Differential co-expression networks of the gut microbiota are associated with depression and anxiety treatment resistance among psychiatric inpatients

Abstract

Background:

Comorbid anxiety and depression are common and are associated with greater disease burden than either alone. Our recent efforts have identified an association between gut microbiota dysfunction and severity of anxiety and depression. In this follow-up, we applied Differential Co-Expression Analysis (DiffCoEx) to identify potential gut microbiota biomarker(s) candidates of treatment resistance among psychiatric inpatients.

Methods:

In a sample of convenience, 100 psychiatric inpatients provided clinical data at admission and discharge; fecal samples were collected early during the hospitalization. Whole genome shotgun sequencing methods were used to process samples. DiffCoEx was used to identify clusters of microbial features significantly different based on treatment resistance status. Once overlapping features were identified, a knowledge-mining tool was used to review the literature using a list of microbial species/pathways and a select number of medical subject headlines (MeSH) terms relevant for depression, anxiety, and brain-gut-axis dysregulation. Network analysis used overlapping features to identify microbial interactions that could impact treatment resistance.

Results:

DiffCoEx analyzed 10,403 bacterial features: 43/44 microbial features associated with depression treatment resistance overlapped with 43/114 microbial features associated with anxiety treatment resistance. Network analysis resulted in 8 biological interactions between 16 bacterial species. Clostridium perfringens evidenced the highest connection strength (0.95). Erysipelotrichaceae bacterium 6_1_45 has been most widely examined, is associated with inflammation and dysbiosis, but has not been associated with depression or anxiety.

Conclusion:

DiffCoEx potentially identified gut bacteria biomarker candidates of depression and anxiety treatment-resistance. Future efforts in psychiatric microbiology should examine the mechanistic relationship of identified pro-inflammatory species, potentially contributing to a biomarker-based algorithm for treatment resistance.

1. Introduction

Anxiety and depressive disorders are among the most common psychiatric illnesses globally. In a 12- month period, 10–20% of all adult primary care visits occur during an anxiety or depressive episode, and 50% of those patients will have a comorbid disorder (Hirschfeld, 2001). Comorbidity is often the rule rather than the exception (Kessler et al., 1994). Approximately two-thirds of patients with depression have a comorbid anxiety disorder and one-third or more of patients with panic disorder or GAD also have depression (Clayton, 1990; Gorman, 1996; Klerman, 1990). STAR*D demonstrated symptom comorbidity with 53% of patients with MDD having anxiety and anxious depression (Fava et al., 2004; Kalin, 2020). Phenomenologically, both are regarded as internalizing disorders with shared symptoms that include: negative affect, impaired emotion regulation, psychomotor disturbance, and neuroticism (Sindermann et al., 2021). Neuroimaging studies consistently implicate structural and functional alterations in the orbitofrontal cortex/middle frontal cortex and in prefrontal limbic pathways across both anxiety and depression (Kalin, 2020; Sindermann et al., 2021; Van Ameringen, 2021). Individuals with co-occurring anxiety and depression have more severe symptoms and poorer prognosis than individuals with either disorder alone (Gorman, 1996; Van Ameringen, 2021). Genetic studies indicate substantial shared variance across depression and anxiety (Kendler et al., 2007). Generalized anxiety disorder (GAD) and major depressive disorder (MDD) are leading contributors to global disability, accounting for one-third to one-half of the global economic burden of mental illness (Franklin et al., 2017; Sindermann et al., 2021; Wittchen, 2002). Comorbidity is also associated with greater social and functional impairment, reduced quality of life, and increased risk for suicide(Van Ameringen, 2021; Zhou et al., 2017). In the last decade, there has been emerging evidence indicating a strong correlation between anxiety/depression and the gut microbiome (Hai-yin Jiang et al., 2018).

