Integrating gut microbiota and violence exposure metrics to classify psychological distress in middle-aged adults

Abstract

Background

Exposure to violence and psychological distress are positively correlated across populations. Microbiota-gut-brain crosstalk research supports that the microbiota is affected by environmental stressors and may influence mental state. Accordingly, we explored how the microbiota relates to exposure to violence and distress in midlife, a pivotal yet underexplored period. This life stage is marked by emerging vulnerability to chronic stress and mental health decline yet offers opportunities for early identification and intervention.

Methods

We characterized the fecal microbiota of a previosly snowball-recruited Israeli-Muslim cohort (n = 305, 40-65 yrs) exposed to ongoing and increasing violence (during adulthood) and examined correlations with subjective reports of exposure to violence and psychological distress. We then used machine learning to leverage microbiota profiles and exposure to violence, classifying individuals into high- and low-distress categories.

Results

We identify unique microbial signatures associated with increasing exposure to violence and distress. Some associated bacteria were previously identified in the literature, while others were not yet described in the context of the gut-brain axis. Microbial profiles associated with violence and distress are largely non-overlapping, yet we are able to classify participants into high- and low distress categories using a combination of microbiota and violence variables. This combined model outperforms those using only microbiota or demographics, but its classification accuracy remains modest (with a median area-under-the-curve of 0.595 (IQR 0.045).

Conclusions

This research sheds light on the microbiota-gut-brain axis, highlighting that psychological distress and exposure to violence are differentially associated with microbiota composition in midlife. These cross-sectional findings, together with moderate classification into distress classes based on the microbiome, suggest that holistic, context-aware approaches would benefit proactive mental health interventions.

Iktilat et al. examine associations between gut microbiota, exposure to violence during adolescence, and psychological distress in a midlife cohort. They identify distinct microbial patterns linked to violence and distress and show that combined microbiome and exposure data modestly predict distress levels.

Plain language summary

This study explored how gut bacteria are linked to psychological distress in middle-aged adults who have been exposed to ongoing violence. Researchers examined 305 Israeli Muslim men and women, aged 40–65, and looked at both their mental health and the makeup of their gut microbiome. Stool and blood samples were collected and analyzed, alongside questionnaires about distress and exposure to violence. The findings revealed that different types of gut bacteria were linked to either exposure to violence or distress, but not both. Using this data, the team was able to moderately predict which participants had high levels of distress. These insights may help researchers develop future tools to identify and support people at risk of mental health challenges related to trauma or stress.

Introduction

Exposure to violence is growing in many Western societies and is associated with lasting psychological distress^1,2. Individuals who have experienced violence are more likely to suffer from mental health issues such as anxiety, depression, and post-traumatic stress disorder (PTSD)^3,4. A study of exposure to community violence among more than 5000 individuals over the age of 18 years, found a positive association between exposure to urban violence and psychiatric disorders: >40% of respondents exposed to three or more traumatic crime-related events reported moderate to severe psychiatric disorders⁵. Psychological distress has also been associated with non-communicable diseases^6–8 and increased inflammation⁹. Yet, the specific mechanism underlying psychological distress embodiment has yet to be fulluciday elted.

Furthermore, despite its long-term consequences, psychological distress during midlife remains understudied. While midlife is broadly associated with rising exposure to stressors and early aging processes^10–14, these challenges are often compounded in disadvantaged populations. Among Arab citizens of Israel, chronic exposure to community violence, structural discrimination, and limited access to mental health care may further intensify psychological vulnerability. Studying this intersection offers a unique opportunity to identify early, modifiable risk factors for later-life health decline.

An emerging research area that may help illuminate how psychological distress develops and potentially shifts over time, particularly in midlife and across the lifespan, is the microbiota-gut-brain axis^15,16. The gut microbiota, composed of trillions of microorganisms residing in the gastrointestinal tract, plays a crucial role in various aspects of human health, including digestion, metabolism, and immune function¹⁷. This microbial community also exerts a profound influence on mental health and well-being via bidirectional gut-brain crosstalk^16,18–20. Additionally, the gut bacteria can modulate neurotransmitter production, neuroinflammation, and stress hormone regulation^21,22, leading to changes in host mood, cognition, and behaviors. Gut microbiota perturbations, termed dysbiosis, have been linked to a range of psychiatric disorders including PTSD^23,24. Moreover, gut microbiota can be used to predict resilience^25,26, but this line of research is still in its infancy.

Although interest in this area is growing, the relevance of the microbiota-gut-brain axis lies not only in its influence on the host but in its potential to reveal specific, modifiable biomarkers and treatment avenues for mental health, particularly in populations facing chronic stress and limited access to psychological care^27–29. While a number of studies have examined the effects of stress on the microbiome^20,30, few have examined how exposure to trauma affects the microbiome and if the microbiome can signal vulnerability to psychological distress development ^24,26.

