.93, ROC-AUC/Accuracy) among other considered protein combinations (Table 3). The addition of A2M and HSPD1 (2 top features in ‘survived vs. lethal’ comparison) and removing of statistically insignificant proteins (PZP, APOC4 and A2M), although slightly reduced some metrics, generally showed similar results to the 10-protein panels; and the LR-9 classifier turned out to be even more effective than LR-10 in the ‘disease vs. control’ comparison (Table 3).

Importantly, the considered proteins with high feature importance values mostly included proteins that were significantly different in at least two different pairwise comparisons (Table 2). In an attempt to improve the model’s performance for classifying both mild vs. severe patients, as well as controls vs. all patients, several other variants of the panel were tested with different combinations of statistically significant proteins. In particular, the 14-protein panel with addition of ALB, APOA1, APOA2, NRP2 and LCP1 led to obtaining effective classifiers with the best result for LR-14 classifier in the ‘disease vs. control’ comparison (Table 3). Nevertheless, the LR-10 classifier on our cohort turned out to be the most versatile among all of the considered variants, as worked best both for the detection of COVID-19-positivity in general, and identification of severe cases with the highest efficiency (Figure 3A). The usefulness of its application further on other cohorts and capability for outcome prediction is yet to be studied.

Figure 3:

Characteristics of the best performing LR-10 classifier distinguishing COVID-19 patients. (A) ROC-AUC metrics for different pairwise comparisons using 4-fold cross-validation. The insert is a ‘Confusion matrix’ for the classifier prediction of mild or severe COVID-19 course; (B) assessment of the severity of the disease for different COVID-19 patients, obtained using the developed LR-10 classifier-10.

For further validation of the developed LR-10 classifier, 70 % of the mild/severe patient’s dataset were used for training, while the remaining 30 % and all samples from the moderate group of patients (which were not used at all during the classifier creation process) were used as a test set. The resulting probabilities assigned by the classifier to specific samples (Figure 3B) show a very good agreement with the diagnosed severity of COVID-19 and confirm the effectiveness of the developed classifier.