The gut microbiome is a complex community that functions to maintain host physiology. Gut microbiota can influence the brain, via the bidirectional communication of the Gut-Brain Axis (GBA) – the central nervous system (CNS), the intestines, and the gut microbiota (Rutsch et al., 2020). This bidirectional communication network has many roles that are essential for homeostasis and physiological functioning, including the modulation of the immune system (Barko et al., 2018). Under normal conditions, the gut microbiota defends the host against pathogens by stimulating the immune system, leading to low-level inflammation (Galland, 2014; Wu and Wu, 2012). Conversely, the composition of the gut microbiota can be influenced by the presence of chronic inflammation present in a series of diseases (Ni et al., 2017; Wang et al., 2020). Studies have shown that psychiatric symptom-atology, such as those found in depression and anxiety, are associated with an imbalance (dysbiosis) in the gut microbial community (Dickerson et al., 2017; Evans et al., 2017; Janssen, 2016; Haiyin Jiang et al., 2015). Dysbiosis of the gut microbiome can be a result of numerous factors including genetics, diet, and exposure to stress. Due to the chronic low-level inflammation, often present in a dysbiotic state (Rutsch et al., 2020), a gut microbial community in dysbiosis can lead to increased permeability of the intestines and the blood-brain barrier, both of which are associated with various psychiatric disorders (Kelly et al., 2015). Previous studies have identified an association between key microbial taxa and severity of depression and anxiety (A. Madan et al., 2020; Valles-Colomer et al., 2019). This identification has led to a significant push to identify bacterial taxa that can be used as predictive biomarkers for various psychiatric disorders (Lv et al., 2017; Misiak et al., 2020). However, the complex nature of the gut microbiome makes it difficult to understand whether the association of depression and anxiety with individual bacterial species is independent or reflective of a broader compositional shift of clusters of bacteria.

Despite the complex nature of the gut microbiome, there exist several strategies to analyze such data to identify microbiome biomarkers relevant to psychiatric conditions. Homeostasis in healthy, complex systems, such as full transcriptome or metabolomes, is maintained by preserving numerous molecular programs that involve yet unappreciated interactions. Conversely, transition to a disease state or exposure to external stressors changes in lock-step complex molecular networks. An unbiased and agnostic methodology to comprehensively identify such changes is Differential Co-Expression Analysis (DiffCoEx). Initially developed to identify gene markers for cancer datasets, DiffCoEx has been used to identify biomarkers for contrasting pairs of clinical conditions in multiple fields. For instance, DiffCoEx has been used to identify non-oncogene addiction (NOA) genes for 15 different cancer types (Hjaltelin et al., 2019), to recognize differing genes between bronchiolitis and emphysema phenotypes of Chronic Obstructive Pulmonary Disease (COPD) (Qin et al., 2019), as well as to investigate the impact of temperature on the metabolome of Drosophila melanogaster (Hariharan et al., 2014). As a proven method to identify pattern changes caused by pairs of contrasting clinical and experimental conditions (Tesson et al., 2010), DiffCoEx analysis can be particularly helpful in identifying bacterial biomarkers that differentiate between psychiatric treatment outcomes based on fecal samples of patients. In this study, we applied DiffCoEx network analysis to identify potential gut microbiota biomarker of depression and anxiety treatment resistance in a cohort of psychiatric inpatients.

2. Methods

2.1. Participants & setting

From January 1, 2015 to December 31, 2016, one hundred (N = 100) adult psychiatric inpatients at a freestanding, nonprofit facility in the Southwestern United States provided self-collected fecal swabs shortly after admission. Patients at the facility received individualized psychiatric and medical care over a prolonged length of stay (average length of stay = 49.7 ± 14.5 days). Psychiatric care included an individualized combination of comprehensive medical evaluation; optimization of psychotropic pharmaceutical drugs (Madan et al., 2015); personal, group, and skills-based psychotherapy; psychoeducation; and family work. All psychiatric and medical mediations occurred within the context of 24-h nursing supervision and a therapeutic milieu. The prolonged length of stay is anomalous for inpatient psychiatric care in the United States but have been established as necessary to achieve the demonstrated progress and remission from anxiety and depression symptomatology (Fowler et al., 2017) as well as multiple cross-cutting clinical features of psychopathology (Allen et al., 2017).

