Molecular Sub‐Classification Suggests IBD Unclassified as a Distinct Subtype of Inflammatory Bowel Disease: Insights From a 3‐Gene Mucosal Signature Authors: Jaslin P James, Ib Jarle Christensen, Boye S Nielsen, Ebbe Langholz, Estrid Høgdall, Lene Buhl Riis Published: 2026 ABSTRACT ABSTRACTDifferentiating Crohn's disease (CD) from ulcerative colitis (UC) is challenging, often resulting as Inflammatory bowel disease unclassified (IBDU). Based on targeted RNA sequencing, we recently reported a 3‐gene signature algorithm that distinguished CD from UC. This study evaluated its ability to classify IBDU cases and explore disease biology. We analyzed 197 mucosal FFPE biopsies (inflamed and non‐inflamed) from 113 individuals at diagnosis: CD (n = 42), UC (n = 41) and IBDU (n = 102; follow‐up: 0–23 years (IQR)). Expression levels of ANXA1, PI3, and VDR mRNAs were measured by RT‐qPCR, and assessed using logistic regression, and ROC curve analysis. The algorithm discriminated CD from UC with AUC values > 0.70 in 6 out of 8 patient subsets, peaking in inflamed adult biopsies (0.78, 95% CI: 0.64–0.92). At 80% sensitivity, specificity ranged from 46% to 58%, replicating previous cohort findings and confirming reproducibility. Application of the algorithm to the IBDU cases showed low classification accuracy, with frequent misclassification of cases that were later received definitive diagnoses. Incorporation of additional IBDU‐specific features did not improve performance. While the 3‐gene signature algorithm demonstrates moderate ability to distinguish CD from UC, its limited performance in IBDU cases highlights fundamental constraints of binary classification approaches in IBD.Keywords: Crohn's disease, diagnosis, gene expression, inflammatory bowel disease, molecular biomarker, ulcerative colitis Abbreviations beta actin anti‐ Saccharomyces cerevisiae antibodies area under the receiver‐operating characteristic curve Crohn's disease cycle threshold female formalin‐fixed parffin‐embedded General Data Protection Regulation gene expression omnibus gastro intestinal healthy controls human genome 19 inflammatory bowel disease inflammatory bowel disease unclassified inter quartile range locked nucleic acid male messenger RNA number not applicable negative predictive values perinuclear antineutrophil cytoplasmic antibodies positive predictive values the National Secretariat for Bio‐ and Genome‐Bank Denmark receiver operating characteristic curve Research Resource Identifiers real time quantitative polymerase chain reaction ulcerative colitis Introduction Crohn's disease (CD) and ulcerative colitis (UC) are the two main subtypes of inflammatory bowel disease (IBD), each with distinct pathological and phenotypical characteristics. CD and UC have significant pathological and phenotypical characteristics: CD typically presents with transmural inflammation, skip lesions, granuloma formation, and can affect any part of the gastrointestinal tract; UC is usually characterized by continuous mucosal inflammation limited to the colon and rectum. However, features such as nonspecific chronic inflammation, absence of granulomas, or inflammation confined to the colon can overlap between the two diseases, especially in early or indeterminate stages, resulting in diagnostic findings that overlap between the two disease categories [1, 2, 3, 4]. In up to 15% of IBD patients, it is difficult to differentiate between CD and UC [5] at the time of initial disease presentation. This problem has existed for decades and has resulted in a third tentatively diagnostic group comprising unclassified IBD (IBDU) patients. IBDU is the rarest of IBD subtypes, representing approximately 5%–15% of all IBD cases [6], and is defined by clinical and endoscopic symptoms of chronic colitis that lack specific markers of CD or UC but have some indirect features of both [6]. The IBDU condition most often establish itself as either CD or UC, which is mandatory for firm treatment decision. The growing use of immune modulators and advanced biological therapies has dramatically changed IBD treatment aiming towards precision medicine resulting in CD‐ and UC‐specific treatment. Notably, the possibility of restorative pouch surgery is reserved for patients with UC, as CD patients have poor outcomes from such surgery [7]. Misclassification may thus be fatal for CD patients, highlighting the importance of solid biomarker signatures. Efforts have been made to establish a diagnostic test that can differentiate