Decoding IBD progression: a dynamic biomarker atlas for personalized disease stratification

Abstract

Background

Accurate staging is pivotal for tailoring treatment intensity, optimizing resource allocation, and improving long-term patient outcomes in IBD. The intestinal microbiota and transcriptional profiles emerge as critical determinants in IBD staging, demonstrating promise as non-invasive biomarkers for predicting disease progression and informing personalized therapeutic strategies.

Methods

We recruited 97 participants (IBD patients and healthy controls) at the First Affiliated Hospital of Chongqing Medical University, collecting fecal and serum samples for integrated multi-omics analysis. Microbial community profiling was performed via 16S rRNA sequencing, and host transcriptomic landscapes were characterized using RNA-seq. Stage-specific microbial signatures were identified using NetMoss, a network-based microbial community analysis tool, while differential gene expression across IBD stages was determined by Boruta feature selection coupled with a recursive SVM classifier (REF-SVM). Cross-omics correlations between gut microbiota abundances and host gene expression were evaluated to map microbiota-host interactions. For predictive modeling, seven machine learning algorithms were trained on microbial and transcriptomic features, with a Stacking Classifier meta-ensemble employed to integrate predictions and optimize classification accuracy for IBD staging. This pipeline enabled the discovery of microbial biomarkers, stage-specific transcriptional markers, and robust multi-omics models for disease stratification.

Results

This study enrolled 97 participants (74 IBD patients, 23 healthy controls) and identified significant clinical dysregulations (P < .05) in albumin, cholesterol, erythrocyte sedimentation rate, CRP, and fecal calprotectin in IBD cohorts. Fecal samples from 57 participants underwent 16S rRNA sequencing (mean depth 42.86 Mbp/sample, 92.4% OTU annotation efficiency), revealing reduced microbial α-diversity (Chao1/Shannon indices, P < .05) in IBD patients, alongside stage-specific taxonomic shifts, including Bacilli/Gammaproteobacteria enrichment and Bacteroidia/Clostridia alterations. NetMoss-based biomarker discovery pinpointed discriminative taxa (Bifidobacterium.catenulatum and Bacteroldes.fragilis in remission, Streptococcus.gallolyticus, Veillonella.atypica and Clostridium.butyricum in mild, Blautia.obeum in moderate, and Bacteroides. Uniformis in severe IBD). Parallel RNA-seq analysis of 72 samples (18,673 genes detected, 97.89% alignment rate) uncovered 161 differentially expressed genes, with stage-specific markers (YIPF4 and GIMAP6 in remission, HBB and FKBP5 in mild, TUBB1 in moderate, ALAS2, MMRN1 and IGFBP2 in severe IBD) enabling 82.4% staging accuracy via a REF-SVM classifier. Multi-omics integration revealed functional microbial-host interactions, such as Bacteroides.fragilis associations with GIMAP6/YIPF4 and Blautia.obeum with TUBB1. Integrated models demonstrated strong predictive performance (AUCs: microbial = 0.79, transcriptional = 0.80), highlighting the clinical utility of multi-omics biomarkers for IBD stratification.

Conclusions

This study establishes a multi-omics framework for non-invasive IBD staging, linking gut microbiota dynamics to host transcriptional reprogramming. The integration of NetMoss-derived biomarkers and machine learning enhances precision in stratifying disease severity, offering actionable insights for therapeutic interventions. These findings underscore the clinical potential of combining microbial and transcriptomic data to transform IBD management.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-025-07024-x.

What is new

1. Prior knowledge on IBD staging mainly focused on clinical and endoscopic criteria. This study adds a new dimension by demonstrating that intestinal microbiota and transcriptional profiles are critical determinants in IBD staging, offering non—invasive biomarkers for predicting disease progression.

2. The study identified IBD staging biomarkers via multi—omics, uncovered microbial-host interactions, and developed predictive models with good performance.

3. The findings suggest that non-invasive multi-omics biomarkers, such as microbial and transcriptional markers, have the potential to inform personalized therapeutic strategies for IBD patients. Clinicians should consider incorporating these biomarkers into their diagnostic and treatment decision-making processes.

Introduction

Inflammatory Bowel Disease (IBD), including Crohn’s disease (CD) and ulcerative colitis (UC), is a chronic, relapsing inflammatory disorder of the gastrointestinal tract driven by dysregulated immune-environmental interactions in genetically susceptible individuals [1, 2]. Over recent decades, the global prevalence of IBD has been steadily increasing, posing significant challenges for healthcare systems and patients alike. The global prevalence of IBD patients surged from 3.3 million in 1990 to 4.9 million by 2019, with projections estimating that Europe and the United States alone may see over 7 million cases by 2030, reflecting a concerning upward trajectory in disease incidence [3, 4]. IBD imposes a significant burden on patients, encompassing substantial economic costs and profound impacts on physical, psychological, and social well-being [5].

The management of IBD significantly depends on its classification as well as its disease stage and severity at diagnosis. Early diagnosis plays a pivotal role in improving treatment outcomes, as it allows for more effective intervention during the initial phases of the disease, potentially preventing further complications [6, 7]. The standard procedures for diagnosing IBD may include colonoscopy, imaging studies, and histological examinations; however, these techniques are invasive and may not always provide a comprehensive picture of disease progression [6]. Moreover, staging methods can be subjective, relying heavily on practitioner expertise and experience, which introduces variability into clinical decision-making. Consequently, there is a pressing demand for non-invasive diagnostic tools that can accurately assess disease status and guide therapeutic decisions [8, 9].

Emerging evidence highlights the important role of gut microbiota in IBD pathogenesis, with dysbiosis strongly implicated. However, the relationship between gut microbiota and distinct IBD stages remains inadequately characterized, emphasizing the need for our study to identify stage-specific microbial biomarkers. Members of the Bacteroidetes, whose reduced abundance represents a well-documented characteristic of gut microbiota dysbiosis in IBD patients, have been particularly implicated in disease pathogenesis. As noted in previous studies, the most prominent changes in the microbiota of IBD patients include decreased bacterial diversity and a reduced abundance of Bacteroidetes [10–12]. Additionally, IBD patients exhibited notable alterations in microbial abundance, community composition, and gene expression profiles. Emerging evidence suggests cytoskeleton-associated gene expression patterns may correlate with IBD disease activity, though their stage-specific variations remain poorly characterized. Previous studies have demonstrated that the TUBB1 gene is associated with cytoskeletal stability and cell proliferation, indicating that patients with moderate IBD may exhibit increased cellular proliferation and an amplified inflammatory response [13]. Despite these advances, the ability to predict disease staging based on specific gut microbiota characteristics or transcriptional markers remains largely unexplored [13]. Further research is required to determine whether distinct microbial profiles or transcriptional signatures can serve as reliable biomarkers for disease staging and progression [14, 15]. Such insights would facilitate early diagnosis and accurate disease stratification, thereby enhancing clinical outcomes for patients.

