Common and Rare Sequence Variants Influencing Tumor Biomarkers in Blood

Tumor biomarkers are substances or processes that can indicate the presence of cancer (1). Several tumor biomarkers are in clinical use for monitoring therapy but all lack the sensitivity and specificity to be used for screening. However, recent advances in the detection of circulating tumor DNA suggest that multi-analyte blood tests that combine an assay of somatically mutated DNA (“liquid biopsy”) and protein and carbohydrate biomarkers in serum have the potential to both find early cancer and to help determine its site of origin (2).

In this work, we focused on six commonly measured biomarkers, namely alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), cancer antigens (CA) 15.3, 19.9, and 125, and alkaline phosphatase (ALP). Measured in serum, these biomarkers are frequently used to monitor status of disease, response to therapy, and recurrence (1). AFP is used as a biomarker of hepatocellular carcinoma (HCC), endodermal sinus tumor of the ovary, and non-seminoma testicular germ cell tumors (TGCT; ref. 3). CEA has been used as a biomarker for colorectal cancer (4). CA-15.3 and CA-125 are mainly used as biomarkers of cancers of the breast and ovary, respectively (5, 6), and CA-19.9 is used as a biomarker for pancreatic cancer (7). We also include ALP in our analysis because its levels are commonly elevated in cancers of the liver and bone and when other cancers metastasize to these tissues (8). However, the measurement of ALP in serum is one of the most common blood tests ordered and we recognize that there are many reasons for ALP measurements other than suspicion of or monitoring of neoplasms.

Despite widespread use of these biomarkers in clinical practice, their low sensitivity and specificity continue to cause controversy over their use (9–11). As their levels are partially determined by genetic factors, one approach to improve their sensitivity and specificity would be to define “normal” values based on age, sex, and genotype (2, 9). We have previously reported how genetic correction for variants affecting levels of PSA results in personalized PSA cut-off value, which is more informative than a general cut-off value when deciding to perform a prostate biopsy (12).

The main goal of this study is to perform a genome-wide association study (GWAS) of the levels of all six tumor biomarkers to identify sequence variants that affect baseline biomarker levels, regardless of cancer diagnosis. We also describe the associations of the six tumor biomarker levels with various cancer diagnoses, obtained from a nation-wide cancer registry, and for comparison, with four nonneoplastic diseases.

Cancer diagnoses, including the date of diagnosis, were extracted from the Icelandic cancer registry (http://www.cancerregistry.is), which contains all diagnoses of solid cancers made in in the country from January 1, 1955 to December 31, 2015 (13). We also assessed four other diseases; inflammatory bowel disease (IBD), liver cirrhosis, and pancreatitis, because these diseases are associated with inflammation in the gastrointestinal organs and fibromyalgia, because patients present with diverse symptoms and often undergo measurements for tumor biomarkers as part of a lengthy diagnosis journey. Tumor biomarker measurements were made in Icelandic laboratories from 1990 to 2015 and linked with disease diagnoses on the basis of encrypted social security numbers.

This study was approved by the National Bioethics Committee of Iceland (reference numbers VSNb2006010014/03.12, 06-007-V3 and VSNb201501033/03.12). A further description of subject recruitment, phenotyping, genotyping, and imputation is available in the Supplementary Materials and Methods.

To identify genetic variants associated with baseline biomarker values, we performed a GWAS of all available data, including both patients with cancer and individuals without cancer diagnosis. When multiple measurements of a biomarker were available for a subject, we used the earliest value recorded. We found this approach to be the most powerful as exclusion of patients with cancer resulted in great loss of power, for residual associations (secondary, tertiary variants etc.) in particular. The first measurement was used as this meant that measurements taken months or years before cancer diagnosis were available for a subset of the patients with cancer.

As indicated in Table 1, the biomarkers all show extremely right-skewed distributions. The data contains a number of extreme outliers but no trends were observed to link these with date of measurement, age of the subject, or even cancer type (Supplementary Fig. S1).

