Microbiome signatures for detection of colorectal lesions in population-based FIT screening Authors: Einar E Birkeland, Ane S Kværner, Ekaterina Avershina, Cecilie Bucher-Johannessen, Vahid Bemanian, Hege S Blix, Anette Hjartåker, Willem M de Vos, Giske Ursin, Geir Hoff, Kristin R Randel, Edoardo Botteri, Paula Berstad, Trine B Rounge Published: 2026 ABSTRACT AbstractThe gut microbiome has been linked to colorectal cancer (CRC) development, with microbe-based classifiers distinguishing between CRC patients and healthy controls. However, there is a lack of studies addressing the utility of the microbiome in screening-relevant settings, including both precancers and CRC. In this Norwegian population-based study, we used fecal immunochemical test (FIT) leftovers from 1034 FIT-positive (i.e. positive for occult blood) screening participants for gut metagenome profiling using shotgun sequencing. Using comprehensive clinical, demographic, and lifestyle data, we modeled gut microbiome associations with CRC screening outcomes. Combining microbial profiles with quantitative FIT values improved detection of premalignant lesions beyond optimizing the FIT value alone, even after incorporating established CRC risk factors. Still, the FIT value maintained superior discriminative ability for CRC. We confirmed enrichment of bacteria such as Fusobacterium nucleatum and Peptostreptococcus stomatis in CRC. In contrast, other bacteria previously associated with the presence of CRC, including Hungatella hathewayi and Clostridium symbiosum, as well as pks-negative Escherichia coli, were enriched in those with no neoplastic findings, suggesting that in a FIT-positive population their presence may reflect other conditions causing intestinal bleeding rather than underlying neoplasia. Microbial profiles were predominantly associated with distal rather than proximal lesions. Together, our findings highlight the potential for microbial markers to improve FIT-based CRC screening, especially by differentiating those with premalignant lesions from those who test FIT-positive for other reasons.Subject terms: Colorectal cancer, Microbiome, Diagnostic markers Introduction The fecal immunochemical test (FIT), detecting fecal occult blood by measuring the concentration of human hemoglobin, has been widely adopted as the primary screening modality for organized colorectal cancer (CRC) screening programs1. In these programs, FITs are used to identify those who are more likely to have CRC or its precursors, thereby allocating constrained screening resources to those who will benefit the most. A positive FIT result is followed up with colonoscopy for detection of CRC and removal of its precursors. The FIT test is effective at identifying CRC, with a sensitivity of 0.75–0.91 for CRC2, and provides a reduction in CRC mortality compared to no systematic screening3. Still, the test has limited sensitivity for detecting precancerous lesions4. Additionally, bleeding in the colon may also be caused by colonic conditions other than CRC or its precursors, such as hemorrhoids5 or inflammatory bowel disease6 (IBD) or other, unexplained reasons. Such conditions may cause FIT-positivity, contributing to lower specificity for the FIT test in CRC screening. To minimize the number of colonoscopy procedures carried out for screening participants who do not have CRC or precancers (false positives), the hemoglobin threshold used to determine colonoscopy referrals may be increased7. Increasing the threshold, however, entails a loss of sensitivity, especially for the identification of individuals with precancerous neoplasia8. Therefore, the development of non-invasive biomarkers to complement or substitute bleeding-related markers may improve CRC screening. The gut microbiome is a potential contributing factor in CRC development. Experimental evidence shows that transplantation of stool from CRC patients can promote carcinogenesis in mouse models9, and that use of antibiotics has been linked to increased risk of CRC10. Bacteria such as pks-positive Escherichia coli have been implicated in colorectal carcinogenesis11, whereas Fusobacterium nucleatum is often found in CRC tissues12–14. Studies of stool microbiomes from CRC patients and healthy controls have shown clear differences in the composition of microbes, indicating that the gut microbiome can be used for improved classification of CRC cases15–18. The risk of developing CRC is affected by modifiable lifestyle factors. Key contributors include a sedentary lifestyle, adiposity, consumption of alcohol and red and processed meat, as well as smoking19–22. These lifestyle factors not only impact CRC risk but several have also emerged as important modulators of microbiome diversity23–25. Additionally, non-modifiable factors such as age and sex impact both CRC risk and the gut microbiome26. Recent studies have shown how associations between the microbiome and CRC may be confounded by modifiable and non-modifiable factors27,28, emphasizing the importance of accounting for lifestyle factors when assessing the relationship between the microbiome and CRC. So far, most microbiome studies have been conducted as case/control studies in the setting of established and symptomatic cancer, contrasting healthy controls, with only a few studies including significant numbers of adenomas or other precancers29,30. In this Norwegian screening cohort based on consecutive recruitment, we demonstrate a potential for microbiome-derived biomarkers in population-based CRC screening31. We focus on differentiating CRC or other colorectal neoplasms from FIT-positive individuals without neoplasia, who have blood in their stools for reasons unrelated to cancer development; a key challenge for evaluation of microbial markers able to complement FIT in real-world screening. We analyze gut microbiomes using shotgun metagenome sequencing of residual buffer from FIT-positive screening samples, integrating comprehensive data on clinical, demographic, and lifestyle factors. Results Dataset description To identify potential microbial biomarkers for the detection of participants at high risk for the development of CRC, we performed metagenome sequencing using leftover buffer from FIT screening samples in the multicenter CRCbiome study in Norway32. The samples were originally collected as a part of a biennial population-based CRC screening pilot, the Bowel Cancer Screening Norway study31 (BCSN; cumulative participation rate 68.4% after three rounds of FIT). Those with a positive FIT test (>15 microgram Hb/g feces) were invited to participate in the CRCbiome study (N = 2700), and 1640 consented. Participants with findings of uncertain clinical significance (1–2 non-advanced adenomas, n = 406, or non-advanced serrated lesions, n = 104), who had polyps without histology-based diagnoses (n = 7), who did not attend follow-up colonoscopy (n = 40), or for which sufficient sequencing data was not obtained (n = 49), were not included in the analyses (Supplementary Data 1). The study population thus consisted of 1034 individuals, where 582 were men (56%), and the median age at invitation was 67.1 years, with participant enrolment occurring in their second to fourth scheduled rounds of screening (113 participants had not attended prior rounds of screening). In total, 584 (56.5%) had neoplastic findings (Table 1), including 66 (6.4%) individuals with CRC (55% Stage I; Supplementary Data 2), 298 (28.8%) with advanced adenoma, 78 (7.5%) with advanced serrated lesions, and 142 (13.7%) individuals with three or more non-advanced adenomas as their most advanced finding (Supplementary Data 3). Participants without neoplastic findings (n = 450; 43.5%) included 9 individuals with “other” findings (inflammatory polyp, granulated tissue polyp), 272 with “non-neoplastic” findings (e.g., diverticulitis, hemorrhoids, IBD), and 169 with no findings. Table 1Participant characteristicsCharacteristicnNo neoplastic findings n = 450aNeoplastic findingsb n = 584ap-valuecAge103466.0 (60.8, 71.1)68.1 (62.8, 72.3)4.2 × 10-4Sex10340.002Female221 (49%)231 (51%)Male229 (39%)353 (61%)Geographic region10341.4 × 10-6Region 1 (Moss)286 (50%)282 (50%)Region 2 (Bærum)164 (35%)302 (65%)FIT value103433 (21, 64)40 (26, 92)9.3 × 10-5Programmatic screening round10340.62nd round123 (45%)150 (55%)3rd round222 (44%)283 (56%)4th round105 (41%)151 (59%)Hemorrhoids10340.001Yes93 (55%)77 (45%)No357 (41%)507 (59%)Diverticulitis10340.6Yes189 (42%)256 (58%)No261 (44%)328 (56%)IBD10348.6 × 10-5Yes21 (84%)4 (16%)No429 (43%)580 (57%)Self-reported gastrointestinal disorders10020.009Yes77 (53%)67 (47%)No359 (42%)499 (58%)National background9890.1Non-Norwegian37 (53%)33 (47%)Norwegian399 (43%)520 (57%)Education10080.7University/college181 (43%)238 (57%)High school168 (43%)225 (57%)Primary school91 (46%)105 (54%)WCRFd9453.50 (3.00, 4.50)3.50 (2.75, 4.25)0.002Smokinge10150.05Non-smoker339 (46%)405 (54%)Smoker104 (39%)165 (61%)Recent antibiotics usef10340.7Yes87 (45%)107 (55%)No363 (43%)477 (57%)FIT Fecal Immunochemical Test, IBD Inflammatory Bowel Disease, n participants in group with dataaValues are median (interquartile range; IQR) for continuous variables and n (row-wise %) for categorical variables.bColorectal neoplasia included individuals with 3 or more non-advanced adenomas, advanced serrated lesions, advanced adenomas, or CRC identified by colonoscopy.cWilcoxon rank sum test; Pearson's Chi-squared test (two-sided; no correction for multiple testing).dWCRF score measures adherence to the 2018 World Cancer Research Fund/ American Institute of Cancer Research (WCRF/AICR) cancer prevention recommendations on a seven-point scale.eSmokers included current smokers (regular or occasional) or those having quit consumption within the last ten years.fRecent antibiotic use was defined based on self-reported use or registry data of prescription drug dispensation indicating antibiotic use within 3 months of sampling. Participant characteristics FIT Fecal Immunochemical Test, IBD Inflammatory Bowel Disease, n participants in group with data aValues are median (interquartile range; IQR) for continuous variables and n (row-wise %) for categorical variables. bColorectal neoplasia included individuals with 3 or more non-advanced adenomas, advanced serrated lesions, advanced adenomas, or CRC identified by colonoscopy. cWilcoxon rank sum test; Pearson's Chi-squared test (two-sided; no correction for multiple testing). dWCRF score measures adherence to the 2018 World Cancer Research Fund/ American Institute of Cancer Research (WCRF/AICR) cancer prevention recommendations on a seven-point scale. eSmokers included current smokers (regular or occasional) or those having quit consumption within the last ten years. fRecent antibiotic use was defined based on self-reported use or registry data of prescription drug dispensation indicating antibiotic use within 3 months of sampling. Individuals with neoplastic findings had higher concentrations of blood in their stools (FIT value) than those without (Table 1, Supplementary Data 3) and were less likely to have hemorrhoids and IBD. In line with the current consensus on risk factors for CRC33, neoplastic findings were more frequently detected in men, participants of advanced age, smokers, and those with lower adherence to cancer prevention recommendations (the WCRF score; see “Methods” section). Previously, we have also shown that participants with colorectal neoplasia have higher intake of alcohol and red and processed meat20,22. Neoplastic findings were more frequent in individuals from region 2 (Bærum; 302/466 (65%)) than from region 1 (Moss; 282/568 (50%)). To evaluate the representativeness of the CRCbiome study population, we compared their use of prescription drugs for common comorbidities with that of the full FIT arm of the BCSN study (N = 46,962) and a nationally representative screening-eligible population (N = 16,500) (see “Methods” section). Drug use for the CRCbiome population within 1 year of initial invitation to BCSN indicated a higher rate of cardiovascular disease (53% vs 46%) and more frequent use of antibiotics (28% vs 26%) for CRCbiome participants than the general screening-eligible population (Supplementary Data 4). Notably, comorbidity-related prescription drug use did not differ substantially from the BCSN FIT arm. Microbial diversity and composition in a CRC screening population Read-based taxonomic and functional profiling resulted in the identification of 2756 microbial species, with a mean of 239 identified per sample (range = 68–498), and a mean of 292,162 (SD = 59,300) gene families per sample, aggregated to a mean of 2678 (SD = 564) KEGG orthogroups (KOs). Gut microbial profiles were dominated by Bacillota (former Firmicutes) species, both in prevalence, relative abundance, and species diversity (Fig. 1a–c; Supplementary Fig. S1a, b). The FIT-derived taxonomic profiles were consistent with publicly available gut microbiomes from stool samples15,34 (Supplementary Fig. S2), and an assembly-based approach resulted in similar microbial profiles (Supplementary Fig. S3). All results are thus reported using read-based profiles, with results for assembly-based profiling available in Supplementary Data 5–10 Microbial diversity (represented by number of observed species and KO groups, the Shannon and inverse Simpson indices), was significantly associated with factors related to CRC risk and socioeconomic status, including physical activity, adherence to cancer prevention recommendations (WCRF score), alcohol consumption, age, smoking status, recent use of antibiotics (within 3 months; self-reported or as indicated by prescription drug records) (Fig. 1d, Supplementary Data 5), and education level. Similar observations were made for microbial composition, where the WCRF score showed the greatest explanatory power, with a PERMANOVA R² of 0.0067 (p < 0.001), followed by smoking status, body mass index (BMI), and education level (Fig. 1e, Supplementary Data 6). Analyses stratified by the presence of neoplastic lesions showed broadly similar patterns, with the exception of participant age, where associations between microbial diversity and composition and older age were more emphasized in those with neoplasia. Microbial diversity in fecal immunochemical tests from screening participants. a Prevalence and b relative abundance of the 10 most prevalent microbial phyla (with remaining phyla grouped in “other”) stratified by participant colonoscopy findings of any colorectal neoplasia (blue) or no neoplasia (gray). For visualization purposes, a relative abundance of zero has been omitted. c The number of species observed per microbial phylum. b, c Boxes encompass the 25th–75th percentile, with the median indicated by a line; whiskers extend to 1.5× the interquartile range from the box, with remaining datapoints plotted as outliers. d Coefficients for associations between microbial alpha diversity measured as observed species and participant lifestyle and demographic factors in the full dataset, and stratified by the presence of neoplastic lesions (n = 450 without neoplasia, and n = 584 with neoplasia; see Supplementary Data 5 for further information). Associations were evaluated using a linear model, adjusting for participant age, sex, geographic region, and sequencing reads. Continuous variables were split into tertiles, with the lowest tertile considered the reference; coefficients denote the difference in observed species associated with each variable level relative to the reference. Participants with unknown or missing data were excluded, as were categories of participants with 10 or fewer observations. Points and lines indicate differences in observed species for the test level relative to the reference with 95% confidence intervals. Significant differences are represented by filled black dots. e Associations between gut microbiome beta diversity (Bray–Curtis distances based on species relative abundance) and lifestyle and demographic factors for the full dataset and stratified by the presence of neoplastic lesions (Supplementary Data 6). Bars show variance explained by participant characteristics in permutational analysis of variance (PERMANOVA) models, in pairwise comparisons between levels of each variable: continuous variables were split into tertiles, and tests were made comparing each level to the lower tertile. Participants with unknown or missing data were excluded, as were categories of participants with 10 or fewer observations. Bars are shaded by PERMANOVA significance from light gray (p = 1) to dark red (p = 0.001); asterisks indicate p < 0.05. * Abundance or diversity values of 0 were omitted from (b) and (c) (prevalence presented in a) indicates the n of presented data). Microbial associations with clinicopathological outcomes were evaluated in models adjusted for lifestyle, diet, antibiotic use, and gastrointestinal conditions (see “Methods” section). Here, each standard deviation increase in Shannon diversity was associated with 16% higher odds of detecting any colorectal neoplasia (OR = 1.16 [1.01–1.33]; Fig. 2a, Supplementary Data 7), and specifically 24% for detection of advanced adenomas relative to no neoplastic findings (OR = 1.24 [1.05–1.47]). Conversely, the number of KO groups was negatively associated with colorectal neoplasia detection (OR = 0.78 [0.68–0.90]), potentially driven by differences in species such as E. coli that have large pangenomes (see below). No significant associations were found for those with CRC or advanced serrated lesions. Associations between outcome groups and microbial diversity, composition, and community states. a Odds ratios and 95% CI of outcome groups associated with one standard deviation increase in microbial alpha diversity, estimated using multinomial logistic regression models adjusted for sequencing depth, age, sex, geographic region, smoking, level of education, the WCRF score, hemorrhoids, IBD, and recent use of antibiotics. b PCoA based on Bray–Curtis distances, categorized by diagnostic groups. Solid points indicate centroids, and ellipses encompass 50% of observations. Boxplots show the distribution of scores for principal coordinates 1 and 2 according to outcome category, with boxes encompassing the 25th–75th percentile, the median indicated by a line, and whiskers extending to 1.5× the interquartile range from the box, with remaining datapoints plotted as outliers. Results from the association between Bray–Curtis distances and outcome groups using PERMANOVA are shown. c PCoA as in (b) categorized by microbial community states derived from Dirichlet multinomial mixture models. d Odds ratios and 95% CI of outcome groups associated with having one microbial community state relative to any other, presented in panels from left to right. Models were adjusted as in (a), except for not including sequencing depth. Compositional differences were found between those with and without neoplastic findings (PERMANOVA R2 = 0.002, p = 0.002), and between diagnostic groups (PERMANOVA R2 = 0.005, p = 0.011; Fig. 2b, Supplementary Data 8). The largest differences in between- versus within-group variance were observed between participants with no neoplastic findings and those with advanced adenomas (Supplementary Fig. S4; Pseudo F = 1.60). In order to further evaluate high-level microbiome features, we used Dirichlet multinomial mixture models (DMMs35; Fig. 2c) to classify the microbial profiles into four community states, with the largest groups enriched by Clostridium spp. (n = 425); Methanobrevibacter smithii (n = 331); Bacteroides spp. (n = 200); or Prevotella copri (n = 78; Supplementary Fig. S5a). Gut community states were strongly associated with alpha diversity (observed species), with Bacteroides-enriched communities having especially low diversity (Supplementary Fig. S5b). The community states were further associated with several host features, including sex, the WCRF score, and recent anti-acid medication use (Supplementary Fig. S5c). Notably, neoplastic findings were less common among those with a Bacteroides-enriched community state (OR = 0.55 [0.39–0.78]; Fig. 2d, Supplementary Data 11); specifically, these participants were less likely to be diagnosed with CRC (OR = 0.33 [0.14–0.81]). An opposite non-significant