One of the few studies on this topic compared microbiota profiles of individuals with PTSD and those that experienced traumatic events but did not develop PTSD³¹. Researchers found very few differences in overall community composition between the groups, yet, a few key phyla were increased among those suffering PTSD, including Actinobacteria, Lentisphaerae, and Verrucomicrobia³¹. A recent study of adolescents exposed to war-associated trauma that did and did not develop PTSD or related pathologies found that some taxa, like Veillonella and Dialister were positively and negatively associated, respectively, with PTSD²⁶. While differentially abundant taxa associated with mental health disorders have been identified in several studies, a strong, specific signature of these disorders remains elusive³². Furthermore, explicit examination of how exposure to violence affects the microbiome is limited; there is mounting evidence that violent behavior is associated with distinct microbial characteristics³³ (and reviewed in refs. ^{34, 35}), but the association with exposure to violence remains unknown.

This study focuses on a cohort of adult Israeli Muslims in midlife (40–65 years old) who have been exposed to increasing violence in recent years³⁶. The few similar studies conducted to date largely focused on adolescents (mean age: 16.13 ± 1.22)²⁶ or had a very small sample of adults. For example, one study assessed trauma related microbiome differences in 30 adults (n = 18 cases and n = 12 controls, age: 42 ± 12.6)³¹, and a subsequent, expanded analysis on this cohort reported additional associations between microbial taxa and PTSD status²⁴. These studies are valuable and underscore the need for larger investigations in midlife populations.

Building on our previous research on this midlife cohort, which revealed a significant correlation between self-reported exposure to violence and psychological distress³⁷, we now examine the association between the gut microbiota and both aspects of this interplay. Here we use a combination of traditional statistical methods and machine learning (ML) techniques. By combining both approaches, we balance explanatory power and predictive performance. Traditional models help us test biologically informed hypotheses, while ML allows us to explore potentially novel patterns and improve prediction accuracy. Together, they strengthen the reliability and depth of our findings, providing a broader understanding of the microbiome’s role in psychological distress.

To guide our analysis, we conducted a targeted literature search to identify microbial taxa that were previously associated with PTSD, anxiety, depression, exposure to violence or aggressive behavior, and similar disorders, in a diverse set of study cohorts and animal models. Then to understand the microbiota-gut-brain axis in the context of exposure to violence and distress, we examined three main hypotheses. As contemporary research has highlighted the importance of the microbiota-gut-brain axis, we hypothesized that¹ there would be an association between microbiota profile and (a) degree of exposure to violence and (b) psychological distress, respectively and² between microbial taxa and these two metrics taken together in a robust model accounting for covariates, including age (reviewed in ref. ³⁸), gender (reviewed in ref. ³⁸), and body mass index (BMI). Similarly, as has been demonstrated with other psychological disorders, like PTSD³⁹ attention deficit hyperactivity disorder⁴⁰, autism spectrum disorders⁴⁰, and depression⁴¹, we hypothesized that³ the microbial signatures of the studied individuals, together with their metrics on exposure to violence, could be used to differentiate between those who do and do not report psychological distress. We accounted for associations with the blood inflammatory markers C-reactive protein (CRP) and fibrinogen to ensure that observed correlations were not proxies for correlations with inflammation.

Here we identify distinct microbial taxa associated with exposure to violence and psychological distress, respectively. We then leverage key microbial taxa abundances and self-reported violence scores to classify participants based on their self-reported levels of distress.

Methods

Study design and sample

This study is a part of larger cross-sectional study that explores the relationships between the environment and biological aging among Muslim adults from northern Israel³⁷. Participants were recruited between March and September 2021 through a website/mobile app, posters in public buildings, and by word of mouth. We included people between the ages 40–65 years old; exclusion criteria were¹ history of neurological disorders (e.g., stroke)² major visual or hearing deterioration and³ acute medical complaints (STROBE, Supplementary Fig. 1). The study was approved by the Ethics Committee of the Faculty of Social Welfare and Health Science at the University of Haifa, Israel, approval number: 237/21, and reporting follows the STROBE checklist for cross-sectional studies. The sample size for this study is n = 305, after excluding those with missing data; this was sufficient based on effect size calculations using the Vanderbilt power and sample size tool ⁴².

Data collection was performed in two sessions, one to two weeks apart: The first was an hour-long session in which participants received an explanation about the study and provided signed informed consent, followed by anthropometric measurements. Participants also completed questionnaires for demographic and background information as well as psychological distress and exposure to violence (below). In the second meeting, a fasting blood sample (to assess inflammation markers) and stool sample were collected and immediately stored at −80 °C.

Psychological distress

Psychological distress was assessed using the Kessler 6 Psychological Distress Scale with a questionnaire containing six items focusing on psychological distress and emotional state⁴³. Each question was rated using a five-point Likert scale ranging from 1 (“never”) to 5 (“always”), and then distress was characterized as the mean number of points across all questions. This is a validated method (e.g., ref. ⁴⁴) with high repeatability, based on a study in an Arab population (Cronbach’s α = 0.81)⁴⁵. For some analyses (below), individuals were classified as either reporting high distress (Kessler 6 ≥ 1.833) or low distress (Kessler 6 < 1.833); this cutoff was chosen since it reflects a (rescaled) commonly used threshold of distress. A raw K6 score of 5 (scale 0–24) is often used as the threshold for low vs. high psychological distress, with scores below 4–5 typically reflecting non-clinical or low distress levels^46,47. As we used a mean rather than sum score and a scale of 1–5 rather than 0–4, the corresponding scaled cutoff is 1.833.