between CD and UC, including serologic testing [8, 9], fecal markers [10, 11], lamina propria T‐helper subsets [12], and using differential gene expression [13, 14, 15, 16]. Nonetheless, follow‐up studies have failed to provide compelling evidence to support the initial conclusions [17, 18]. The technologies utilized, as well as variations in control groups, disease activity, biopsy selection, and data standardization, may all contribute to a lack of reproducibility across these studies. We previously developed a 3‐gene signature algorithm comprising mucosal mRNA expression rates of ANXA1, PI3 and VDR, which together with the patient's age and sex is able to discriminate CD from UC in 80% of random IBD cases [19]. Our study represented a step forward in addressing an important longstanding challenge of accurately differentiating between CD and UC by biomarker analysis. In this study, we evaluated the performance of the algorithm [19] in an independent cohort of IBD patients, while placing particular emphasis on its application to IBDU cases. By analyzing biopsies from both inflamed and non‐inflamed regions of the colon in treatment‐naïve patients (biopsies collected at or within 3 months of diagnosis), we aimed not only to assess diagnostic performance but also to investigate whether molecular classification approaches can adequately capture the biological complexity of IBD. Specifically, we sought to determine whether the behavior of the algorithm in IBDU cases could provide insight into whether this group represents a pre‐determined transitional diagnostic category. We hypothesized that precise molecular differentiation of CD and UC would facilitate prediction of early classification of IBDU. In this study we obtained a sample/patient cohort of CD, UC and IBDU cases independent of the previously reported cohort comprising CD and UC patients, only. We first validated the algorithm in its performance to discriminate CD from UC and then submitted the algorithm to the IBDU cohort. Materials and Methods Patient Samples The study included 196 biopsies from 113 individuals composing 107 IBD patients and 6 healthy controls (HCs). Individuals were chosen based on their IBD subtype and the availability of mucosal biopsies at the time of diagnosis. As part of the initial diagnostic phase, the samples were collected and stored as formalin‐fixed paraffin‐embedded (FFPE) biopsies in clinical biobanks at the hospitals in Greater Copenhagen region. To significantly reduce the possibility of misclassification, all patients underwent a full clinical review at the time of diagnosis, which included gastroduodenoscopy, colonoscopy, and small bowel MRI. IBD was diagnosed using either the Porto criteria [20] (for children) or the Copenhagen criteria [21] (for adults). Patients had different disease parameters such as age at diagnosis, degree of inflammation, treatment history, and disease location. Biopsies from definite cases of CD and UC patients were considered the validation cohort (24 IBD_CD patients, 24 IBD_UC patients). Patients diagnosed as IBDU at the time of diagnosis and patients with double diagnosis (with CD and UC diagnosis during the disease course) were included as a separate cohort (59 IBDU patients) to assess algorithm performance in diagnostically uncertain cases. IBDU patients were followed for a range of 0–23 (IQR) years from the first diagnosis with IBD. To capture spatial and inflammatory heterogeneity, and to evaluate the robustness of the molecular signature across different anatomical locations, up to two biopsies per patient were included where available. One biopsy was from the inflamed area in the colon, as close to the rectum as possible. The other biopsy was from a non‐inflamed section of the colon, as far away from the inflamed area as feasible. In the absence of a colon biopsy, a rectal biopsy was chosen. For 28 patients there was only one biopsy available. Because there were no non‐inflamed biopsies available for 31 patients, two inflamed biopsies were chosen that were widely apart to confirm robustness of the algorithm. Biopsies collected from the descending and sigmoid colon were considered as biopsies from the left side. Biopsies collected from ileum, caecum, ascending, and transverse colon were considered as biopsies from the right side. In this study “location” of biopsy refers to the site of the tissue, and “side” refers to left versus right. This sampling approach was implemented to evaluate the influence of anatomical and