¹H-NMR spectroscopy metabonomics of reactive, ovarian carcinoma and hepatocellular carcinoma ascites

Human peritoneal effusion harvesting

Seventy two AEs were prospectively collected between 2018 and 2019 at the Surgical Pathology and Cytopathology Unit at Padua University, 31 patients had HGSOC, 16 HCC and 25 AEs derived from patients with reactive peritonitis (clinical and pathological data are reported in Table 1).

All the AE samples were centrifuged within 1 h after withdrawal at 2000×g for 10 min at room temperature. The precipitate was used to prepare two slides for cytology, while 4 mL of each supernatant were divided in two parts (to produce two identical samples of the same AE, for repeated analysis), transferred in 2 mL sterile tubes, and centrifuged at 2000×g for 20 min at 4 °C. The AE supernatants were then stored at −80 °C until use.

Two cytopathologists (LN and FC) re-examined all cases independently and defined the three cohorts of HGSOC, HCC and bAE. Diagnosis was confirmed considering effusion morphology, ancillary test as immunohistochemistry on cell block where feasible, and considering clinical history. In the HGSOC and HCC cohorts we included only cases with unequivocal history of malignancy provided by histological confirmation; patients were enrolled regardless the percentage of tumor cells, since we focused on liquid composition of confirmed neoplastic effusions (Figure 1).

Figure 1:

Representative images of a positive effusion from a patient included into the HGSOC cohort (A) and a negative effusion from a patient with HCC (B). Corresponding histology is shown in (C) for HGSOC and in (D) for HCC.

Chemicals and reagents

All chemicals were purchased from Sigma-Aldrich: Na₂HPO₄ anhydrous and NaH₂PO₄ anhydrous (≥99%), sodium azide (NaN₃), D₂O 99.98%, 2,2,3,3-d4-3-(trimethylsilyl) propionic-acid sodium salt (TSP-d4, 98%D atom) and methanol-d4 99.8%. TSP-d4 was used as internal chemical shift reference for NMR spectra, while methanol-d4 was used for NMR temperature calibration.

Buffer for NMR analysis

Phosphate buffer NaPi was prepared in H₂O:D₂O (1:1) to a final concentration of 1.5 M, pH 7.4. TSP-d4 and NaN₃ (5 and 15 mM final concentration, respectively) were added to the buffer solution as internal chemical shift reference (TSP-d4) and for antibacterial purposes (NaN₃). All the AE samples were diluted in this buffer after their de-freezing, immediately before the NMR analysis.

AE sample treatment for NMR investigation

We followed a rigorous protocol for NMR sample preparation. To minimize the variability in pH that potentially characterize the ascitic effusions from different individuals, a AE sample was defrosted from −80 °C immediately before the NMR analysis, quickly centrifuged at 12,000×g for 5 min at 4 °C and rapidly diluted in phosphate buffer (80:20 V/V), final pH 7.30 ± 0.02. Finally, 600 μL of the buffered sample are transferred into 5 mm NMR tube and immediately inserted in the spectrometer for the metabonomic assay.

NMR equipment and ¹H-NMR data acquisition

¹H-NMR spectra were acquired using a Bruker Advance III spectrometer (Bruker Biospin, Rheinstetten, Germany) operating at 300.13 MHz proton Larmor frequency. A 5 mm outer diameter ¹H – ¹³C probe was used, equipped with Z gradient facility. For each sample three different one-dimensional (1D) NMR spectra were acquired, namely 1D Noesy sequence (“noesygppr1d” in Bruker library), 1D Carr–Purcell–Meiboom–Gill (CPMG) spin-echo sequence, (“cpmgpr” in Bruker library), and 1D Diffusion-edited sequence (“ledbpgppr2s1d” in Bruker library). The 1D Noesy sequence was used to collect each ¹H-NMR spectrum as it is, without applying any digital filter that generally is used to attenuate or enhance the signals from low or high molecular weight species, thus allowing a more reliable determination of the metabolites’ concentrations. The water peak suppression was obtained by pre-saturation using a standard zgpr pulse sequence. The 1D CPMG sequence was used to attenuate the broad signal from macromolecules (e. g. proteins) that may be present in the sample, thus permitting a better evaluation of low molecular weight species (metabolites). This sequence was preferred to de-proteinization procedures, in order to minimally perturbate the biological fluid [20]. As for the NOESY sequence, for CPMG sequence the water peak suppression was also obtained by pre-saturation using a standard zgpr pulse sequence. We used the 1D Diffusion-edited sequence to better visualize the signal from higher molecular weight species (e. g. lipids, lipoproteins). The spectra were recorded with 64 or 2048 scans, 64 K data points, a spectral width of 30 ppm (8971 Hz), 4 dummy scan and a relaxation delay of 4 s. All 1D spectra were processed with line broadening of 0.3 Hz before the Fourier Transform. Spectra acquisition and processing were both performed by using TopSpin 3.6.1 software (Bruker Biospin). The chemical shifts were internally calibrated to the TSP peak at 0.0 ppm. All NMR experiments were acquired at the constant temperature of 310 K, with temperature stabilization of approximately 0.1 K. For temperature equilibration, the samples were kept for 5 min inside the NMR probe head before starting the spectrum acquisition. The assignments of various resonances were made using metabolite NMR databases (BBIO-REFCODE 2 database, Bruker Biospin), Human Metabolome Database HMBD (www.hmdba.ca) and literature data based on chemical shifts and split patterns of proton signals [23].

