Identification of an autoantibody panel to separate lung cancer from smokers and nonsmokers

Early detection of lung cancer is critical for impacting the poor 5-year survival of 15 percent that has persisted for decades. If the cancer is detected in Stage I, survival can exceed 80 percent [1]. CT scans of the chest have greater sensitivity compared to posterior-anterior chest x-rays in detecting small non-calcified nodules that may represent early lung cancers; however, this technique has poor specificity because of the high prevalence of non-calcified and ground glass pulmonary nodules [2, 3]. These nodules may be due to granulomatous disease, fibrosis, atypical adenomatous hyperplasia (AAH), bronchoalveolar carcinoma (BAC), adenocarcinoma, or slowly resolving inflammatory lesions. Ground glass opacities are hazy nodules without obscuration of underlying vascular markings. The National Lung Cancer Screening Trial has been undertaken by the National Institutes of Health to test the hypothesis that CT-scan screening reduces mortality by increased detection of Stage I and II lung cancers.

Blood tests such as serum autoantibodies may identify individuals with early lung cancer and distinguish high-risk smokers with benign noncalcified lesions from those with lung cancer. Malignant cells can activate both the cellular and humoral immune systems, leading to autoimmunity to autologous cellular antigens. Autoantibodies to p53, a tumor suppressor protein mutated in 70% of smokers' lung cancer, were detected in the serum of patients with breast cancer in 1982 [4]. The NHLBI Lung Health Study had 23 cancers in the 5-year trial and 5 (23%) had p53 autoantibodies including sera from 2 lung cancers drawn 6 and 7 months prior to diagnosis [5]. In 133 lung cancer patients, antibodies to p53 were detected in 25 (18.8%), with significant associations with squamous cell type, smoking, and advanced stage [6], and a similar study found p53 autoantibodies in 17.9% of heavy smokers with lung cancer [7]. Five percent of NSCLC sera contained antibodies to c-myc, p53, and eukaryotic translation initiation factor 4 G (eIF4G) [8]. Antibodies with specificity for antigens initially recognized by cytotoxic T lymphocytes, e.g. MAGE, tyrosinase, and NY-ESO-1 have been found in few lung cancer sera [9]. Other immunogenic tumor-associated antigens (TAAs) have been identified through the use of autoantibodies in cancer sera to immunoscreen cDNA expression libraries in order to identify cDNA clones encoding those TAAs [10–12]. Most of these TAAs are involved in cellular functions, including DNA replication and transcription, and pre-messenger RNA splicing and translation [12]. TAAs are often membrane receptors such as HER-2/neu oncoprotein, tumor suppressor gene proteins, cell cycle associated proteins (cyclin B1), centromere protein F (CENP-F), or onconeural antigens [13, 14]. Autoantibodies to some of these antigens have been found in patients with breast, hepatocellular, ovarian, neuroendocrine, colorectal and lung cancers [4, 15–19]. A proteomic approach using solubilized proteins from lung adenocarcinoma cell line A549 and lung tumors were subjected to two-dimensional PAGE followed by Western blot analysis in which individual sera were tested for antibodies [20]. Glycosylated Annexin I was detected in 12/30 (40%) of lung adenocarcinomas and 3/18 squamous cell carcinomas but none in 51 healthy subjects; Annexin II was similar but was more specific for lung cancer. Annexin I is expressed in bronchial epithelium and Annexin II is found in type I and II alveolar epithelial cells. Another autoantibody referred to as PGP 9.5, an ubiquitin COOH-terminal hydrolase, was also detected in 9/64 lung cancer patients and as an antigen in 2 more [21]. Chapman and colleagues recently showed that screening for antibodies in lung cancer using a panel of seven TAAs, p53, c-myc, HER2, NY-ESO-1, CAGE, MUC1, and GBU4-5, resulted in a sensitivity of 76% and specificity of 92% [22].