inflammatory variation on gene expression profiles and algorithm performance. HCs were included to provide a baseline reference for gene expression levels and to support quality control of the RT‐qPCR measurements. Although not included in the classification model, these samples enabled contextualization of disease‐associated expression patterns. Patients who were diagnosed with IBD under the age of 18 years were considered as pediatric. There were 23% and 51% of pediatric patients among definite IBD and IBDU cohorts, respectively. Table 1 provides description of patient characteristics. The rate of inflammation was scored on hematoxylin and eosin‐stained slides using a three‐tiered system where the presence of ulcerations was graded as severe inflammation; cryptitis, and crypt abscesses but no ulceration as moderate inflammation; and cryptitis alone (no ulceration, no crypt abscesses) as light inflammation. Besides these features the biopsies presented the histological characteristics of IBD: architectural distortion, basal plasmacytosis and chronic inflammation in lamina propria. TABLE 1Clinical and demographic characteristics of the study population, including patients with definitive Crohn's disease (CD), ulcerative colitis (UC), and IBD unclassified (IBDU) cases used to assess algorithm performance in both well‐defined and diagnostically ambiguous cohorts.IBD_CDIBD_UCIBDUHCTotalNumber of patients (N)2424596113Number of biopsies (N)424110211196Age group, adult/pediatric (N)19/520/433/263/375/38Age at diagnosis (median, IQR, range)26.529.51921.510–639–714–7613–8319–4720–4513–4015–70Female (N, %)13, 54%11, 46%22, 37%2, 33%Rate of inflammation (N), non‐inflamed/light/moderate/severe13*/11/6/1112/9/16/426/36/32/811/0/0/062/56/54/23Inflamed/non‐inflamed biopsies (N)30/1229/1276/260/11135/61IBDU duration in years (median, IQR, range)NANA4NA0–231–9Involvement of the disease at the time of diagnosis* (number of patients)** Terminal ileum only1Colon only132448Terminal ileum and colon75Colon and upper GI3Terminal ileum, colon, and upper GI21Maximum involvement during disease course* (number of patients)Terminal ileum only1Colon only62140Terminal ileum and colon12312Colon and upper GI23Terminal ileum, colon, and upper GI22Site of tissue (number of biopsies)Ileum9724444Caecum529218Ascending and transverse colon516012Descending and sigmoid colon101534362Rectum101516041Colon (unspecified)3113017Abbreviations: IQR = inter quartile range, N = number, NA = not applicable.*One IBD_CD patient without information on rate of inflammation and involvement of disease.**Two IBDU patients without information on involvement and maximum involvement. Clinical and demographic characteristics of the study population, including patients with definitive Crohn's disease (CD), ulcerative colitis (UC), and IBD unclassified (IBDU) cases used to assess algorithm performance in both well‐defined and diagnostically ambiguous cohorts. Abbreviations: IQR = inter quartile range, N = number, NA = not applicable. One IBD_CD patient without information on rate of inflammation and involvement of disease. Two IBDU patients without information on involvement and maximum involvement. Ethical Considerations The study has been approved by the National Ethical Committee (j. no. H‐20032221) and the Danish Data Protection Agency (P‐2020‐737). The need for consent was waived by Danish National Ethical committee as tissue used for the project is FFPE tissue previously collected and stored in the clinical biobank (tissue archive) at the Department of Pathology, Herlev Hospital, Denmark and Pediatric Department or Gastro unit at Hvidovre Hospital, Denmark. The study was conducted in accordance with the ethical standards, according to the declaration of Helsinki, and according to national and international guidelines. RNA Extraction Total RNA was isolated from two 10‐μm thick sections of FFPE tissue samples. AutoLys M Tubes (Applied Biosystems) were used for FFPE sample lysis and nucleic acid separation from debris. Automated RNA extractions were performed on KingFisher Duo Prime Purification System (ThermoFischer Scientific) using MagMAX FFPE DNA/RNA Ultra Kit from Applied Biosystems according to the manufacturer's instructions [22]. RNA quantity and purity were assessed using the NanoDrop spectrophotometer (ThermoFisher Scientific, RRID: SCR_015804). RNA samples were kept at −80°C until further use. RT‐qPCR Gene expression for ANXA1, PI3, and VDR were measured using RT‐qPCR (QuantiNova SYBR Green