This study leverages an integrated multi-omics framework to systematically dissect microbial and transcriptional networks, revealing biomarkers predictive of distinct IBD progression stages. Utilizing paired fecal microbial profiling (16S rRNA sequencing) and peripheral blood transcriptomic analysis, we implemented network-based computational frameworks to identify stage-associated biological signatures. The analytical pipeline incorporates feature selection algorithms optimized for high-dimensional biological data, coupled with machine learning validation protocols to ensure biomarker robustness. A key clinical innovation of our approach lies in its capacity to concurrently assess dynamic shifts in microbial communities alongside host transcriptional pathway activation—a critical advancement given the well-established interplay between gut microbiota composition and intestinal barrier function. This study addresses a critical gap in conventional IBD staging paradigms by introducing integrated predictive models that prioritize molecular pathology over symptom-based criteria, thereby enabling a mechanistic stratification of disease progression rooted in microbial-host interplay. This study pioneers three transformative innovations that redefine IBD stratification: (1) Concurrent profiling of microbial and transcriptional biomarkers across the full disease spectrum, capturing dynamic cross-talk between microbial dysbiosis and host immune dysregulation; (2) Network-driven feature selection algorithms that model biological system interactions, prioritizing pathogenic mechanisms over individual omic markers; and (3) Scalable computational frameworks enabling longitudinal disease tracking, where temporal omic trajectories inform real-time progression monitoring. By integrating these elements, we establish a blueprint for converting multi-omics complexity into precision medicine tools—bridging molecular pathogenes with actionable clinical decision-making to enable stratified therapeutic strategies tailored to individual disease trajectories.

Methods

Study design and sample collection

This study was conducted at The First Affiliated Hospital of Chongqing Medical University between August 2023 and June 2024, involving a prospective cohort of patients with clinically confirmed IBD and age/sex-matched healthy controls. All participants provided written informed consent prior to enrollment. The study protocol was approved by the Institutional Review Board of The First Affiliated Hospital of Chongqing Medical University (Approval No. 2023–278) and executed in strict compliance with the ethical standards outlined in the Declaration of Helsinki. This rigorous framework ensures methodological transparency and clinical relevance while addressing potential confounders through matched control selection.

Inclusion criteria: (1) Age 18–70 years; (2) Diagnosis of IBD by a gastroenterologist based on the “2018 Chinese Consensus on Diagnosis and Management of Inflammatory Bowel Disease”; (3) Capacity to provide informed consent and adhere to study protocols. Exclusion criteria: (1) History or current diagnosis of malignancy (including solid tumors or hematological malignancies); (2) Evidence of infectious or non-IBD inflammatory colorectal conditions, including intestinal tuberculosis, HIV-associated enteropathy, schistosomiasis, amoebic colitis, infections by Yersinia, Campylobacter, clostridioides difficile, or Cytomegalovirus (CMV), ischemic colitis, radiation proctitis, intestinal lymphoma, or diverticulitis; (3) Severe nutritional deficits, malabsorption syndromes, or systemic disorders affecting gut microbiota stability.

We used the Mayo score and the Crohn’s Disease Activity Index (CDAI) score to assess the disease activity of IBD patients, respectively. Specifically, the activity of CD was defined as remission (CDAI score of < 150), mild (CDAI scores of 150–220), moderate (CDAI scores of 221–450) and severe (CDAI scores of > 450). UC activity was defined as remission (Mayo scores of 0–2), mild (Mayo scores of 3–5), moderate (Mayo scores of 6–10) and severe (Mayo scores of 11–12).

Analysis of 16S rRNA sequences

Microbiota gDNA was extracted from fecal samples. For the DNA extraction, 180 to 220 mg aliquots of each frozen fecal sample were processed using the QIAamp® DNA Stool Mini Kit (Hilden, Germany), following the manufacturer’s protocol. Polymerase chain reaction (PCR) was performed using primers targeting the V3–V4 region of 16S rRNA (338F: 5′-ACTCCTACGGGAGGCAGCAG-3′ and 806R: 5′-GGACTACHVGGGTWTCTAAT-3′). After purification and quality assessment, final products were sequenced on an Ion S5™ XL platform (Thermo Fisher Scientific, USA). Clean data were generated from the preliminary quality control of raw data (removing barcodes, primers and chimeras, and splicing and filtering low quality sequences). The clean data of 16S rRNA gene sequencing were analyzed using the QIIME237 platform (v2020.2). In brief, the DADA2 plugin was used to filter the sequencing reads and construct an ASVs feature Table [16]. The taxonomy information for the ASVs was assigned against the Silva Database (https://www.arb-silva.de) (v138.1) using the classify-sklearn algorithm in the feature classifier plugin. Low-abundance ASVs, whose relative abundance did not reach 0.1% in at least 10% of the samples, were excluded. The analysis of alpha diversity for individual samples was performed using mothur (v.1.31.2), including the Chao1 index, ACE index, Shannon index, and Simpson index [17]. The differences in species diversity between samples were compared using QIIME (v1.80) [18].

Netmoss

The NetMoss score was used to measure the driving force of every stage in the transition of the network structure [19]. This was defined as follows. First, the correlation matrix of the health state was and the correlation matrix of the disease state was , where is the correlation coefficient of stage i and stage j:

To obtain the optimized module structure, linear transformation was implemented to convert cij to sij

Thus, the correlation matrices of the health and disease states after transformation were and , respectively.

The soft threshold β was calculated based on the WGCNA algorithm, and the weighted network was

Thus, the weighted correlation matrices of health and disease states were and , respectively.

The topological overlap matrix of stage i and stage j was calculated as .

where , ; u represents other stages besides stage i and stage j in the network.