We performed a rank-based inverse normal transformation adjusting for age at measurement and time to death for deceased subjects for each gender separately. Adjustment for time to death was performed as we have observed large changes close to the time of death for many quantitative measurements. In this case, a high biomarker value shortly before death might indicate a high tumor burden. We tested the association between biomarker value and genotype by a generalized form of linear regression. To assess significance of primary associations, we used different P value thresholds depending on the annotation class of the variant as described in Sveinbjornsson and colleagues (14). We consider loss-of-function variants (frameshifts, stop codon gained/lost, initiator codon variants, and splice acceptor/donor variants) significant at 3.6 × 10⁻⁷, in-frame insertions/deletions, missense, and splice region variants at 7.4 × 10⁻⁸, synonymous variants, up/downstream variants, variants that resulted in a stop codon being retained, and variants in 3′/5′ untranslated regions at 5.3 × 10⁻⁹, intronic and intergenic variants within DNase hypersensitivity sites at 3.3 × 10⁻⁹, and intergenic and intronic variants outside DNase hypersensitivity sites at 1.1 × 10⁻⁹. We used the Variant Effect Predictor release 80 (15) to annotate variants, considering only protein-coding transcripts from RefSeq release 67 (16).

Loci harboring variants associated by these criteria underwent further analyses to check for the presence of other, independent variants affecting the trait. While any variant passed the significance threshold at a locus, we tested all variants flanking the primary signal by 1–13 Mb, depending on the strength of association and the recombination pattern at the locus, by sequentially adding the top variant from previous steps as covariate in the regression. The use of wide windows was occasionally necessary to avoid spurious associations arising from very low, but not negligible, linkage disequilibrium (LD) between an extremely significant variant and distant variants. Residual associations were generally close to the primary signal, as can be seen in Tables 2 and 3. We considered significant variants that passed a simple Bonferroni correction for the number of variants in the respective region. This process was repeated, adding the top variant as covariate for the next round until no significant associations remained. The gene expression and colocalization analyses are described in the Supplementary Materials and Methods.

Many of the subjects have undergone several biomarker measurements, prior to, during, and/or after being diagnosed with cancer, with most measurements being from a few months before official diagnosis date or later. Other subjects have undergone measurements without being diagnosed with cancer. The latter subjects show a trend for being younger than the patients with cancer. Supplementary Figure S2 shows the age distribution of each subject group for each biomarker. Some subjects have had more than one type of cancer. In this case, the first diagnosis was used.

We used a two-sided Wilcoxon rank-sum test to compare the highest values of biomarkers recorded for individuals with specific diseases to the highest values of subjects without cancer diagnosis (or diagnosis of any of the other diseases; Fig. 2). We considered significant P values that passed Bonferroni correction for 144 tests (six biomarkers × 24 disease conditions).

The cohorts for each biomarker moderately overlap. Supplementary Figure S3A shows the overlap among subjects measured for AFP, CA-15.3, CA-19.9, CA-125, and CEA. Almost all subjects also had ALP measured. Supplementary Figure S3B shows the overlap among extreme outliers, defined as subjects having biomarker values >1,000 for AFP, CA-15.3, CA-125, and CEA and >5,000 for CA-19.9.

We performed a GWAS on biomarker levels with subjects ranging from 7,107 to 162,774 (Table 1) and identified 84 associations between sequence variants and biomarker levels. We found three sequence variants associated with AFP levels, six with CA-15.3, nine with CA-125, four with CA-19.9, 11 with CEA, and 51 with ALP (Fig. 1; Tables 2 and 3). To assess whether any of the identified variants associates with risk of developing cancer, we tested the association between all the variants and the diagnosis of 20 cancer types in the Icelandic material. With the exception of the association of rs760077 with gastric cancer described below, none of the variants associates with the risk of developing cancer in our cohorts (P > 1 ×10⁻⁶; Supplementary Fig. S6). We also assessed the effect of the variants on gene expression in whole blood in a sample of 2,528 Icelanders who were whole-genome RNA sequenced (Table 3). Finally, we performed a colocalization analysis to link the identified variants with a range of biochemical traits for which summary statistics are available at deCODE Genetics (Supplementary Table S2). We summarize the results for each biomarker below.

We found three sequence variants associated with AFP levels that collectively explain 2.3% of its variance. Two are independent missense variants in the SERPINA1 gene on chromosome 14 [rs28929474-T (Glu366Lys, allele Z) and rs17580- A (Glu288Val, allele S)] and one is a common intergenic variant at the same locus. SERPINA1 encodes the serine protease inhibitor |\alpha $|-1-antitrypsin and we found the minor alleles of both missense variants to be associated with decreased quantity of |\alpha $|-1-antitrypsin in blood in 6,452 Icelandic subjects for which this measurement was available (⁠|\beta = - 1.71\ {\rm{SD}},\ P = 1.0\ \times {10^{ - 83}}$| and |\beta \ = \ - 0.61\ {\rm{SD}},\ P\ = \ 2.0\ \times {10^{ - 38}}$| for rs28929474-T and rs17580-A, respectively). Reduced levels of this protein are known to cause emphysema, cirrhosis, and HCC (17). The same alleles that reduce SERPINA1 also associate with reduced levels of AFP. In addition, rs28929474-T associates with higher levels of three liver enzymes (ref. 18; Supplementary Table S2).