RT‐qPCR Kit [Qiagen]). QuantiNova LNA PCR Custom Panel (GeneGlobe ID: SBCA00967) was designed with the three target genes and two reference genes (ACTB and RPLP0) in addition with recommended controls from the manufacturer. HS_PPC_2467741, HS_HGDC_2467744 and HS_QIC_2467742 were used as reference assays according to custom panel manufacturer's recommendation. Real‐time PCR amplifications were done using the Light Cycler 480 system (Roche, RRID: SCR_018626). RT‐qPCR data was analyzed using the 2−(∆Ct) method and ACTB as reference gene to yield relative fold expression levels. Samples with missing Ct values were substituted the higher limit of detection (40). Statistics The analysis of binary IBD outcome data was done using logistic regression with the predicted probability for CD calculated using the coefficients from the published algorithm [19] based on the expression levels of the 3 genes, sex and age of the patient. The coefficients for this model assumed in the current study that the ACTB level was < 31. The algorithm was applied to IBDU samples with a final diagnosis of UC or CD as done for definite IBD cases. Exploratory analysis using mixed modeling was performed, with multiple samples for patients including inflammation/non‐inflammation and the site of biopsy. Based on the mixed models, variance components were estimated and shown as coefficients of variation (CV) for the odds of CD. Statistical analysis was done using SAS (SAS Institute, Cary, NC, USA, ver 9.4, RRID: SCR_008567) and R (The R Foundation for Statistical Computing Platform, RRID: SCR_001905). Results Performance of the 3‐Gene Signature Algorithm in IBD Cohorts The 3‐gene signature algorithm was evaluated using logistic regression models to estimate the probability of CD across multiple biopsy subsets listed in Table 2. Across these analyses, the algorithm demonstrated moderate discriminatory capacity, with AUC values consistently around 0.70. This performance level is similar to our previously published cohort, in which the algorithm achieved an overall classification accuracy of approximately 80%, indicating reproducibility of the model in an independent dataset, albeit with slightly lower discrimination in the current study. TABLE 2Performance of the 3‐gene signature algorithm in patients with definitive IBD. The table summarizes the number of samples (N), area under the ROC curve (AUC) with 95% confidence intervals, specificity at approximately 80% sensitivity, and estimated positive and negative predictive values (PPV and NPV) across different biopsy subsets. Results highlight moderate discrimination and variability across clinical and sampling conditions.ModelIBD CD (N)IBD UC (N)AUCLower limitUpper limit p SensitivitySpecificityPPVNPVAll biopsies42410.700.590.820.00783466173Inflamed biopsies30290.730.590.860.01380486270Adult24240.780.640.920.00479586674Pediatric650.470.070.860.979830500Biopsies from patients with pure colonic involvement of the disease at the time of diagnosis23410.700.570.830.01987294180Inflamed biopsies18290.760.610.900.01078595481Adult biopsies22350.680.540.820.03686204070Adult inflamed biopsies17240.740.580.900.02082174157 Performance of the 3‐gene signature algorithm in patients with definitive IBD. The table summarizes the number of samples (N), area under the ROC curve (AUC) with 95% confidence intervals, specificity at approximately 80% sensitivity, and estimated positive and negative predictive values (PPV and NPV) across different biopsy subsets. Results highlight moderate discrimination and variability across clinical and sampling conditions. Performance was slightly better in adult (AUC = 0.78) inflamed biopsies compared to pediatric samples (AUC = 0.47), likely reflecting limited sample size. At a clinically relevant sensitivity threshold of 80%, the corresponding specificities would range from 46% to 58%, suggesting a considerable rate of false positive classifications. Restricting the analysis to colonic disease and in inflamed colonic biopsies, only, the specificity reached 59% at 78% sensitivity, suggesting that inflamed biopsies contain more information in disease than the less inflamed biopsies. Receiver operating characteristic (ROC) curves confirmed modest but consistent discrimination across most subsets, although performance was notably reduced in pediatric cases (Figure 1). Importantly, incorporation of biopsy location (left vs. right colon) into the model significantly improved discrimination between CD and UC in all, whereas