Exposure to violence

The Screen for Adolescent Violence Exposure (SAVE) questionnaire was used to quantify current, adulthood exposure to violence⁴⁸. It consists of thirty items describing different violent situations, such as “I saw someone carrying a gun”, or “I heard someone being beaten”. For this study, the original questionnaire was translated into Hebrew; cross-translation was performed by the research team, following the recommended guidelines⁴⁹, and respondents were asked to rate each situation on a five-point Likert scale ranging from 1 (“never”) to 5 (“always”), representing the degree of exposure. Each participant’s total exposure to violence was determined by calculating the mean of the number of points per question. Factor analysis (below) was also used to divide scores into sub-sections, with each presenting a different type of exposure; sub-section means were also recorded. The original SAVE questionnaire was validated both objectively with local crime data and theoretically, and was found to be highly reliable; the SAVE results were significantly correlated with high and low violence based on local data, and the mean Cronbach’s α for internal consistency was previously determined to be 0.81 across all sub-fields examined⁴⁸. While the SAVE questionnaire has not been formally validated in Hebrew or specifically for this ethnic group, its multidimensional structure was deemed appropriate for capturing the types of exposures to violence relevant to our study. Cross-cultural translation procedures were followed, and internal factor analysis supported its use within our sample.

Factor analysis of SAVE questionnaire components

Towards more fully understanding the nuances in the SAVE questionnaires, we performed factor analysis to identify possible subcategories that describe different aspects of exposure to violence, as has been done in other contexts (e.g., ref. ⁴⁸). This analysis was done using factor analysis (IBM SPSS v.27.0.0.0) and yielded five categories, each characterizing a different type of exposure. Identified categories include: (Viol-1) “tertiary exposure”; including seven questions focusing on participants mainly hearing about violence, (Viol-2) “eye-witness exposure”; including six questions focusing on eye witnessing violence, (Viol-3) “first person exposure”; including questions on directly being attacked by others, (Viol-4) “severe violence exposure”; including seven questions about witnessing and exposure to cases of violence and weapons in the immediate environment, and (Viol-5) “life-threatening violence”; including exposure or personal injury as a result of physical violence or the use of weapons.

Microbiome sample collection and characterization

Fixed protocols were followed for collection and processing of stool samples as described by Uzan-Ulzari et al.⁵⁰. Briefly, following DNA extraction, the V4 region of the 16S rRNA gene was^51–54 amplified in a one-step PCR, and amplicons were purified and sequenced (single-end reads, 250 base pairs) using the Illumina MiSeq platform.

Blood sample collection and processing

Blood samples were collected by a registered nurse and processed in Holy Family Hospital, Nazareth, Israel. Fibrinogen, an acute-phase protein involved in clot formation, and C-reactive protein (CRP)—two inflammatory markers—were quantified for each participant.

Statistics and reproducibility

This study included fecal samples and meta-data from 305 participants highlighting robustness of the findings. Demographic variables and inflammatory markers were compared between individuals with high and low levels of self-reported distress using two-tailed independent samples t tests, Mann–Whitney U tests, and chi-square tests, as appropriate; the level of significance was set at P < 0.05. Variables are presented as median ± IQR (interquartile range). Bivariate correlations between exposure to violence and psychological distress were examined by calculating Spearman correlation coefficients for the total SAVE score and for each component separately.

Microbiome biostatistical analysis

All microbiome data were processed using the YaMAS pipeline (https://github.com/YarinBekor/YaMAS). YaMAS converts raw sequencing data to FASTQ format and then processes it using QIIME2-2023.2⁵⁵ for demultiplexing using DADA2⁵⁶. Within QIIME2, we applied quality control measures to filter out low-quality reads and trim sequences, ensuring the reliability of the data. High-quality sequences were then processed to identify Amplicon Sequence Variants (ASVs), which were used to generate a taxonomic abundance table. Taxonomy was assigned to the identified features using a trained GreenGenes2 classifier⁵⁷. While species-level assignments are not definitive with 16S rRNA metabarcoding, we present them here, alongside the raw sequence data (in the EBI repository, accession number: ERP173874)⁵⁸, which readers can independently verify using BLAST, as we feel this adds value by illustrating within-genus diversity. Next, we used the MIPMLP pipeline⁵⁹ to merge similar features based on the taxonomy, scale distributions, standardize to z-scores, and perform dimension reduction at the species level. Specifically, we applied the sub-PCA method to merge species-level features within each genus. This involved performing a local PCA on ASVs collapsed to the same genus and retaining the top principal components that explained at least 50% of the variance to represent that group.

Alpha diversity was computed for each sample using Shannon’s diversity index⁶⁰. Beta diversity was computed using the Jaccard method⁶¹. We used Permutational Analyses of Variance (PERMANOVAs)⁶², to assess whether there is a statistically significant difference between groups of each focal parameter: psychological distress, the violence exposure components, and total (mean) violence, based on the distance matrices. The threshold for distinguishing between low and high psychological distress was set at 1.833 (as described above). For exposure to violence, the mean of each measure was used as a threshold.