trend was observed for the community state dominated by P. copri (OR = 2.0 [0.78–5.44]). Prediction of colorectal neoplasia using microbial models To assess the utility of the microbiome as a potential biomarker for colonoscopy need, we constructed machine learning models for the identification of individuals with neoplastic findings. Here, 20% of the dataset was set aside as a test set for validation purposes. The random forest (RF) machine learning algorithm with relative abundance of microbial species as predictors outperformed other classification methods and gene content (KOs) as predictors, whether used alone or in combination with demographic and lifestyle factors (see Supplementary Fig. S6). The best-performing microbe-based RF models predicted neoplastic findings with a mean AUROC across 20 times repeated 5-fold cross-validation of 0.610 (SD = 0.036) in the training set and 0.56 (95% CI = 0.48–0.64) in the test set (Table 2). The microbe-based model demonstrated improved predictive performance for identifying colorectal neoplasia relative to raising the FIT threshold above 15 µg Hb/g feces in the training set (FIT AUROC = 0.564, SD = 0.042), but more ambiguous results in the test set (FIT AUROC = 0.60, 95% CI = 0.52–0.68). We employed an ensemble approach for the prediction of neoplastic findings, incorporating both microbial predictions and the FIT value, taking advantage of the FIT accuracy for CRC, achieving an AUROC of 0.615 (SD = 0.040) in the training set and 0.61 (95% CI = 0.54–0.69) in the test set. In ensemble models, the microbial predictions contributed to model accuracy (median LRT p-value < 0.001 [max p-value = 0.004]) and were generally more accurate in the 20% cross-validation sets for the combined FIT-microbial models than for the FIT alone. However, these differences in accuracy did not consistently reach statistical significance (median AUROC difference = 0.05, median p-value for AUROC difference = 0.25, significant differences in 20% of repeats). Table 2Training and test set prediction accuracy for neoplastic findingsModelAUROCaSensitivityaSpecificityaTraining setFIT alone0.564 (0.042)0.594 (0.133)0.496 (0.141)Microbial RF model alone0.610 (0.036)0.810 (0.115)0.322 (0.153)Microbial RF model + FIT0.615 (0.040)0.754 (0.126)0.398 (0.163)Microbial RF model + FIT, sex, age, and WCRF score0.639 (0.037)0.835 (0.089)0.328 (0.117)FIT, sex, age, and WCRF score0.607 (0.037)0.806 (0.122)0.332 (0.144)Test setFIT alone0.60 (0.52–0.68)0.81 (0.74–0.88)0.32 (0.22–0.42)Microbial RF model alone0.56 (0.48–0.64)0.78 (0.70–0.84)0.31 (0.22–0.41)Microbial RF model + FIT0.61 (0.54–0.69)0.82 (0.75–0.89)0.29 (0.20–0.38)Microbial RF model + FIT, sex, age, and WCRF score0.64 (0.56–0.72)0.82 (0.74–0.89)0.38 (0.27–0.49)FIT, sex, age, and WCRF score0.61 (0.53–0.69)0.77 (0.69–0.85)0.31 (0.21–0.41)aMetric uncertainty given as standard deviation across 20 times repeated 5-fold cross-validation for the training set, and as 95% confidence intervals for 1000-fold bootstrap for the test set. Training and test set prediction accuracy for neoplastic findings aMetric uncertainty given as standard deviation across 20 times repeated 5-fold cross-validation for the training set, and as 95% confidence intervals for 1000-fold bootstrap for the test set. When including participant age, sex, and the WCRF score as predictors, the accuracy for prediction of neoplastic findings improved, with a mean AUROC in the training set of 0.639 (SD = 0.037) and 0.64 (95% CI = 0.56–0.72) in the test set. Notably, omitting the microbial score from the predictors resulted in a decline in prediction accuracy (training AUROC = 0.607; SD = 0.037, test AUROC = 0.61, 95% CI = 0.53–0.69). Microbial predictions contributed to ensemble model accuracy also in this case (median LRT p-value < 0.001). At a threshold where 90% of CRC cases were correctly classified, microbial models exhibited higher sensitivity for the detection of precancerous lesions than the FIT value alone, although use of the FIT value alone had higher specificity (Fig. 3a, right panel; Table 2). Prediction of colorectal neoplasia based on gut microbial species abundance. a–e Random forest models based on microbial species abundance were constructed and combined with either the FIT value or the FIT value, sex, age, and WCRF score using logistic regression. Results are also presented for the FIT value alone as a predictor, or for a model including the FIT value, participant age, sex, and WCRF score, with colors indicating models. a Classification results from the training set are summarized as the mean (±SD) of 20-times repeated 5-fold cross-validation (CV) of AUROC, sensitivity, and specificity, in panels from left to right. Sensitivity and specificity were defined as the probability score resulting from setting the threshold for positivity to include 90% of CRC cases. b Receiver operating characteristic curves for model predictions in the test set (20% of the dataset; n = 206). c Model performance in the training set for participant subsets (as in a). The number of participants in the training set for each subset is indicated. d Model performance for classification of colorectal neoplasia by colonoscopy-based diagnosis relative to no neoplastic findings. Sensitivity was measured using the same cut-off as in (a), and specificity is omitted, as it is identical to that reported in (a). e Predictive performance for colorectal neoplasia (as in a) stratified by lesion localization. Only advanced lesions were treated as cases, and the comparison was made only for those without a corresponding lesion in the opposite compartment (Those with advanced distal lesions were excluded from the comparison of proximal lesions to those with no neoplastic findings (nnf), and vice versa). aa advanced adenoma, as advanced serrated, crc colorectal cancer, naa non-advanced adenoma, nnf no neoplastic findings. Classification accuracy using microbiome-based models was relatively stable across subsets of participant characteristics (Fig. 3c, Supplementary Fig. S7a, and Supplementary Data 9). Overall, there was a slight improvement in prediction accuracy among participants without recent use of antibiotics, who belonged to region 1 (Moss), were female, and in the lower ranges of FIT values. The largest impact on prediction accuracy was associated with microbial community state, where among those with enrichment of P. copri, the microbial model had an AUROC of 0.72 (SD = 0.19). Combining the microbiome-based predictor with either the FIT value or with the FIT value, sex, age, and WCRF score, consistently improved prediction accuracy for colorectal neoplasia across participant stratifications except for those with drug use indicating diabetes, or those who had recently used antibiotics (Fig. 3c and Supplementary Fig. S7a). While no comparable external validation dataset was available, we performed a geography-based hold-out evaluation (Supplementary Data 12), showing test set performance to be similar regardless of which region was used for training. For classification of neoplastic findings, feature selection and subsampling of the training set showed modest improvements in predictive potential associated with inclusion of more features and increased sample size (Supplementary Fig. S7b). We investigated whether microbiome-informed models could improve prediction of specific types of premalignant lesions (Fig. 3d). For CRC, the FIT value had slightly higher predictive power (mean AUROC = 0.742) than the combined microbial/FIT model (mean AUROC = 0.715). Notably, both the microbial model alone, and in combination with the FIT value were better able to detect precancerous neoplastic findings than the FIT alone at varying thresholds for positivity (Supplementary Fig. S8). Evaluation of lesion detection by localization showed increased accuracy associated with addition of the microbial model for both distally and proximally localized lesions (Fig. 3e). Evaluation of neoplasia-associated microbes To identify individual microbes associated with colorectal neoplasia, we carried out multivariate adjusted differential abundance analyses (see “Methods” section) including species present in at least 10% of participants (n = 606). Here, 13 species were differentially abundant between those with and without colorectal neoplasia (Fig. 4a, left panel; Supplementary Data 10), and 20 species were differentially abundant when contrasting any of the diagnostic groups to those without neoplastic findings (Fig. 4a, right panel), for a total of 24 unique species. Of these, 19 species were also associated with at least nominal significance when employing a centered log-ratio approach to differential abundance (Supplementary Fig. S9). In unadjusted models, only 9 species were associated with neoplastic findings, and 10 with any specific outcomes, indicating that statistical adjustment was necessary to uncover significant associations with outcome groups. For KOs, 25 and 11 were significantly associated with colorectal neoplasia or any diagnostic group, respectively (Supplementary Data 10). Differential abundance of species between diagnostic groups. a Effect sizes of differential abundance of microbial species for those with relative to those with no neoplastic findings (left panel) or for each diagnostic group relative to those with no neoplastic findings (right panel). Only species differentially abundant with regards to colorectal neoplasia, or for CRC, advanced adenomas, or advanced serrated lesions relative to no neoplastic findings are included in the plot (a full list of differentially abundant species is available in Supplementary Data 10). Differential abundance models included adjustment for age, sex, geographic region, sequencing reads, smoking status, level of education, the WCRF score, hemorrhoids, IBD and recent use of antibiotics. Circles and lines indicate log2 fold change ± standard error. Opaque symbols indicate significant associations at FDR < 0.05, and semi-transparent ones indicate an FDR value > 0.05. For colorectal neoplasia, log2 fold changes derived from unadjusted models are shown as red points. b