inflammation status did not contribute significantly. Receiver operating characteristic (ROC) curves illustrating the performance of the 3‐gene signature algorithm in distinguishing Crohn's disease (CD) from ulcerative colitis (UC) across defined IBD subgroups. The curves demonstrate moderate discriminatory capacity across most subsets, with variation depending on biopsy characteristics and patient groups. Subgroups include all biopsies (in red), inflamed biopsies (in black), adult inflamed biopsies (in blue), pediatric inflamed biopsies (in green), colon only involvement (in orange), inflamed and colon only involvement (in light blue), adult biopsies inflamed and colon only involvement (in pink), colon only involvement and adult biopsies (in gray). Variance component analysis revealed substantial inter‐individual variability (CV 117%), compared to relatively low intra‐individual variability (CV 26%), suggesting that patient‐specific biological differences are a dominant source of variation in gene expression profiles. This degree of variability may partly explain the differences observed between the current and previously published cohorts. Together, these findings indicate that while the algorithm captures disease‐associated signals, and shows reproducible performance across cohorts, its discriminatory capacity remains moderate and is influenced by both biological heterogeneity and sampling‐related variability. Application to IBDU Cohort Reveals Classification Limitations The algorithm was subsequently applied to biopsies from patients with inflammatory bowel disease unclassified (IBDU). In contrast to the performance level in the defined CD and UC cohort, the algorithm demonstrated poor classification accuracy in IBDU cases, failing to reliably stratify patients into CD or UC categories (Table 3). TABLE 3Performance of the 3‐gene signature algorithm in IBD unclassified (IBDU) cases. The table presents the number of samples (N) and AUC values with 95% confidence intervals for predicting eventual diagnosis (CD or UC). The results demonstrate limited classification accuracy, reflecting challenges in applying CD/UC‐based molecular models to diagnostically ambiguous cases.ModelIBDU CD (N)IBDU UC (N)AUCLower limitUpper limitAll biopsies36200.580.420.73Inflamed biopsies27150.560.360.75Adult2040.700.371.00Pediatric7110.700.450.95Biopsies from patients with pure colonic involvement of the disease at the time of diagnosis22190.560.380.74Inflamed biopsies16140.500.280.72Adult biopsies2070.700.470.93Adult inflamed biopsies1540.700.341.00 Performance of the 3‐gene signature algorithm in IBD unclassified (IBDU) cases. The table presents the number of samples (N) and AUC values with 95% confidence intervals for predicting eventual diagnosis (CD or UC). The results demonstrate limited classification accuracy, reflecting challenges in applying CD/UC‐based molecular models to diagnostically ambiguous cases. Among patients who later received a definitive diagnosis, the algorithm frequently misclassified cases at the time of initial biopsy. These misclassifications were particularly prominent in cases exhibiting overlapping clinical and molecular characteristics, highlighting the difficulty of forcing IBDU cases into a binary CD/UC framework. Attempts to improve classification performance through incorporation of additional variables—including inflammation status, disease duration, and biopsy location—did not yield significant improvements. Similarly, re‐fitting the model specifically on IBDU cases did not produce statistically significant predictive performance. Collectively, these findings suggest that the poor performance of the algorithm in IBDU is not merely due to technical limitations but may reflect fundamental biological differences that are not adequately represented within a CD/UC classification framework. Discussion The present study demonstrates that while a 3‐gene signature algorithm provides moderate discrimination between CD and UC in well‐defined cases, its performance is substantially reduced in IBDU patients. These findings highlight fundamental limitations in current molecular diagnostic approaches and provide important insights into the biological complexity underlying IBD. The Binary Prediction Classification of CD/UC Patients The moderate performance observed across CD and UC cohorts, with AUC values around 0.70 and relatively low specificity, indicates that the algorithm captures disease‐associated signals but provides only moderate