The distance between stage i and stage j was defined as

Then, the distance matrices of the health and disease states were D = [dij] and D′ = [d′ij], respectively.

dule division was conducted in matrices D and D′, with matrix Dcontaining n modules and matrix D′ m modules. The number of intersection modules of D and D′ was K (K ≤ mn). The average distance of every stage from intersection modules to health distance matrix D and disease distance matrix D′was calculated, obtaining N-by-K-order matrices Dmod and D′mod, respectively. The differential module distance matrix was defined as

The NetMoss score of stage i in any intersection module is

where A and B represent the health and disease networks, respectively; NeighborsA represents all neighboring modules in the health network, and NeighborsB represents all neighboring modules in the disease network.The intersection modules represent the stable elements during the transition from health to disease, where the transited modules resulted in alteration of the network structure. The NetMoss algorithm measures the driving force in the transition of the network structure to evaluate the importance of every stage.

Transcriptome sequencing

Library preparation is performed using Optimal Dual-mode mRNA Library Prep Kit (BGI-Shenzhen, China). A certain amount of RNA was denatured at suitable temperature to open the secondary structure, and mRNA is enriched by oligo (dT) attached magnetic beads. After reacting at a suitable temperature for a fixed time period, RNAs are fragmented with fragmentation reagents. Then First-strand cDNA is generated using random hexamer-primed reverse transcription, followed by a second-strand cDNA synthesis. The synthesized double strand cDNA is subject to end repairment reaction. After cDNA end repairment, a single ‘A’ nucleotide is added to the 3’ ends of the blunt fragments through A tailing reaction. Then the reaction system for adaptor ligation configured to ligate adaptors with the cDNAs, and finally, the library products are amplified through PCR reaction and subjected to quality control. Next, the single-stranded library products are produced via denaturation. The reaction system for circularization is set up to get the single-stranded cyclized DNA products. Any uncyclized single stranded linear DNA molecules will be digested. The final single strand circularized library is amplified with phi29 and rolling circle amplification (RCA) to make DNA nanoball (DNB) which carries more than 300 copies of the initial single stranded circularized library molecule. The DNBs are loaded into the patterned nanoarray and PE 100/150 bases reads are generated on G400/T7/T10 platform (BGI-Shenzhen, China).

The sequencing data was filtered with SOAPnuke [20] by (1) Removing reads containing sequencing adapter; (2) Removing reads whose low-quality base ratio (base quality less than or equal to 15) is more than 20%; (3) Removing reads whose unknown base (‘N’ base) ratio is more than 5%, afterwards clean reads were obtained and stored in FASTQ format. The subsequent analysis and data mining were performed on Dr. Tom Multi-omics Data mining system (https://biosys.bgi.com). Expression level of gene was calculated by RSEM (v1.3.1) 5). The heatmap was drawn by pheatmap (v1.0.8) 4 according to the gene expression difference in different samples. Essentially, differential expression analysis was performed using the DESeq2(v1.4.5) with Q value ≤ 0.05 (or FDR ≤ 0.001). We used Boruta and Recursive Feature Elimination Support Vector Machine (REF-SVM) algorithms to identify marker genes associated with IBD staging.

Constructing prediction models

To identify microbial features and marker genes that are predictive of IBD staging, we developed a comprehensive machine learning framework. All numerical features were standardized using StandardScaler to mitigate scale-related biases, with the training set processed via fit_transform to compute mean and standard deviation parameters, while the test set was transformed using the same parameters via transform. We employed seven classical and advanced models: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbors (KNN), eXtreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGBM). To enhance generalizability, a Stacking Classifier integrated predictions from seven base models using logistic regression as the meta-learner. Models were independently trained on standardized data via fit methods and evaluated using accuracy, confusion matrices, and classification reports (precision, recall, F1-score). The stacking ensemble demonstrated superior performance and was serialized as final_model.pkl using joblib for downstream applications. For robustness validation, receiver operating characteristic (ROC) curves and area-under-the-curve (AUC) values were computed after binarizing target variables. Model interpretability was enhanced via LIME, which identified and visualized the top ten predictive features, providing actionable insights into microbial and genetic determinants of IBD progression.

Statistical analysis

All statistical analyses were performed using R (version 4.3.2). Continuous variables were compared between groups using two-tailed Student’s t-tests when parametric assumptions were satisfied. For correlation analyses between microbiota composition and gene expression profiles, Pearson correlation coefficients were computed to assess linear associations. Differential abundance analysis of microbial features was conducted using a genewise negative binomial generalized linear model (GLM) with quasi-likelihood ratio tests, accounting for overdispersion and repeated measures where applicable. To control false discovery rates in high-dimensional omics datasets, the Benjamini–Hochberg procedure was applied as the sole adjustment method for all P-values. Significance thresholds were defined as follows: non-significant (NS), P > 0.05; “^*”, P < 0.05; “^**”, P < 0.01; “^***”, P < 0.001.

Results

Population

This study enrolled participants with IBD (encompassing UC and CD), and healthy controls. An overview of the study workflow is provided in Fig. 1. A total of 97 participants were recruited, including 74 IBD patients and 23 healthy controls. Clinical and laboratory parameters were compared between groups, revealing statistically significant differences (P < 0.05) in albumin (g/L), cholesterol (mmol/L), erythrocyte sedimentation rate (mm/h), C-reactive protein (CRP; mg/L), and fecal calprotectin (μg/g). No significant disparities (P > 0.05) were observed for age, body mass index (BMI), leukocyte counts (× 10³ cells/μL), sex distribution, or transcriptomic profiles (Table 1). Following completion of sequencing analyses, the cohort was stratified into two subgroups based on data availability: a 16S rRNA gene sequencing cohort (n = 57) and an RNA-sequencing (RNA-seq) cohort (n = 73). Notably, both subgroups demonstrated comparable demographic and clinical baseline characteristics, ensuring robust comparative analyses across multi-omics datasets (Tables 2 and 3).

Fig. 1.

Schematic representation of the microbial composition analysis and predictive modeling for Inflammatory Bowel Disease (IBD) staging

Table 1.