2.2. Clinical procedures

Data were gathered as part of the hospital’s Adult Outcomes Project to evaluate treatment response (Allen et al., 2009; Fowler et al., 2013) as well as a broader overlapping study aimed at determining biomarkers of psychiatric illness and treatment response (Alok Madan et al., 2017; Sharp et al., 2016). All baseline psychiatric measures were collected within 72 h of admission. The primary outcome assessments (including depression and anxiety severity) were repeated every 14 days during the hospitalization as well as at discharge. A hospital-wide web survey on laptops was used to conduct all assessments. Given pharmacologic agents’ potential to alter the gut microbiota, data related to drugs administered at the hospital were extracted from the internal pharmacy database. A finite number of medication classes were chosen in an a priori manner to minimize the possibility of spurious findings. The medication classes were antibiotics, probiotics, antidepressants, anti-psychotics, and opioids.

2.3. Clinical measures

Demographics, history of psychiatric care, and health-related information were assessed through a standardized patient information survey (Allen et al., 2009; Fowler et al., 2013). Research adaptations of the Structured Clinical Interview for DSM-IV Disorders (SCID-I/II) were employed to diagnose psychiatric disorders, such as personality disorders. Researchers at the master’s level conducted the SCID-I (First et al., 1997) and SCID-II (First et al., 2002) interviews. History of suicidality was measured using the Columbia Suicide Severity Rating Scale (CSSRS) (Madan et al., 2016; Posner et al., 2011). Trauma history was assessed with the Stressful Life Events Screening Questionnaire (SLESQ) (Allen et al., 2015; Goodman et al., 1998). The primary patient-reported outcome measures utilized in this study were the Patient Health Questionnaire – Generalized Anxiety Disorder screener (GAD-7) (Löwe et al., 2008) and the Patient Health Questionnaire – 9 (PHQ-9) (Kroenke and Spitzer, 2002); they were used to evaluate depression and anxiety severity, respectively. These measures also were selected as primary outcome measures in light of our recent report (Fowler et al., 2017) that demonstrated these measures’ sensitivity to alter among patients with SMI receiving multimodal inpatient treatment at the study psychiatric facility. Finally, the Acceptance and Action Questionnaire-II (AAQ-II) (Bond et al., 2011) was selected for these analyses given previous assessments that have demonstrated the AAQ-II to abate depression and anxiety treatment response in the patient cohort (Fowler et al., 2017). All the previously mentioned measures have sound psychometric properties and are commonly used in clinical and research settings.

2.3.1. Microbiota-related procedures

Self-collected fecal swabs from study participants shortly after admission to the hospital (20.1 ± 12.8 days) were initially stored in a 80 °C freezer. Bacterial genomic DNA then were extracted from fecal samples using the MO BIO PowerSoil DNA Isolation Kit. The 16S rDNA V4 region was PCR-amplified and sequenced using an Illumina MiSeq system. The amplification primers used (515F-806R) contain MiSeq adapters and single-end barcodes supporting sample pooling and PCR products sequencing. The 16S rRNA pipeline uses both phylogenetic and alignment-based methods. Read pairs were first demultiplexed using unique molecular barcodes, then merged using USEARCH v7.0.1090. The UPARSE algorithm was employed to cluster 16S rDNA sequences into Observed Operational Taxonomic Units (OTUs) at the 97% similarity cutoff value. Detected OTUs were further annotated by mapping to an optimized SILVA Database version that contained only the 16S v4 region. Abundance was quantified by mapping demultiplexed reads to OTUs. Using a minimum rarefaction depth of 4815, no samples were omitted due to insufficient reads. The final OTU table was then used for alpha-diversity, beta-diversity, and phylogenetic analysis.