Prevalence (left panel) and mean relative abundance (right panel) of differentially abundant species as in (a). c Species with significant associations with diagnostic groups were evaluated for associations with participant characteristics in separate models (adjusted for participant age, sex, and region). Colors show coefficient magnitude and direction, with significant associations indicated by asterisks. Participant characteristics measured on a continuous scale were split in tertiles (low, mid and high), where low was considered the reference value. Reference values for categorical variables were as follows: antacid/antibiotic use – No; smoking status – non-smoker; region – Region 2; and employment status – employed. Only variables with any significant associations are displayed; a full table of results is available in Supplementary Data 13. Comparing species abundance between pairs of diagnostic groups indicated incremental changes in microbial communities from those with no neoplastic findings through stages of CRC precursors, to CRC, with the clearest contrast between those with no neoplastic findings and those with CRC (Supplementary Fig. 10). Species with differential abundance did not display any consistent pattern in overall prevalence or abundance (Fig. 4b), and those species contributing to community states (Supplementary Fig. S5a) did not show abundance differences corresponding to the associations observed for the community states overall. Differentially abundant species were, however, frequently associated with host characteristics (Fig. 4c; Supplementary Data 13). Microbes positively associated with colorectal neoplasia also tended to be positively associated with beneficial factors such as the level of physical activity, whereas microbes associated with no neoplasia were typically associated with factors indicating an unhealthy lifestyle. E. coli abundance was associated with having no neoplastic findings (Fig. 4a), and the presence of E. coli (n = 568; 55%) was associated with a 43% lower odds of colorectal neoplasia (OR = 0.57 [95% CI = 0.43–0.75]; Fig. 5a). As previous research has shown pks-positive E. coli to be associated with CRC36, we stratified samples with E. coli by the presence of the pks gene cluster, finding 167 samples to be pks positive (>10% coverage of the colibactin gene cluster by read mapping), and 401 to be pks negative. Interestingly, pks-negative E. coli was associated with a lower rate of colorectal neoplasia (OR = 0.53 [0.39–0.71]) and a contrasting trend toward a higher rate of colorectal neoplasia and of CRC in particular among those with pks-positive E. coli (OR = 1.33 [0.89–1.98] and OR = 1.99 [0.94–4.20], respectively). CRC-related depletion of non-pks-encoding E. coli and evaluation of species previously associated with CRC. a Odds ratios (with 95% CI) for detection of neoplastic lesions associated with the detection of E. coli overall (left panel), or stratified by pks status (right panel). b Effect sizes of differential abundance relative to no neoplasia for species identified as being associated with the presence or absence of CRC (n = 122) across populations in a large case-control study as observed in the current study. Those with significant differences (p < 0.05; no correction for multiple testing) are shown here, where species reported to be more abundant in controls are shown in the top panel, and those reported to be more abundant in CRC are shown in the bottom panel. Six species were differentially abundant in CRC cases. Four uncharacterized bacteria (one belonging to the genus Clostridia, one to the family Oscillospiraceae, and two within the phyla Bacillota and Tenericutes) were positively associated with CRC, whereas two Streptococcus species, S. salivarius and S. parasanguinis, were less abundant in CRC cases. Levels of both S. salivarius and S. parasanguinis were higher in individuals with recent use of proton pump inhibitors (PPIs) or other anti-acid medications (Fig. 4c), as has been previously reported37. To assess whether the association of these species with CRC was driven by drug use, we conducted a sensitivity analysis stratifying participants by recent antacid/PPI use, showing significant associations also for those without recent use (Supplementary Fig. S11). Similarly, differential abundance results in general were consistent for those without recent antibiotic use, and both with and without non-neoplastic findings, although associations were diminished among recent antibiotic users. Evaluation of 122 species recently identified as associated with CRC18 showed that 32% (n = 38) of these had low prevalence (<10%) in our dataset (Supplementary Data 14). Differential abundance analysis, including species with at least 2 observations in our dataset, resulted in 98/118 showing no evidence of differences in abundance between those with no neoplastic findings and those with CRC. Although nominal significance was observed for 20 species (Fig. 5b), 11 were enriched in the opposite diagnostic group (no neoplastic findings rather than CRC (n = 5), or vice versa (n = 6)). Species reproducibly enriched in CRC (p < 0.05) included Peptostreptococcus stomatis, Dialister pneumosintes, Gemella morbillorum, Porphyromonas somerae, Alistipes ihumii, Fusobacterium nucleatum sensu stricto (SGB6011), Fusobacterium nucleatum polymorphum (SGB6001), Porphyromonas uesonis and the uncharacterized bacterium GGB9342 SGB14306 (belonging to phylum Bacillota), whereas the species Enterocloster spp. aldensis and boltae, Eisenbergiella tayi, Ruthenibacterium lactatiformans, and Clostridium symbiosum, which were previously reported to be associated with CRC, were here found to be enriched in those with no neoplastic findings. Site-specific associations of colorectal lesions and the microbiome Using detailed information about lesion localization available from endoscopist records, we evaluated differences in the microbiome relative to the localization of advanced lesions in the gut. CRC and advanced adenomas were more commonly detected in the recto-sigmoid regions and the cecum, with a lower degree of site-specificity for serrated lesions and non-advanced adenomas (Fig. 6a, Supplementary Data 15). Site-specific association between the microbiome and colorectal lesions. a Localization of lesions detected at colonoscopy by diagnostic group. Lesions were assessed by a pathologist as part of clinical routine, and if more than one lesion was present in a patient, the most severe diagnosis was used (see “Methods” section). Note that, due to discontinuation of the colonoscopy procedure when CRC was detected, no subsequent lesions (that is, more proximal ones) were registered. b Number of observed species according to whether advanced colorectal lesions (advanced serrated lesions, advanced adenoma, or CRC) were detected at colonoscopy. Lesions located on the right side of the transverse colon up to the splenic flexure were considered “proximal”, and those distal to this were considered “distal”. c Species alpha diversity (as in b) by distance from the anal verge to the most severe lesion. The distance was measured using the insertion depth of the colonoscope. Trend lines are based on linear regression with 95% confidence intervals indicated in shaded areas. d Species alpha diversity and colorectal lesion burden (the sum of lesion diameters for each colonic region). e PCoA based on Bray–Curtis distances, categorized by findings of advanced colorectal lesions by site (as in b). Solid points indicate centroids, and ellipses encompass 50% of observations. f Differentially abundant species according to lesion localization and lesion burden according to five alternative models. In panel one, advanced lesions were coded as present or absent, using one variable for distal lesions (blue) and one for proximal lesions (green); in panels two and three, the distance to the most severe lesion was included. In models reported in panels four and five, the localization-specific burden of lesions was included as an independent variable, using separate models for distal and proximal lesions. Filled markers represent significant associations, with marker size denoting nominally (small) and FDR-corrected (large) significant results. Overall, individuals with distal lesions had higher alpha diversity (observed species) than those without (Fig. 6b; linear model: p = 8.6 × 10−4, t(1018) = 3.34, mean difference = 16.4 [95% CI: 7.2,25.5]), which increased the closer lesions were located to the anal verge (Fig. 6c; p = 0.04, t(286) = −2.05, observed species by distance from anal verge (cm) = −0.59 [95% CI: −1.16,−0.02]). There was also a positive association between distal lesion burden and observed species (Fig. 6d; p = 0.04, t(428) = 2.08, observed species per sum of lesion diameters (mm) = 0.38 [95% CI: 0.02,0.74]), strengthening the inference of a localization-dependent association between colorectal lesions and microbial diversity. Regarding proximal lesions, we observed no association between diversity and lesion presence, nor with their distance to the anal verge or lesion burden (Fig. 6b–d). Beta diversity estimates mirrored the alpha diversity findings, with significant differences between distal, but not proximal lesions, when compared to no lesions (Fig. 6e; PERMANOVA p = 0.001, R2 = 0.002 for distal; p = 0.10, R2 = 0.001 for proximal). Further, 26 species were associated with the presence of distal lesions, whereas only 2 were associated with proximal lesions (Fig. 6f; left panel). Species associated with distal lesions included several of those also associated with neoplastic findings overall (5/13; Supplementary Fig. S12), while this was not the case for proximal lesions (0/13). Interestingly, four out of 12 Faecalibacterium species (SGBs) were positively associated with the presence of distal lesions and negatively associated (nominally significant) with the distance of lesions to the anal verge. Two of the F. prausnitzii SGBs were also associated with the burden of distal lesions (the sum of lesion diameters; nominal significance). Still, no KOs encoded by members of the Faecalibacterium