We hypothesized that cumulative reactivity to a selected panel of TAAs would distinguish individuals with lung cancer from those with benign pulmonary nodules, and ground glass opacities caused by inflammation, fibrosisor neoplasia [23]. Using a mini-array of TAAs (c-myc, p53, cyclin B1, p62, Koc, IMP1 and survivin), we previously showed a greatly increased frequency of positive immune reactivity in breast, lung, prostate, gastric, colorectal and hepatocellular carcinoma [23–26]. Our goal was to develop a panel of TAAs to guide CT scan screening of high risk smokers who have noncalcified nodules to determine whether they are at risk for lung cancer.

The characteristics of the subjects enrolled in the five groups are shown in Table 1. We noted that the lung cancer patients were older (median age 64.5) compared with other subjects and had increased pack years of smoking (52) compared with other subjects; these lung cancer cases also had the lowest median FEV₁/FVC (68.5). Gender and median pack-years smoking were comparable among the four smoking groups. The lung cancer patients tended to have slightly lower FEV₁/FVC ratios compared with individuals without known cancer (lung cancer median 68.5% vs. smokers with normal CTs 71 ± 8%). Table 2 provides a summary of medical history for these groups of subjects. The nonsmoking controls had been screened to be in good health. Note that 5 (10.9%) subjects with ground glass opacities developed lung cancer on follow-up (13% resolved); 9.1% of patients with solid nodules also had resolution of these nodules on follow up (Table 2). Only the proportion of patients with emphysema and pneumonia differed among the groups with the highest proportions with these diseases on CT scan observed in the ground glass group (p ≤ 0.001, p = 0.04, respectively, chi square tests).

All of the patients with lung cancer and most of the individuals with ground glass opacities demonstrated reactivity to one or more of the 10 TAAs, (using a cut-off value of > mean + 3 SD of non-smoking normal sera). Figure 1 provides box plots that summarize the distributions of age and of autoantibodies to each of the 10 TAAs for the five groups of subjects. The distributions of each of the 10 biomarkers differed among the five classification groups (p = 0.008 for cyclinB1; p < 0.001 for all others). After a conservative adjustment for multiple comparisons, we note that the healthy nonsmoking controls have significantly lower median levels of each of these biomarkers compared with the four groups of smokers including lung cancer cases. Table 3 summarizes the results of the pairwise Wilcoxon Rank Sum tests (bolded values are significant with a conservative Bonferroni adjustment for multiple testing) for the comparisons among the four groups of subjects in the NYU cohorts. In particular, the median levels of autoantibodies to Cyclin A, Cyclin D1, and survivin are lower in the group with ground glass opacities compared with the groups with cancer, no nodules, or solid nodules. The cancer group has significantly higher levels of autoantibodies to c-myc, Cyclin A, and CDK2 compared to the solid nodule group.

		p53	c-myc	imp1	p62	imp3	cycA	cycB1	cycD1	cdk2	survivin
No Nodules	Solid Nodules	1.000	0.031	1.000	0.001	1.000	1.000	0.228	0.351	1.000	0.022
No Nodules	GGO	0.750	0.810	0.128	1.000	0.680	<0.001	1.000	<0.001	1.000	<0.001
No Nodules	Cancer	0.630	0.049	1.000	1.000	1.000	0.258	1.000	0.001	0.025	1.000
Solid Nodules	GGO	0.500	<0.001	<0.001	1.000	0.170	0.003	0.007	<0.001	0.423	<0.001
Solid Nodules	Cancer	0.630	<0.001	0.009	1.000	1.000	<0.001	1.000	0.018	0.001	0.139
GGO	Cancer	1.000	0.532	1.000	1.000	0.350	<0.001	1.000	<0.001	0.262	<0.001

One hundred forty nine of the 158 high-risk smokers and cancer cases had complete data on all 10 biomarkers. A multiple logistic model to predict cancer versus no cancer (no nodules, solid nodules, or ground glass opacities) yielded an AIC of 86.6 with 91% accuracy using 10 fold cross validation (Table 4A). The best stepwise model reduced the AIC to 79.4 with the same cross validation prediction accuracy (Table 4B). C-myc, cyclin A, cyclin B1, cyclin D1, CDK2, and survivin contributed to the separation between the cancer and non cancer groups. The resulting Receiver Operating Characteristic (ROC) curve is shown in Figure 2 along with the sensitivity and specificity associated with the optimal cutpoint (91%, 99% respectively). When this model, which includes c-myc, cyclin A, cyclin B1, cyclin D1, CDK2, and survivin, was applied to the healthy non-smoking control group as a test set, we correctly classified 31/36 control patients as non-cancer using a cutoff value of 0.085 which maximizes the sensitivity and specificity of the logistic function at 81% and 97% respectively. Removal of outliers yielded similar results.

