Combined Performance of Fecal Immunochemical Tests and a Genetic Risk Score for Advanced Neoplasia Detection

Fecal immunochemical tests (FITs) show overall high accuracy for colorectal cancer detection with sensitivities of approximately 70%–80% (1) and specificities of approximately 90%–95%. However, detection of its precursors, such as advanced adenomas (AAs), remains quite limited, with sensitivities typically around 25%–40% only (1). In recent years, genome-wide association studies have identified an increasing number of SNPs that are associated with the risk of colorectal cancer (2, 3). Although risk information conveyed by single SNPs is small, the combination of multiple SNPs in polygenic risk scores (PRS) has shown potential to enhance colorectal cancer risk stratification (4). In addition, numerous risk factor–based risk scores were proposed in the past years and showed overall rather low to moderate performance for detecting advanced neoplasia (AN), that is, colorectal cancer or AAs (5).

PRS may discriminate between AN cases and controls by themselves to some extent and may also improve risk discrimination by risk factor–based risk scores (6). However, if and to what extent the combination with PRS may also enhance diagnostic performance of FIT for detecting colorectal neoplasia has, to our knowledge, not been systematically investigated. Such improvement could, however, be of utmost importance given the wide and rapidly increasing use of FITs for colorectal cancer screening and selecting those who should undergo colonoscopy.

We therefore analyzed diagnostic performance of marker combinations comprising FIT and a PRS results in detecting AN in a large colorectal cancer screening population from South-Western Germany. Furthermore, we examined whether diagnostic performance of the marker combination would be better than FIT or a PRS alone.

This article follows the STARD (Standards for Reporting of Diagnostic Accuracy; ref. 7), FITTER (Fecal Immunochemical test for Hemoglobin Evaluation Reporting; ref. 8), and TRIPOD (Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis; ref. 9) guidelines.

All analyses are based on data from the BliTz study, a large ongoing study among participants of screening colonoscopy that has been described in detail elsewhere (10–15). The BliTz study has been approved by the Ethics Committees of the Medical Faculty Heidelberg (178/2005) and of the responsible state physicians’ chambers [Baden–Württemberg, M118–05-f; Rheinland-Pfalz, 837.047.06(5145); Saarland, 217/13; Hessen, MC 254/2007] and is registered in the German Clinical Trials Register (DRKS-ID: DRKS00008737) and is conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent. In brief, participants of the German screening colonoscopy program are recruited in gastroenterology practices in Southern Germany. Before bowel preparation, stool and blood samples are taken for evaluating diagnostic performance of noninvasive tests compared with screening colonoscopy.

For the current analysis, we included participants recruited from November 2005 to January 2019. During this recruitment period, two different FITs were applied. From November 2005 to 2008, participants collected raw native fresh stool samples in small containers, stored them in the freezer in provided plastic bags, and brought them to the practice visit for colonoscopy. At the practice visit, the stool-filled containers were immediately stored in the freezer (−15°C to −40°C), shipped to a central laboratory on dry ice within one to few days, stored at −70°C and analyzed with Ridascreen Hemoglobin (r-biopharm, Darmstadt, Germany). Stool collection was done in the same way as described above from November 2008 to February 2012, but another quantitative FIT was applied (FOB Gold, Sentinel Diagnostics). From February 2012 to January 2019, participants collected stool samples in collection tubes containing hemoglobin stabilizing buffers (10-mg stool in 1.7-mL buffer, Sentinel Diagnostics; Ref. 11561H). The tubes were to be sealed in envelopes and mailed to the study center at the German Cancer Research Center, where they were kept at 2°C to 8°C in the refrigerator before transporting in a cold chain to the central laboratory (Labor Limbach) for analysis with FOB Gold. One stool sample was collected per participant. The analytic range of Ridascreen Hemoglobin was 0.65-40 μg hemoglobin (Hb) per gram of stool whereas FOB Gold had an analytical range of 0.2-132 μg Hb/g stool.