genus were differentially abundant in individuals with distal lesions (Supplementary Data 16). Discussion Despite extensive efforts to characterize the CRC-associated metagenome, few studies have examined screening-relevant populations. In this CRC screening cohort, where inclusion was based on national registry linkage, we employed clinically relevant diagnostic categories and integrated microbiome findings with FIT test values. Our results show that the microbiome of those with colorectal neoplasia differs from that of those without and contributes to a modest increase in predictive capacity for premalignant lesions compared to the FIT value alone. Previous studies have achieved high levels of accuracy for the identification of both CRC15,16,29 and adenomas29,38,39 using microbiome profiling. However, these have often included symptomatic cases or lacked well-described participant identification procedures, which, as shown in a recent systematic review, exposes them to a high risk of bias40. Evaluating microbiome-based biomarkers for CRC screening requires recruitment strategies that are transparent and minimize bias, but also predictive modeling supporting the dual aims of CRC screening, of both early CRC detection and its prevention through detection of premalignant lesions. Our results show that gut microbiome-based predictions made in a relevant screening setting can improve identification of those who are likely to benefit from colonoscopy, especially for those with premalignant lesions. We addressed the problem of distinguishing neoplasia from other causes of bleeding in FIT-positive participants, most of whom had tested FIT-negative in previous rounds and were asymptomatic, making accurate classification particularly difficult. Additionally, most neoplasms were non-malignant, which has been shown to be harder to identify through analysis of the gut microbiome41,42. The utility of new diagnostic tests depends on how they compare to existing procedures. Participants recruited in the CRCbiome study were FIT-positive at a threshold of 15 μg hemoglobin per gram of feces – a low threshold compared to other screening programs7,43, allowing us to compare how the microbiome can contribute to improved prioritization of colonoscopy referrals relative to increases in the FIT value threshold. Here, we show how adding predictions from a microbiome-based model to FIT results could improve classification for premalignant lesions beyond what was achievable by using either the FIT alone or the FIT together with lifestyle and demographic factors. While higher accuracy for CRC and more efficient tools for measuring microbial signatures would be needed for the implementation of a microbial marker as a screening tool, our results suggest that the microbiome may be a source of biomarkers, with the potential to improve the prediction of colonoscopy outcomes in FIT-based screening programs. In a screening setting, individuals with occult blood in their stools, but without neoplasia, are more likely to have other gastrointestinal conditions or have unhealthy habits associated with an increased rate of intestinal bleeding44,45. Several species negatively associated with neoplasia in our study have previously been found to be more abundant in CRC, including E. coli, H. hathewayi, E. lenta, C. symbiosum, E. boltae, and F. plautii either in general18,46–48, or as site-49, stage-18, or age-specific associations50. We also found a positive association between microbes that are commonly considered beneficial and the presence of distal lesions, including several Faecalibacterium species51,52. Our contradictory findings suggest that enrichment of H. hathewayi, E. lenta, and C. symbiosum might be caused by bowel conditions other than CRC development; rather, functioning as indicators of an unhealthy gastrointestinal community in general. Indeed, several of these species have been reported by others as being associated with poor gut health53. This could also explain the somewhat unexpected finding that our neoplasia-associated microbes were associated with beneficial lifestyle and demographic features – and vice versa- and with our observation of a lower microbial diversity – which has typically been associated with poorer gut health – among those without colorectal neoplasia. We identified E. coli as being negatively associated with the presence of colorectal neoplasia, and that this association was restricted to pks-negative E. coli. In contrast, there was a trend toward enrichment of pks-positive E. coli in individuals with CRC or any colorectal neoplasia. The genotoxic effect of colibactin produced by pks-positive E. coli has been demonstrated11, with mutations caused by colibactin likely being an early event in carcinogenesis54,55. While a recent large study of Dutch screening participants cast doubt on the utility of pks-positivity alone as a marker in CRC screening56, our results suggest that FIT-positivity should also be considered when evaluating the role of pks-positivity for E. coli in CRC screening. Supporting such a conclusion, Chénard et al., while not stratifying by pks status, showed that E. coli was overrepresented among neoplasia-negative screening participants with fecal occult blood29, although such enrichment was not reported in another study57. The enrichment of oral bacteria in the gut microbiome of individuals with CRC has been shown in multiple reports15,18,39,58. We reproduce several of the most consistently reported associations in this screening population, including enrichment of F. nucleatum and P. stomatis. Although the prevalence of these species was limited, our results confirm these species to be specifically enriched in CRC. We also found a depletion of some oral bacteria in those with CRC, including several Streptococcus species. The Streptococcus genus comprises both pathogenic, commensal, and probiotic species and strains. The association of Streptococcus species with CRC is also mixed: while bacteremia with S. gallolyticus (previously known as S. bovis) has a long-established link to the presence of CRC59, other members of the genus, including S. salivarius, have been found to be depleted in CRC60. Interestingly, S. salivarius, which was among the CRC-depleted streptococci in our study, has been shown to inhibit the growth of F. nucleatum through secretion of nisins61, with further evidence for a protective effect for other Streptococcus species62. While the abundance of several Streptococcus species has previously been associated with PPI use37,63–65, we found that the depletion of streptococci in CRC was independent of recent anti-acid medication use. Classification of gut microbiomes as belonging to discrete community types, typically defined by enrichment or dominance of specific taxa, has been linked to host health66. Here, microbiome profiles were clustered into four community types, dominated by, e.g., P. copri and Bacteroides species, and characterized by differences in alpha and beta diversity, consistent with other reports28,34,65. We found community states to have different associations with CRC-related outcome groups that were not attributable to single taxa, where a Bacteroides spp. enriched state was strongly associated with a lack of neoplasia. Moreover, prediction accuracy for colorectal neoplasia differed between the microbial community types, suggesting that such community types could serve as a basis for tailored screening. Still, larger studies are needed to validate such an approach. Our results showed microbial diversity to be elevated among those with advanced distal lesions. This is in line with results from a recent study also identifying elevated diversity in distal CRC compared to proximal cases and healthy controls18. Additionally, we found an increase in diversity associated with both distal lesion proximity to the anal verge and distal lesion load, providing further support to the conclusion that distal lesions are associated with increased diversity. Similarly, we found most differentially abundant species to align with the presence or absence of distal lesions, but not of proximal ones. These findings could indicate that lesion-associated microbial alterations are more readily measured distally, likely due to dilution or degradation of signals during colonic transit, or alternatively, reflecting cellular differences between distal and proximal lesions. Recent reports have highlighted the importance of confounding in CRC-microbe associations28. We observed several cases where associations come to light only after making statistical adjustments. As such, depletion of F. plautii in advanced serrated lesions became evident only after statistical adjustments for lifestyle and demographic factors. This study has notable strengths. We employed standardized participant recruitment strategies, sample collection procedures, and laboratory protocols to ensure uniform capture of microbiome signatures representative of a colorectal screening population. Study participants were deeply characterized using validated questionnaire data and complete registry-based prescription drug use histories, the latter indicating that participants are largely comparable to CRC screening participants in general when it comes to comorbidities. Outcome measures were based on detailed clinicopathological records from centralized screening centers. There are also several important limitations. Although metagenomics of FIT samples yields microbial profiles comparable to those recovered from stool samples67–69, using surplus FIT buffer likely limited the sensitivity for very low-abundance species. Second, while the use of a single type of sampling kit ensured uniform processing, it may constrain the generalizability of the findings to settings that employ other kits. Further, the study design exclusively included FIT-positive participants, limiting the generalizability of our findings to FIT-negative populations. Lastly, basing recruitment on consecutive screening cases resulted in few CRC cases, likely causing CRC-directed models to be underpowered. Even modest improvements in the accuracy