An examination of personality traits and the taxonomic composition of the gut microbiome among psychiatric inpatients

Abstract

Objective:

Through the brain-gut-microbiome axis, myriad psychological functions that impact behavior share a dynamic, bidirectional relationship with our intestinal microbiome. Little is known about the relationship between personality – a stable construct that influences social and health-related behaviors – and our bacterial ecosystem. This exploratory study examined the relationship between general and maladaptive personality traits with the composition of the gut-microbiome.

Methods:

105 psychiatric inpatients provided clinical data and fecal samples. Personality traits were measured using the five-factor model of personality, the Schedule of Clinical Interview for Personality Disorders, and Personality Inventory of the DSM-5. 16S and whole genome shotgun sequencing methods were used to process fecal samples. Machine learning was used to identify personality traits associated with bacterial variability and specific taxa.

Results:

Using supervised learning techniques, we successfully classified traits of social detachment (max AUROC = 0.944, R² >0.20), perceptual disturbance (max AUROC =0.763, R² = 0.301), and hoarding behaviors (max AUROC = .722) using limited sets of discriminant bacterial species or genera.

Conclusions:

Observations from this study are consistent with recent findings demonstrating person-to-person interactions as a mode of gut microbiome transmission. Established bacterial genera associated with psychosis (e.g., Peptococcus, Coprococcus) were associated with traits of perceptual disturbance. Consistent with the existing literature, hoarding behaviors were associated with a defined gut microbial composition that included Streptococcus – a known contributor to the development of pediatric autoimmune neuropsychiatric disorders. This study adds to the emerging literature on the intricate connections between brain and gut function, expanding the interdisciplinary field of psychiatric microbiology.

Introduction

The gut microbiome and associated by-products influence various physiological systems including immune function, metabolism, and organ development (1) and extend to interactions with the central nervous system, influencing functioning in a dynamic, bidirectional exchange referred to as the microbiome-gut-brain axis (2). Through this axis, the human intestinal microbiome influences psychological functioning and behavior, including mood, stress response, anxiety, cognition, perception, and pain experience (3). Given the relevance of these factors to mental health, researchers have demonstrated a link between gut microbiome composition and severity of psychiatric symptoms (4).

The composition of the gut-microbiome is largely derived from maternal seeding of microbial taxa in utero and at birth (5). By the age of 2 years old, expanded microbial communities in the gut achieve stability, with correspondence into adulthood (1, 6). However, despite core stability, microbiome composition varies based on several factors, especially diet, lifestyle (e.g., physical activity, smoking status, work habits), hygiene, sex (assigned at birth), and medication use (7). The microbiome also varies based on the social environment. Recent studies have demonstrated that gut microorganism strains are transmitted interpersonally, with substantial strain sharing found within-household and within-population (6, 8).

Recent efforts have begun to explore personality traits that influence social engagement and interaction, as well as other health-related behaviors, that have been found to influence the structure of the gut microbiome. In a pilot study of sixty healthy Korean adults, personality traits were found to distinguish between different stable gut microbiome clusters, referred to as enterotypes (9), with the traits of novelty seeking and reward dependence being higher in an enterotype characterized by abundances of Bacteroidaceae. (10). In a larger study (n = 672) of healthy Korean adults (11), high neuroticism was linked to increased abundances of Gammaproteobacteria, high conscientiousness was associated with higher prevalence of Lachnospiraceae, and low conscientiousness was connected to increased Proteobacteria. Those with high trait openness had greater genus richness than those with low openness, and higher agreeableness was related to increased diversity though the effect size was small. In a study from an international sample of adults (12), higher abundances of Akkermansia, Lactococcus, and Oscillospira were noted in individuals with higher sociability, while those lower in sociability had increased Desulfovibrio and Sutterella. Individuals with higher neuroticism had lower abundances of Streptococcus and Corynebacterium. In support of the social transmission hypothesis (6), people with larger social networks had more gut microbiome diversity.