Of the 10,362 study participants, genotyping was done in all colorectal cancer and AA cases and an age- and sex-matched random subset of controls. Data on genotyping information in addition to FIT and colonoscopy were available from 6,919 participants. To ensure representativeness of the samples of an average-risk screening population and to minimize the number of screening colonoscopies with missed neoplasms, the following exclusion criteria were applied (Fig. 1). <50 or ≥80 years of age (N = 228), history of colorectal cancer or inflammatory bowel disease (N = 47); colonoscopy in the preceding 5 years (N = 456); inadequate bowel preparation (N = 623); incomplete colonoscopy (cecum not reached; N = 53); most advanced finding was a “non-defined polyp” (N = 206). Thus, 5,306 participants were retained for the analysis.

Blood cell DNA was genotyped using Illumina OncoArray-500k V1.0 BeadChip (Illumina) for 957 subjects and Global Screening Array (Illumina) for 4,349 subjects. Genotyping quality was assessed as described previously (16). Imputation of missing genotypes was performed uniformly using the Haplotype Reference Consortium (version r1.1.2016) as reference panel. PLINK v1.90 was then used to extract SNPs for the required regions of interest.

For the PRS, we considered a recently reported set of 140 common risk variants that were associated with a higher risk of colorectal cancer in the world's largest colorectal cancer genome-wide association study (GWAS) in populations of European descent (ref. 17; see Supplementary Table S1 for details). For sensitivity analyses, we used a weighted PRS (“wPRS”) in which numbers of risk alleles were weighted by the strength of their association (in terms of the natural logarithm of the odds ratio) with colorectal cancer. Blinding from colonoscopy results was ensured when the PRS was constructed and applied to the study participants.

Laboratory personnel were blinded with respect to colonoscopy findings. Stool samples tested with Ridascreen hemoglobin (R-Biopharm AG) were analyzed on a Tecan Freedom Evolyzer. For samples tested with FOB Gold (Sentinel Diagnostics), Abbott Architect c8000 was used.

The median time (IQR) between fecal sampling and laboratory analysis was 9 (6–13) days.

Colonoscopy and histology reports were reviewed by trained investigators who were blinded with respect to genetic and stool testing results. Every participant was classified to one of the following categories according to the most advanced finding at colonoscopy: Colorectal cancer, AA, non-AA, other or no finding. Adenomas were defined as advanced if they matched any of the following features: size ≥1-cm, tubulovillous or villous architecture, high-grade dysplasia. Sessile-serrated polyps were not considered as outcome because they were rarely diagnosed and reported as such during the earlier years of the recruitment period.

Boxplots with data points for individual observations, stratified by colonoscopy result (AN or no AN) and by sex, were used to describe the distribution of PRS values (numbers of risk alleles). Our focus was on comparison of partial AUCs (pAUCs), treating both predictors, FIT and PRS, as continuous variables, rather than selecting specific cutoff values. We used logistic regression models to obtain pAUCs of FIT, PRS and their combination in predicting presence of AN. For PRS, we used the actual number of risk alleles as continuous measure. The regression equations thus had the general form “ln p/(1-p) for having AN = β₀ + β₁f(PRS) + β₂g(FIT),” where β₀ is the intercept, β₁ is the regression coefficient for PRS, β₂ is the regression coefficient for FIT, and f and g are positive monotonous transformations of the PRS and FIT values, respectively (f = g = 1 in the standard models). Our main outcomes were pAUCs that were computed for a restricted range of specificities (80%–100%) supposed to be relevant for population-based screening strategies (18) and rescaled to take values of 1.0 and 0.5 for perfectly discriminating and non-discriminating predictors. AUCs across the entire range of specificities from 0% to 100% were additionally computed to obtain goodness-of-fit and results of likelihood ratio test for model comparison. Model calibration was assessed with calibration plots. In addition to pAUCs, AUCs corrected for overfitting were computed for a combination of FIT and PRS and their 95% confidence intervals were derived by applying the 0.632 bootstrap method with 1,000 replicates. Sensitivities and their 95% “exact” Clopper–Pearson CIs were derived for FIT alone and combinations of FIT with PRS at cutoff values yielding pre-selected specificities of 85%, 90%, and 95% among individuals free of AN and at cutoff values recommended by the FIT manufacturers (2 μg/g for Ridascreen Hb and 17 μg/g for FOB Gold, yielding specificities of 86.5% and 93.2%, respectively). In addition, we calculated sensitivity and specificity for various further conceivable combinations of FIT and PRS (unweighted and weighted) that yielded the same specificity.