of screening models can save lives and reduce the burden on healthcare systems. Especially in settings where colonoscopy resources are limited, a more accurate pre-screening tool may improve CRC screening by making it more efficient. By combining microbiome-based models with the FIT test and lifestyle exposure, we could distinguish those with a need for colonoscopy due to the presence of premalignant lesions or CRC more accurately than FIT alone. These results indicate that the microbiome may be used to improve screening-based early detection – which is precisely where the potential for improvement is greatest. Still, refinement of markers and external validation are warranted. In conclusion, we show that the gut microbiome can contribute to more accurate classification of colorectal neoplasia in a CRC screening population than FIT alone, highlighting the potential of microbial biomarkers to improve early detection when combined with FIT. Our findings reveal specific microbial signals associated with CRC screening outcomes. We reproduce known CRC-associated enrichments of bacteria such as F. nucleatum and P. stomatis and show that pks-negative E. coli is enriched in participants without neoplastic findings. In contrast, bacteria such as F. prausnitzii, H. hathewayi, E. lenta, and C. symbiosum appear to be associated with other conditions causing FIT-positivity rather than with CRC development. Both overall microbial community structure and specific taxa were associated with the presence of distal lesions, while associations with proximal lesions were limited. Methods The CRCbiome project complies with all relevant ethical regulations, and all participants provided informed consent. The CRCbiome project has been approved by the Regional Committees for Medical Research Ethics South East Norway (REK approval no. 63148). Study design and sample collection Participants in the CRCbiome study were recruited from FIT-positive participants in the Bowel Cancer Screening in Norway (BCSN) study, a randomized controlled trial (ClinicalTrials.gov: NCT01538550) comparing the effectiveness of once-only sigmoidoscopy and biennial FITs, testing for fecal occult blood31. In the FIT arm of BCSN, men and women in the age range 50–74 years in 2012 (born between Jan 1, 1938, and Dec 31, 1962), residing in one of two closely situated, but geographically distinct regions in South-East Norway (Moss (region 1), or Bærum (region 2)), were invited to self-collect stool samples using provided sampling kits suitable for FIT analysis (OC-Sensor, Eiken Chemical, Japan), with no dietary or medication restrictions prior to sampling. Participants were instructed to ship stool samples by mail using a prepaid envelope on the day of sampling to a central laboratory for FIT testing (estimated 3–10 days shipping time). If same-day shipment was not feasible, they were to store the sample in the refrigerator until shipment. Hemoglobin concentration (FIT value) was measured using the OC-Sensor Diana system (Eiken Chemical, Tokyo, Japan), and the sampling tube with remaining buffer (containing about 10 mg feces) was stored at −80 °C. Those with a positive FIT result (>15 µg hemoglobin/g feces, i.e., fecal occult blood detected) were scheduled for a follow-up colonoscopy, and further FIT screening rounds were discontinued. BCSN participants who tested FIT-positive between October 2017 and March 2021 were invited to take part in CRCbiome. Invitation letters were sent two to 14 days after FIT test results, with invitations to follow-up colonoscopy. CRCbiome recruited men and women partaking in their second, third, or fourth scheduled round of FIT screening as part of the BCSN trial. Out of 2700 individuals invited to CRCbiome, 1640 provided informed consent. FIT-positive BCSN participants consenting to participation in CRCbiome were excluded based on the following criteria (see sections below for further details): (1) did not undergo colonoscopy, (2) no histological examination of lesions, (3) diagnosis of 1–2 non-advanced adenomas or non-advanced serrated lesions, and (4) insufficient sequencing data (<1 Gb). The final study population included 1034 individuals. Further details on recruitment procedures and sample size considerations can be found in Kværner et al.32. Investigators were not blinded to participant outcomes during analyses. Clinical examination, diagnosis, and outcome definitions After a FIT-positive screening result, participants were scheduled for colonoscopy at a screening center (one for each region), preceded by a bowel cleansing procedure. During colonoscopy, localized lesions (mostly polypous lesions) were removed and submitted for histopathological examination. Participants with no findings during colonoscopy were categorized as “negative” based on the endoscopist’s assessment. All other lesions were classified by a pathologist according to standard clinical routine and resulted in categories based on their most severe finding, ranging from “non-neoplastic lesions”, “other findings”, “non-advanced serrated polyps”, “1–2 non-advanced adenomas”, “3 or more non-advanced adenomas”, “advanced serrated lesion”, “advanced adenoma” and “cancer”. Those with non-neoplastic findings included endoscopy-detected presence of hemorrhoids, IBD, diverticulitis, or other conditions not related to CRC, and those with other findings included inflammatory polyps and granulated tissue polyps. We considered those with non-advanced serrated lesions and 1 or 2 non-advanced adenomas to be of uncertain clinical significance, since while potential premalignant lesions were removed, these findings did not indicate further colonoscopy surveillance70. This group was therefore not included in the metagenome analyses seeking to evaluate the gut microbiome as a source of biomarkers for CRC screening. CRC was defined as adenocarcinoma of the colon or rectum (ICD10 codes C18-20), and advanced adenomas were defined as adenomas ≥10 mm, or those with ≥25% villous histology, or with high-grade dysplasia. Advanced serrated lesions included any serrated lesions (hyperplastic polyp, sessile serrated lesion, or traditional serrated adenoma) with a size of ≥10 mm or dysplasia. Lesion localization for all lesions detected during colonoscopy was recorded by the endoscopist and categorized according to colonic region as proximal when it was located proximally to, or in the splenic flexure, or as distal. Due to discontinuation of colonoscopy upon detection of CRC, subsequent lesions (more proximal ones) were commonly not registered. Lesion location was also registered as its distance from the anal verge, measured by insertion depth of the colonoscope. For each participant, the distance from the anal verge to the most severe lesion was recorded. As a measure of lesion burden, we calculated the sum of lesion diameters. The latter two measures were calculated separately for the proximal and distal regions of the colon. DNA extraction, library generation, and metagenome sequencing FIT cartridges were retrieved and thawed, and leftover buffer was used as input material for DNA extraction and shotgun metagenome sequencing library preparation, as previously described32. Briefly, 500 µl aliquots (approximately 30% of remaining buffer) were used for DNA extraction using the QIAsymphony DSP Virus/Pathogen Midikit (Qiagen), with an offboard lysis protocol based on bead-beating. Extracted DNA was eluted in 60 µl AVE buffer (Qiagen). For samples with a DNA concentration of less than 1.5 ng/µl, DNA extraction was repeated using a second aliquot. Screening samples with a DNA concentration of 0.7 ng/µl or more were considered eligible for metagenome library preparation. Sequencing libraries were generated according to the Nextera DNA Flex Library Prep Reference Guide, with the exception of scaling down the reaction volumes to ¼ of the reference. Library pools of 240 samples were combined and size-selected to a fragment size of 650–900 bp. Sequencing was performed using the Illumina NovaSeq system using S4 flow cells with a lane divider, with each pool sequenced on a single lane, resulting in paired-end 2 × 151 bp reads. Sequencing output has been reported previously71. Six negative controls were included for DNA extraction and two for library preparation. The DNA extraction controls resulted in DNA concentrations below the level of detection and generated 32 QC sequencing reads in total, whereas the library preparation controls produced only 3 reads. Further evaluation of potential contamination was done by comparing taxonomic profiles for samples with low DNA concentration (<1.5 ng/µl, n = 37; Supplementary Fig. S13). Given the negligible read counts observed in negative controls and the lack of evidence for species enrichment in low-input samples, no additional contaminant filtering steps were deemed necessary. Taxonomic classification, determination of functional potential, and establishment of metagenome-assembled genomes (MAGs) Sequencing reads were processed for removal of adapters, low-quality bases, and reads mapping to the human genome using kneaddata (v0.12) with standard settings. Read-based taxonomy and gene content were assessed using MetaPhlAn4 (v4.0.6) and HumanN3 (v3.7)72,73, respectively, with the mpa_vOct22_ChocoPhlAn_202212 pangenome database, using the UniRef90 database to assign gene families to KEGG orthogroups (KOs). Read-based taxonomic abundance was evaluated at the level of species-level genome bins (SGBs; referred to as “species” in the text) using relative abundance determined by MetaPhlAn4. Gene content abundance derived from HumanN3 analysis was scaled by the number of quality-controlled reads per million. Separately, the metagenome processing framework Metagenome-ATLAS74 (v2.4.3) was used for sequencing QC, assembly, binning, generation of representative genomes (MAGs), and cataloging of gene content. In brief, low-quality bases and reads were filtered, and human and phiX sequences were removed using BBMap (v37.99)75. Remaining reads passing QC were assembled using MetaSpades76 (v3.13) with default settings. Assemblies were grouped into genomes using