The significance and robustness of a plasma free amino acid (PFAA) profile-based multiplex function for detecting lung cancer

The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the ethics committees of the Chiba Cancer Center, the Osaka Medical Center for Cancer and Cardiovascular Diseases, the Gunma Prefectural Cancer Center, the Kanagawa Health Service Association, the Kameda Medical Center Makuhari, and the Mitsui Memorial Hospital. All subjects gave their written informed consent for inclusion before participating in the study. All data were analyzed anonymously throughout the study.

The participants in this study consisted of Japanese patients who had previously been histologically diagnosed with lung cancer at the Chiba Cancer Center (n=171) between 2007 and 2009. Control subjects (n=3849) without apparent cancers who were undergoing comprehensive medical examinations at the Kanagawa Health Service Association, the Kameda Medical Center Makuhari, or the Mitsui Memorial Hospital, Japan between 2008 and 2010 were recruited to participate in the study. Among the participants, 85 cancer patients (P1) and 421 gender- and age-matched controls (C1) were used as the study dataset for two preliminary studies (Table 1) [32, 34]. The remaining 86 cancer patients (P2) and 323 gender- and age-matched controls (C2) were used as a test dataset and were not used to derive the discriminating functions in previous studies (Table 1) [32, 34]. The remaining 3427 unmatched controls (C3) were also included and were not used to derive the discriminating functions in previous studies (Table 1) [32, 34].

Using these subjects, four data sets were evaluated in this study. Dataset 1 includes P1 and C1, Dataset 2 includes P2 and C2, Dataset 3 includes all of the subjects involved in this study (P1, P2, C1, C2, and C3), and Dataset 4 includes all of the patients involved in this study (P1 and P2) (Table 1).

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 stored at −80°C until analysis. 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. The analytical methods used have previously been described [29‐31]. Among the 20 genetically encoded amino acids, glutamate (Glu), aspartate (Asp), and cysteine (Cys) were excluded from the analysis because they are unstable in blood. Citrulline (Cit) and ornithine (Orn) were measured instead because they are relatively abundant in blood and are known to play important roles in metabolism. The following 19 amino acids were measured and analyzed: alanine (Ala), arginine (Arg), asparagine (Asn), Cit, glutamine (Gln), glycine (Gly), histidine (His), isoleucine (Ile), leucine (Leu), lysine (Lys), methionine (Met), Orn, phenylalanine (Phe), proline (Pro), serine (Ser), threonine (Thr), tryptophan (Trp), tyrosine (Tyr), and valine (Val). The concentrations of amino acids in the plasma were expressed as μM values. For analysis of the PFAA profile, two measurements were conducted for each of the 19 amino acids. The absolute concentration of each amino acid and the ratios of the amino acid concentrations expressed by the follow equation as previously described were used [32, 34]. The concentrations of the amino acids in the plasma were expressed in μM, and the ratios of the amino acid concentrations were expressed by the follow equation:

where X2 _i,j is the ratio of the amino acid concentration of the j-th amino acid of the i-th subject, and X _i,j is the plasma concentration (μM) of the j-th amino acid of the i-th subject.

Using serum samples from lung cancer patients, the levels of the following five tumor markers were measured: CEA (chemiluminescence immunoassay, normal range ≦ 5.0 ng/ml), CYFRA (electrochemiluminescence immunoassay, normal range ≦ 3.5 ng/ml), ProGRP (enzyme-linked immunoadsorbent assay, normal range ≦ 46 pg/ml), SCC (enzyme immunoassay, normal range ≦ 1.5 ng/ml), and NSE (radioimmunoassay, normal range ≦ 10 ng/ml) [39].

The PFAA profiles of subjects were substituted into the discriminating functions obtained from the results of three independent preliminary studies [32, 34, 35]. Both Discriminant- 1 and Discriminant- 3 were logistic regression functions, whereas Discriminant- 2 was a linear discriminating function using plasma concentrations (expressed in μM) as explanatory variables.

Table 1 summarizes the characteristics of the subjects in this study. No significant differences in body mass index (BMI) were observed between patients and matched controls (Table 1). Weight loss due to malnutrition was therefore not expected to influence the results. Although significant differences in average age were observed between the data sets, the effects appeared to be relatively minor because the absolute values of these differences were small (Table 1). Disease stages were determined according to the Sixth Edition of the International Union Against Cancer (UICC) Tumor–Node–Metastasis (TNM) Classification of Malignant Tumors [38]. The fractions of patients at each stage according to the type of cancer were as follows: ~40% stage I, ~5% stage II, ~30% stage III, and ~25% stage IV (Table 1).

The cancer patients were also further subdivided based on histological tumor type; approximately 65% of the patients were classified as having adenocarcinoma, 15% as having squamous cell carcinoma, and 10% as having small cell lung cancer (SCLC) (Table 1).

First, the PFAA profiles of the study data set in previous studies and of the test data set, which was never used for analysis, were used to verify the changes in PFAA profiles observed in cancer patients. Interestingly, the PFAA profiles of the test data set were quite similar to those of the study data set, especially for the ratios of the amino acid concentrations (Figure 1 and Table 2), indicating that the alteration in PFAA profiles observed in cancer patients is robust. Significant increases in both the concentration and ratio of Pro and Orn and significant decreases in His were observed in both the study and test data sets compared to controls (Figure 1 and Table 2). The ratios more clearly reflected the alterations in the PFAA profile than the concentrations; the profiles of five additional amino acids were altered in the ratio data (Gln, Met, and His were decreased in patients, and Ile was increased in patients), while significant changes in concentration were detected in only one direction (Figure 1 and Table 2).