Genomic bacterial DNA (gDNA) extraction methods were refined for Whole Genome Shotgun Sequencing (WGS) to improve the bacterial DNA yield and reduce the background amplification (The Human Microbiome Project Consortium, 2012a, 2012b). Paired-end sequencing reads were filtered for low quality sequences and for Illumina phix sequences. Subsequently, Illumina adapters were removed using bbduk (BBMap version 37.58; Bushnell et al., 2017); resulting sequences were mapped to the human hg38 reference database using bowtie2 v.2.3.4.3 (Langmead and Salzberg, 2012) completely and with high-stringency to remove sequence-level host contamination. Taxonomic profiles were inferred using MetaPhlAn2 (Segata et al., 2012), and functional profiling of the microbial community was done using HUMAnN2 (Franzosa et al., 2018). The final output files were represented in biom ormat (McDonald et al., 2012).

2.4. DiffCoEx and network analysis

DiffCoEx was used to identify modules of microbial features, whose abundance differed significantly due to treatment resistance states of depression and anxiety (Tesson et al., 2010). Following the construction of the dissimilarity matrices based on pairwise correlations of bacterial and pathway abundance data, DiffCoEx used hierarchical clustering to identify clusters of microbial features that were significantly different based on presence of treatment resistance as indicated by distinct co-expression patterns. We incorporated permutation testing in the DiffCoEx algorithm to assess the significance of within and module-to-module co-expression changes associated with each pair of differing phenotypes. More specifically, for each module and between each pair of modules of microbial features, dispersion values were calculated using 1000 permutations to assign P-values for within-module, and module-to-module differentiation. For each pair, only those modules with within-module dispersion values smaller than 0.05 (or < 50/1000) were selected.

DiffCoEx algorithm’s outputs revealed an overwhelming amount of overlap between microbial biomarkers identified for depression treatment resistance and anxiety treatment resistance. Thus, common outputs of both phenotypes were used for further network analyses. Using hierarchical clustering within the WGCNA framework (Langfelder and Horvath, 2008), we generated an undirected, weighted correlation network of these biomarkers. With a weight cut-off of 0.3, we visualized the network using the Cytoscape (Shannon et al., 2003) software.

2.5. Pubmed text-based knowledge search

To extract manuscripts that contained species and MESH terms chosen in an a prior manner, we used the PubTator Central text mining platform (Wei et al., 2019) to search PubMed and PMC databases (through February 17, 2022). The search was not limited to a publication timeframe, however, reviews, meta-analyses, and publications not written in English were excluded.

2.6. Ethics

The present study adheres to the guidelines and ethical principles outlined in the Declaration of Helsinki. Participants provided written informed consent following a full explanation of all procedures. Through informed consent, study participants attested to the voluntary nature of their participation. Study design was approved by the Institutional Review Board (IRB) at Baylor College of Medicine.

3.1. Patient characteristics

In this study, the final cohort (n = 100) consisted primarily of young, Caucasian adults. At admission, the majority were unemployed, never married, and many had significant trauma histories, as well as some degree of suicidal ideation. Individuals presented with multiple Axis I disorders, and many additionally had high levels of functional impairment and service utilization consistent with current definitions of severe mental illness (SMI) (Ronald C Kessler et al., 2010). Most individuals in this cohort had extensive histories of outpatient and inpatient psychiatric treatment before this study (Table 1). At discharge, 54% of individuals had a PHQ-9 score of >5, signifying treatment resistant depression, and 43% of individuals had treatment resistant anxiety, with a GAD-7 score of >5. A small population of individuals exclusively had treatment resistant depression (5%), or exclusively had treatment resistant anxiety (3%).

Table 1.

Baseline demographic and burden of illness characteristics of sample.