DAStool77 (v1.1), based on genomic bins identified by MetaBat78 (v2.2) and MaxBin79 (v.2.14). Representative genomes were defined by dereplication using dRep80 (v2.2) with a 95% sequence identity for 60% genome overlap. Only genomes with a quality score Q ≥ 50 were retained, where the quality score was defined based on CheckM metrics as Q = completeness-5×contamination. Taxonomy was assigned using the GTDB-Tk81 (v1.3), with the GTDB database (v95). The abundance of MAGs was determined by calculating the median depth of reads mapping to 1000-bp bins of each genome and was scaled by the number of reads per million. For evaluation of consistency between read-based (MetaPhlAn4) and assembly-based taxonomic characterization, we matched species names as assigned using GTDB annotations for both datasets. Species uniquely identified in both datasets were compared across samples by prevalence in each dataset and by the cosine similarity measure of their abundance across samples. A publicly available dataset of gut metagenomes34 derived from conventionally collected stool samples was retrieved and subjected to characterization of species abundance using MetaPhlAn4 using the procedure described for CRCbiome. Comparisons were made of mean relative abundance for each species and species-based beta diversity using a principal coordinates analysis (PCoA) visualization. Evaluation of microbial diversity and community states Microbial alpha diversity was measured as the number of observed species (or KOs in the case of gene data), and by the Shannon and inverse Simpson diversity indices based on the relative abundance of species or KOs and used in analyses as continuous variables. Species- and KO-based beta diversity was calculated using the Bray–Curtis dissimilarity measure. Microbial community states were defined using Dirichlet Multinomial Mixture35 (DMMs) models as implemented in the R package ‘DirichletMultinomial‘, based on the relative abundance of species with a minimum abundance of 0.1% in 10% of individuals. The number of community states (components; k) was determined by selecting the number of components from 1 to 8, resulting in the lowest Laplace approximation score. Species contribution to the community state definition was defined as the difference in contribution between a single-component model and each community of a four-component model. Naming of the community states was done manually, denoting the taxa, or combination of taxa, that contributed most strongly relative to other states. Detection of pks-positive E. coli Presence of the pks gene cluster in E. coli genomes was determined by mapping reads to the pks gene cluster (GenBank AM229678.1). Here, any individual with E. coli detected using MetaPhlAn4 and with reads covering at least 10% of the pks module was classified as pks positive, unless read mapping showed the colibactin gene cluster to be incomplete. Questionnaire data Data on regular diet, lifestyle factors, and demography were collected using a lifestyle and demography questionnaire (LDQ) and a validated food frequency questionnaire (FFQ) designed to register the habitual diet over the last year. Details on the two questionnaires have been described previously32. For the FFQ, those with low-quality questionnaires (two or more missing pages, too many missing food items and/or portion sizes, or inconsistent reporting; n = 21), too low energy intake (<600 and <800 kcal/day for women and men, respectively; n = 11), or too high energy intake (>3500 and >4200 kcal/day for women and men, respectively; n = 46) were excluded. No exclusion criteria were applied for the LDQ, and missing data was handled on a question-by-question basis. The questionnaires provided information on factors with strong evidence for an increased risk (consumption of red and processed meat and alcohol, body fatness, adult attained height, and smoking) or a decreased risk of CRC (intake of dairy products, calcium, wholegrains, and fiber, and level of physical activity). It further provided information on the consumption of main food groups (fruit and berries, vegetables, fish, and sugar-sweetened beverages), as well as total energy intake, being positively correlated to many of the included foods and nutrients82. Variables derived from questionnaires that were used in statistical models (see below) included smoking status, level of education, recent antibiotic use, and a composite score defining adherence to cancer prevention recommendations. Smoking status was categorized as smoker or non-smoker, where smokers included current smokers (any level), or those who reported having ceased smoking within 10 years, or who had smoked for 30 years. Non-smokers included never-smokers and current non-smokers who ceased smoking at least 10 years prior to study inclusion and reported having smoked for less than 30 years. Education level was recorded as the highest level of completed education (primary school, high school, or college/university). Recent antibiotic use was defined using self-reported use and registry data (see below). Based on the questionnaire data, we composed a seven-point score measuring adherence to the recommendations in the 2018 World Cancer Research Fund/American Institute for Cancer Research (WCRF/AICR) Cancer Prevention Report(referred to as the WCRF score, operationalization described in a separate paper21). The score measured adherence to the following recommendations: (1) be a healthy weight, (2) be physically active, (3) consume a diet rich in wholegrains, vegetables, fruits, and beans, (4) limit consumption of “fast foods” or other processed foods high in fats, starches, and sugars, (5) limit consumption of red and processed meat, (6) limit consumption of sugar-sweetened drinks, and (7) limit alcohol consumption. Adherence to each recommendation was assigned a score of 0, 0.5, or 1, with increasing degree of adherence; the WCRF score was defined as the sum of these scores. The 2018 WCRF Score has been linked to colorectal carcinogenesis in several longitudinal studies83–86, including the current study population21. A complete list of the diet, lifestyle, and demography variables employed in the study can be found in Supplementary Data 17. Self-reported medication use and prescription drug dispensation The LDQ included questions about the use of antibiotics or anti-acid drugs (antacids or proton pump inhibitors (PPIs)) within 3 months of answering the questionnaire. For each participant, we linked records of dispensed prescription drugs provided by the Norwegian Prescribed Drug Database (NorPD), which maintains records of dispensed prescription drugs in Norway, classified according to the Anatomical Therapeutic Chemical (ATC) system87. Drugs that are purchased without prescriptions or are supplied to hospitals or nursing homes are not included. For antibiotics and anti-acid drug use, recent use was recorded if participants had self-reported use, or registry data showed dispensation of drugs of the same classes (ATC codes J01, A07AA, or P01AB for antibiotics, and A02A or A02B for antacids/PPIs) within 4 months. Data from NorPD was also used for the evaluation of the study population representativeness. Here, data included NorPD records for CRCbiome participants, as well as for two comparison populations: (1) the remaining participants of the FIT arm of the BCSN trial, and (2) a nationally representative population: a random draw of 16,500 Norwegians eligible for bowel screening in 2017 (born in 1946–1962). We obtained information on the dispensation of prescription drugs, indicating common comorbidities, or that were of relevance to microbiome analyses. These included drugs for diabetes (≥2 dispensations of any drug with ATC code A10); cardiovascular disease (≥1 of any with ATC codes C, and/or B01); chronic obstructive pulmonary disease (≥1 of any with ATC codes R03AC, R03AK, R03AL, R03BB, R03DA, R03DC, R03DX, R03BA); and antibiotics (≥1 of any with ATC codes J01, A07AA, or P01AB). Prescription drug dispensation was assessed within 12 and 24 months prior to participant inclusion. For BCSN participants, inclusion was defined as the date of invitation to the first round of biennial screening. For the population controls, a fictitious inclusion date was set as a random date in 2017. For CRCbiome participants, two definitions of inclusion were used: (1) the invitation date for the first round of biennial screening (i.e., their inclusion in BCSN), and (2) the invitation date for the FIT screening round in which they tested positive and were invited to CRCbiome. Statistical analysis For descriptive statistics, categorical variables were summarized as counts and row-wise percentages, and continuous variables as medians with interquartile range (Q1, Q3). Associations between microbial variables and participant characteristics were evaluated in models adjusted for age, sex, and region, and continuous variables were encoded as tertiles (low, mid, and high), where the lower level was treated as reference. Results of associations for the groups “unknown”, “missing”, or those with fewer than 10 observations were included in statistical analyses but not reported separately. Evaluation of associations between microbial variables and outcomes (any colorectal neoplasia, diagnostic groups, or localization-dependent measures of colorectal lesions) was modeled with adjustments for age, sex, region, the WCRF score, smoking status, level of education, findings of hemorrhoids, findings of IBD, and recent antibiotic use unless noted otherwise. Participant age and the WCRF score were included as continuous variables, whereas sex, region, smoking status, level of education, findings of hemorrhoids and IBD, and recent antibiotic use were included as categorical variables. The selection of covariates was largely based on the latest Expert Reports from WCRF/AICR of 2018, summarizing the available evidence on diet, physical activity, body weight, and CRC risk19,88. Assessment of associations with site-specific lesions was performed in separate models for each measure, employing the presence, burden, or distance to the anal verge of lesions