Statistical analyses were conducted using R version 4.1.2 (19) and SAS version 9.4. The R package “pROC” was used to construct ROC curves, compute pAUCs and test for significance in differences between pAUCs (bootstrapping) and AUCs (DeLong test). In addition, models were compared by likelihood ratio test. To correct AUCs for FIT+PRS for overfitting, we used 0.632 bootstrapping (SAS “proc surveyselect” and R package “rms”). Boxplots together with points for the individual observations (“jitter”) were created with the R package “ggplot2”. All statistical tests were two-sided and statistical significance was set at an α level ≤0.05.

The data that support the findings of this study are not publicly available due to concerns regarding participants’ privacy but are available upon reasonable request from the corresponding author (T. Niedermaier).

Characteristics of the study participants are shown in Table 1. Of the 5,306 participants, 2,756 (51.9%) were male and 2,550 (48.1%) were female. Mean age was 62.1 years. The sample included 48 participants with colorectal cancer (28 men), 647 with AA (401 men), and 4,611 controls (2,327 men), thereof 1,322 (826 men) with non-AAs as most advanced finding during colonoscopy. Distributions of sex, age and most advanced findings were similar for the subgroups in whom the two FITs were used. The distribution of numbers of risk alleles according to most advanced finding is shown in Fig. 2 and indicates that participants with AN had on average slightly higher numbers of affected risk alleles (mean, 137.4) than those without AN (mean, 135.6), although distributions widely overlapped. Distributions were very similar when stratifying by sex. The weighted PRS showed almost identical differences between participants with versus without AN compared with the unweighted PRS (Supplementary Fig. S1).

Regression coefficients, odds ratios and P values for mutually adjusted logistic regression models are shown in Supplementary Table S2 for the unweighted PRS and in Supplementary Table S3 for the weighted PRS. Because associations between PRS and AN risk were not altered notably by age and sex, we subsequently focused on models comprising only FIT and PRS. Results for pAUCs and AUCs of FIT and PRS alone and apparent and corrected AUCs for FIT and PRS combined are shown in Table 2. Both Ridascreen and FOB Gold were superior to PRS regarding pAUC and AUC for AN (all P < 0.006).

PRS alone had poor discriminatory power for AN detection, with pAUCs of 0.524 (95% CI, 0.499–0.550; Ridascreen group) and 0.530 (95% CI, 0.516–0.545; FOB Gold group). pAUCs did not improve by adding PRS to FIT were even slightly lower for FIT+PRS combined compared with FIT alone (0.659; 95% CI, 0.622–0.697) vs. 0.661 (95% CI, 0.625–0.698) for Ridascreen and 0.667 (95% CI, 0.650–0.687) vs. 0.682 (95% CI, 0.661–0.701) for FOB Gold (Fig. 3). Results also did not improve when using the weighted PRS instead of the unweighted PRS: Also the weighted PRS did not improve AUCs in the region of high specificity, neither for males nor females, and neither for Ridascreen (Supplementary Fig. S2) nor for FOB Gold (Supplementary Fig. S3).