Preliminary studies have established connections between the gut microbiome and personality; however, the studies are few and findings are discrepant, underscoring the need for more research. This study builds upon seminal literature to explore the relationship between the gut microbiome and personality traits in a psychiatric inpatient sample. We hypothesize that personality traits affecting social and health related behaviors will differentially impact gut microbiome composition. Specifically, we posit that participants with greater extraversion/lower detachment will exhibit differential genus or species signatures given they are more likely to interact with others, including novel others, and potentially have more diverse dietary choices. Additionally, we hypothesize that people with significant personality pathology (namely cluster B traits) will also have notable differences in gut microbiome composition related to adverse health behaviors and outcomes.

Methods

Participants & Setting

Data were collected over the course of two years, drawn from individuals (N=105) who were admitted to a psychiatric inpatient hospital with longer than typical lengths of stays, averaging 49.7 days (± 14.5 days) for stabilization and treatment. During this time, the participants received a variety of personalized psychiatric and psychological interventions, as well as medical intervention as indicated. The treatment setting offered a multimodal approach to addressing symptomology, which included individual and group psychotherapy, family work, milieu engagement, and 24-hour nursing support.

Procedures

After hospital admissions procedures were completed, psychological measures were administered within 72 hours. Self-collected fecal swabs were gathered on average 20.1 days after admission (±12.8 days) and were held in a −80 °C freezer while waiting for analysis. Data were collected as part of a larger study aimed at gathering treatment outcome data and as part of separate study designed to identify the role that biomarkers may play in mental illness and treatment outcome (13). Information about current medications that may impact the gut microbiome were drawn from the hospital’s pharmacy database, with particular classes of medications selected a priori, namely, opioids, antipsychotics, antidepressants, probiotics, and antibiotics. We previously demonstrated that these medication classes did not influence the structure or function of the gut microbiome in this study sample (4). Thus, they were excluded from subsequent analyses.

Measures

A standardized survey was used to gather information about patient demographics, history of psychiatric illness, and pertinent health information. To identify personality disorders (PD), master’s level research assistants administered the Structured Clinical Interview for DSM-IV Disorders (SCID-II) (14). The SCID-II provides data on item level endorsements of PD criteria, dimensional raw scores for each PD, and categorical information about whether participants met criteria for a PD. Information was included for six of the ten traditional PD categories: avoidant, obsessive-compulsive, schizotypal, narcissistic, borderline, and antisocial. Due to low base rates of less than 1% over the course of the two-year data collection, other PDs were not included. The five-factor model of personality traits (i.e., openness, extraversion, conscientiousness, agreeableness, and neuroticism) were measured with the Big Five Inventory (BFI) (15). The BFI has strong reliability, converges with the NEO Five Factor Inventory, and has demonstrated good internal consistency in other studies using the sample (Cronbach’s α = 0.81) (16). To measure personality pathology, the Personality Inventory for the DSM-5 (PID-5) (17) was used. The PID-5 is composed of 220-items that load onto discrete personality trait facets. These 25 facets form the subscales for five overarching personality domains, namely, detachment, disinhibition, negative affect, psychoticism, and antagonism. The PID-5 has demonstrated good psychometrics and has provided incremental validity over and above the five-factor model (16).