Multivariate statistical analysis

A total of 72 spectra, 31 from HGSOC, 16 from HCC and 25 from control patients, have been processed with multivariate analysis. The PCA followed by the PLS-DA were performed by AMIX 4.0.1 software package (Bruker Biospin). Each 1D spectrum area was bucketed into frequency windows of 0.05 ppm (rectangular bucketing amplitude) in the range between 0.50 and 10.00 ppm. The region corresponding to water (4.50–5.00 ppm) was excluded during bucketing to avoid artifacts due to the pre-saturation of water. The spectral content of each bucket was converted to the numerical value corresponding to the area subtended by the portion of the spectrum enclosed in the bucket. Thus, each bucket represents the original variable to which the statistical analysis is applied. All original variables were preprocessed with mean-centered and Pareto-scaling procedures before being evaluated by the multivariate statistical analysis. As first step, the PCA analysis was performed to obtain the principal components of the variance (PC). Principal component (PCi) is a linear combination of all the original variables, and each successive PCi (progressively at a lower extent) describes the maximum amount of variance possible in the data set. PCA was used to find if the spectra could be clustered in separate groups containing similar spectra. Based on these PCA results, PLS-DA was performed to assess the strength of such a separation, using the condition control/cancer (either HGSOC or HCC, depending on the case) as classifier for the clustering.

Metabolites spectra

Figure 2 shows typical ¹H-NMR 1D CPMG spectra of AE from a patient with HGSOC (top), a patient with HCC (middle) and from a patient from the Control group (bottom). The metabolites identifiable in the above mentioned spectra are listed in Table 2 and were found to give signals in four spectral regions: (1) the region from 0.8 to 1.80 ppm: lipids, triglycerides, branched-chain amino acids (leucine, isoleucine and valine), aliphatic amino acids (alanine, lysine), organic acids (lactate, butyrate) and β-hydroxybutyrate (BHB); (2) the region from 2.0 to 2.50 ppm: other aliphatic amino acids (glutamine, glutamate), other organic acids (pyruvate, citrate), other ketone bodies (acetoacetate) and N-acetyl groups; (3) the region from 3.0 to 4.0 ppm: sugar (glucose) and alcohol (glycerol), creatine, creatinine, taurine and choline; (4) the region from 6.90 to 7.50 ppm: aromatic amino acids (tyrosine, histidine, phenylalanine).

Figure 2:

Typical ¹H-NMR spectra of AE from a patient with HGSOC (top), from a patient with HCC (middle) and from a patient of the control group (bottom).

Multivariate statistics

The results from the PCA analysis of the data embedded in the ¹H-NMR spectra were visualized in graphical mode, by plotting the PC scores and PC loadings for any PCi–PCj pair: each point in the scores plot represents an individual patient (i. e. an individual NMR spectrum), and each point in the corresponding loadings plot represents one spectral region. The points that in the loadings plot move away from the center (coordinate 0,0) represent the spectral regions that more strongly influence patterns in the scores plot and, consequently, the molecules that mostly influence patterns in the NMR spectra, or, in other words, the molecules that mostly influence the clustering of NMR spectra/patients in separate groups. The reliability of the qualitative evaluation of the clustering highlighted by the graphical presentation of the results is given by the percentage of the variance explained by the two PCi to which the plots refer: the higher the percentage of total variance explained, the higher the reliability of the clustering. Performing a PLS-DA analysis on the same data set as PCA can give further support to the reliability of the separation or weaken it: if the clustering shown by PCA is actually reliable, the PLS-DA analysis must confirm it, with a separation between groups that is generally at least equal if not better. On the contrary, a poor clustering getting out from PLS-DA makes debatable the validity of the separation obtained with the PCA.