To examine the associations of specific microbial features with exposure to violence and psychological distress, we first constructed regression models for each focal parameter separately, using microbial taxa identified in a literature review (Supplementary Table 1), alpha diversity, and age, BMI, and gender as predictors. Then to detect potentially novel microbiota associations, we performed regression models of all bacterial taxa. We applied Benjamini–Hochberg corrections to account for multiple comparisons⁶³.

Next, we applied Canonical Correlation Analysis (CCA)⁶⁴ to explore the relationships between two sets of variables and extract a linear combination of the original variables in each set while maximizing the correlation between each other (scikit-learn library)⁶⁵. The first set included the metrics on exposure to violence, and the second included psychological distress and blood inflammatory markers previously associated with distress (fibrinogen and CRP) to ensure that the microbiota-distress correlations were not proxies for correlations with inflammation. We then performed regression of the linear combination of the first variable set generated in the CCA process and the variables age, BMI, and gender, along with the microbiome data on the linear combination of the second set. Similar to the previous regression models, we first applied the model with only bacterial taxa identified in a focused literature search (Supplementary Table 1), and afterward used broader models with all bacterial taxa, applying Benjamini–Hochberg false discovery rate (FDR) corrections for multiple comparisons. All bacteria that were significant in the first model (targets from the literature) remained significant when applying FDR in the full microbiota examination, so only the full, FDR-corrected model is presented.

Lastly, we used XGBoost⁶⁶ to classify samples based on self-reported psychological distress levels. We used the SHAP library⁶⁷ for explainability and feature selection. Five separate models were examined aiming to predict psychological distress using¹ the full microbiome², only participant features: violence metrics, age, BMI, and gender³, the set of focal microbes identified in the literature review (Supplementary Table 1)⁴, the participants’ ten most important microbes identified from the whole microbiota model (first model, determined by feature selection using the SHAP algorithm), and⁵ a combination of the features included in the second model, and the microbes included in the fourth model.

For each of the five models, we used an 80-20 training-test set with a random split. We first performed 10-fold cross-validation on the training set to tune hyperparameters, using the GridSearchCV method in the Python scikit-learn library⁶⁵ (Supplementary Table 4). Following this, we performed a final one-time prediction on the held-out 20% test set, using the optimized hyperparameters. To ensure the robustness of our results and minimize the influence of a specific random split, we repeated this entire process 100 times, each with a different split. We then reported the median area under the curve (AUC) of the test sets across these 100 iterations. This procedure was applied identically across all five inputs models.

The predictions were evaluated using the AUC and a Receiver Operating Characteristic (ROC) curve. To determine if the mixed microbiota-demographics model (fifth model) was significantly better than the other models, a one-way ANOVA followed by Dunnett’s post-hoc test was applied.

Results

Cohort description

A total of 305 participants (180 women, 125 men), with mean age of 50.82 ± 6.94 (40–65 years) were included in the final analysis. Demographics and distributions of the descriptive characteristics are presented in Table 1. The were no significant differences in gender, BMI, or age across distress categories, yet there were significant differences in some of the violence components, namely Viol-1 (tertiary exposure) and the total violence score. We found significant correlations between nearly all pairs of the five identified components of violence and the total violence score, as well as correlations between these metrics and psychological distress (Fig. 1). Exceptions were Viol-1 with Viol-5 (life-threatening violence) and psychological distress with Viol-2 (eye-witness exposure), Viol-4 (severe exposure), and Viol-5, respectively. Interestingly, no correlations between inflammatory markers and self-reported distress or exposure to violence were identified.

Table 1.

Descriptive statistics of the cohort

Characteristic	Total population	Low distress	High distress	P value
Sample size (N)	305	157	148
Male (%)	40.98%	42.68%	37.16%	0.3890
Age	5043–54	5044–55	49 (44.75–55.00)	0.1934
BMI	29.56 (26.70–32.70)	29.41 (26.51–32.08)	29.76 (26.88–33.25)	0.2006
Exposure to violencea	1.80 (1.61–2.03)	1.73 (1.52–1.98)	1.86 (1.70–2.10)	0.0026
Viol-1	3.29 (2.86–3.79)	3.14 (2.71–3.71)	3.29 (3.00–3.86)	0.0294
Viol-2	1.00 (1.00–1.33)	1.00 (1.00–1.33)	1.00 (1.00–1.33)	0.3241
Viol-3	1.00 (1.00–1.20)	1.00 (1.00–1.20)	1.00 (1.00–1.20)	0.0710
Viol-4	2.00 (1.50–2.67)	2.00 (1.37–2.50)	2.17 (1.67–2.67)	0.0572
Viol-5	1.00 (1.00–1.00)	1.00 (1.00–1.00)	1.00 (1.00–1.00)	0.1394
Psychological distressb	1.67 (1.33–2.17)	1.33 (1.17–1.50)	2.17 (2.00–2.67)	<0.0001
Fibrinogen	430.08 (390.81–469.67)	421.96 (356.18–470.05)	436.09 (404.05–469.62)	0.3522
C-reactive protein (CRP)	3.03 (1.62–6.02)	2.98 (1.63–5.81)	3.02 (1.57–6.09)	0.8166