Analysis of basic characteristics between healthy and IBD patients

Full Cohort	Total (n = 97)	Healthy (n = 23)	IBD (n = 74)	P
Sex, n(%)				0.354
Female	34 (35.42)	10 (43.48)	24 (32.88)
Male	62 (64.58)	13 (56.52)	49 (67.12)
Age, M (Q1, Q3)	38.50 (29.00, 51.00)	41.000 (30.00, 45.00)	35.000 (28.00, 52.00)	0.770
BMI, M (Q1, Q3)	22.32 (20.19, 24.74)	22.85 (20.88, 25.08)	22.14 (20.00, 24.68)	0.334
Albumin(g/L), Mean ± SD	37.65 ± 5.11	35.22 ± 4.57	38.42 ± 5.06	0.008*
Cholesterol (mmol/L), M (Q1, Q3)	4.00 (3.30, 4.53)	4.90 (4.50, 5.65)	3.70 (3.10, 4.20)	< .001*
White Blood Cells (× 103/μL), M (Q1, Q3)	6.98 (5.11, 9.00)	7.53 (5.15, 9.37)	6.01 (5.05, 8.11)	0.111
Erythrocyte Sedimentation Rate(mm/h), M (Q1, Q3)	4.68 (1.64, 24.86)	3.18 (1.29, 5.66)	43.37 (25.21, 50.11)	< .001*
CRP(mg/L), M (Q1, Q3)	2.78 (1.28, 7.94)	1.92 (1.13, 3.61)	12.58 (9.70, 15.59)	< .001*
Fecal Calprotectin(µg/g), M (Q1, Q3)	103.98 (51.66, 172.06)	83.71 (43.82, 140.07)	255.36 (203.04, 367.54)	< .001*
Remission, n(%)			33(44.60)
Mild, n(%)			18(24.32)
Moderate, n(%)			17(22.97)
Severe, n(%)			5(6.76)
Unknown, n(%)			1(1.35)

SD: standard deviation, M: Median, Q₁: 1st Quartile, Q₃: 3st Quartile

Table 2.

Analysis of basic characteristics between healthy and IBD patients in 16S rRNA cohort

16S rRNA Cohort	Total (n = 57)	Healthy (n = 19)	IBD (n = 38)	P
Sex, n(%)				1.000
Female	21 (36.84)	7 (36.84)	14 (36.84)
Male	36 (63.16)	12 (63.16)	24 (63.16)
Age, M (Q1, Q3)	41.00 (31.00, 49.00)	38.00 (30.00, 42.50)	43.00 (32.25, 53.50)	0.150
BMI, M (Q1, Q3)	22.66 (20.00, 24.88)	22.70 (20.59, 25.07)	22.54 (19.78, 24.19)	0.498
Albumin(g/L), Mean ± SD	36.63 ± 5.03	37.70 ± 4.98	34.50 ± 4.52	0.022*
Cholesterol (mmol/L), M (Q1, Q3)	4.00 (3.30, 4.70)	4.80 (4.35, 5.65)	3.60 (3.10, 4.07)	< .001*
White Blood Cells (× 103/μL), M (Q1, Q3)	6.63 (4.86, 8.43)	6.39 (5.27, 8.11)	6.71 (4.80, 8.56)	0.847
Erythrocyte Sedimentation Rate(mm/h), M (Q1, Q3)	15.46 (2.30, 35.93)	3.65 (1.53, 13.08)	45.94 (29.38, 50.62)	< .001*
CRP(mg/L), M (Q1, Q3)	4.04 (1.77, 9.69)	2.29 (1.27, 3.98)	13.31 (9.31, 15.59)	< .001*
Fecal Calprotectin(µg/g), M (Q1, Q3)	133.17 (72.71, 213.58)	84.12 (54.32, 139.06)	255.36 (181.57, 367.54)	< .001*

SD: standard deviation, M: Median, Q₁: 1st Quartile, Q₃: 3st Quartile

Table 3.

Analysis of basic characteristics between healthy and IBD patients in RNA-seq cohort

RNA-seq Cohort	Total (n = 73)	Healthy (n = 19)	IBD (n = 54)	P
Sex, n(%)				0.492
Female	26 (35.62)	8 (42.11)	18 (33.33)
Male	47 (64.38)	11 (57.89)	36 (66.67)
Age, M (Q1, Q3)	38.00 (30.00, 51.00)	41.00 (33.50, 48.00)	33.50 (28.25, 51.50)	0.333
BMI, M (Q1, Q3)	22.14 (20.19, 25.24)	22.70 (20.59, 24.14)	22.08 (20.19, 25.53)	1.000
Albumin(g/L), Mean ± SD	37.86 ± 5.03	38.51 ± 5.16	36.02 ± 4.26	0.064
Cholesterol (mmol/L), M (Q1, Q3)	4.00 (3.20, 4.60)	4.90 (4.45, 5.60)	3.60 (3.10, 4.20)	< .001*
White Blood Cells (× 103/μL), M (Q1, Q3)	6.73 (4.97, 8.90)	6.01 (4.92, 8.36)	6.95 (5.09, 9.38)	0.285
Erythrocyte Sedimentation Rate(mm/h), M (Q1, Q3)	4.36 (1.64, 24.69)	3.04 (1.34, 5.42)	44.99 (26.66, 50.62)	< .001*
CRP(mg/L), M (Q1, Q3)	2.69 (1.27, 7.89)	1.76 (1.08, 3.24)	12.58 (9.70, 15.96)	< .001*
Fecal Calprotectin(µg/g), M (Q1, Q3)	89.28 (46.94, 168.82)	63.25 (36.29, 107.26)	255.36 (203.04, 342.90)	< .001*

SD: standard deviation, M: Median, Q₁: 1st Quartile, Q₃: 3st Quartile

Gut microbiota characterization and RNA sequencing

Fecal samples from 57 participants underwent 16S rRNA gene sequencing targeting the V3–V4 hypervariable region. Sequencing generated an average of 77,348 reads per sample, aligned against the SILVA reference database. Following rigorous quality control, an average of 42.86 Mbp of clean data per sample was retained, with a data utilization rate exceeding 99% across all samples. Raw reads were processed using cutadapt and FLASH to trim low-quality sequences and merge paired-end reads, yielding high-quality contigs (concatenation efficiency: 99.18–99.73%). Operational Taxonomic Units (OTUs)/Amplicon Sequence Variants (ASVs) were taxonomically annotated using the RDP classifier against the SILVA v138 database (confidence threshold: 0.6). After exclusion of non-bacterial taxa (e.g., Archaea) and unclassified sequences, 92.4% of OTUs were resolved to bacterial lineages.