A: All Biomarkers.
	Estimate	Std. Error	z value	Pr(>|z|)
(Intercept)	6.39	2.52	2.54	0.01
P53	-0.11	0.48	-0.22	0.82
C-myc	0.93	0.56	1.66	0.10
IMP1/Koc	-0.21	0.56	-0.37	0.71
P62/IMP2	-0.30	0.49	-0.60	0.55
IMP3	0.14	0.67	0.21	0.84
Cyclin A	2.69	0.79	3.41	<0.01
Cyclin B1	-0.84	0.69	-1.22	0.22
Cyclin D1	-2.70	0.83	-3.27	<0.01
CDK2	1.32	0.67	1.95	0.05
Survivin	2.39	0.91	2.62	0.01
AIC	86.60
10 fold Cross Validation	91%
B: Stepwise Multiple Logistic Model.
	Estimate	Odds Ratio	Std. Error	z value	Pr(>|z|)
(Intercept)	6.95	1043.15	2.25	3.09	<0.01
C-myc	0.80	2.22	0.53	1.53	0.13
Cyclin A	2.59	13.32	0.71	3.64	<0.01
Cyclin B1	-0.87	0.41	0.63	-1.38	0.17
Cyclin D1	-2.73	0.06	0.69	-3.94	<0.01
CDK2	1.27	3.56	0.60	2.11	0.04
Survivin	2.44	11.47	0.89	2.75	0.01
AIC:	79.40
10 fold Cross Validation	91%

A biomarker for the early detection of lung cancer will likely require a panel of 4-10 markers to maximize sensitivity and specificity. High throughput technological advances allow for the rapid evaluation of a huge number of possible markers, for example, by microarray analysis of gene expression, or proteomic analysis using SELDI or MALDI technology. 2 D-PAGE has identified antioxidant enzymes, ATP synthases, β1,4-galatosyltransferase, glutathione-S-transferase, and ubiquitin thiolesterase among others to be increased up to 10-fold in lung cancer compared to uninvolved tissue and correlated with their gene expression [29]. Elevated levels of phosphoglycerate kinase 1 in the serum were significantly correlated with poor outcome in lung adenocarcinoma [30]. Protein expression profiling using MALDI-TOF of lung adenocarcinoma has identified unique peaks that were sequenced and found to match macrophage migration inhibitory factor and cyclophilin A that also were over-expressed in tumor tissue using immunohistochemistry [31]. In contrast, using tissue microarrays and immunohistochemistry for these two proteins in 234 lung cancer patients, Howard and colleagues found no correlation with outcome although both proteins were over-expressed in most NSCLC tumors [32]. We identified 7 of 84 antibodies using two-color rolling-circle amplification protein microarrays that gave a significant (p < 0.01) difference for 24 lung cancer patients compared to 24 controls and 32 COPD subjects [33]. Proteins identified were C-reactive protein, serum amyloid A, mucin 1, and α-1-antitrypsin. C- Reactive Protein and Serum Amyloid A may reflect the inflammatory milieu surrounding the lung tumor, or may reflect over-production of these as actual biomarkers. C-Reactive Protein along with lung function and pack-years of smoking predict progression of bronchial dysplastic lesions [34]. Recently 14-3-3 Theta and LAMR1 were added to Annexin I and PGP 9.5 achieving a ROC of 0.73 (0.69 to 0.81) with significant p values for three out of four autoantibodies comparing the samples drawn at 0-6 months to 7-12 months prior to diagnosis in 85 lung cancer and control paired samples from the Carotene and Retinol Efficacy Trial serum bank [35].

