Possibility of multivariate function composed of plasma amino acid profiles as a novel screening index for non-small cell lung cancer: a case control study

Blood samples (5 ml) were collected from forearm veins, after overnight fasting, in tubes containing ethylenediaminetetraacetic acid (EDTA; Termo, Tokyo, Japan), and were immediately placed on ice. Plasma was prepared by centrifugation at 3,000 rpm and 4°C for 15 min, and then stored at -80°C until analysis. After plasma collection, all samples were stored and processed at the Life Science Institute of Ajinomoto Co., Inc. (Kawasaki, Japan). To reduce any bias introduced prior to analysis, samples were analyzed in random order. The plasma samples were deproteinized using acetonitrile at a final concentration of 80% before measurement. The amino-acid concentrations in the plasma were measured by HPLC-ESI-MS, followed by precolumn derivatization [17‐19]. The analytical methods were described in detail previously [17]. The concentrations of amino acids in the plasma were expressed as μM.

The mean amino-acid concentrations ± standard deviations (SDs) were calculated. Differences between the plasma amino-acid concentrations in NSCLC patients and controls were assessed using the Mann-Whitney U-test and receiver-operator characteristic (ROC) curve. The area under the curve (AUC) for each ROC curve (the ROC_AUC) was calculated for each amino acid.

Principal component analysis (PCA) was also used to assess differences in the plasma amino-acid profile between the controls and the NSCLC patients, with linear combinations of all of the amino acids included as explanatory variables. In PCA analysis the plasma amino-acid concentrations were transformed using the following equation:

where z _i,j was transformed concentration of of the i-th sample of the j-th amino acid, x _i,j was concentration of the i-th sample of the j-th amino acid, n was sample size, and X ¯ j was the average concentration of j-th amino acid.

First, an unconditional multiple logistic regression analysis with variable selection was used to construct a criterion for distinguishing NSCLC patients from controls using the study data set with the raw plasma concentrations of 21 amino acids as explanatory variables. The candidate variables of most appropriate logistic regression model, which had the minimum Akaike's information criterion (AIC) value, were selected from among all of the possible combinations in which the number of variables was below seven. A leave-one-out cross-validation (LOOCV) was performed to correct potential over-optimization for all models in parallel. Briefly, one sample was omitted from the study data set, and the logistic regression model was calculated for the remaining samples, to estimate coefficients for each amino acid. The logistic regression function values for the left-out sample were calculated based on the model. This process was repeated until every sample in the study data set had been left out once, and the function values generated were then used for AIC calculation. Finally, a case-control study was utilized for our study, and so a conditional logistic regression analysis, conditioned on the matching factors (i.e., gender, age, and smoking status), was performed in order to evaluate the association between the combination of amino acids obtained above and NSCLC. The discriminant score, which was defined as a logit of the conditional logistic regression function value, was constructed as a criterion. The degree of discriminancy of this score between NSCLC patients and controls was evaluated through the ROC curve. A distinct test data set, which had not been used in the model generation, was also used to confirm the stability of the obtained model, and to calculate the ROC_AUC values for the discriminant scores.

To assess the effects of cancer stage and histological type, both the study data set and the test data set was stratified according to the analysis parameters. To assess the effects of cancer stage and histological type on the discriminant scores of NSCLC patients, a subgroup analysis was performed using the ROC curve, in each data set. A two-sided P value of less than 0.05 was considered to indicate statistical significance.

All statistical analyses were performed using MATLAB (The Mathworks, Natick, MA), LogXact (Cytel, Cambridge, MA), and GraphPad Prism (GraphPad Software, La Jolla, CA).

The study data set comprised 141 patients with NSCLC, and 423 age-matched, gender-matched, and smoking status-matched control subjects, whereas there were 162 patients and 3,917 controls in the test data set; a further 15 SCLC patients were also included (Table 1). Among the patients, 28% and 36% were non-smokers in the study and test data sets, respectively, whereas almost 50% of the control subjects were non-smokers (Table 1). There were no significant differences in body mass index (BMI) between the patients and the control subjects (Table 1). In both the study and test data sets ~50% of the patients were categorized as having stage I disease, ~5% as stage II, ~25% as stage III and ~20% as stage IV (Table 1). The Eastern Cooperative Oncology Group performance status (ECOG) score of most patients was 0 or 1; hence, the majority of the patients were asymptomatic or symptomatic but completely ambulatory (Table 1). The histological type was adenocarcinoma in almost 75% of the patients and squamous cell carcinoma in almost 25%, the other types present included large-cell carcinoma, adenosquamous carcinoma, pleomorphic carcinoma and mucoepidermoid carcinoma (Table 1).