Additionally, 72 samples were analyzed via the DNBSEQ platform for metagenomic profiling. The average genome-wide alignment rate was 97.89%, indicative of high mapping consistency. Metagenomic analysis revealed an average gene set alignment rate of 75.06%, with a total of 18,673 genes detected, reflecting comprehensive microbial functional coverage.

Distinct microbial composition in IBD patients when compared with healthy participants

The analysis of the OTU Rank Curve and species accumulation curve reveals that healthy individuals have higher microbial diversity than those with severe disease, indicating a correlation with disease severity. As OTU numbers increase, relative abundance declines across all groups, with the healthy group showing higher abundance at lower OTU counts (Fig. 2A). The species accumulation curve demonstrates that the number of detected OTUs rises with sample size, emphasizing the need for larger samples to identify additional species, especially in lower diversity groups. These findings support the use of microbial communities in disease diagnosis and treatment, particularly for assessing severity and monitoring therapeutic outcomes (Fig. 2B).

Fig. 2.

Gut Microbiota Diversity and Composition Across IBD Stages. A OTU rank curve illustrating the relative abundance of microbial diversity across varying states: Healthy, Mild, Moderate, Severe, and Remission. The curve highlights the decreasing microbial diversity with increasing IBD severity. B Species accumulation curve showing the incremental detection of OTUs as the number of sequenced samples increases. This curve demonstrates the robustness of the sequencing depth and its ability to capture microbial diversity across all stages. C Bar chart comparing the gut microbiota composition between healthy individuals and individuals with IBD. The relative abundance of major bacterial taxa is depicted, showing clear differences between the two groups. D Bar chart presenting the relative abundance of microbial taxa across different IBD conditions: Healthy, Remission, Mild, Moderate, and Severe. The chart highlights shifts in microbial community structure as the disease progresses, emphasizing the increasing dominance of certain taxa and the reduction of others

The species composition bar charts (Fig. 2C and D) highlight significant differences between healthy individuals and those with IBD. The healthy exhibited a distinct microbial signature characterized by significantly higher relative abundances of Bacteroidia and Clostridia, whereas the IBD group demonstrated a more heterogeneous taxonomic profile, with a marked increase in pathogenic Bacilli during the severe stage and Gammaproteobacteria during the moderate stage. These findings suggest that the onset and progression of IBD are associated with a dysbiosis of multiple microbial communities in the gut. Furthermore, stratification by disease severity (healthy → moderate → severe) revealed a progressive taxonomic rearrangement: a pronounced increase in Bacilli accompanied by a marked depletion of Fusobacteria in the severe IBD subgroup. These findings highlight a dynamic restructuring of the gut microbiome aligned with disease progression, underscoring potential diagnostic and therapeutic targets. The remission and mild disease groups exhibited microbiota profiles closely resembling those of the healthy cohort, suggesting a trajectory toward microbial restoration as symptoms abate. This observation underscores a critical association between gut microbiota composition and disease severity, wherein microbial community structure appears to mirror clinical status. Notably, the attenuation of dysbiotic features in milder cases highlights the dynamic interplay between the microbiota and host physiology, implying that targeted modulation of the gut ecosystem could influence disease progression. Collectively, these findings not only illuminate microbiota-driven mechanisms in IBD pathophysiology, but also reinforce the therapeutic potential of restoring microbial homeostasis, warranting further investigation into microbiota-based biomarkers and interventional strategies.

The analysis of the Alpha diversity box plots (Fig. 3 A–D) reveals that the Chao index (P = 0.002) and Shannon index (P = 0.024) are significantly higher in the healthy group compared to those with IBD, indicating that healthy individuals exhibit markedly greater microbial diversity than IBD patients. However, as the severity of the disease increases, a more pronounced declining trend is observed in the Chao index compared to the Shannon index, although the severe group does not appear to have the lowest Chao index value. The remission and mild groups display diversity levels similar to those of the healthy group, suggesting a potential recovery trend. These findings underscore the importance of microbial diversity in assessing gut health and disease severity, providing valuable insights for future therapeutic strategies. Figure 3E and F illustrate that the IBD group, as well as its various stages, exhibit higher β-diversity compared to the healthy control group, indicating a decrease in the stability of the gut microbial community structure in IBD patients.

Fig. 3.

Gut microbiota diversity analysis in healthy controls and ibd patients across disease stages. A Box plot showing the alpha diversity of gut microbiota between healthy individuals and IBD patients, measured using the Chao diversity index, indicating a decrease in species richness associated with disease progression (P = 0.002). B Box plot illustrating the alpha diversity between healthy individuals and IBD patients, calculated using the Shannon diversity index (P = 0.024). C Alpha diversity across all stages of IBD and healthy controls, calculated using the Chao diversity index. This box plot reveals a decrease in species richness across disease stages, P = 0.024. D Alpha diversity of gut microbiota across all stages of IBD and healthy controls, measured using the Shannon diversity index. A decline trend in diversity is observed with increased disease severity (P = 0.031). E Beta diversity group difference box diagram of healthy people and IBD patients calculated by Weighted unifrac, p < 0.001. F Beta diversity group difference box diagram of all stages calculated by Weighted unifrac, p < 0.001

The analysis combining the LEfSe clustering tree and LDA scores reveals significant changes in microbial communities across healthy, remission, mild, moderate, and severe states (Fig. 4, and supplementary Figure 1). Specific bacteria, such as Lactobacillus and Bifidobacterium, exhibit increased abundance in the remission state, potentially correlating with health. In contrast, the abundance of Peptoniphilus, Parabacteroides, Faecalibacterium, and Ruminococcus in mild and moderate states may reflect early disease progression and exacerbation of dysbiosis. In the severe state, the significant enrichment of Bacilli and Lactobacillaceae may be associated with disease severity and inflammatory responses. These changes in microbial abundance may serve as potential biomarkers for disease states, aiding in the assessment of disease progression and providing insights for future intervention strategies, such as the use of probiotics or dietary modifications to regulate the microbiome and improve patient health outcomes.

Fig. 4.

Stage-specific microbial biomarkers revealed by LEfSe analysis. A–D Cladograms showing LEfSe-based clustering trees of microbial biomarkers across different IBD stages compared to healthy controls. Each panel highlights the distinctive taxa enriched in a specific disease stage, with bacterial taxa represented by nodes and branches

The microbial distribution in the IBD group is more diverse, particularly with an increased abundance of Bacilli and Gammaproteobacteria, suggesting that microbial dysbiosis is associated with inflammation.