Data Analyses

For analysis of fecal samples, the MO BIO PowerSoil DNA Isolation Kit was used to identify bacterial genomic DNA. The Illumina MiSeq System sequenced and PCR-amplified the 16S rDNA V4 region using the 515F-806R amplification primer. The 515F-806R primer uses single-end barcodes and MiSeq adapters, which support sequencing of PCR products and sample pooling. Alignment-based methods, as well as phylogenetic methods, were used in the 16S rRNA pipeline. The identifying molecular barcodes were demultiplexed and merged. The software used for this task was USEARCH v7.0.1090. Demultiplexed read pairs underwent an initial quality filtering to remove Illumina adapters, PhiX reads and reads with a Phred quality score lower than 15 and length lower than 100 bp after trimming. Quality controlled reads were merged then further filtered to remove reads exceeding a max expected error rate and not matching the expected length. To cluster the 16S rDNA into Observed Operational Taxonomic Units (OTUs), the UPARSE algorithm was used. The rarefaction depth was selected automatically via a bootstrapping method using the R software Agile Toolkit for Incisive Microbial Analyses (ATIMA). The cut value was set at 97% similarity. In addition, the SILVA database version 138.1 of just the 16S v4 region was also employed to annotate OTUs. Abundance values were calculated by mapping demultiplexed reads to OTUs. No samples had to be excluded for insufficient reads. The minimum rarefaction depth was set at 4815. The result was an OTU table suitable for phylogenetic analysis, as well as diversity (alpha and beta) analysis. A Whole Genome Shotgun Sequencing (WGS) approach was used for genomic bacterial DNA (gDNA) in order to minimize background amplification and increase the yield of gDNA (18, 19). Data were scanned to identify any Illumina PhiX or low quality sequences in the paired-end sequencing reads. The software, bbduk, was then used to remove the Illumina adapters (BBMAP v37.58) (20). Following the corrections, bowtie2 (v.2.3.4.3e) (21) software mapped the final sequences to the human hg38 reference database. To prevent host contamination at the sequence level, the mapping process was done with high stringency. The tool used to identify taxonomic profiles was MetaPhlAn2 (22). HUMAnN2 (23) was responsible for functional profiling. The data was put into final format according to The Biological Observation Matrix (24).

For supervised machine learning analyses, 16S rRNA relative abundance at genus level was used. For WGS, relative abundance at species level, computed with MetaPhlAn, and pathway abundance, computed using HUMAnN, were used. Machine learning (ML) models were built using Random Forest (RF) (25) Support Vector Machines (SVM) (26), and k-nearest neighbor (kNN), either for classification (yes/no personality traits) or for regression (numerical personality traits). Data were split into 80% training, models were computed, then tested in the remaining 20%; the R package caret was used to implement the cross-validation iterations (27). Classification performance was measured using the Area Under the ROC curve (AUROC), whereas for regression performance was measured using the goodness-of-fit R². The training/testing split was repeated over 100 cross-validation iterations, and the median and distribution of the performance metrics were collected. Feature importance was assessed using the Interpretable Machine Learning R package (28). Further, only features determined as important in at least 70% of the cross-validation iterations were reported. For the best performing model for each machine learning problem, the top 10 increased and top 10 decreased informative features were reported.

Preliminary ML analyses of potential confounding associations of age and gender to the gut microbiome (16S and WGS) revealed weak associations and thus were excluded from subsequent analyses.

Ethics

This study followed guidelines and ethical principles as outlined in the Declaration of Helsinki. Following a full explanation of all procedures, each participant provided written informed consent. Study participants attested to the voluntary nature of their participation and were allowed to withdraw from the study at any time without consequence to the clinical care they were receiving at the study hospital. The Institutional Review Board (IRB) at Baylor College of Medicine (Houston, TX) approved this study.

Results

Sample characteristics

The study sample (N=105) consisted primarily of young, Caucasian adults with a relatively even split between sexes. The study sample age was 36.4 ± 13.8 years; other descriptors of this sample are provided in Table 1. Most had completed some college, though were unemployed. Individuals presented with multiple Axis I disorders, and many had high levels of functional impairment and service utilization. Greater than a third of the sample met diagnostic criteria for a formal personality disorder. Participants had engaged in significant outpatient and inpatient psychiatric treatment prior to the index hospitalization. 16S microbiome analysis was performed for 105 donors, whereas WGS analysis was performed only for 100 donors. SCID2 traits were available for 105 donors, whereas PID5 traits were available only for 83 donors.

Table 1.