HGSOC vs. control group

In Figure 3 (top) are shown the scores and the loadings plots of the PCA performed on HGSOC and Control groups. The plotted scores indicate a clear clustering of the spectra, with PC1 and PC2 explaining 65.38% of the total variance (45.76% and 19.62%, respectively). The spectral regions that in the correlated loadings plot appear to be responsible for such a separation are those which contains the signals from lipids, triglycerides, lactate, leucine, isoleucine, valine, alanine, lysine, phenylalanine, β-hydroxybutyrate (BHB), acetate, acetoacetate, N-acetyl groups, pyruvate, taurine, choline, glutamine and glycerol. Figure 3 (bottom) shows the scores and the loadings plots resulting from the PLS-DA analysis performed on the same dataset as PCA, using as classifier for the Discriminant Analysis (DA) the condition control vs. cancer (with 0 assigned to control condition and 1 assigned to ovarian carcinoma condition). The graphical separation getting out the PLS-DA was as clear as that obtained by PCA, thus confirming the reliability of the latter. The low value of 0.174 obtained for the root mean square error of calibration (RMSEC) of the PLS-DA confirmed on a numerical basis the good clustering graphically observed (the lower RMSEC, the more reliable the clustering).

Figure 3:

Top: scores plot (left) and loadings plot (right) obtained from PCA performed on m-AEs from HGSOC patients and from control group. Bottom: scores plot (left) and loadings plot (right) obtained from PLS-DA performed on the same dataset as PCA.

HCC vs. control group

The same statistical approach was used to evaluate the metabolic information embedded in the ¹H-NMR spectra of ascitic effusions collected from HCC patients. The scores and the loadings plots of the PCA performed on HCC and Control groups are shown in Figure 4 (top). The scores plot shows a weaker, but still clear separation of the spectra in two groups, with PC3 and PC4 explaining 20.31% of the total variance (14.41% and 5.90%, respectively). The spectral regions responsible for such separation in the correlated loadings contain signals from triglycerides, leucine, isoleucine, valine, lactate, lysine, phenylalanine, β-hydroxybutyrate (BHB), acetate, acetoacetate, pyruvate, taurine, choline and glycerol.

Figure 4:

Top: scores plot (left) and loadings plot (right) obtained from PCA performed on m- AEs from HCC patients and b-AEs from control group. Bottom: scores plot (left) and loadings plot (right) obtained from PLS-DA performed on the same dataset as PCA.

The reliability of the clustering found by PCA on HCC vs. Control group was indeed confirmed by the PLS-DA analysis performed on the same data set: the scores and loadings plots calculated by PLS-DA (Figure 4, bottom) showed a good separation of the NMR spectra in two groups, with a RMSEC value (0.157) as good as that found for HGSOC vs. Control group.

HGSOC vs. HCC

To assess the ability of ¹H-NMR metabonomics to discriminate a mAE (HGSOC, HCC) from controls (bAE), and to recognize specific malignancies (e. g. HGSOC vs HCC), we submitted the entire ensemble of recorded spectra to PCA and PLS-DA analysis. The choice of considering all the ascites together and not simply HGSOC vs. HCC, was driven by the fact that the separation from Controls is stronger for HGSOC than for HCC, and there is the risk of finding an artificial discrimination between the two specific cancers due essentially to the weak separation between HCC and controls, and therefore due to the statistical similarity of Controls and HCC with respect to HGSOC. Figure 5 shows the scores plot obtained by PCA (left) and PLS-DA (right) performed on the entire ensemble of spectra. The clear clustering shown by the PCA result (with PC1 and PC2 explaining 65.60% of the total variance) highlights the presence of three groups of spectra. The reliability of the clustering is confirmed by the good graphical separation shown by the PLS-DA. Moreover, having used as classifier for the Discriminant Analysis the triple condition Control – Cancer A – Cancer B (with 0 assigned to control condition, 1 assigned to HGSOC condition and 2 assigned to HCC condition), the PLS-DA confirmed the presence of three different classes of AE.

Figure 5:

Scores plot obtained from PCA (left) and PLS-DA (right) performed on the entire ensemble of AEs.