Identification of microbial biomarkers at different stages of IBD using NetMoss

The network diagram illustrates the changes in microbial communities across different disease severities, highlighting the relationship between microbial diversity and health status (Fig. 5A–D). Using the NetMoss method, we identified key microbial taxa with scores greater than 0.9 at various stages. We found that bifidobacterium.catenulatum and bacteroides.fragilis were identified as biomarkers distinguishing patients with IBD in remission, streptococcus.gallolyticus, veillonella.atypica, and clostridium.butyricum were biomarkers distinguishing patients with mildly active IBD, blautia.obeum was identified as a biomarker distinguishing patients with moderately active IBD, bacteroides.uniformis was identified as a biomarker distinguishing patients with severely active IBD (Fig. 5E–H).

Fig. 5.

microbial co-occurrence networks and NetMoss-derived biomarkers across IBD stages. A–D Network maps illustrating co-occurrence relationships among microbial species at different stages of IBD, including Remission (A), Mild (B), Moderate (C), and Severe (D) stages. Nodes represent bacterial taxa, and edges indicate co-occurrence links, with line thickness proportional to the strength of the association. Taxa are color-coded by phylum: Bacteroidota, Actinobacteriota, Firmicutes, Proteobacteria, Fusobacteriota, and others. The network complexity and composition vary significantly between stages, reflecting stage-specific microbial interactions and dysbiosis progression. E–H Top five bacterial taxa identified as biomarkers for each stage of IBD based on NetMoss scores, with higher scores indicating a stronger relevance to the disease stage. Scores greater than 0.9 are considered significant

Transcriptional signatures and staging of IBD revealed through peripheral blood profiling

Transcriptomic analysis of peripheral blood samples identified 161 genes exhibiting significant upregulation in IBD patients relative to healthy controls (log² fold change ≥ 0.5, FDR < 0.05; Fig. 6A). Hierarchical clustering of these differentially expressed genes (DEGs) revealed distinct transcriptional patterns between healthy participants and IBD patients (Fig. 6B). To dissect stage-specific transcriptional signatures, we implemented a dual-algorithm approach combining Boruta feature selection and REF-SVM classification. This framework enabled systematic prioritization of genes whose expression dynamics most discriminately characterize disease progression and remission states.

Fig. 6.

Transcriptomic signatures of IBD staging. A Volcano plot illustrating 161 upregulated genes identified through differential expression analysis. B Heatmap depicting hierarchical clustering of DEGs between healthy participants and IBD patients. C-F Stage-specific gene expression patterns identified via Boruta (left panels) and REF-SVM (right panels, only top 10 were presented): C remission (GIMAP6, YIPF4), D mild (FKBP5, HBB), E moderate (TUBB1), and F severe (ALAS2, MMRN1, IGFBP2). The genes highlighted in red are those that are highly expressed only in this stage

The Gene Ontology (GO) analysis results demonstrate the biological processes associated with different stages of IBD. In remission, terms such as “cell activation” and “wound healing” are prominent, indicating partial restoration of immune homeostasis (Supplementary Figure 2A). In mild to moderate stages, immune-related processes become more pronounced (Supplementary Figure 2B–C). The severe stage shows enhanced activation of immune responses, including “granulocyte activation” and “regulated exocytosis”, reflecting heightened immune dysregulation (Supplementary Figure 2D). Similarly, Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis showed that key pathways in remission involve “Chemokine signaling pathway” and “Platelet activation”, reflecting immune response and coagulation process activation (Supplementary Figure 2E). In the mild stage, pathways such as “Chemokine signaling pathway” begin to emerge, indicating the onset of inflammation (Supplementary Figure 2F). The moderate stage highlights pathways associated with innate immunity, such as “NOD-like receptor signaling pathway” and “Complement and coagulation cascades”, suggesting an intermediate inflammatory response (Supplementary Figure 2G). Severe IBD is marked by enrichment in pathways such as “Platelet activation”, “ECM-receptor interaction”, and “Transcriptional misregulation in cancer”, indicating extensive tissue remodeling and immune dysregulation (Supplementary Figure 2H).

Notably, remission-stage samples were characterized by elevated expression of YIPF4 and GIMAP6, which exhibited robust discriminative power (Fig. 6C). Mild activity was associated with upregulation of FKBP5 and HBB, while moderate activity featured overexpression of TUBB1 (Fig. 6D, E). Severe disease demonstrated a unique signature dominated by ALAS2, MMRN1, and IGFBP2 (Fig. 6F).

The integrative analysis further revealed dynamic transcriptional reprogramming across the disease continuum, where genes implicated in cytoskeletal remodeling (TUBB1), erythroid differentiation (ALAS2), and immune regulation (GIMAP6) emerged as central nodes in stage-specific networks. These findings underscore the utility of blood-derived transcriptional markers for non-invasive IBD stratification.

Multi-omics integration identifies robust biomarkers and predictive models for IBD staging

Multi-omics integration leveraging 16S rRNA sequencing, RNA-seq, and clinical metadata revealed a dual-axis framework for precision staging of IBD. Microbiome profiling via NetMoss algorithm identified seven stage-specific bacterial taxa with distinct abundance patterns across remission, mild, moderate, and severe disease phases. In a paired-omics subset (n = 40), cross-correlation analysis demonstrated significant associations between microbial dysbiosis indices and host transcriptional signatures (Fig. 7A–B), suggesting bidirectional communication between gut microbiota and immune-metabolic pathways during IBD pathogenesis.

Fig. 7.

Identification and Evaluation of Microbial and Transcriptomic Biomarkers for IBD Staging. A Correlation matrix displaying pairwise relationships between key microbial species across IBD stages. The colour gradient represents Spearman correlation coefficients, with darker blue indicating stronger negative correlations. Microbiota and genes from the same stage are marked with red boxes. B Heatmap showing the correlationtrends between selected bacterial taxa (rows) and stage-specific transcriptomic markers (columns) identified through Boruta feature selection and REF-SVM analysis. The colour gradient represents the expression levels, ranging from low (blue) to high (red). ^*: p < 0.05. C-D ROC curves for classifiers trained on different stages of IBD with stacking model, using microbial features (C) and combined transcriptomic-microbial features (D). Class 0: healthy. Class 1: remission. Class 2: mild. Class 3: moderate. Class 4: severe. The pink dashed lines depict the micro-average AUC, reflecting the model’s overall performance across all samples

To operationalize these findings, we developed a hybrid machine learning pipeline integrating microbial and clinical features. Seven candidate algorithms—Logistic Regression, Random Forest, Support Vector Machine, Decision Tree, K-Nearest Neighbors, XGBoost, and LightGBM—were benchmarked on the 16S rRNA cohort (n = 57). A meta-ensemble model combining these classifiers through stacking achieved robust stratification performance (AUC = 0.79, Fig. 7C). Parallel analysis of RNA-seq data identified eight host genes whose expression trajectories mirrored disease progression, with a comparable stacking ensemble yielding AUC = 0.80 (Fig. 7D).