Categorical baseline demographic and burden of illness characteristics of the sample

	N	%
Age, years
Sex, female	57	54
Ethnicity, White	94	90
Single/never married	46	44
Some college or greater	91	87
Unemployed (past 30 days)	56	53
Alcohol use disorder	40	38
Substance use disorder	38	36
Major depressive disorder	59	56
Bipolar disorder	6	6
Generalized anxiety disorder	23	22
Any personality disorder	38	36
≥ 1 year of psychotherapy	57	54
≥ 2 outpatient psychotherapists (lifetime)	84	80
≥ 2 psychopharmacologists (lifetime)	71	68
Hospitalizations for acute psychiatric care (lifetime)	53	50
Hospitalizations for extended psychiatric care (lifetime)	40	38

Machine Learning Results

Supervised machine learning results across SCID2 items using separately the 16S and WGS data revealed three items with acceptable classification performance (AUROC > .70). Each of the top ten PID5 items and facets evidenced meaningful percentage of variance explained (R² range: 0.145 – 0.301), though the performance of the machine learning models varied based on whether they were derived using 16S or WGS data. See Table 2 for details.

Table 2.

Machine Learning Classification and Regression Results for the Top Performing Models

Trait	16S	WGS
SCID2 Items: Machine Learning – Classification (AUROC)
Lack of close friends or confidants other than first-degree relatives	0.789	0.944
Ideas of reference (excluding delusions of reference)	0.763	0.588
Is unable to discard worn-out or worthless objects even when they have no sentimental value	0.722	0.471
Shows arrogant, haughty behaviors or attitudes	0.684	0.500
Impulsivity	0.679	0.579
Chronic emptiness	0.635	0.610
Has a sense of entitlement (i.e., unreasonable expectations of especially favorable treatment or automatic compliance with his or her expectations)	0.605	0.611
Is preoccupied with details, rules, lists, order, organization, or schedules to the extent that the major point of the activity is lost	0.592	0.522
Is unwilling to get involved with people unless certain of being liked	0.583	0.412
Is inhibited in new interpersonal situations because of feelings of inadequacy	0.578	0.575
PID5 Items & Facets: Machine Learning – Regression (R-squared)
I have seen things that weren’t really there	0.301	0.052
Detachment domain average score	0.060	0.232
I am better than almost everyone else	0.228	0.078
I don’t hesitate to cheat if it gets me ahead	0.220	0.126
I prefer to keep romance out of my life	0.043	0.197
I avoid social events	0.030	0.163
Values artistic, aesthetic experiences	0.162	0.029
Detachment domain total score	0.058	0.155
I just skip appointments or meetings if I’m not in the mood	0.056	0.151
Withdrawal raw score	0.032	0.145

There are two notable overlapping patterns of items/domains from SCID2 and PID5 data: social detachment and perceptual disturbance. For participants who endorsed that they lack close friends or confidents other than first-degree relatives from the SCID2, the best performing classification model was derived using SVM on the WGS data (median AUROC=0.94, IQR=0.833–1.00; with 17 informative species); conversely, for the Detachment Domain Score from the PID5 the best performing regression model was RF on the WGS data (R²=0.23, IQR=0.105–0.345, with 35 informative species) . An examination of the informative features for the previous two machine learning problems revealed one in common, i.e., C. bartlettii (see Figure 1). Participants who endorsed ideas of reference from the SCID2 were best modeled using SVM on the 16S data (median AUROC=0.76, IQR=0.408–0.947, with 10 informative genera); for a single item from the PID5 related to perceptual disturbance the best regressor was RF (median R²=0.30, IQR=0.058–0.490, with 14 informative genera). An examination of the top informative genera associated with the previous SCID2 and PID5 traits revealed two in common, i.e., Desulfovibrio and Eubacterium, as shown in Figure 2.

Figure 1. Microbiome to predict social detachment.

(A) Violin plot of classification performance after machine learning (ML) classification (Random Forest [rf], Support Vector Machines [svm], and K-nearest neighbor [knn]) between whole genome sequencing (WGS) samples from true and false responses, with median Area Under the Receiver Operating Characteristic curve (AUROC) indicated. Top informative species from the svm method are shown along with the log2 fold change between true and false samples. (B) Violin plot of classification performance after machine learning regression using WGS samples from detachment domain average scores, with median R2 indicated. Top informative species from the rf method are shown along with the correlation of average score with species expression. (C) Overlap of the species informative in at least 70% of the ML iterations from (A) and (B).