	Mean (SD)	N (%)
Age, years	37.0 (13.8)
Sex, female		47 (47.0)
Ethnicity, White		90 (90.0)
Single/Never Married		41 (41.0)
Some college or greater		70 (70.0)
Unemployed (past 30 days)		51 (51.0)
Total Axis I Disorders	2.7 (1.4)
Total Axis II Disorders	0.6 (0.8)
Suicidal ideation, past month		65 (65.0)
Suicide attempt(s), lifetime	0.9 (1.7)	35 (35.0)
Substance Use Disorders		50 (50.0)
Major Depressive Disorders		74 (74.0)
Bipolar Spectrum		14 (14.0)
Anxiety Spectrum		72 (72.0)
Psychotic Spectrum		2 (2.0)
Borderline Personality Disorder		18 (18.0)
Traumatic life event(s), lifetime	2.4 (2.2)	81 (81.0)
Any Personality Disorder		38 (38.0)
Outpatient therapists (lifetime)	3.8 (2.6)
Psychopharmacologists (lifetime)	2.9 (2.2)
Hospitalizations for acute psychiatric care (lifetime)	1.0 (1.4)
Hospitalizations for extended psychiatric care (lifetime)	1.0 (2.0)

3.2. Modules of bacterial features are differentially co-expressed between depression treatment outcomes

We used DiffCoEx to detect patterns of differential co-expression between the depression treatment responder group (n = 55 patients) and depression treatment resistant group (n = 45 patients). Of the 10,403 bacterial features analyzed by DiffCoEx, 105 were assigned to one of three modules, which were arbitrarily assigned colors (Fig. 1A). Features in all three modules identified in DiffCoEx show significant differential co-expression in the binary variable (depression treatment resistance = 1 vs 0) (dispersion statistics p value<0.05). The turquoise module, containing 44 bacterial features, indicated a stronger unidirectional differential correlation in microbial relative abundance associated with treatment response; specifically, the 44 features were strongly correlated only in the depression treatment resistant patient group. A less distinctly unidirectional correlation difference was seen in the blue and red modules, containing 42 and 19 bacterial features respectively; interestingly, for both modules the depression treatment resistant group showed a loss of correlation compared to the depression treatment response group.

Fig. 1.

Differential Co-Expression (DiffCoEx) plots for clinical depression treatment resistance and anxiety treatment resistance. Patients were stratified based on their treatment resistance status for each of the two conditions, depression, and anxiety. For each condition, bacteria species and pathways were clustered into modules with strong correlation within module and patient group, but with differential correlation between the distinct treatment resistant patient groups. Modules are depicted by arbitrarily chosen colors. Inside each module, pathways and species are significantly differentially co-expressed between the treatment resistant (indicated with 1) and lack of treatment resistance (indicated with zero) for each clinical condition. Significance was achieved for p < 0.05 based on dispersion statistics.

3.3. Modules of bacterial features are differentially co-expressed between anxiety treatment outcomes

We used DiffCoEx to detect patterns of differential co-expression between the anxiety treatment responder group (n = 54 patients) and anxiety treatment resistant group (n = 46 patients). DiffCoEx analysis revealed 157 out of 10,403 bacterial features that were assigned to one of two modules (Fig. 1B). The turquoise module, containing 114 bacterial features, indicated an increased correlation of microbial relative abundance in patients that had treatment resistant anxiety (Anxiety_treatment resistance = 1). A less distinctly unidirectional correlation difference was seen in the blue module, containing 43 bacterial features, with a higher correlation of the 43 features in the anxiety treatment responder patient group.

3.3.1. Detection of a commonly induced microbial features modules in depression and anxiety treatment resistance

Due to the high comorbidity between depression and anxiety, we determined common and concordant microbial features associated with both anxiety and depression treatment resistance. We focused on common features in the turquoise modules of depression and anxiety inferred by DiffCoEx due to their concordant direction of correlation difference. 43 out of the 44 microbial features that were associated with treatment resistant depression overlapped with 43 out 114 microbial features associated with treatment resistant anxiety. Interestingly, the same 43 microbial features showed decreased co-expression in both the depression treatment resistant and the anxiety treatment resistant groups. The integration of differential co-expression results for these two comorbidities has enabled us to prioritize 43/10,403 microbial features we want to investigate in the future.