Discussion

This study advances the understanding of IBD pathogenesis through a multi-layered omics framework that integrates gut microbiome profiling, host transcriptomic characterization, and machine learning-driven predictive modeling. By applying the NetMoss algorithm for microbial feature extraction and implementing rigorous cross-omics feature selection, we discovered distinct microbial and transcriptional signatures that exhibit stage-specific stratification across the IBD disease continuum—from remission to severe inflammation. Notably, the identified biomarkers demonstrate dual utility: serving as diagnostic indicators of disease severity and providing mechanistic insights into microbiota-host interplay during IBD progression. For instance, microbial dysbiosis patterns were found to correlate with dysregulated immune-metabolic pathways at the transcriptional level, revealing potential therapeutic targets at the host-microbe interface. To translate these biological insights into actionable tools, we developed an ensemble machine learning pipeline that harmonizes microbial, transcriptomic, and clinical features. This model achieves robust classification performance in distinguishing IBD stages, outperforming single-omics approaches and highlighting the synergistic power of multi-modal data integration. Collectively, these findings establish a paradigm for precision IBD staging and pave the way for stratified therapeutic strategies that address both microbial and host-driven disease mechanisms.

Our findings highlight distinct microbial and transcriptional landscapes across IBD stages, underscoring the characteristic changes in microbiota and gene expression associated with each stage of the disease. Healthy controls exhibited significantly higher alpha diversity and enriched taxa such as Clostridia and Bacteroidia, consistent with a balanced microbiota-immune axis. Certain strains within the Clostridia family, such as clostridium.butyricum, are renowned for their ability to produce butyrate, a short-chain fatty acid vital for the health of intestinal epithelial cells and essential in regulating immune responses [21]. Bacteroides.fragilis is an important strain within the Bacteroidia, not only serving as a vital component of human colonic symbiotic microorganisms but also playing a role in regulating gut immune responses [22]. In contrast, IBD patients demonstrated reduced diversity and dysbiosis marked by increased Bacilli, Fusobacteriia and Gammaproteobacteria, correlating with inflammatory burden [23, 24]. However, conventional alpha/beta diversity metrics and LEfSe-identified taxa exhibited limited specificity, as overlapping distributions between remission/mild groups and interindividual variability obscured clear diagnostic boundaries [25]. These limitations highlight the challenges of relying solely on relative abundance or diversity indices for precise staging. While 16S rRNA analysis provides foundational insights into ecological disturbances, its standalone clinical utility remains constrained by noise and redundancy. Integrating NetMoss-derived biomarkers with transcriptional and clinical data offers a robust framework for precision staging, bridging the gap between microbial ecology and host pathophysiology in IBD management [19].

This study pioneers the application of the NetMoss algorithm for IBD staging, introducing a network-centric paradigm that deciphers stage-resolved microbial interaction patterns and co-occurrence architectures. By transcending conventional diversity metrics, which provide limited insight into community structure, NetMoss quantifies topological rearrangements within microbial ecosystems, enabling precise characterization of dysbiosis dynamics across disease progression. Through NetMoss-driven network deconvolution, we identified stage-discriminative microbial biomarkers with exceptional specificity. Notably, the remission-stage enrichment of bifidobacterium.catenulatum and bacteroides.fragilis suggests their role in fostering microbial resilience during clinical recovery, potentially via immunomodulatory metabolite secretion. These network-derived biomarkers were validated through a multi-model machine learning framework, achieving robust classification performance. This outperformed conventional diversity-based stratification, underscoring the superiority of network-based approaches in resolving complex microbial interdependencies.

Transcriptomic profiling identified GIMAP6 and YIPF4 as remission-specific biomarkers, corroborating their established roles in immune modulation and cellular homeostasis as documented in prior investigations [26, 27]. Conversely, mild activity was marked by FKBP5, a glucocorticoid receptor co-chaperone linked to steroid resistance, and HBB, indicative of compensatory erythropoiesis triggered by chronic inflammation [28, 29]. The moderate-stage dominance of TUBB1, a tubulin isoform involved in cytoskeletal remodeling, aligns with enhanced platelet activation and endothelial dysfunction observed in active IBD [30, 31]. Severe disease was characterized by ALAS2, a heme biosynthesis enzyme upregulated during anemia of chronic disease, alongside MMRN1 (platelet α-granule protein) and IGFBP2, an oncogenic driver of epithelial-mesenchymal transition, reflecting systemic inflammation, coagulopathy, and tissue repair dysregulation [32–34]. By integrating these findings into a predictive framework, our study addresses the limitations of previous research, which often focused on isolated omics layers without exploring their combined diagnostic potential [35, 36].

Incorporating GO and KEGG pathway analyses further contextualized these findings: moderate-stage enrichment of NOD-like receptor and chemokine signaling pathways underscores aberrant inflammasome activation and leukocyte recruitment [37, 38], while severe disease-associated secretory vesicle exocytosis and neutrophil activation highlight unbridled release of proteolytic enzymes and neutrophil extracellular traps (NETs), perpetuating mucosal damage [39, 40]. Remission-phase enrichment of ECM-receptor interaction pathways implicates fibroblast-mediated tissue remodeling in inflammation resolution, whereas mild-stage stress response pathways (e.g., HIF-1 signaling) suggest hypoxia-driven adaptation to early inflammatory insults [41, 42].

The innovative integration of multi-omics datasets within a machine learning architecture represents a critical advancement over previous IBD research paradigms. By harmonizing microbial metagenomic and host transcriptomic signatures, our framework enables a systems-level understanding of IBD pathogenesis, simultaneously resolving localized intestinal dysbiosis and systemic immune perturbations. This dual-pronged approach achieves unprecedented precision in disease stratification while generating mechanistic insights into dynamic host-microbiota crosstalk, offering translational opportunities for microbiome-targeted interventions.

