Multi-omics analysis reveal clinical-gut-brain interactions in female ibs patients with adverse childhood experiences

Premenopausal female participants with IBS or HCs were recruited by the G Oppenheimer Center for Neurobiology of Stress and Resilience at University of California Los Angeles (UCLA). All IBS patients were evaluated by a gastroenterologist or nurse practitioner with expertise in IBS for presence of a Rome IV diagnosis of IBS [23]. Rome IV diagnostic criteria for IBS is the current international standard for symptom-based diagnosis of IBS, developed by a multinational expert consensus including. The criteria require the presence of recurrent abdominal pain, on average, at least 1 day per week in the last 3 months, associated with at least two of the following: pain related to defecation, associated with a change in stool frequency, or associated with a change in stool form. The symptom onset must have occurred at least 6 months before diagnosis. Participants were excluded for the following reasons: pregnant or lactating, substance use, abdominal surgery, tobacco dependence (half a pack or more daily), extreme strenuous exercise (> 8 h of continuous exercise per week such as marathon runners), and major medical or neurological conditions. Participants taking medications that interfere with the central nervous system (unless on a steady dose for more than 3 months) or regular use of analgesic drugs were excluded. To avoid potential cofounders in microbiome analyses, included participants were required to not have taken antibiotics for at least 3 months and probiotics for at least 1 month prior to enrolling in the study. Only premenopausal females were enrolled and were scanned during the follicular phase of their menstrual cycles as determined by the self-report of their last day of the cycle. No participants exceeded 400lbs due to MRI scanning weight limits.

Participants underwent multimodal brain-imaging studies at UCLA and provided fresh stool samples for 16 s ribosomal RNA gene sequencing collected within a week of the scan. All procedures complied with the principles of the UCLA Institutional Review Board and informed consent was obtained from all participants.

Multiple validated questionnaires were used to measure baseline clinical and behavioral characteristics. We utilized multiple complementary constructs, including measures of general health, somatic symptoms, visceral sensitivity mood, anxiety, stress, and IBS-related symptoms and quality of life, as relevant to a multidimensional analysis of IBS and ACE exposure.

Participants’ stool samples were collected using standardized procedures to ensure sample integrity and minimize contamination. To avoid cross-contamination, participants were instructed to collect the sample without allowing contact with urine or toilet water. Following collection, the samples were immediately sealed, labeled with the participant’s unique study identifier, and stored at 4 °C until transportation to the laboratory. Samples were processed within 24 hours of collection, aliquoted into sterile cryovials, and stored at −80°C for subsequent microbiome analyses.

DNA from stool was extracted using the DNA Fecal Microbe Miniprep Kit (Zymo Research). The V4 region of 16S ribosomal RNA was amplified and underwent paired end sequencing on an Illumina HiSeq 2500. Sequences were processed through the DADA2 pipeline to generate exact amplicon sequence variants (ASVs) and taxonomy was assigned based upon the SILVA 138 database. Predicted metagenomics was performed using PICRUSt2 in QIIME2 with the default settings to predict abundances of bacterial gene families annotated as KEGG orthologs (KO) based on nearest reference genomes to 16S sequences.

Acquisition: Whole brain structural and resting-state scans were acquired at the Ahmanson-Lovelace Brain Mapping Center on a 3.0 Tesla Siemens Prisma MRI Scanner (Siemens, Erlangen, Germany). Comprehensive information on the standardized acquisition protocols, quality control measures, and image preprocessing can be found in previously published studies [41–43]. High-resolution T1-weighted images were obtained with the following parameters: echo time/repetition time (TE/TR) = 3.26 ms/2200 ms, field of view = 220 × 220 mm, slice thickness = 1 mm, 176 slices, a 256 × 256 voxel matrix, and a voxel size of 0.86 × 0.86 × 1 mm. Whole-brain resting-state images were acquired with participants’ eyes closed using an echo-planar sequence with the following parameters: echo time/repetition time (TE/TR) = 28 ms/2000 ms, flip angle = 77°, scan duration = 10 minutes 6 seconds, field of view (FOV) = 220 mm, 40 slices, and a slice thickness of 4.0 mm. A diffusion weighted image was acquired to assess white matter anatomical connectivity (64 noncollinear directions, b = 1000 s/mm², 9 b = 0 s/mm² images, TR: 9500ms, TE: 88ms, field of view: 2304 x 2304, acquisition matrix: 128 x 128, slice thickness: 2 mm, spacing between slices: 2 mm).