3.4. Literature search of microbial features

Once the overlapping microbial features between treatment resistance in both depression and anxiety were identified, we employed a knowledge-mining tool, Pubtator, to determine the relevance and importance of our findings. Specifically, Pubtator was used to comprehensively explores PubMed abstracts and PMC full-text articles using the 43 significant features associated with resistance in both anxiety and depression. We searched for articles mentioning any of the 43 features and a list of previously determined MESH terms associated with dysbiosis, inflammation, short chain fatty acids, and with different presentations of depression and anxiety. Conducting text mining allows us to corroborate some of our findings to present DiffCoEx as an effective tool for biomarker identification. The results of the Pubator search revealed inflammation, as well as dysbiosis, were the most annotated mesh terms. Our literature search results further showed that 42% of the microbial features identified with our methods have >15 annotations related to any mesh terms, corroborating our findings. The remaining 58% microbial features that had fewer than 15 annotations related to any mesh term are potentially newly identified biomarker candidates for psychiatric disorders, pointing to a direction for future research.

3.5. Network analysis

Cytoscape (Shannon et al., 2003), a software used to visually build and analyze biological networks, in conjunction with the weighted gene co-expression network analysis (Langfelder and Horvath, 2008), was used to identify the biological interactions in the gut microbiome that could impact depression and treatment outcomes. The overlapping 43 microbial features previously identified in the DiffCoEx analysis were qualified to be included in the network analysis. At a feature correlation weight cutoff of 0.3, eight network connections were identified (Fig. 2). Of those connections, three of them had a weight >0.7, indicating high connection strength. Clostridium perfringens and Prevotella bivia had the strongest connection strength (0.95). Facklamia hominis and Anaeroccocus prevoti (0.79) had the second strongest connection, with a connection strength of 0.79, andErysipelotrichaceae bacterium 6_1_45 and Clostridium innocuum had the third strongest connection, with a connection strength of 0.71. Supporting data for the microbial features network analysis are shown in Table 2.

Fig. 2.

Microbial features networks for the pathways and bacteria species identified by DiffCoEx.

A. Correlation was assessed for each pair of the 43 overlapping features between the turquoise modules identified; only feature pairs with correlation cutoff >0.3 were indicated by edges.

B. Detailed view of the microbial subnetwork of strongly correlated features. Node names correspond to either species or pathways. Edge thickness properties represent the weights of the edges; connected components of nodes are labeled with the same color.

Table 2.

WGCNA network edge-weight table with a weight cut-off of 0.3. Among all possible edges (connections) between the 43 bacterial features, 8 of the edges have a weight cut-off of higher than 0.3.

Network Edge	Weight
Clostridium perfringens: Prevotella bivia	0.95
Anaerococcus hydrogenalis: Facklamia hominis	0.79
Clostridium innocuum: Erysipelotrichaceae bacterium_6_1_45	0.71
Marvinbryantia formatexigens: Synergistes_sp_3_1_syn1	0.51
Facklamia hominis: Prevotella bergensis	0.38
Anaerococcus hydrogenalis: Anaerococcus prevotii	0.38
Anaerococcus hydrogenalis: Prevotella bergensis	0.31
Anaerococcus prevotii: Facklamia hominis	0.31

The results of the literature search were further truncated to only highlight the microbial features that had a connection strength of at least 0.3 (Table 3 and Table 4).

Table 3.

Literature search results for bacterial species in network analysis. 15 mesh terms related to anxiety and depression were searched in Pubtator in conjunction with the identified species. The numbers represent the number of manuscripts associated with the search terms.