The machine learning framework employed in this study utilized seven classical and advanced algorithms, including LR, RF, SVM, DT, KNN, XGBoost, and LightGBM. Among these, a stacking ensemble model integrating predictions from all base models achieved the highest performance, with an AUC of 0.79 and 0.80 for microbial and transcriptomic datasets. The integration of microbial and transcriptomic features further implies the predictive power of the two models. In a subset of participants with paired 16S rRNA and RNA-seq data, we observed significant correlations between microbial abundance and transcriptional profiles, suggesting a mechanistic interplay between gut microbiota dysbiosis and host immune reprogramming [43, 44]. This multi-omics approach not only improved staging accuracy but also provided insights into the biological underpinnings of IBD progression. This finding is consistent with previous studies, one study developed a microbiome-based diagnostic test by analyzing 5,979 stool samples from individuals with different geographic and racial backgrounds, identifying bacteria associated with IBD and constructing diagnostic models, which achieved an AUC exceeding 0.90 in the discovery cohort for distinguishing IBD from controls and maintained good performance in validation cohorts from eight races [45]. Additionally, another study found significant differences in the microbiome and metabolites between IBD patients and healthy individuals by analyzing the fecal microbiome and metabolites of 183 subjects, indicating that combining assessments of metabolites and fecal microbiome can effectively differentiate between healthy individuals and IBD patients [46]. A prospective pediatric cohort study evaluated the diagnostic utility of fecal calcium-binding protein (FCBP) in 53 children presenting with suspected IBD. Through correlation analysis of FCBP levels with endoscopic and histopathological findings, the investigation demonstrated that FCBP measurement exhibited robust discriminative power in distinguishing IBD from non-IBD inflammatory gastrointestinal conditions (P < 0.001), outperforming conventional fecal biomarkers in this pediatric population [47]. While prior investigations have collectively underscored the diagnostic promise of microbiome-related biomarkers in IBD, methodological limitations remain evident. Specifically, these studies frequently exhibit insufficient granularity to resolve disease-specific stages and overlook the integration of complementary multi-omics datasets, which are critical for capturing the multidimensional nature of IBD pathogenesis. This study systematically implemented an integrated multi-omics framework that combines microbial and host transcriptomic profiling. This innovative approach enables high-resolution identification of stage-specific biomarker signatures while demonstrating comparable or superior diagnostic performance compared to conventional strategies. Central to this advancement is the deployment of sophisticated feature selection algorithms, including Boruta and REF-SVM, which prioritize biologically meaningful features through rigorous statistical validation. These algorithms inherently mitigate noise propagation and enhance model generalizability, contrasting sharply with prior studies that frequently employed less robust univariate methods or heuristic approaches, which may compromise feature stability and clinical translatability.

Limitations

This study has several limitations. First, the sample size was relatively small, particularly for the paired multi-omics cohort, which may limit the generalizability of our findings. Future studies with larger, multi-center cohorts are needed to validate our results. Second, the use of 16S rRNA sequencing restricts taxonomic resolution to the genus level, potentially overlooking strain-specific differences that may be critical for IBD pathogenesis. Metagenomic and metatranscriptomic approaches could address this limitation in future research. Third, while our predictive model demonstrated promising performance, its applicability to diverse populations and IBD subtypes remains to be tested.

Conclusion

This study presents a pioneering integrated multi-omics analytical framework that elucidates the dynamic crosstalk between gut dysbiosis and host immune responses during the progression of IBD. By decoding actionable microbial-transcriptomic biomarker signatures, we have constructed a molecular bridge linking localized intestinal microenvironment perturbations to systemic immune reactions, thereby establishing a translational medicine roadmap for precision-stratified diagnosis and personalized IBD therapies. Notably, the stage-specific biomarkers identified in this study not only reflect the pathophysiological characteristics of the disease but also hold promise as therapeutic targets. Future validation in large-scale, multicenter cohorts may facilitate the clinical adoption of non-invasive multi-omics-based diagnostic tools, poised to revolutionize the monitoring and management paradigms of this complex chronic condition, marking a paradigm shift from empirical to stratified IBD care.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

UC

Area under the curve

ASVs

Amplicon sequence variants

BMI

Body mass index

CD

Crohn’s disease

CDAI

Crohn’s disease activity index

CMV

Cytomegalovirus

DEGs

Differentially expressed genes

DNB

DNA nanoball

DT

Decision tree

GLM

Generalized linear model

GO

Gene ontology

IBD

Inflammatory bowel disease

KEGG

Kyoto encyclopedia of genes and genomes

KNN

K-nearest neighbors

LGBM

Light gradient boosting machine

LR

Logistic regression

NETs

Neutrophil extracellular traps

NetMoss

Network-based microbial composition similarity score

OTUs

Operational taxonomic units

PCR

Polymerase chain reaction

RCA

Rolling circle amplification

RF

Random forest

RNA-seq

RNA sequencing

ROC

Receiver operating characteristic

REF-SVM

Recursive feature elimination support vector machine

SVM

Support vector machine

UC

Ulcerative colitis

XGB

EXtreme gradient boosting

Author contributions

YT designed the study and provided academic supervision, LFW and PL implement the study. YT, LFW conceived the analysis and drafted the paper, YT, LFW and PL contributed substantially to study conception, analysis and interpretation of the data. JJJ and XNZ monitored data during the period of implementation. PL provided clinical supervision. RS, YJL, LJZ, MHS and LHY are investigators and contributed to data acquisition. RS and YJL disposed data. XNZ, JJJ, and YT had full access to all the data in the study and verified the data. XNZ revised the manuscript, All authors read critically for important intellectual content and approved the final version of the manuscript for publication.

Funding

This study was funded by National Natural Science Foundation of China (National Natural Science Youth Fund) (82404767).

Availability of data and material

Deidentified individual participant data from this study cannot be shared publicly due to the sensitive nature of the data and the terms outlined in the informed consent. Data requests can be made via this email: taoziyi77@hotmail.com.

Declarations

Consent for publication

All authors agreed to the content of the manuscript and its submission for publication in Nature Communications.

Competing interests

All authors declare that they have no competing interests. Generative AI was not used in any portion of the manuscript writing.