Figure 2. Microbiome to predict ideas of reference or perceptual disturbance.

(A) Violin plot of classification performance after machine learning (ML) classification between 16S ribosomal RNA (16S) samples from true and false responses, with median Area Under the Receiver Operating Characteristic curve (AUROC) indicated. Top informative genera from the svm method are shown along with the log2 fold change between true and false samples. (B) Violin plot of classification performance after machine learning regression using 16S samples from Personality Inventory for DSM-5 (PID5) scores, with median R2 indicated. Top informative genera from the rf method are shown along with the correlation of average score with genera expression. (C) Overlap of the genera important in at least 70% of the ML iterations from (A) and (B).

The remaining top performing items and their associated distinguishing features are presented in Figure 3. Notable among these is the observation that participants who endorsed an item from the SCID2 associated with hoarding behaviors differed from those who did not based on a unique bacterial pattern that included Streptococcus as the top feature; the best performing model was SVM on 16S data (median AUROC=0.72, IQR=0.521–0.833, with 76 informative genera). For the PID5 item “I’m better than almost everyone else” the best model was RF on 16S (median R²=0.23, IQR=0.106–0.344, with 21 informative genera). Finally, for the PID5 item “I don’t hesitate to cheat if it gets me ahead” the best model was achieved with RF on 16S data (median R²=0.22, IQR=0.064–0.329, with 22 informative genera).

Figure 3. Microbiome to predict other top performing traits.

(A) Violin plot of classification performance after machine learning (ML) classification between 16S ribosomal RNA (16S) samples from true and false responses, with median Area Under the Receiver Operating Characteristic curve (AUROC) indicated. Top informative genera from the svm method are shown along with the log2 fold change between true and false samples. (B), (C) Violin plot of classification performance after machine learning regression using 16S samples from Personality Inventory for DSM-5 (PID5) scores, with median R2 indicated. Top informative genera from the rf method are shown along with the correlation of average score with genera expression.

Discussion

We hypothesized that levels of extraversion or detachment would differentially impact the composition of the gut microbiome at either the genus or species level. In support of this hypothesis, findings from this nascent effort to examine personality traits and taxonomic composition of the gut-microbiome suggest that individuals with higher social detachment have a microbiome structure that is distinguishable from those with higher social integration. Individuals high in detachment are known to avoid social engagement, likely reducing microbial input from social exchange with implications on gut microbiome composition. These results are consistent with recent findings that demonstrate that transmission of gut microbiome occurs through person-to-person interactions within a shared environment (6, 8), and gut-microbiome diversity is associated with social network size (12). It is compelling to hypothesize that this decreased microbial input may render the gut microbiome vulnerable to compositional or functional imbalances known to play a role in gut dysbiosis; the gut microbiome may partially mediate the link between social isolation and loneliness with adverse health effects (12). Of note, proinflammatory immune responses and chronically activated hypothalamic pituitary adrenocortical (HPA) axis activity are known to drive the relationship between isolation and poor health (29). These mechanisms appear to converge with our growing understanding that gut dysbiosis, possibly worsened by social detachment, plays are role in activating inflammatory responses with far reaching physiological and psychological consequences mediated through the gut-brain axis (30).

Additionally, those who endorsed items associated with visual hallucinations and ideas of reference were found to have distinct gut-microbial composition. Although these findings were based on item-level analysis and cannot be assumed to reflect personality trait domains, these items are often associated with perceptual dysregulation, psychosis, and schizotypal personality disorder. Those who endorsed visual hallucinations had microbiomes defined by Peptococcus, Prevotella, and Coprococcus. Although Peptococcus is not typically associated with psychosis or schizophrenia, there is some evidence that it is able to distinguish between people with schizophrenia and healthy controls (31). It has also been found in those with treatment-resistant depression (32) and in mice demonstrating OCD-like behaviors (33). A clearer signal emerges regarding Prevotella and Coprococcus. Increased relative abundances of Prevotella and decreased relative abundances of Coprococcus have been consistently linked with psychosis (34). Prevotella has also been associated with childhood sexual trauma (35), as well as with anxiety and depression (36). Coprococcus appears to be consistently underrepresented in depressed individuals and associated with greater levels of psychiatric severity (4). Those who endorsed ideas of reference had microbiome most notably characterized by Dialister, Unclassified_Lachnospiraceae, Tyzzerella, Ruminococcus. All of these genera, aside from Tyzzerella, have been previously linked to psychosis (34). Although Tyzzerella has not previously been connected with psychosis, it has been implicated in arthritis, suggesting a role in immune-mediated inflammatory disease (37).