It is of note that we find no association of variants within the AFP gene or its promoter with serum levels of AFP. We replicate neither of the two variants reported to associate with AFP levels in the Chinese population (ref. 19; P = 0.84 for rs12506899 and P = 0.01 with effect in the opposite direction for rs2251844).

We found six sequence variants associated with CA-15.3 levels that collectively explain 42.3% of its variance. Five of these variants are within 1 Mb from the MUC1 gene on chromosome 1, which encodes CA-15.3, and one lies within an intron of the ABO gene on chromosome 9. A single missense variant in MTX1, rs760077-A, explains 33.5% of the trait variance. We have previously shown that rs760077-A is associated with a higher number of tandem repeats within exon 2 of MUC1, and protection against gastric cancer (20). It is likely that the increased number of epitopes on Muc1 leads to higher apparent CA-15.3 levels in the carriers. A second variant at the locus, rs41264915-G, is associated with decreased levels of MUC1 mRNA in blood (⁠|\beta \ = \ - 0.33\ {\rm{SD}},P\ = \ 1.5\ \times \ {10^{ - 16}})$|⁠, a surprising finding given that the same allele is associated with elevated CA-15.3 levels.

We found nine sequence variants associated with CA-125 levels that collectively explain 10.5% of its variance. One variant is within an intron of the MUC16 gene on chromosome 19, which encodes CA-125. One of the variants lies upstream of the MSLN gene on chromosome 16, which encodes mesothelin, a cell surface molecule known to bind CA-125 on the mesothelial lining (21). We further found two low frequency missense variants rs9927389 and rs150425699 (Ala72Val and Arg557His, respectively) in MSLN, suggesting that the observed effect on CA-125 levels is mediated through structural changes in mesothelin.

Five variants are in or close to GAL3ST2 on chromosome 2. Of these, three are low frequency [minor allele frequency (MAF) < 2%] missense variants (Pro143Leu, Leu383Pro, and Tyr153Cys), implicating the encoded galactose-3-O-sulfotransferase in regulation of CA-125 levels. It is currently not clear how this enzyme affects CA-125 levels. We note that the binding site of the antibody used for CA-125 detection has not been reported by the antibody's producer and speculate that variants in GAL3ST2 may influence the glycosylation of Mucin16 and thus affect the outcome of CA-125 measurements but not necessarily the quantity of the protein in blood. Notably, these variants explain a larger fraction of the variance than does the cis-variant in MUC16 (Table 2).

We found four sequence variants associated with CA-19.9 levels that collectively explain 27.4% of its variance. Some are moderately correlated with variants found to associate with CA-19.9 levels in the Chinese population (19). CA-19.9 levels are defined by antibody binding to the cell surface carbohydrate Sialyl-Lewis A and all the associated variants implicate genes known to function in the production and secretion of this molecule. We found a stop-gained variant, rs601338, in FUT2 (secretor gene of the Lewis antigen pathway), a variant upstream of FUT6, rs708686, and a variant in an intron of the FUT3 (Lewis gene), rs2608894, to associate with CA-19.9 levels. In addition, a variant upstream of B3GNT3, rs34262244-A, which greatly affects the expression of the gene (⁠|\beta = -1.04, P = \ 1.3\ \times \ {10^{ - 296}}$|⁠) associate with CA-19.9. B3GNT3 encodes UDP-GlcNAc-|\beta $|-Gal |\beta $|-1,3-N-acetlyglucosaminyltransferase 3, which plays a role in the formation of the backbone of sulfo-sialyl Lewis X tetrasaccaride structures (22).

The signals characterized by rs601338 (stop-gained in FUT2) and rs708686 (upstream of FUT6) colocalize with biochemical traits in blood (Supplementary Table S2). Among these are associations of rs708686-T with levels of galactoside 3(4)-l-fucosyltransferase (⁠|$\beta = - 0.84\ {\rm{SD}},\ P = 2.5\ \times {10^{ - 22}})$|⁠, an effect consistent with reduced levels of CA19.9), and replicated associations of rs601338 with lipase levels reported previously (23).