Zhong and colleagues used fluorescent microarray technology to identify immunogenic phage-expressed proteins from T7 phage NSCLC tumor libraries, and to assess the presence of corresponding antibodies in the plasma of patients with NSCLC [36, 37]. Their five most predictive phage-expressed proteins, combined in a logistic regression analysis, achieved 90 percent sensitivity and 95 percent specificity in their population of advanced stage NSCLC. The leave-one-out statistical analysis achieved 88.9 percent diagnostic accuracy. A subsequent study using similar antibody profiling in patients with early stage lung cancer from a CT screening study identified a slightly different panel of five antibody markers that was able to correctly classify stage I NSCLC with 91 percent accuracy, and predicted the presence of lung cancer up to five years before diagnosis with a sensitivity of almost 83 percent [18]. However, this panel of markers did not perform as well in identifying bronchoalveolar cell carcinoma. Proteins of interest were the heat shock proteins, proto-oncogenes, ras-associated oncogene, transcriptional regulators, DNA mismatch repair, tumor necrosis factor receptor, paxillin (focal adhesion protein), and several unknown proteins. Plakophilin 3 and ubiquilin 1 have been found to be humoral response targets using lung cancer phage display libraries [38, 39]. Serum tumor markers such as carcinoembryonic antigen have been reported to be positive in 46 percent of 200 lung adenocarcinoma patients, as well as CA-125 and cytokeratin 19, but these markers correlate better with tumor burden or advanced stage [40].

We have taken a different approach, and focused on known proteins that play a role in the development of lung cancer. We included three proteins containing similar RNA-binding motifs, that we and others have shown to be over-expressed in cancers and where antibodies to these TAAs have been identified in cancer patients: IMP1 (insulin-like growth factor 2 mRNA binding protein), IMP2/p62, and IMP3/Koc [12, 19]. We also included the tumor suppressor gene product p53 and oncogene c-myc, proteins known to be aberrantly expressed in early lung cancers. Survivin is an anti-apoptotic protein that we have shown to be antigenic in cancer patients [23]. The cell cycle proteins cyclin A, cyclin B1 (shown to be antigenic in several malignancies), cyclin D1 and CDK2 were also included. Zhang and colleagues studied sera of 777 cancer patients with 20/84 (24 percent) of lung cancer patients having one or more autoantibodies to this panel of antigens [19].

The objective of this study was to distinguish biomarker measurements among five classifications based on the CT scan: Non-smoking controls, no nodules, solid nodules, ground glass opacities, and cancer. Based on the nonsmoking and healthy smoker controls, patients with lung cancer had increased levels of immune reactivity to the TAAs. Smokers with benign solid nodules and smokers with no nodules had equivalent lower levels of reactivity. Smokers with GGOs, which may represent a preneoplastic or neoplastic condition in some patients, had an intermediate level of reactivity between patients with cancer and smokers with normal CTs or benign solid nodules. We found that reactivity to two tumor associated antigens, c-myc and p62, correlated most closely with the presence of either GGOs or lung cancer. From these analyses, it appears that it is possible to distinguish cancer and ground glass opacity patients based on their biomarker measurements with some measure of certainty. Cyclin A, Cyclin D1, CDK2, and survivin appear most often in models, irrespective of outcome. However, this study is limited by small sample size, of which only 22 patients had cancer and extreme biomarker measurements could not be repeated or validated.

Persistent GGOs are more likely to represent preneoplastic or neoplastic lesions. Nakata and colleagues reported that 100 percent of ground glass opacities ≤ 2 cm in size that persisted for at least three months were found to be malignant or pre-malignant lesions on biopsy or resection [41]. Many persistent GGOs are likely to be adenocarcinoma, bronchoalveolar carcinoma, or atypical adenomatous hyperplasia, and assessing these versus benign nodules or smoker controls using autoantibodies to TAAs is unique. Combinations of antibodies against TAAs are most likely to achieve the necessary sensitivity and specificity for early detection in CT scan screening trials when small noncalcified solid or ground glass nodules are discovered in the >8 mm size range. These prediction models require validation in larger patient groups with samples collected prospectively.