Participant characteristics, for the total cohort and divided by level of self-reported psychological distress (Kessler 6 survey). Low distress was defined as a total mean score less than 1.833, and high distress was a total score greater than or equal to 1.833. Values are given as median (25th–75th percentile). Fibrinogen and CRP are measures of inflammation. Characteristics of low and high distress groups were compared using two-tailed independent samples t tests for age and BMI, Mann–Whitney U test for violence exposure components, psychological distress and inflammation measures (fibrinogen and CRP), and a Chi-square test for gender. Statistically significant values are bolded. Viol-1: tertiary exposure, Viol-2: eye-witness exposure, Viol-3: first person exposure, Viol-4: severe exposure, Viol-5 life-threatening violence.

^aTotal SAVE: Screen for Adolescent Violence Exposure.

^bKessler 6.

Fig. 1. Self-reported exposure to violence metrics, psychological distress, and blood inflammatory markers.

a Histograms of levels of self-reported exposure to violence based on five SAVE (Screen for Adolescent Violence Exposure) components identified using factor analysis, total SAVE score (mean), mean Kessler 6 questionnaire score for self-reported psychological distress, and blood inflammatory markers fibrinogen and C-reactive protein (CRP). b Spearman correlations between all metrics. Color bar and lower half of the matrix represent Spearman’s rho. Outlined boxes are correlations whose p-value is significant (two-tailed, uncorrected P < 0.05); the upper half of the matrix represents the p-values of the correlations. VIO1: tertiary exposure, VIO2: eye-witness exposure, VIO3: first person exposure, VIO4: severe exposure, VIO5 life-threatening violence, Viol_total: measure of exposure to violence (SAVE). N = 305.

Microbiota characteristics associated with exposure to violence

We did not find a significant correlation between alpha diversity and each of the conditions. Similarly, the community composition (beta diversity) did not significantly differ between participants exposed to violence above versus below the mean reported levels. We did, however, identify 15 bacterial taxa from the list of identified targets (Supplementary Table 1) significantly associated with the various components of exposure to violence in a model controlling for BMI, age, and gender (p-value < 0.05; Fig. 2). The species Bifidobacterium bifidum, Lactobacillus ruminis, and Parabacteroides distasonis, were positively correlated with three violence components and the total violence score. Similarly, Clostridium citroniae and Ruminococcus torques were correlated with two violence components, and Ruminococcus torques was also correlated with total violence. Two taxa, Bacteroides uniformis and Clostridium ruminantium were negatively correlated with total exposure to violence or one of its components and a single taxon—an unclassified Paraprevotella species, was positively correlated with Viol-1 and negatively correlated with Viol-5. The rest of the significant taxa (Fig. 2) were correlated with only a single component of violence. When expanding the analysis to include all microbial features (and including FDR corrections), many of the above-described relationships were still significant. In addition, Desulfovibrio D168 was also positively associated with two violence components and the total violence score, and Dehalobacterium was negatively correlated with Viol-2 (Fig. 2). A full statistical summary is presented in Supplementary Table 2.

Fig. 2. Bacterial taxa associated with self-reported exposure to violence (Screen for Adolescent Violence Exposure, SAVE) and psychological distress (Kessler 6).

Bacterial taxa that were identified in a regression model as significantly correlated with focal metrics when age, BMI, and gender are included as covariates. Significant negative and positive correlations between the taxa and the violence/distress metrics are colored in a blue-red heatmap, respectively, reflecting the beta coefficient values. Underlined are significant taxa from the pool of target microbial taxa that were identified in a non-exhaustive literature review (Supplementary Table 1). Values in bold were also significant following FDR correction (P and Padj <0.05). Significant taxa not found in the literature search were identified using regression models testing all bacterial taxa and are included only if they met the stricter significant threshold (p and Padj <0.05). N = 305. See Supplementary Table 2 for the full statistical summary of results. Psych_Dist: psychological distress measure (Kessler 6), VIO1: tertiary exposure, VIO2: eye-witness exposure, VIO3: first person exposure, VIO4: severe exposure, VIO5 life-threatening violence, viol_total: measure of exposure to violence (SAVE). N = 305.

Microbiota characteristics associated with psychological distress

We performed analyses parallel to those described above for the self-reported psychological distress score as well. Again, we did not find a significant correlation between alpha diversity and distress or significant community differences between groups of high and low distress. Three taxa from the literature search were significantly positively associated with distress: Bacteroides barnesiae, Lactococcus garvieae, and an unclassified Prevotella species (Fig. 2). When expanding our analyses to include the total microbiota, these three taxa remained significant and several other taxa were also identified: an unclassified member of the family Anaeroplasmataceae, Bulleidia moorei, and the genera Coprobacillus and Desulfovibrio (Fig. 2). Interestingly, none of the species associated with distress were also associated with measures of exposure to violence, though several genera, including Bacteroides and Desulfovibrio overlapped between the metrics. In the case of Desulfovibrio, abundance was positively correlated with both exposure to violence and psychological distress. For Bacteroides, however, patterns were mixed across species of the genera, both within exposure to violence categories and between violence and distress measures. A full statistical summary is presented in Supplementary Table 2.