Structural Processing: Cortical reconstruction and volumetric segmentation was done using FreeSurfer 7 [44]. FreeSurfer-processed structural images were parcellated using the Destrieux cortical atlas [45], Harvard-Oxford subcortical atlas [46] and Ascending Arousal atlas [47–49], and values of cortical thickness, surface area, mean curvature and volume for cortical regions, and volume for subcortical regions, were extracted.

Functional Image Processing. All functional images were preprocessed using a pipeline for volume-based resting-state functional connectivity (rs-FC) analyses in CONN [50]. All functional images underwent realignment and unwarping, slice-timing correction, and ART-based identification of outlier scans for scrubbing. Using structural image data, functional images were normalized and segmented into grey matter, white matter, and CSF tissue [51]. Functional images were then denoised by using ordinary least squares (OLS) regression of potential confounders and temporal band-pass filtering. Specifically, we used the default anatomical component-based noise correction procedure (aCompCor), which includes noise components from white-matter, cerebrospinal fluid [52], estimated subject-motion parameters [53], scrubbing of outlier scans based on framewise displacement [54], and removal of potential ramping effects at the start of the session [55]. A temporal band-pass filter of 0.008–0.09 Hz after regression was used to minimize the influence of physiological and head motion, and other noise sources [56]. Fisher-transformed correlations (Z) between the functional time series of each pair of regions, for all parcellated regions, were computed in CONN to derive a 165 x 165 matrix for each participant. The bottom half of the undirected matrix was then concatenated into one vector for each participant, representing the resting-state functional connectivity between every ROI pair.

Diffusion Processing: All diffusion weighted images were corrected using DiPy’s Median Otsu algorithm for skull stripping and denoising. Each scan underwent translation, rigid, and affine registration using the aforementioned atlases. The resulting homography was also used to transform the b-vectors and b-values. Finally, diffusion tensors were fit, and fractional anisotropy was calculated. This process resulted in a 188x76 matrix, with each row representing a participant and each column representing a diffusion feature.

To capture the complex relationship between ACEs and IBS, the model integrated multiple omics and behavioral assessments through Stacked Ensemble Modeling [57–59], with Extreme Gradient Boosting (XGBoost) as base models and a Support Vector Machine (SVM) as the meta-learner. Given the variability in range and dimensionality across the 5 datasets, hyperparameters such as L1 and L2 regularization were optimized for each omic. Regularization penalized redundant or weak predictors, enhancing interpretability and reducing overfitting. Importantly, decision tree-based methods, including XGBoost, are inherently robust to multicollinearity because splits are determined by maximizing information gain rather than relying on linear correlations, thereby minimizing bias introduced by overlapping constructs.{Piramuthu, 2008 #4051} This approach ensured that overlapping behavioral measures did not disproportionately influence the models while retaining complementary predictive information.

The data underwent 5-fold cross validation. Each iteration involved an 80% training set (n=150 or 151) and a 20% testing set (n=37 or 38). Within each training fold, the data was further partitioned into five sub-folds to facilitate meta-model training and hyperparameter optimization. During the training of each base model, optimal training parameters were identified through an iterative tuning process, ensuring the selection of parameters that maximized model performance prior to their application in the final base models.

Each data block was examined for response variables that could disrupt the analysis. From the clinical dataset, columns ACE Total Score and Group (IBS or HC) were removed as they were directly related to the response variable. IBS-QOL and IBS-SSS metrics were only measured for participants labeled as IBS and were imputed with a 0 for HCs. For the metagenomic dataset, to mitigate skewed distributions of bacterial counts, log transformation was applied, and outputs were then z-normalized across subjects for each bacterial strain. Through the base-model training step of the architecture, the features were reduced in the following way: 52 clinical variables down to 50, 2661 metagenomic variables down to 183, 658 structural brain features down to 168, 76 diffusion fractional anisotropy (dti.fa) features down to 60, and 15753 resting-state functional features (rspw) down to 111.

