Magnetic Resonance Spectroscopy-based Metabolomic Biomarkers for Typing, Staging, and Survival Estimation of Early-Stage Human Lung Cancer

Abstract

Low-dose CT has shown promise in detecting early stage lung cancer. However, concerns about the adverse health effects of radiation and high cost prevent its use as a population-wide screening tool. Effective and feasible screening methods to triage suspicious patients to CT are needed. We investigated human lung cancer metabolomics from 93 paired tissue-serum samples with magnetic resonance spectroscopy and identified tissue and serum metabolomic markers that can differentiate cancer types and stages. Most interestingly, we identified serum metabolomic profiles that can predict patient overall survival for all cases (p = 0.0076), and more importantly for Stage I cases alone (n = 58, p = 0.0100), a prediction which is significant for treatment strategies but currently cannot be achieved by any clinical method. Prolonged survival is associated with relative overexpression of glutamine, valine, and glycine, and relative suppression of glutamate and lipids in serum.

Figure 1

Serum MRS identification of LuCa from controls and differentiation among LuCa types and stages. LuCa groups vs. healthy controls (central column) and various LuCa types and stage groups (right column) were compared with two-tailed Student’s t-test or Mann-Whitney-Wilcoxon test. “Increase” indicates that the values are higher in the lower group compared to the upper group; “Decrease”, indicates the opposite. For instance, the first red square in the table shows that the group of all LuCa samples presented significantly higher serum lactate (Lac, 4.11–4.10 ppm) than the group of healthy controls. See Supplementary Table S1 for sample numbers in each group, and see text for significance notation details. Abbreviations: Ala, alanine; Asp, aspartate; Cho, choline; Glc, glucose; Gln; glutamine; Glu, glutamate; GPC, glycerophosphocholine; GSH, glutathione; Lac, lactate; Lip: lipids; M-Ino, myo-inositol; PC, principal component; St, stage; Tau, taurine; Thr, threonine; Tyr, tyrosine; Val, valine. The notations of statistical significance levels are: “*”p < 0.05; “**”p < 0.005; and “***”, Bonferroni-corrected thresholds of statistical significance of p < 0.0016 or p < 0.0063 for 32 individual regions or 8 PCs, respectively. The star symbols denote statistical significances after FDR calibration.

Figure 2

Comparisons between analyses with serum MRS metabolites and analyses with metabolomics. (a) Panels of three metabolites that can significantly differentiate LuCa from control groups both for all the tested Stage I, SCC (n = 27) and Adeno (n = 31), and control (n = 29) cases, and for the Training and Testing cohorts, separately. (b) The 3D ellipsoids generated from the three panels in (a) that cover the volume of 3D Mahalanobis distance ≤1 (one standard deviation from the centroid along each axis) from the class means for LuCa and control groups, according to the covariance matrix of all class-mean-corrected samples. (c) Two-class (LuCa vs. control) and three-class (Adeno vs. SCC vs. control) linear discriminations calculated from 19 spectral regions in Fig. 1. Misclassifications are indicated by X’s. (d,e) 2D projections for three-class and two-class LD results from (c). The ellipses represent the 2D Mahalanobis distances 1 and 2 standard deviations from the class means, respectively. Solid lines on projection planes represent active decision boundaries. Classification accuracies shown are after leave-one-out cross-validation. (f) Differentiations between LuCa and control cases calculated with LD canonical correlation analysis for the Testing cohort by using analytic parameters obtained from the Training cohort.

Figure 3

Tissue MRS differentiations among LuCa types and stages. After pathology calibrations for the tissue MRS spectral data, various LuCa types and stage groups were compared with two-sided Student’s t-test (for normal spectral region distributions) or Mann-Whitney-Wilcoxon test (for non-normal spectral region distributions). “Increase” indicates that the values are higher in the lower group compared to the upper group; “Decrease”, indicates the opposite. For instance, the first blue square in the table shows that the group of SCC samples presented significantly lower spectral intensity in the glycerophosphocholine and tyrosine region. See text for significance notation details, and see Supplementary Table S1 for sample numbers in each group. Abbreviations: Ala, alanine; Glu, glutamate; GPC, glycerophosphocholine; Lip: lipids; PC, principal component; St, stage; Tyr, tyrosine; Val, valine.

Figure 4

The effects of pathology calibrations on the measured tissue MRS data. Before a least squares regression of an over-determined linear model (LSR-ODLM) pathology calibration for tissue MRS spectral data, the spectral region containing alanine could not distinguish the two groups in a Mann-Whitney-Wilcoxon comparison between Stage 1 SCC (n = 27) 15.234 ± 2.287 and Adeno (n = 31) 6.273 ± 2.134 (relative intensity) cases. After calibrating the MRS spectral data for different percentages of pathology as described by Eqs 2–5, the alanine spectral region, among others, can significantly differentiate between the groups, 3.272 ± 0.241 (SCC) and 2.166 ± 0.225 (relative intensity) (p = 0.0004).

Figure 5

Differentiation between LuCa types and stages with tissue-derived serum metabolomic profiles. (a) Training cohort: ANOVA comparison between canonical scores (arbitrary units, a.u.) of Stage I LuCa cases SCC (n = 27; Mean = 0.033 ± 0.208) vs. Adeno: (n = 31; Mean = −0.503 ± 0.194). (b) Testing cohorts: (1) ANOVA comparison between canonical scores (a.u.) of tissues from LuCa cases of Stages II, III, and IV SCC (n = 15; Mean = 0.495 ± 0.211) vs. Adeno (n = 20; Mean = −0.038 ± 0.182) and (2) ANOVA comparison between canonical scores (a.u.) of serum PCs calculated with tissue PCA loadings for Stage I cases SCC (n = 27; Mean = −0.010 ± 0.213) vs. Adeno (n = 31; Mean = 0.682 ± 0.199) vs. controls (n = 29; Mean = −0.436 ± 0.206). Given our hypothesis from the tissue training cohort result that mean canonical score values were higher for SCC than for Adeno, a one-tailed test was used for the tissue testing cohort only. (c,d) Linear discriminant canonical correlation analyses with the 19 spectral regions (cf. Fig. 1), for tissue and serum MRS data of the Training cohort, respectively. The capabilities of the resulting canonical scores in differentiating SCC from Adeno groups were presented with the results obtained from the Testing cohort. (e) Linear discriminant canonical correlation analyses including both tissue and serum results (from c,d), increased statistical significance for SCC and Adeno differentiations.

Figure 6

Evaluations of overall survival with serum MRS metabolomics. (a) Results of leave-one-group-out validation show the ability of canonical scores to differentiate short living (red, n = 29) from prolonged living (green, n = 25) groups for all 93 cases in 8 cohorts (7 groups = training; left-out group = testing). (b,c) All 93 cases were combined. Nine spectral regions were identified (d) which differentiated survival groups. CCA scores for these regions significantly differentiated 10-year survival for both (b) the entire LuCa population and (c) Stage I cases alone (red, n = 14; green, n = 18). (d) As mentioned, all 93 cases were combined, and the serum spectral regions were evaluated by canonical analysis. Nine spectral regions could differentiate between controls (C), prolonged survival (P), and short survival (S) groups, and one region could identify overall survival. The heatmap presents log spectral intensities to illustrate that the prolonged survival group more closely resembles the short survival group or controls, but that the short survival group and controls are noticeably different.