Microbiota characteristics in the context of the exposure to violence—psychological distress—inflammation interplay

Psychological distress was previously associated with an inflammatory response⁶⁸. To understand the triple interaction between microbiome, distress, and violence, we first performed a CCA between the different components of violence (set 1) and psychological distress and inflammatory markers (set 2, see methods). There was a clear positive association (Fig. 3a) between violence components 1,3,4,5—mainly 1 and 3—and distress, and a clear negative correlation between component 2 with distress levels, as determined by the CCA coefficients. There was only a very minor effect of CRP and fibrinogen levels on the combined measure of violence as demonstrated by the CCA coefficients (Fig. 3a, b). We examined the microbes associated with the combination of distress and inflammation in a full multiple regression model. Interestingly, eight microbes were found to be positively associated with distress in this model, and five were negatively associated with distress (Fig. 3). Only one of the significant taxa matched those identified in the above analyses (Fig. 2), though at the genus level, there was some overlap. Furthermore, none of the specific species identified here were previously found to be associated with distress or similar outcomes in the literature, but five of the genera were previously associated with distress (Supplementary Table 1). A full statistical summary is presented in Supplementary Table 3.

Fig. 3. Canonical Correlation Analysis (CCA) examining correlations between exposure to violence and distress + inflammation and the relationship with the microbiome.

a Canonical coefficients fitted to each set of variables¹: Violence exposure metrics Viol-1, Viol-2, Viol-3, Viol-4, Viol-5 and² psychological distress combined with inflammatory measures C-reactive protein (CRP) and fibrinogen. The coefficients maximize the correlation between the two variable sets. b The linear model fitted to the generated combinations of¹ the violence metrices, and² psychological distress and inflammation measures. c Bacterial taxa that were identified in the model as being significantly correlated with focal metrics. Taxa identified as targets in a non-exhaustive literature review (Supplementary Table 1) are underlined. Significant positive and negative correlations between the taxa and the CCA output: a linear combination of psychological distress and inflammatory measures are colored in blue and red heatmap, respectively, representing the beta coefficient. See Supplementary Table 3 for the full statistical summary of results. N = 305.

Classification into high and low distress classes based on microbiota and exposure to violence

After identifying taxa associated with the connection between exposure to violence and psychological distress, we were curious to see if the microbiome signature and self-reported exposure to violence data were jointly related to distress state and could be used to classify individuals into high and low distress groups. We built five models based on the microbiome, demographics, or a combination of the two (Fig. 4a), and using the best model, we were able to classify individuals with a median AUC of 0.595 (IQR 0.045). This model considered self-reported exposure to violence, gender, age, BMI, and a set of 10 key taxa (fifth model, Fig. 4b, c), and performed slightly better than random classification. This model performed significantly better than all other tested models (Fig. 4b, ANOVA and Dunnett post-hoc tests, all p-values < 0.05). More than half of the key taxa belonged to genera that were previously identified in the body of literature we reviewed (Fig. 3c, Supplementary Table 1), showing the generalizability of markers across studies.

Fig. 4. Classification of participants by psychological distress level based on exposure to violence and microbiota characteristics.

(a). The variables included in each model and associated median (IQR) AUC score for predicting high versus low distress (100 iterations, 80-20 training-test set). High distress was defined as a score ≥ 1.833. b AUC distribution plot for each of the 5 models. ** denotes models with a significant improvement compared to the “Combine: Features & SHAP” model⁵. c SHAP plot of the top ten microbial taxa based on model 1 and included in models 4 and 5. Dots colored in blue represent low levels of the given microbe, and dots colored in red depict high levels of the microbe. The x-axis represents the impact of the microbe on the model. N = 305.

Discussion

Here, we aimed to unravel microbiota-gut-brain crosstalk in the context of exposure to violence and psychological distress in a unique midlife population of adult Muslims exposed to ever-increasing violence. We identified several taxa associated with each self-reported variable, supporting our first hypothesis. Interestingly, there was very little overlap in the taxa correlated with exposure to violence and those correlated with varying components of distress. When considering the measures together, to address our second hypothesis, we identified a nearly unique set of bacteria significantly associated with psychological distress in midlife when including violence metrics in the models. Finally, in line with our third hypothesis, we found that a small subset of the microbiome (10 taxa) and measures of self-reported exposure to violence can be used to classify individuals into high- and low- distress groups with better-than-random performance. Taken with our previous findings, we present mounting evidence of an interplay between the gut-brain axis, the environment, and the microbiota in midlife psychological distress. Integrating microbiota composition with survey-based and demographic data may help researchers identify new avenues for intervention in midlife.