Gain analysis was used to identify the features that contributed the most to the model’s loss reduction and were visualized through bar plots. To further elucidate the associations between each high-gain variable and subject classes, SHAP and SHAP dependence analysis were utilized. Associations visualized through tile plots, highlighting the contribution of selected features to each subject class [60]. To report significant pairwise Pearson correlations between each data modality, connectograms were used to depict the relationships between features with correlation r >=0.4. For reader interpretability purposes, only the top variables responsible for 30% of each base model’s gain are provided in all visualizations.

To assess the predictive performance of our model, we report the Cohen’s kappa coefficient as a measure of agreement between predicted and actual classifications [61]. Kappa values serve as more interpretable alternatives to error rate in the context of multiclass classification, though for further interpretability, simple accuracies were also calculated and reported.

We also utilized the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC) metric. Since our task involved multiclass classification, we employed a one-vs-rest (OvR) approach, where an ROC curve was generated for each class against the remaining classes. An ROC curve was also generated for the meta model.

To identify clinical measures that demonstrated statistically significant differences between groups, analyses of variance (ANOVA) and chi-squared tests were applied. To further elucidate between-group differences, results of pairwise linear contrasts were also calculated.

The overall model exhibited an estimated kappa of 0.67, signifying high levels of agreement between the true and predicted labels.

The overall model yielded a ROC with an AUC of 0.70 for identifying HC, high ACE participants; 0.82 for IBS, low ACE participants; and 0.87 for IBS, high ACE participants. The accuracy for all base models were also calculated at 78.19%.

In descending order of importance, the top clinical variables included were VSI Score, PHQ-15 Score, ACE Parental Divorce/Separation, and IBS QOL Social Reaction. All three high-gain variables were related to GI symptoms, and their SHAP values were strongly associated with patient classes exhibiting IBS (Table 3, Fig 2a).

Bacterial transcriptomes in order of importance are listed in Table 3, and their associations to each of the 4 groups are depicted in Fig 2b. For example, the most significant bacteria contributing to the model included Akkermansia, Bifidobacterium, Intestinimonas, Subdoligranulum, Christensenella, and Burkholderia.

Structural, diffusion, and resting-state pairwise brain signatures identified as the most significant for prediction of the 4 participants groups are listed in Table 3 in descending order. Their contributions to the prediction of each participant group are depicted in Figures 2c, 2 d, and 2e, respectively. The key brain regions contributing to the model included those from the central autonomic, (CAN), salience (SAL), sensorimotor (SMN), default mode (DMN), and emotion regulation (ERN) networks.

The greater levels of depression, anxiety, GI symptom-related anxiety (VSI), perceived stress (PSS), somatic symptom severity (PHQ-15), and poorer physical and mental health scores (SF-12) observed in IBS participants compared to healthy controls align with the biopsychosocial model of IBS. This model suggests that central dysregulation of affective and pain circuits plays a role in symptom expression [62].

Although ACEs have been associated with greater symptom severity in IBS, the present study only found trends toward higher IBS-SSS and lower QoL in IBS participants with high ACE scores. The absence of statistical significance may indicate that ACEs primarily influence extraintestinal domains such as psychological distress and stress sensitivity. This is supported by statistically significant stronger differences in somatic symptom burden.

Further, our results suggest that the influence of ACEs may be partially mediated by heightened GI symptom-related anxiety (VSI), rather than by direct physiological alterations detectable through neuroimaging or microbiome analyses. In high ACE participants, GI symptom-related anxiety (VSI) was positively correlated with connectivity between the postcentral sulcus and parahippocampal gyrus (R PosCS to L PaHipG), a region involved in interoceptive processing and perception of internal bodily sensations. Individuals with high ACE scores may exhibit heightened interoceptive awareness, amplifying their perception and anxiety related to gastrointestinal symptoms. The parahippocampal gyrus, part of the limbic system, plays a crucial role in emotional processing and stress responses. Increased connectivity in this region may reflect an exaggerated emotional response to visceral sensations, further contributing to gastrointestinal symptom-related anxiety. Studies have demonstrated that altered resting-state functional connectivity in regions responsible for sensory processing and emotional regulation is linked to visceral hypersensitivity in IBS patients [63]. This hypersensitivity, which can be exacerbated by early life stress, may lead to increased anxiety surrounding gastrointestinal symptoms.