discrimination between the two diseases. These findings are consistent with previous studies evaluating serological and gene expression‐based biomarkers, which have similarly demonstrated variable diagnostic performance across patient populations [9, 16, 18]. The significant effect of presence of inflammation, cell composition in the tissue samples, and biopsy location further underscores the influence of spatial heterogeneity within the intestinal mucosa. Previous studies have suggested that altered mucosal gene expression due to inflammation [18], and intersegmental gene expression differences [23] contribute to variability in molecular IBD diagnostics [19, 24]. Consistent with these observations, inclusion of biopsy side (left versus right colon) in addition to the 3‐gene signature algorithm significantly improved discrimination between CD and UC, whereas inflammation status did not. Furthermore, intra‐individual variability was considerably lower than inter‐individual variability (CV 26% vs. 117%), indicating that biological differences between patients outweigh variability between biopsies from the same individual. IBDU as a Distinct Molecular Subtype Based on the validation of the 3‐gene algorithm's ability to sufficiently predict CD and UC patients, we considered the sample cohort as representative for the IBD samples, and therefore found it reasonable to submit the algorithm to the IBDU sample cohort. However, we found that the 3‐gene signature algorithm was unable to reliably classify (or predict) the IBDU cases even after considering IBDU‐specific features. Rather than reflecting purely technical limitations, this observation may indicate underlying biological differences between IBDU and the traditionally defined CD and UC categories. Importantly, misclassification patterns were not random but were concentrated among cases with intermediate prediction probabilities, suggesting the presence of overlapping or hybrid molecular features. These findings support the hypothesis that IBDU occupies a distinct position within the IBD spectrum, rather than representing a transient or unresolved form of CD or UC. This interpretation is further supported by our previous work demonstrating unique molecular signatures in IBDU, particularly at the level of short noncoding RNA expression, which has shown improved discriminatory performance compared to gene‐based models [14]. In addition to the previous study [14], we further tested the short noncoding RNA signature algorithm in a prospective study including the current cohort of IBDU patients (unpublished data, ongoing study). The preliminary observations have shown that the short noncoding RNA signature algorithm could differentiate the IBDU cases with an AUC of 70.7% (53.7%–87.8%) for all samples and 76.9% (51.1–100) among adults, indicating better performance on IBDU cases than the algorithm developed using definite CD and UC cases. Together, these observations suggest that the molecular architecture of IBDU differs from that of classical CD and UC and may involve regulatory pathways that are not captured by gene expression signatures performed in this study. Nevertheless, alternative explanations should also be considered, including the inherent molecular heterogeneity of IBDU and the possibility that a limited three‐gene panel cannot adequately capture disease complexity. Consequently, our findings so far should be interpreted as supporting, rather than proving, the hypothesis that IBDU may represent a distinct molecular subtype. Clinical Implications of Misclassification The limited performance of the algorithm in IBDU cases has important clinical implications. Misclassification at the time of diagnosis may lead to suboptimal treatment decisions, particularly in cases where therapeutic strategies differ between CD and UC. Given the moderate specificity and poor performance in ambiguous cases, reliance on simplified molecular classifiers should be approached with caution in routine clinical practice. These findings highlight the need for diagnostic strategies that incorporate longitudinal assessment and multiple layers of molecular data, rather than relying on single‐time‐point gene expression measurements. Such approaches may be particularly valuable for patients presenting with atypical or overlapping disease features. Future research may also embrace artificial intelligence (AI)—and deep learning (DL), which have already shown promise in the diagnostic assessment of IBD [25, 26]. Strengths