In this study, we identified 11 bacterial taxa that were significantly and positively associated with the total SAVE score or its components. Of these taxa, 10 genera were targets from our literature review. Veillonella was previously associated with PTSD in adolescents²⁶, which is in line with the increased abundance observed here among participants exposed to traumas like violence, but some of our findings seem counterintuitive when compared to previous trends. Species in five genera—Akkermansia, Bifidobacterium, Clostridium⁶⁹, Parabacteroides, and Ruminococcus—were found to be protective in the context of anxiety and/or PTSD^70–73, raising questions about our findings of increased prevalence in those exposed to more violence. There are also some inconsistencies in the literature. For example, a study of PTSD in adolescents found Bifidobacterium to be increased among those with the condition, and in the case of Parabacteroides, the protective effect may be species-specific⁷². Similarly, Ruminococcus gnavus was associated with generalized anxiety disorder in one study⁷⁴ highlighting how genus-level associations may be a bit reductive. In line with these nuances, a recent study also reported reduced levels of Bifidobacterium in adults with self-reported anxiety, alongside elevated Bacteroides, a genus that had mixed species-level associations to violence components in our cohort ⁷⁵.

Lactobacillus, which we found to be positively correlated with exposure to violence in our cohort, has been shown, in animal models, to modulate aggressive behavior, with some strains protective against antibiotic-induced aggressive behavior³³. While exposure to violence in humans reflects victimization rather than perpetration, both experiences may involve overlapping neurobiological and microbiota-related pathways linked to stress response and behavioral regulation. Similarly, Desulfovibrio, here positively associated with exposure to violence, was previously observed at higher abundance in more aggressive mice (unpublished data). Further supporting the need to look beyond genus-level associations, one species of Bacteroides in our study was positively associated with violence, one was negatively associated, and a third was positively associated with distress. This is similar to reports in the body of literature (Supplementary Table 1) which found species-specific patterns within this genus; two studies found the whole genus associated with anxiety⁷¹ and major depressive disorder⁷⁶ and a third found B. massiliensis protective against PTSD⁷². The species associated with distress in our study (B. banesiae), is different from that found to be protective against PTSD, further highlighting the large, species-specific variance in associations. Lastly, an unidentified Paraprevotella species was positively associated with Viol-1 but negatively associated with Viol-5. It was the only taxon without a consistent correlation across exposure to violence components. Previous research found that a member of this genus was associated with PTSD. Viol-5 represents exposure to extreme violence, but interestingly, the relationship between taxon abundance and this component was negative, seemingly at odds with the previous research. On the whole, there is a unique microbiota profile associated with exposure to violence, and in most cases, the patterns of bacterial abundance are conserved across exposure components. Some patterns of differential abundance align with previous microbiota-gut-brain axis studies and others do not, suggesting that the role of exposure to violence in this axis is highly nuanced, either related to distinct exposure types or species- and strain-level interactions rather than genus-level ones. Future studies should be further resolved, allowing species-strain analyses and considering microbial functional properties rather than taxonomy alone.

When examining the correlation between relative abundances of bacterial taxa and self-reported psychological distress, seven key taxa were identified, three of which were previously associated with PTSD, anxiety, and other mental disorders in adult populations (Supplementary Table 1). In line with a study of general anxiety disorder in Chinese adults (46.08 ± 12.09)⁷¹, we found the genus Bacteroides to be positively associated with psychological distress. In contrast, this same study found a negative correlation between Prevotella and anxiety⁷¹, whereas in our study, the association with psychological distress was positive. This genus is known to have diverse roles in the human host, though, suggesting more resolved sequencing may be needed to mechanistically understand the role of Prevotella in gut-brain interactions. Here, Lactococcus was also positively associated with distress, in contrast to previous findings from a study of PTSD in adolescents²⁶, suggesting that either this genus has diverse functions, or its role in younger hosts may differ from those at the onset of aging. The dearth of taxa correlated with distress in this study compared to those associated with exposure to violence here, or other mental disorders in the literature, could result from the relatively low self-reported levels of distress (Table 1, Fig. 1). Also of interest is the absence of overlap between taxa associated with exposure to violence and those associated with distress suggesting that the primary effects are largely independent.

The current study takes our previous research³⁷ one step further, demonstrating that regressing both a violence variable and the microbiome on a linear combination of inflammation and psychological distress reveals eight and five taxa that were positively and negatively associated with distress, respectively. This highlights that the gut microbiota may be involved in pathways underlying distress etiology in this population. Interestingly, only one of the taxa here was significant when examining correlations with violence alone, and none correlated with distress, suggesting that there is a complex interaction between the three components. Of the thirteen significant taxa, five belong to genera identified in our literature review. A different species of Lactobacillus was previously positively associated with violent behavior, and here, there is a positive association with distress when also considering violence. Similarly, SMB53 and Veillonella were positively associated with psychological distress here, but in a previous study of PTSD in Israeli adolescents, only SMB53 was positively associated with the disorder. Veillonella was negatively associated with PTSD²⁶. Bacteroides and Fusobacterium were negatively associated with distress here when including exposure to violence in our model, in contrast to studies on generalized anxiety which found a positive correlation between the condition and both taxa^71,74. In line with our findings, a previous study of PTSD in women found a species of the Bacteroides taxa protective against the disorder ⁷².