Bacterial Species	Dysbiosis	Inflammation	Short Chain Fatty Acids	Animal Models & Depression	In vitro & Depression	Research Subjects & depression	Dysbiosis & Depression	Inflammation & Depression	Short Chain Fatty Acids & Depression	In vitro & Anxiety Disorders	Animal Models & Anxiety Disorders	Research Subjects & Anxiety Disorders	Dysbiosis & Anxiety disorders	Inflammation & Anxiety Disorders	Short Chain Fatty Acids & Anxiety Disorders
Enterobacteriaceae bacterium_9_2_54FAA	627	6350	362	35	157	0	7	31	5	0	0	0	1	2	3
Bacteroides sp3_1_40A	723	1177	495	5	5	0	9	8	4	0	0	0	1	0	1
Klebsiella pneumoniae	41	548	13	2	5	0	0	4	0	0	0	0		0	0
Lachnospiraceae bacterium_4_1_37FAA	312	323	203	2	1	0	10	6	4	0	0	0	1	0	0
Lachnospiraceae bacterium_5_1_57FAA	312	323	203	2	1	0	10	6	4	0	0	0	1	0	0
Lachnospiraceae bacterium_6_1_63FAA	312	323	202	2	1	0	10	6	4	0	0	0	1	0	0
Fusobacterium nucleatum	103	370	14	0	0	0	0	1	1	0	0	0	0	0	0
Streptococcus agalactiae	15	256	1	0	0	0	0	0	0	0	0	0	0	0	0
Clostridium perfringens	33	153	63	2	6	0	0	1	0	0	0	0	0	0	0
Erysipelotrichaceae bacterium_3_153	70	67	29	0	0	0	0	0	0	0	0	0	0	0	0
Erysipelotrichaceae bacterium_6_145	70	67	29	0	0	0	0	0	0	0	0	0	0	0	0
Enterococcus faecium	27	77	14	0	1	0	0	1	0	0	0	0	0	0	0
Enterobacter cloacae	7	41	1	0	0	0	0	0	0	0	0	0	0	0	0
Peptostreptococcus anaerobius	7	17	3	0	0	0	0	0	0	0	0	0	0	0	0
Enterobacter aerogenes	1	17	0	0	0	0	0	0	0	0	0	0	0	0	0
Prevotella bivia	8	14	0	0	0	0	0	0	0	0	0	0	0	0	0
Citrobacter freundii	1	13	0	0	0	0	0	0	0	0	0	0	0	0	0
Acidaminococcus sp_D21	9	10	5	0	0	0	0	1	0	0	0	0	0	0	0
Bifidobacterium catenulatum	4	4	2	0	0	0	0	0	0	0	0	0	0	0	0
Parabacteroides goldsteinii	3	10	1	0	0	0	0	0	0	0	0	0	0	0	0
Clostridium scindens	4	4	2	0	0	0	0	0	0	0	0	0	0	0	0
Coprococcus eutactus	3	3	2	0	0	0	1	0	0	0	0	0	0	0	0
Synergistes sp_3_1_syn1	1	2	0	0	0	0	0	0	0	0	0	0	0	0	0
Blautia hydrogenotrophica	0	1	1	0	0	0	0	0	0	0	0	0	0	0	0
Catenibacterium mitsuokai	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0
Clostridium innocuum	1	1	1	0	0	0	0	0	0	0	0	0	0	0	0
Facklamia hominis	0	1	0	0	0	0	0	0	0	0	0	0	0	0	0
Lachnospiraceae bacterium_1_1_57FAA	1	1	1	0	0	0	1	0	1	0	0	0	0	0	0
Marvinbryantia formatexigens	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0
Solibacillus silvestris	1	0	0	0	0	0	0	0	0	0	0	0	0	0	0
Anaerococcus hydrogenalis	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0
Anaerococcus prevotii	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0
Bacteroides clarus	1	1	0	0	0	0	0	0	0	0	0	0	0	0	0
Eremococcus coleocola	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0
Prevotella bergensis	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0
Pseudomonas fragi	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0

Table 4.

Literature search results for bacterial pathways in network analysis. Mesh terms related to anxiety and depression were searched in Pubtator in conjunction with the identified species and pathways. The numbers represent the number of manuscripts associated with the search terms.