When investigating various alternative functional forms of a combination of FIT (FOB Gold) and PRS, none of them showed clinically relevant or statistically significant improvements compared with FIT alone (Supplementary Table S4). Furthermore, no combination of FIT and PRS cutoff values resulted in both higher sensitivity and specificity compared with FIT (Supplementary Table S5). Neither the weighted nor the unweighted PRS could be combined with FIT in a way that overall accuracy exceeded that of FIT alone to a clinically relevant extent. This result was confirmed by classification tables in which no combination approach achieved more correctly classified individuals than FIT alone (Supplementary Table S6). Finally, for proximal AAs as outcome (which are more frequently missed by FIT than other advanced neoplasms), no combination of FIT and PRS resulted in significantly higher AUCs (Supplementary Table S7). In particular, the clinically more relevant pAUCs were lower than for FIT alone.

Calibration plots for FOB Gold indicated that there was no systematic over- or underestimation or predicted risks (Supplementary Fig. S4).

Table 3 shows sensitivities of FIT and PRS alone and combined for the detection of AN at cutoff values yielding specificities of 85%, 90%, and 95%. In addition, results for all participants screened with Ridascreen at 86.5% specificity and with FOB Gold at 93.2% specificity are reported, which are the specificities achieved at the cutoff values recommended by the respective FIT manufacturer. Overall, results were consistent with comparisons of pAUCs: Sensitivities for AN were very similar with FIT and PRS combined and with FIT alone, with a range from 5%-units lower to 1%-units higher sensitivities. No relevant differences by sex were observed.

In this study, we analyzed measures of diagnostic accuracy [sensitivity, specificity, (p)AUC)] of two FITs and a PRS individually and combined for the detection of AN, that is, colorectal cancer and AA. PRS showed very limited ability to discriminate between AN cases and controls, with pAUCs of 0.524 (among individuals screened with Ridascreen Hb) and 0.532 (among individuals screened with FOB Gold). Adding the PRS to FIT did not improve pAUC for AN detection compared with FIT alone, neither among males nor among females. In line with this, when investigating sensitivities of FIT+PRS at specificities achieved by FIT at cutoff values recommended by the manufacturers (2 μg/g or 86.5% specificity for Ridascreen, 17 μg/g or 93.2% specificity for FOB Gold), improvement in sensitivity was not clinically relevant with Ridascreen (+1.2%-units) and sensitivity with FOB Gold even slightly decreased (−2.1%-units).

To our knowledge, ours is the first study to investigate the potential for improving diagnostic accuracy by combining FIT with a PRS. Almost 700 screen-detected colorectal cancer and AA cases and a random sample of the AN-free participants of screening colonoscopy were included, ensuring representativeness of the results for the target population of screening and avoiding spectrum effects/bias (20). We calculated pAUCs for a clinically relevant specificity range (100%–80%), and sensitivities for AN at selected specificities within this range for two different FITs. Combinations of FIT and PRS were investigated at various functional forms (e.g., PRS+square-root of FIT). Finally, adherence to the TRIPOD guidelines is another strength.

Our study also has limitations. AN cases were predominantly male, which however reflects the higher incidence and prevalence of colorectal neoplasms among men. Our results pertain to a predominantly Caucasian population. Further research would be needed to assess generalizability of results to non-Caucasian populations. Our study did not enable reliable assessment of diagnostic performance for detecting sessile serrated polyps that were rarely diagnosed and reported as such during the earlier years of the recruitment period. Our study population comprised mainly participants ages 50–79, which is the relevant age for colorectal cancer screening. Thus, future studies are needed to address usefulness of FIT and PRS combined among younger older participants. Finally, we cannot rule out that a combination of FIT and PRS would perform differently if other FITs were used, although performance of FITs is supposed to be very similar when adjusting them to yield equal specificities (21). However, given the limited discriminatory power of PRS, larger accuracy gains in combination with other FITs are not expected.

Some studies examined combined use of FIT and other risk scores (though not PRS). Stegeman and colleagues (22) combined FIT with age, colorectal cancer family history, and additional modifiable risk factors such as smoking, BMI, use of NSAID and physical activity. AUC for AN improved significantly, from 0.69 to 0.76. Sensitivity at 93% specificity increased from 32% (FIT alone) to 40% (FIT+risk score) that further suggests substantial improvement in discrimination. Chiu and colleagues (23) investigated sensitivity of an algorithm combining FIT and the Asia-Pacific Colorectal Screening (APCS) scoring system. They found that a combination of FIT and APCS could detect the majority of AN and reduce the number colonoscopies to detect one AN by 40%.