ACE-induced chronic stress disrupts the brain-gut-microbiota axis, leading to long-term alterations in gut microbiota composition, reduced microbial diversity, and a shift in the balance between beneficial and potentially pathogenic bacteria [64–66]. This dysbiosis contributes to immune dysregulation through altered pro-inflammatory and anti-inflammatory cytokines, fostering the chronic low-grade inflammation characteristic of IBS [66, 67]. In our cohort, IBS participants with high ACE scores demonstrated distinct microbial signatures, notably a negative correlation of beneficial genera Akkermansia and Bifidobacterium with somatic symptom severity (PHQ-15) and ACE parental divorce/separation, respectively. Akkermansia is linked to gut barrier integrity, immune regulation, and pain reduction [68, 69]. Previous studies have shown that treatment with pasteurized Akkermansia in IBS mouse models improved anxiety-like behavior, reduced colonic hypersensitivity,and reinforced intestinal barrier function [70], suggesting its potential role in modulating the gut-brain axis and reducing inflammation as a therapeutic approach. Additionally, as Bifidobacterium is a well-documented genus shown to improve intestinal permeability [71], alleviate IBS symptoms [72], reduce inflammatory cytokines (e.g., IL-6 and TNF-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}) [73], and modulate the gut microbiota to increase the abundance of beneficial bacteria and short-chain fatty acids (SCFAs) that are crucial for gut health [74], its depletion in high ACE IBS participants suggests a potential mechanism linking early life adversity to somatic symptom burden in IBS. This is consistent with prior research highlighting chronic stress from ACEs as exacerbating psychological distress and contributing to visceral hypersensitivity, a hallmark of IBS [75, 76].

Early-life stress has been shown to contribute to central nervous system dysregulation through maladaptive neuroendocrine responses to stress, which manifest as gastrointestinal symptoms [77]. Additionally, IBS patients with a history of childhood trauma exhibit increased visceral hypersensitivity and altered brain responses to pain [78]. Our neuroimaging findings revealed significant associations between ACEs, IBS, and alterations in key neural circuits, including the salience network (SAL), sensorimotor network (SMN), default mode network (DMN), central autonomic network (CAN), and central executive network (CEN), which all play crucial roles in stress and emotional regulation and pain processing. In high ACE IBS participants, Bifidobacterium was negatively linked to CAN/SAL. Hyperconnectivity in the salience network is linked to hypervigilance and hyperalgesia in IBS [63] while CAN regulates stress responses by integrating physiological and emotional signals across brain regions to modulate autonomic, neuroendocrine, and behavioral adaptations [79]. Lower levels of Bifidobacterium may exacerbate these network dysfunctions, contributing to more severe IBS symptoms. In contrast, in low ACE IBS participants, there was a positive association between medial occipito-temporal (collateral) sulcus and lingual sulcus to calcarine sulcus and the CEN, which is involved in high-level cognitive functions [80]. This may suggest a potential compensatory mechanism within cognitive control regions to help manage IBS symptoms. Additionally, Subdoligranulum, a butyrate-producing beneficial bacterium, was negatively linked to the DMN, highlighting a possible role of microbiome-brain interactions in modulating pain and emotional processing in low ACE IBS participants.

In high ACE HC participants, salience (SAL) connectivity was negatively associated with beneficial gut microbial genera, including Akkermansia, Subdoligranulum, Intestinimonas, and Christensenella, suggesting a potential link between gut microbial composition and stress-related brain network activity. The salience network is crucial for detecting and integrating salient stimuli, including pain and emotional stress. In IBS patients, particularly those with a history of ACE, increased SAL connectivity likely contributes to heightened pain perception, emotional dysregulation, and hypervigilance [81, 82]. Additionally, Intestinimonas, a butyrate producing bacteria, was positively associated with the subcentral sulcus and sulci (G and S subcentral) part of the sensorimotor network connectivity. As the subcentral area encompasses parts of the primary motor and somatosensory cortices, these connectivity changes may relate to visceral sensorimotor integration and interoception, which may contribute to a latent risk for IBS. Individuals with a history of ACEs are twice as likely to have IBS than those without an ACE [6].