as independent variables. For the presence of advanced distal and proximal lesions, one dummy variable was included as an independent variable for distal lesions, and one for proximal lesions. Modeling of associations with the distance to colorectal lesions, or of the burden of colorectal lesions, only included participants with lesions, and for models of one colorectal region, those with lesions in the other region were excluded. Models including alpha diversity were also adjusted for sequencing depth. Differences in alpha diversity between groups of participants defined by host characteristics were assessed using linear regression models, with alpha diversity as the dependent variable. Differences in alpha diversity or other microbially defined groups (community states, pks-stratified E. coli) related to outcome categories were assessed using multinomial logistic regression models with outcome as the dependent variable and the standard adjustment variables. Differences in beta diversity between groups were evaluated using permutational analysis of variance (PERMANOVA) as implemented in the R package vegan89 using the “adonis2” function with 999 permutations. For evaluation of differences between levels of categorical variables, pairwise tests were performed. PERMANOVA tests were run using marginal testing, with adjustment for age, sex, and region. Machine learning/classification Machine learning for classification of participants by the presence of colorectal neoplasia was carried out using three categories of datasets: (1) species abundance estimated using either read-based taxonomic profiling or MAGs; (2) abundance of KEGG orthologous (KO) groups using read-based (HumanN3) quantification; or (3) a set of diet, lifestyle, and demographic factors capturing broad features of demography and participant lifestyle (Supplementary Data 17). Models were trained and evaluated using the tidymodels90 framework with a 20-fold repeated 5-fold cross-validation (CV) within an 80% training set, with final models additionally evaluated in a 20% test set. The test set was randomly selected, being matched with the training set by age, sex, and screening result. Algorithm selection, parameter tuning, dataset specifications, and participant stratifications were explored exclusively within the training dataset. Model hyperparameters were predefined and are available in Supplementary Data 18. Evaluations of model performance in assigning class probabilities were made by assessing the mean area under the receiver operating characteristic curve (AUROC) across CV partitions. Features with low prevalence (<70%) or with a low degree of variation (the most common value is 19 times more frequent than the second most common) were excluded from prediction models. For unbalanced datasets (one outcome category constituting more than 2/3 of the dataset), the adaptive synthetic algorithm91 was used (within the resampling loop to avoid data leakage) for upsampling the minority group. Classification using subsets of participants and a feature selection strategy based on selection of informative taxa (using the ‘ranger’ engine with ‘permutation’ mode of variable importance) was employed to evaluate the effect of sample size, number of taxa used for model training, and feature selection procedure. Ensemble models were set up using stacking with logistic regression models, including predictions made with microbial models and a set of demographic and clinical variables, including (1) log-transformed FIT values, and (2) log-transformed FIT values, sex, age, and the WCRF score. Additionally, logistic regression models were set up for the latter set of variables without the microbial predictions. Ensemble models were trained on datasets limited to those without neoplastic lesions or with advanced adenomas or CRC. Differential abundance analyses Differential abundance was assessed using Microbiome Multivariable Associations with Linear Models (MaAsLin2)92. To account for differences in KO abundance largely explained by taxonomy, KOs with high correlation to at least one species (Pearson’s product-moment correlation coefficient above 0.5) were excluded. Further, only features present in at least 10% of samples were retained for analyses. Analyses were performed using a linear model with log transformation and total sum scaling normalization. Differential abundance analyses evaluating associations with outcome categories were adjusted for age, sex, region, sequencing depth, smoking status, level of education, the WCRF score, hemorrhoids, IBD, and recent use of antibiotics. Sensitivity analyses were performed, stratifying by recent use of antibiotics (see definition above), recent use of anti-acid medication, or the detection of any non-neoplastic lesions during colonoscopy. Characterization of host-microbe associations was carried out using MaAsLin2 analyses contrasting levels of host characteristics (as described for alpha- and beta diversity), adjusting for age, sex, region, and sequencing depth, with results reported for species with significant outcome associations. To evaluate associations reported in the literature, we extracted a list of microbial species identified as being associated with CRC or controls across populations in a recent study18. We restricted the set of microbes to be evaluated to those that were significant in their models, adjusting for sex, age, and BMI (FDR < 0.05), which included 122 SGBs. For evaluation of these species, we carried out a differential abundance analysis using MaAsLin2, including those observed in at least two individuals and considering statistical significance without FDR correction. To evaluate Faecalibacterium functional potential, we extracted from species resolution HumanN3 data all KOs annotated as belonging to Faecalibacterium species KOs and normalized read counts per 1 million reads. Differential abundance for each of these KOs was assessed using a Kruskal-Wallis test with FDR correction. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Acknowledgements We would like to express our gratitude to Jan Inge Nordby from the Department of Medical Biochemistry at Oslo University Hospital and Erik Natvig and Anita Jørgensen from the Department of Colorectal Cancer Screening at the Cancer Registry of Norway (CRN), National Institute of Public Health (NIPH), for their contributions to biobanking, laboratory work, and data management. We also wish to thank Harri Kangas and Pekka Ellonen at the Finnish Institute of Molecular Medicine (FIMM) for their assistance in preparing sequencing data; library preparation and sequencing were conducted at the FIMM Technology Centre, supported by HiLIFE and Biocenter Finland. Our thanks extend to Martin Solbrekken and Stine Langbråten at the CRN, NIPH, along with Anne Marte Wetting Johansen at the Department of Nutrition, University of Oslo (UiO), for their essential contributions to scanning and processing the questionnaire data. Additionally, we acknowledge the support of current and former group members for their administrative and scientific contributions, including Elina Vinberg, Maja Sigerseth Jacobsen, Even Sannes Riiser, Jonas Thy, Astrid Riseth Andersen, Paula Istvan, and Lisi Zhan. Author contributions T.B.R., P.B., G.H., and G.U. contributed to study conception and design. P.B., T.B.R., A.S.K., E.E.B., C.B., and V.B. participated in data collection. E.E.B., A.S.K., E.A., and C.B. performed the data analysis. H.B., A.H., E.B., and T.B.R. supervised data analysis. E.E.B., A.S.K., E.A., W.M.dV., G.H., K.R., and T.B.R. interpreted the results. E.E.B., A.S.K., and T.B.R. prepared the manuscript. All authors have read and approved the final manuscript. Peer review Peer review information Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Funding Research reported in this publication was supported by the Norwegian Cancer Society, projects 190179 (T.B.R.) and 198048 (P.B.), and the South-East Norway Regional Health Authority projects 2022067 (T.B.R.) and 2020056 (T.B.R.). Data availability The DNA sequencing data generated in this study have been deposited in the Federated EGA under accession code EGAS50000000170. The data are available under controlled access for per-participant consent; access can be obtained by following the detailed procedure described here: https://www.mn.uio.no/bils/english/groups/rounge-group/crcbiome/. The data are protected and are not publicly available due to EUs General Data Protection Regulation (GDPR); access requires ethical and legal approvals and a Data Protection Impact Assessment (DPIA). Provided approval of the required documents by the data controller, the data will be made available within a month. Code availability All scripts are available at https://github.com/Rounge-lab/Microbiome-signatures-CRCbiome ref. 93. Competing interests The bowel preparation used for colonoscopy in the BCSN trial was provided free of charge by Ferring Pharmaceuticals; the company had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors report no other conflicts of interest. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Einar E. Birkeland, Email: einareb@uio.no Trine B. Rounge, Email: trinro@uio.no Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-75962-1. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data Reporting Summary Transparent Peer Review file Data Availability Statement The DNA sequencing data generated in this study have been deposited in the Federated EGA under accession code EGAS50000000170. The data are available under controlled access for per-participant consent; access can be obtained by following the detailed procedure described here: https://www.mn.uio.no/bils/english/groups/rounge-group/crcbiome/. The data are protected and are not publicly available due to EUs General Data Protection Regulation (GDPR); access requires ethical and legal approvals and a Data Protection Impact Assessment (DPIA). Provided approval of the required documents by the data controller, the data will be made available within a month. All scripts are available at https://github.com/Rounge-lab/Microbiome-signatures-CRCbiome ref. 93.