We used three different discriminating functions to distinguish lung cancer patients from controls (Table 3). Discriminant 1 was derived from the PFAA profiles of cancer patients recruited from the Osaka Medical Center for Cancer and Cardiovascular Diseases and controls recruited from the Center for Multiphasic Health Testing and Services of the Mitsui Memorial Hospital [32]. Discriminant 2 and Discriminant 3 were derived from patients from the Osaka Medical Center for Cancer and Cardiovascular Diseases, the Chiba Cancer Center, the Kanagawa Cancer Center, and the Gunma Prefectural Cancer Center and controls recruited from the Center for Multiphasic Health Testing and Services of the Mitsui Memorial Hospital, the Kameda Medical Center Makuhari, and the Kanagawa Health Service Association [34, 35]. Discriminant 3 is commercially used in the “AminoIndex® Cancer Screening” service in Japan (Ajinomoto, CO., Inc.) [35]. Both Discriminant 1 and Discriminant 3 were logistic regression models, whereas Discriminant 2 was a linear discriminating function. Explanatory variables used in these functions are listed in Table 3.

Three different data sets (Dataset 1, Dataset 2, and Dataset 3) were used to verify the performance of the discriminating functions (Table 4 and Figure 2). Notably, the discrimination abilities of each data set were evaluated using the AUC of the ROC of the discriminate score and were found to be > 0.7 in all cases, indicating that the discrimination functions were both reproducible and robust using independent data sets (Figure 2, Table 4). Specifically, AUCs for the discrimination of lung cancer patients were estimated as follows: 0.731 (95% CI: 0.668 - 0.794) for Dataset 1, 0.822 (95% CI: 0.768 - 0.875) for Dataset 2, and 0.777 (95% CI: 0.718 - 0.836) for Dataset 3 for Discriminant- 1; 0.797 (95% CI: 0.738 - 0.856) for Dataset 1, 0.775 (95% CI: 0.714 - 0.836) for Dataset 2, and 0.761 (95% CI: 0.700 - 0.823) for Dataset 3 for Discriminant 2; and 0.805 (95% CI: 0.767 - 0.846) for Dataset 1, 0.806 (95% CI: 0.767 - 0.843) for Dataset 2, and 0.795 (95% CI: 0.755 - 0.831) for Dataset 3 for Discriminant 3 (Figure 2, Table 4).

Selected explanatory variables partially overlapped for the discriminating functions (Table 3); therefore, the discriminant scores were highly mutually correlated as presented in Table 5. The correlation coefficients were as follows: 0.609 (Discriminant- 1 and Discriminant- 2), 0.552 (Discriminant- 1 and Discriminant- 3), and 0.719 (Discriminant- 2 and Discriminant- 3) for all of the data in Dataset 3. For the patients in Dataset 3 (i.e., P1 and P2), the correlation coefficients were 0.559 (Discriminant- 1 and Discriminant- 2), 0.506 (Discriminant- 1 and Discriminant- 3), and 0.686 (Discriminant- 2 and Discriminant- 3), and the correlation coefficients for the controls in Dataset 3 (i.e., C1, C2, and C3) were 0.674 (Discriminant- 1 and Discriminant- 2), 0.645 (Discriminant- 1 and Discriminant- 3), and 0.810 (Discriminant- 2 and Discriminant- 3) (Table 5).

For further investigation of the clinical applicability of PFAA profiles, the combinatorial use of both the discriminating function from PFAA profiles as explanatory variables and existing tumor markers generally used for lung cancer detection and monitoring (CEA, CYFRA, ProGRP, SCC, and NSE) was assessed [39, 41]. In this analysis, Dataset 4 (P1 and P2) was analyzed using discriminant scores obtained from Discriminant- 3. Subgroup analysis was also performed using patient data stratified into cancer stages (stages I and II). For all patients, significantly higher sensitivities were observed upon combinatorial use of Discriminant- 3 and the tumor markers than upon single use of either Discriminant- 3 or the tumor markers (Figure 3). Similar results were observed among stage I and II patients using the combination of Discriminant- 3 and three tumor markers (CEA, SCC, and NSE), while no significant improvement of sensitivity was observed using Discriminant- 3 and CYFRA or ProGRP (Figure 3). These results suggest that the combinatorial use of Discriminant 3 and other tumor markers is effective for lung cancer detection and monitoring, and an increase in sensitivity was indeed confirmed (Figure 3).

Among the tumor markers, CYFRA and SCC are specific to squamous cell carcinoma (SqCC), ProGRP and NSE are specific to small cell lung cancer (SCLC), and CEA is not specific to any particular histological type of lung cancer [39]. Clinically, the combinatorial use of multiple independent tumor markers is effective for detecting lung cancer. Notably, a low correlation was observed between Discriminant 3 and the tumor markers; the correlation coefficients were 0.304 for CEA, 0.481 for CYFRA, -0.228 for ProGRP, 0.346 for SCC, and 0.102 for NSE (data not shown).