The findings did not support the hypothesis that cluster B personality pathology significantly affects the composition of the gut microbiome. However, a notable finding from this study was that those who endorsed a criterion of obsessive-compulsive personality disorder (i.e., unable to discard worn-out or worthless objects), had a defined gut microbial composition. The most important genera driving this result were Streptococcus, Erysipelatoclostridium, uncultured Clostridiales Family XIII bacterium, and Tyzzerella. This finding is consistent with previous research linking abundances of Streptococcus with reduced neuroticism (12). Reduced abundances of Streptococcus have also been linked with brain development, with depletions being evident in infants who were able to pass a joint attention task (38). Interestingly, Streptococcus is known to be associated with pediatric autoimmune neuropsychiatric disorders (PANDAS), a condition in which obsessive compulsive or tic disorders emerge directly following a Streptococcus infection in childhood. The onset of symptoms is thought to be driven by antibodies from the immune response to Streptococcus that interact with neurons in the basal ganglia (39). We are not aware of other research linking the remaining bacteria with obsessive or obsessive-compulsive features, although Clostridiales at the order level has been associated with OCD (40).

The role of these taxa individually and interactively in the context of interpersonal functioning should be explored in greater detail in future studies. Dysbiosis is causally linked to atypical immune responses, which are often accompanied by an upregulated production of inflammatory cytokines that are related to the pathophysiology of depression and anxiety (30).

Evaluation of study elements

Merits of the current study include the comparatively large number of donors and the systematic assessment of personality traits and pathology in a controlled physical setting. However, several limitations must be recognized and should be addressed in follow-up initiatives. Key findings revealed correlations at both the item and personality trait domain levels. While item-level endorsements were derived from measures of personality functioning, it is possible that some may reflect clinical syndromes not typically associated with personality pathology, such as schizophrenia or obsessive-compulsive disorder. The study population consists of psychiatric inpatients with a wide range of comorbid conditions, including somatic distress with both organic and psychological etiologies. Discordant results may be observed in a patient population with a greater diagnostic homogeneity, e.g., exclusively participants with borderline personality disorder. Fecal samples were collected early during the hospitalization but not systematically. The variability in number of days between admission and fecal sample collection may have affected the findings. Participants were variably exposed to a controlled environment with scheduled and prepared meals as well as observed abstinence for alcohol/drugs and may have affected findings. Additionally, the study sample had a limited number of nutritional options in the hospital, but they were free to select meals from a menu of options. This variation in dietary consumption was not consistently monitored and could have influenced our findings. Future studies should take nutritional intake more into account. Future studies should also measure and quantify adverse health behaviors, especially substance use, to control for possible cofounding effects. For certain personality traits the prevalence was 8–14%; this study used a stringent cross-validation approach, but future studies should be conducted to validate our findings in larger donor samples.

Conclusions

Our knowledge of the gut microbiome’s mechanistic significance in the onset and maintenance of interpersonal functioning, as well as psychiatric illness, must be further developed. This study contributes to the emerging body of research on the intricate connections between gut and brain function, expanding the interdisciplinary field of psychiatric microbiology. As the relationships between gut and brain function are elucidated, these data form the foundation for interventional studies. In the future, it may be possible to directly and indirectly manipulate bacterial structure and function to treat the psychological and physiological consequences of psychiatric symptomatology and illness.