Eleven sequence variants were associated with CEA levels in our study, explaining 11.9% of its variance. Again, some show moderate correlations with variants associated with CEA levels in the Chinese population (19). Five of the variants are within 250 kb from CEACAM5 gene on chromosome 9, which encodes CEA. The variants rs601338 and rs708686 in FUT2 and upstream of FUT6, respectively, that associate with CA-19.9 levels also associate with CEA but the effect of rs708686 is in the opposite direction to that on CA-19.9.

We also observed an association with an intergenic variant, rs7041150, and two independent variants upstream and downstream of ABO. One of these, rs10901252-C, associates with increased expression of ABO in whole blood (⁠|\beta \ = \ 1.61\ {\rm{SD}},\ P = 1.8\ \times \ {10^{ - 210}}$|⁠). The variants in ABO colocalize with several blood-related traits such as hematocrit, hemoglobin, and granulocyte levels (24) and with cholesterol levels (ref. 25; Supplementary Table S2).

We identified 51 variants associated with ALP levels, collectively explaining 6.9% of its variance. We confirm 13 of 14 associations reported in a recent study of ALP in European populations (26), which partially overlaps with our material. Among these are several associations with the ALPL gene on chromosome 1, driven by familial clusters carrying very rare coding variants (Table 2). The remaining variants are scattered throughout the genome, with variants in or near genes involved in glycoprotein biology (ABO, FUT2, ST3GAL4, GPLD1, and ASGR1) and carbohydrate and lipid metabolism (JMJD1C, TREH, HNF1A, B4GALNT3, PCK1, DHRS9, and GCKR) being prominent. Many of these variants also colocalize with other biochemical traits, such as triglyceride concentrations (ref. 25; Supplementary Table S2).

Out of the 51 variants, 19 are coding and three associate with changes in gene expression (Table 4). We found associations between rs8736-T and decreased expression of TMC4 (⁠|\beta \ = \ - 0.97\ {\rm{SD}},\ P\ \lt\ {10^{ - 300}}$|⁠), between rs4654748-T and decreased expression of NBPF3 |(\beta = - 0.94\ {\rm{SD}},\ P = 8.6\ \times \ {10^{ - 311}}$|⁠) and between rs174564-G and increased expression of FADS2 in whole blood (⁠|\beta = 0.23\ {\rm{SD}},\ P = 7.9\ \times \ {10^{ - 23}}$|⁠).

Given the lack of specificity of the cancer biomarkers, it is of interest to explore which conditions are likely to affect the levels of a particular marker. To search for conditions associated with elevated levels of the selected biomarkers, we compared the highest values for all individuals in a patient group with subjects without cancer diagnosis and identified several diseases where patients showed elevations in biomarker levels (Fig. 2; Supplementary Table S1). We also compared the first measurements done on each individual (Supplementary Fig. S4), as well as the proportion of the highest values falling into bins (Supplementary Fig. S5) defined using the reference values used at Landspitali, the National University Hospital in Iceland (Table 1).

In general, the biomarker levels agree with the indicated clinical use of the respective biomarker but there are some interesting departures from the expected. High AFP levels are overwhelmingly found in HCC with increased levels also seen in TGCT and cirrhosis, a risk factor for HCC. The largest CA-125 levels are strongly associated with ovarian/peritoneal and pancreatic cancers as expected, but a highly significant association is also found between CA-125 levels and cancers of unknown primary site and a more moderate increase in a number of other malignancies. In addition to pancreatic cancer, CA-19.9 levels are also high in cholangiocarcinoma, but the increase in other cancers is much less. CA-15.3 shows a moderate increase in several cancer types in addition to breast cancer, which does not stand out with respect to this biomarker. The levels of CEA and ALP are highest in cancers of the gastrointestinal tract, but both markers show increased levels in many cancer types and ALP levels are increased in all the noncancer phenotypes tested as well.

Tumor antigens are often measured in individuals with unknown ailments in search of diagnostic clues, as well as being used for monitoring the progression of tumors and response to treatment. In our study, we sought to identify variants that influence tumor biomarker values regardless of cancer diagnosis. We did not remove patients with cancer from the cohort before association testing. Their inclusion adds variance to the dataset and results in more conservative estimates of the variance explained by each variant. We repeated the GWAS of the cancer antigens, excluding patients with breast cancer, peritoneal and ovarian cancer, pancreatic, colorectal, and gastric cancer for CA-15.3, CA-125, CA-19.9, and CEA, respectively. We observed all the same primary associations but generally larger effect estimates and higher P values, as a result of reduced sample size. This, together with the observation that only rs760077 is associated with cancer status in the Icelandic cohorts (Supplementary Fig. S6) suggests that these variants generally do not act through effect on cancer risk or aggressiveness.