and Limitations The present work incorporated several methodological improvements, including standardized biopsy selection from defined anatomical sites (inflamed and non‐inflamed biopsies from defined areas of the colon), inclusion of treatment‐naïve patients to minimize the potential confounding effects of treatment on gene expression, and automated extraction of high‐quality RNA from routinely available FFPE tissue. Together, these refinements reduced potential sources of biological and technical variability and strengthened the robustness of the molecular analyses, important if implemented in a routine diagnostic setting. One of the study's limitations, however, was the relatively small sample size, particularly for IBDU patients (total of 59 patients), which limits the statistical power for more detailed analyses. Although we stratified IBDU patients by adult and pediatric status and by the location of left vs. right colon biopsies, the smaller numbers in each subgroup reduce the strength of the findings. It may therefore be relevant to build an algorithm based on IBDU patient material, only, to obtain evidence of a predictive molecular profile. Expanding the cohort in future studies will be crucial to increasing the generalizability and accuracy of the algorithm. Comparison With Existing Methods Similar to previously proposed serological and transcriptomic biomarkers, the present 3‐gene signature achieved moderate diagnostic performance but did not overcome the challenges associated with diagnostically ambiguous IBD cases. Our 3‐gene algorithm aims to address some of the limitations of existing methods. Several existing methods have been employed to distinguish CD from UC, including serologic testing, fecal markers, lamina propria T‐helper subsets, and differential gene/microRNA expression. While each of these methods have shown promise, they have also faced significant challenges and limitations. Serologic markers, such as anti‐ Saccharomyces cerevisiae antibodies (ASCA) and perinuclear anti‐neutrophil cytoplasmic antibodies (pANCA), have been used to differentiate CD from UC [9]. However, their sensitivity and specificity vary widely across studies, often leading to inconsistent results. Additionally, almost 50% of these patients do not develop ASCA or P‐ANCA antibodies [9]. The analysis of lamina propria T‐helper cell subsets have shown potential in differentiating CD from UC. However, this method requires invasive biopsy procedures and complex laboratory techniques, limiting its practicality in routine clinical practice [13]. Studies exploring the use of differential gene and microRNA expression profiles to distinguish CD from UC showed promising results, however, often with challenges in reproducibility and required advanced technology and expertise, limiting their accessibility for widespread clinical use [17, 18]. The persistent difficulty in classifying IBDU across different molecular approaches show that current binary biomarker models may not fully capture the biological diversity of inflammatory bowel disease. Nevertheless, the simplicity of a three‐gene RT‐qPCR gene expression assay using routinely available FFPE tissue represents a practical advantage over more complex molecular platforms at least for handling of CD and UC. Implications for Future Biomarker Development Although the 3‐gene algorithm identified CD and UC patients with a positive predictive value (PPV) of 61%, our findings underscore the limitations of minimal gene signatures in capturing the complexity of IBD, particularly in the context of IBDU. While such approaches offer practical advantages in terms of simplicity and cost, they may lack the depth required for accurate classification across diverse patient populations. To translate these discoveries into routine clinical practice, future studies should identify the most stable and informative features across multiple biological layers combined with clinical information. AI/DL approaches may further enhance this process by integrating complex multi‐omics and clinical datasets to identify robust diagnostic signatures and improve disease classification. These features could then be incorporated into targeted, cost‐effective assays, such as multiplex PCR or focused sequencing panels, suitable for routine diagnostic laboratories. Funding This work was supported by Aage Louis Hansen fonden [j. no. 23‐2B‐13469]; Colitis‐Crohn Foreningen [2024]. Conflicts of Interest The authors declare no conflicts of interest.