Some studies using risk factor scores for risk stratification showed more promising results than those investigating biomarker combinations: Jung and colleagues (24) assessed prevalences of AN in FIT-positive and -negative groups ages 40–49 years (N = 12,420 overall) with APCS <2 and ≥2 and found prevalence of AN to be much higher (∼13%) in FIT positives with APCS ≥2 compared with those with APCS<2 and a negative FIT (1%). Aniwan and colleagues (25) examined prevalences of AN groups with positive versus negative FIT and APCS in 945 individuals and also found a strong gradient in prevalence of AN, ranging from 6.4% (FIT-negative, moderate risk according to APCS) to 44% (FIT-positive, high risk). Like our study, all aforementioned studies were conducted among asymptomatic individuals.

Overall, it might thus appear that a PRS is of very limited (if any) use as a noninvasive colorectal cancer screening test and would not contribute to any enhancement when combined with a FIT. Further studies, including studies at younger ages, should evaluate to what extent PRS might still be useful, by themselves or in combination with risk factor information, in risk stratification for screening for colorectal cancer (26, 27) or other cancers (28–31), such as defining risk adapted, personalized starting ages (32), or screening or surveillance intervals (14, 33) despite their limited diagnostic value. Furthermore, in clinical practice, positive or negative predictive values might even be more relevant parameters of diagnostic performance than sensitivity and specificity. Those predictive values critically depend on the prevalence of findings among risk groups. Despite limited diagnostic performance of PRS, prevalence of AN varies according to levels of PRS (14). It has been demonstrated recently that clinically relevant differences in predictive values exist between FIT positives in different risk groups, and accordingly, risk-adapted cutoff values would be needed to achieve comparable positive predictive values across different risk groups (34). Those risk-adapted cutoff values might help to optimize FIT-based screening in terms of avoiding unnecessary colonoscopies among those at low risk and reducing the number of missed AN among those at high risk: Instead of using uniform FIT cutoff values to decide who should undergo colonoscopy work-up, a more “targeted” approach could be based on the probability of screening participants having AN. For such a risk-based model, one might consider further variables such as age, sex, previous colonoscopies, family history of colorectal cancer, smoking and other colorectal cancer risk factors such as those examined in previous studies (APCS score, NSAIDs, smoking etc.). Furthermore, it is conceivable that future studies might discover further SNPs that are potentially more predictive for AN, alone and/or in combination with FIT.

In conclusion, PRS did not enhance diagnostic accuracy for AN detection in combination with FIT compared with FIT alone, and FIT+PRS cannot be recommended as combined screening test. Of note, no improvement in detection of proximal advanced neoplasms could be achieved, which are more frequently missed by FIT than other advanced neoplasms (35). Nonetheless, the potential for improving predictive values by PRS in FIT-based screening should be further assessed in future studies. Future studies might also investigate if there is potential for improved screening with a more targeted focus on early stage cancers, that is, the most relevant finding frequently missed by FIT (36, 37).

H. Brenner reports grants from German Cancer Aid (Deutsche Krebshilfe) during the conduct of the study. No disclosures were reported by the other authors.

T. Niedermaier: Formal analysis, investigation, methodology, writing–original draft. F. Guo: Data curation, writing–review and editing, construction and imputation of the updated polygenic risk score. K. Weigl: Writing–review and editing, construction and imputation of the original polygenic risk score. M. Hoffmeister: Writing–review and editing. H. Brenner: Conceptualization, resources, data curation, supervision, funding acquisition, methodology, project administration, writing–review and editing.

This study was partly funded by grants from the German Research Council (DFG, grant No. BR1704/16-1), the Federal Ministry of Education and Research (BMBF, grant no. 01GL1712), and the German Cancer Aid (No. 70113330). Recipient of the grants (to H. Brenner).

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.