In the study data set, the plasma concentration of His was significantly lower, and those of Ser, Pro, Gly, Ala, Met, Ile, Leu, Tyr, Phe, Orn, and Lys were significantly higher, in NSCLC patients than in controls (Table 2).

Amino acids in the human body undergo interdependent regulation; comparing single amino-acid concentrations between controls and patients might thus be insufficient to elucidate any changes in plasma amino-acid profiles associated with cancer development. Changes in the balance of the plasma amino acids in the study data set were therefore investigated using principal component analysis (PCA) in the current study. Five PCs with eigenvalues >1 were identified (Table 3). To evaluate their performance, the Mann-Whitney U-test was used to compare each PC score between the controls and NSCLC patients. Three of the PCs showed significant p values (< 0.001): PC1, PC3, and PC5 (Table 3). The contributing amino acids for the PCs that had a variance of >0.05 were then extracted; the results identified Ala, Val, Met, Ile, Leu, Tyr, Phe, Trp, and Lys as contributing factors for PC1, Cit, His, Trp, Orn, and Arg as contributing factors for PC3, and Ser, Gly, Cit, His, and Arg as contributing factors for PC5 (Table 3). As a result, fifteen amino acids (Ser, Gly, Ala, Cit, Val, Met, Ile, Leu, Tyr, Phe, His, Trp, Orn, Lys, and Arg) were identified as whose profile in plasma were associated with NSCLC (Table 3).

The results described so far suggested that it should be possible to improve the discrimination between cancer patients and normal controls by deriving multivariate functions, using the raw plasma amino-acid concentrations as explanatory variables, which would summarize the changes in metabolic status. Multiple logistic regression analyses by unconditional and conditional likelihood methods were therefore performed with variable selection and LOOCV cross-validation, using the study data set (as described in the Methods). The resulting conditional logistic regression model included six amino acids: Ala (p = 0.007), Val (p < 0.001), Ile (p < 0.001), His (p = 0.035), Trp (p = 0.027) and Orn (p < 0.001). The area under the curve (AUC) of the ROC for the discriminant score was 0.817 in the study data set (Figure 1).

Bild vergrößern

Furthermore, to verify the robustness of the resulting model, a ROC curve was generated using the split test data set, which had not been used to construct the model. A ROC_AUC of the ROC for the discriminant score was 0.812 in the test data set (Figure 1), again demonstrating that the obtained model performed well.

From the point of view of cancer screening, attention might be paid to whether or not the obtained model also provides sufficient discriminating power to extract effectively patients with early-stage cancer and for all histological types. Thus, to investigate the consistency of the results based on the discriminant scores among different subpopulations defined by cancer stage and histological type, a subgroup analysis was performed using both study data set and the test data set. The discriminant scores of the SCLC patients were also calculated to verify whether the obtained model could discriminate them from the controls.

Interestingly, it was suggested that the model could discriminate lung cancer patients regardless of cancer stage or histological type. Using the discriminant scores, the ROC_AUCs were 0.796 (study data set) and 0.817 (test data set) for stage I patients, 0.906 (study data set) and 0.801 (test data set) for stage II patients, 0.823 (study data set) and 0.843 (test data set) for stage III patients, and 0.836 (study data set) and 0.713 (test data set) for stages IV patients (Figure 2A, B). The model would thus be expected to be effective in detecting early, as well as advanced, cancers. We also demonstrated that the model could detect both adenocarcinomas and other histological types of cancer equally well: the ROC_AUCs were 0.795 (study data set) and 0.796 (test data set) for adenocarcinoma, and 0.860 (study data set) and 0.892 (test data set) for squamous cell carcinoma (Figure 2C, D). Furthermore, the distribution of the discriminant scores for SCLC patients was similar to that for NSCLC patients, with a ROC_AUC of 0.877 (Figure 2D).

Bild vergrößern