In summary, IBS and ACE can influence key neural circuits part of a complex interplay between early-life adversity and gut microbiota in shaping IBS pathophysiology, emphasizing the need for a more integrative approach to understanding and managing this disorder. Despite the lack of significant differences in symptom severity between IBS participants with high and low ACEs, we found unique microbiome and brain signatures. These signatures may be relevant to personalized medicine/treatment response differences. Additionally, it is possible that resilience or other protective factors buffer the effect of ACEs. However, it remains unclear whether these signatures arise from IBS itself or reflect increased vulnerability in individuals with ACEs. Future research with larger samples and a longitudinal design, following individuals with ACEs to see how ACEs can increase the vulnerability to develop IBS symptoms, is needed. Overall, these findings emphasize the critical influence of early-life stress on gut microbiota composition and support the potential of microbiome-targeted interventions in alleviating and preventing IBS symptoms in individuals with a history of ACE.

The ensemble model outperformed the individual base models, indicating that integrating multiple biological modalities enhances predictive power. Among the base models, the clinical model had the highest independent predictive ability (kappa = 0.67), suggesting that questionnaire-based assessments of IBS severity and symptom impact remain critical for classification. In contrast, individual neuroimaging and metagenomic models had relatively low kappa values (ranging from 0.07 to 0.18), indicating that while these features contribute to prediction, their standalone predictive power is limited. This aligns with prior work suggesting that neurobiological and gut microbial alterations in IBS are subtle and context-dependent rather than deterministic. Labus et al. (2017) indicated that while gut microbial composition differences correlate with regional brain volumes in IBS patients, these alterations are subtle and vary among individuals. Notably, AUC values varied across groups, with the highest classification performance for HCs with low ACEs (AUC = 0.98), followed by IBS patients with high ACE (AUC =0.87) and IBS patients with low ACE (AUC = 0.82). The HC high ACE group was the most difficult to classify (AUC = 0.70). The exceptionally high AUC for HC low ACE suggests that healthy individuals with minimal early-life stress exhibit the most distinct biological signatures, making them the easiest to differentiate. In contrast, the lower AUC for HC high ACE indicates that early-life adversity in healthy individuals may lead to more variable biological profiles, with heterogenous coping mechanisms, resilience factors, or subclinical symptomatology blurring distinctions from IBS groups.

Although this was a large, well-characterized cohort with rigorous classification into four distinct subgroups, certain limitations should be acknowledged. The sample size distribution across subgroups was uneven, with a particularly smaller representation of high ACE participants, which may impact the generalizability of findings. Additionally, the cross-sectional design precludes causal inferences regarding the relationship between ACE and multi-omic alterations. Further validation in larger, independent cohorts is necessary, including functional validation of key biomarkers. Moreover, the study was limited to female participants, restricting the ability to evaluate sex-specific differences in multi-omic profiles. The reliance on self-reported menstrual history to determine follicular phase status may be subject to recall bias and imprecise phase classification, potentially affecting the interpretation of hormonally sensitive outcomes. Further, it is possible that having a condition such as IBS in adulthood may bias the recall of ACEs. Finally, future research should compare IBS with other chronic gastrointestinal or pain conditions to further delineate biological alterations related to ACEs in IBS.

This study highlights the magnitude of chronic stress from ACEs, which are adverse SDoH, on the multifaceted nature of IBS including psychological, neural, and gut microbiome factors that contribute to gastrointestinal symptoms in women. The study’s multi-omics approach builds on emerging research demonstrating the value of integrating clinical, neuroimaging, and microbiome data to capture the heterogeneity of IBS phenotypes. Gut microbiome and brain imaging markers can potentially be used as predictive tools for therapeutic outcomes. It can also be the key to identify therapeutic targets tailored to modulate specific microbial profiles or brain circuits to address the long-term consequences of early-life stress in individuals with IBS. Further research should comprise longitudinal studies to track dynamic changes in multi-omic markers over time, mechanistic studies to investigate causal pathways linking ACEs to brain and gut alterations, and most importantly, interventional studies to assess impact of therapies targeting microbiomes or brain connectivity on IBS outcomes. Since it is known that there are significant sex-specific differences in clinical presentation, microbiome composition, and brain-gut interactions, it would be valuable to do similar multi-omic studies with both male and females to explore these differences further.