Despite the nuances in microbiota-gut-brain patterns identified here and differences compared to the body of literature (discussed above), we were able to classify individuals into psychological distress groupings (low/high) based on their fecal microbiota and self-reported exposures to violence with modest success, similar to the classifications made for Israeli adolescents in the study of PTSD²⁶. While preliminary, this finding suggests that future, more sensitive, cross-cohort microbiota-based models may offer complementary insights into identifying at-risk populations or selecting therapeutic leads in a data-driven manner. Taxa that significantly contributed to the model’s performance (Fig. 4c) include some of the genera highlighted above—Bifidobacterium, Lactobacillus, Ruminococcus, Clostridium and Roseburia—all taxa known to be associated with host state, and particularly the gut-brain axis.

It is important to note that this is a cross-sectional study considering microbiota, exposure to violence, and psychological distress at the same time point. While we have shown the utility of exposure to violence and microbiota composition in classifying participants into distress groups, predictive ability before distress onset would need to be refined and confirmed prospectively using a longitudinal study in which microbiota and violence measures recorded at an earlier time point would be used to predict psychological distress that developed later in life. The current model should be viewed as exploratory and hypothesis-generating, rather than clinically actionable. Our primary aim is to advance basic scientific understanding of the microbiota-gut-brain axis in midlife populations exposed to chronic stress, rather than to provide a diagnostic or screening tool. Furthermore, including a baseline group, also in midlife, but exposed to little or no violence, would allow us to better understand the contribution of exposure to violence itself rather than natural microbiota shifts and distress fluctuations in aging populations. If this can be confirmed, there would be broad implications for public health schemes, which could more precisely identify target populations for early interventions. In addition, while the SAVE questionnaire provided useful multidimensional data on exposure to violence, it was originally developed for adolescents and has not been formally validated in Hebrew or in this ethnic group. Although we followed cross-translation guidelines and conducted internal factor analysis, this remains a methodological limitation to consider in future studies.

Beyond sampling timeline considerations, this study could be expanded upon using metagenomic sequencing or including metabolomics analyses to obtain more accurate species-level classifications and strain-level information and better understand how the microbiome is functionally entwined in the violence exposure-psychological distress interaction. Here we have demonstrated that there is a significant microbial signature associated with each variable and also when considering the two together in a single model; future studies are needed though to understand how this signature affects host state and to validate it in a preclinical setting towards developing microbially associated interventions.

Conclusion

In this research, we studied a Muslim population in midlife, living in Israel and exposed to increasing violence. We identified unique microbial signals associated with exposure to violence and with psychological distress, when controlling for participant age, BMI, and gender. Some of our findings match patterns previously identified for similar mental health states and others are unique. We could also classify participants into low-distress and high-distress groups based on microbiota and exposure to violence, with significantly greater accuracy than a microbiome model or a demographic model alone. Future research is needed to determine if this classification method can be used to identify midlife adults at high risk of developing psychological distress in the future when exposed to heighted violence. If confirmed, there can be far-reaching public health benefits, reducing some of the burden of aging in high-risks populations. Furthermore, filling gaps in the literature, particularly regarding the link between exposure to violence and the microbiome in midlife, is essential for advancing our understanding of the complex interplay between environmental stressors, gut microbiota, and mental health outcomes at the onset of aging. Our findings suggest that therapeutic interventions targeting the microbiome might be relevant to mitigate the adverse mental health effects of traumatic experiences. Although we did not directly assess aging outcomes, our findings may inform future research on biological mechanisms that influence healthy aging, particularly in chronically stressed midlife populations.

Author contributions

Khalil Iktilat conceived the central research idea, coordinated the project, led the manuscript writing and revisions, oversaw data curation, and ensured overall integration across teams and analyses. Maayan Agmon and Omry Koren jointly supervised the study, contributed to conceptualization and study design, assisted with interpretation of the findings, and provided critical revisions to the manuscript. Sondra Turjeman contributed to the microbiome analysis design and interpretation and edited the manuscript. Yoram Louzoun and Gali Levin performed the statistical and machine-learning analyses, conducted robustness and reproducibility checks, and contributed to figures and methods. Michal Isacson contributed to the literature review, manuscript editing, and assisted in data interpretation. Roy Tzemah-Shahar assisted in data collection, literature review, manuscript revisions, and supported the statistical analysis. Gila Gamliel contributed to data organization, documentation, and database management.

Peer review

Peer review information

Communications Medicine thanks Francesca R. Querdasi and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

Microbiome data have been made publicly available on the EBI repository: Accession number: ERP173874. https://www.ebi.ac.uk/ena/browser/view/ERP173874. The source data for Figs. 1–4 are provided in Supplementary Data files 1–4, respectively.

Code availability

All analyses were conducted using publicly available tools; no new software was created for the analysis of this data. Step-by-step, annotated analyses are provided in a public github entry with detailed notes on data format and model parameters. https://github.com/louzounlab/Distress_Microbiome_Analyzing⁷⁷.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-025-01319-2.