We report 84 associations between sequence variants and tumor biomarker levels in the Icelandic population. While we are unaware of a study having been published on CA-15.3 and CA-125, variants affecting the other biomarkers have been reported (19, 26). We confirmed most of those associations and discover many additional variants, some of which are rare variants with large effects. In an attempt to shed light on their biology, we tested all the variants for association with expression of genes in their vicinity and for colocalizations with biochemical traits in blood. Blood is not the most relevant tissue for any of the biomarkers and we would not be able to detect tissue- or cell type–specific effects. Some of the variants may also assert their influence on genes farther away than the 250 kb cutoff.

Our comparisons across cancer types show that tumor antigens are strongly associated with diagnosis of several cancers while also highlighting the lack of specificity of these tests. As demonstrated in Fig. 2, many cancers other than those for which the biomarkers are most commonly used showed significant elevation of biomarker values.

A limitation and a potential source of bias in our study is that subjects were not randomly selected for biomarker assaying but rather measurements were done because of suspicion of a particular ailment, to monitor the progression of an existing disease, or for some other clinical reason. In particular, the subject labeled as “without cancer diagnosis” in Fig. 2 are individuals seeking medical assistance and biomarker levels may not reflect those found in healthy individuals.

The utility of genetic correction in prediction models depends on the fraction of the trait's variance explained by sequence variants. The relatively high variance explained by identified genetic factors, for some of the cancer antigens in particular, suggests that improvements could be achieved by the inclusion of these variants in such models. A single variant, rs760077, explains >33% of the variance in CA-15.3. This biomarker is generally considered of low specificity but this high fraction of variance explained highlights the potential for improvement by correcting values for genotype. However, CA-125 is perhaps of most interest in this regard, as effective screening tools for ovarian cancer are currently lacking. Recent screening trials of ovarian cancer reported no mortality benefit from screening with CA-125 and trans-vaginal ultrasound (9, 27). A retrospective study of these cohorts based on genotypically corrected CA-125 levels may show greater benefit of CA-125 screening. Furthermore, corrected biomarker values should be useful when combining liquid biopsies and more traditional biomarkers to detect early tumors (2).

No potential conflicts of interest were disclosed.

Conception and design: S. Olafsson, O. Gunnarsson, K. Olafsson, P. Sulem, T. Jonsson, T. Rafnar, D.F. Gudbjartsson, K. Stefansson

Development of methodology: S. Olafsson, D.F. Gudbjartsson

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Olafsson, J. Gudmundsson, S.N. Stacey, E.S. Bjornsson, S. Olafsson, S. Bjornsson, K.B. Orvar, A. Vikingsson, G. Bjornsdottir, T.E. Thorgeirsson, S. Sigurdsson, O.T. Magnusson, H. Holm, I. Jonsdottir, G.I. Eyjolfsson, I. Olafsson, U. Thorsteinsdottir, T. Jonsson, T. Rafnar

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S. Olafsson, K.F. Alexandersson, J.G.K. Gizurarson, K. Hauksdottir, J. Gudmundsson, S.N. Stacey, G. Sveinbjornsson, G.H. Halldorsson, G. Masson, P. Sulem, D.F. Gudbjartsson

Writing, review, and/or revision of the manuscript: S. Olafsson, O. Gunnarsson, K. Olafsson, J. Gudmundsson, S.N. Stacey, A.J. Geirsson, G. Bjornsdottir, T.E. Thorgeirsson, S. Sigurdsson, H. Holm, I. Jonsdottir, P. Sulem, U. Thorsteinsdottir, T. Rafnar, D.F. Gudbjartsson, K. Stefansson

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J. Saemundsdottir, S. Arinbjarnarson, G.H. Halldorsson, I. Jonsdottir, I. Olafsson

Study supervision: T. Rafnar, D.F. Gudbjartsson, K. Stefansson

Other (responsible partly for the data used from the hospitals and responsible for methods and external quality control at the laboratory of the hospital in Akureyri): O. Sigurdardottir

The authors would like to acknowledge the work of the staff of the genotyping and informatics facilities in deCODE Genetics and of the Icelandic Cancer registry, without whom this study would not have been possible. This study was funded by deCODE Genetics/Amgen and supported in part by the National Institute of Dental and Craniofacial Research of the National Institutes of Health, under award number R01DE022905, awarded to K. Stefansson.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.