Bacterial Pathways	Dysbiosis	Inflammation	Short Chain Fatty Acids	Animal Models & Depression	In vitro & Depression	Research Subjects & depression	Dysbiosis & Depression	Inflammation & Depression	Short Chain Fatty Acids & Depression	In vitro & Anxiety Disorders	Animal Models & Anxiety Disorders	Research Subjects & Anxiety Disorders	Dysbiosis & Anxiety disorders	Inflammation & Anxiety Disorders	Short Chain Fatty Acids & Anxiety Disorders
l-lysine biosynthesis I	22	1691	39	12	92	0	1	8	0	0	1	0	0	0	1
Pentose phosphate pathway	7	175	8	1	6	0	0	0	0	0	0	0	0	0	0
aerobic respiration I (cytochromec)	0	66	0	4	4	0	0	1	0	0	0	0	0	0	1
Rubisco shunt	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0
L-1,2-propanediol degradation	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0
superpathway of acetyl-CoA biosynthesis	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0
superpathway of purine nucleotide salvage	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0

4. Discussion

The present study identified co-expression patterns of gut bacteria and pathways associated with depression and anxiety treatment resistance. As hypothesized there was a strong and concordant directional co-expression pattern for both. 43 out of the 44 bacterial features associated with treatment resistant depression were also associated with treatment resistant anxiety. A comprehensive search of the existing literature using inputs associated with microbial features, psychiatric features, and broadly brain-gut-axis dysregulation features identified taxa and pathways previously implicated in brain-gut dysfunction as well novel features that are uniquely associated with the study sample. Moreover, key biological interactions among bacterial species from the overlapping features were discovered through the construction of an undirected, weighted correlation network analysis. Findings from this study narrow the scope of bacterial features associated with brain-gut dysfunction from >10,000 to fewer than 50 – features that may enable more focused research efforts in the future.

Notable taxa from this study to highlight include: Erysipelotrichaceae bacterium 6_1_45 (most related to existing literature) as well as Clostridium perfringens and Prevotella bivia (which demonstrated the highest connection strength from our network analysis). These taxa are associated with inflammation, globally as well as specifically related to histotoxic and intestinal infections (Freedman et al., 2015; Li et al., 2013; Shrestha et al., 2018) and pelvic inflammatory disease and bacterial vaginosis (Aroutcheva et al., 2008; Gilbert et al., 2019; Mikamo et al., 2014). The role of these taxa individually and interactively in the context of anxiety and depression 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 (Schirmer et al., 2016). Previous studies have found an association between an increase in pro-inflammatory cytokines and the pathophysiology of depression and anxiety (Alshammari et al., 2020; Carlessi et al., 2021; Dantzer et al., 2008; Nobis et al., 2020; Schirmer et al., 2016). Notably, toxins present in Clostridium perfringens can cause a dramatic increase in cytokine production, leading to cytokine cascades during sepsis (Hifumi, 2020; Takehara et al., 2019). Thus, examining the mechanistic relationship of these inflammatory species in depression and anxiety is a promising future direction.

Our work quantifies and identifies disruption in the correlation of different microbiota genera. As such, it poses multiple interesting questions: which mechanisms are responsible for such disruptions and how can one potentially intervene? To strengthen a positive correlation between two genera that may be lacking or limited among patients with treatment resistance, a potential intervention would be probiotic supplementation with both genera, informed by preclinical studies in animal models of depression and/or anxiety. Additionally, future efforts may build upon models that use host transcriptomes. LINCS/Connectivity Map (Subramanian et al., 2017) is one such approach that represents a fertile resource to determine potential mechanisms at the level of gene suppression or over-expression, or to stratify over 2300 chemical compounds based on their positive or negative correlation with a novel gene signature. Similar well-powered resources have not been developed for microbiome disruptions. However, with the continuing commoditization of 16S sequencing, and lowering prices for WGS metagenome sequencing, we expect that such resources will eventually be developed and can be used to inform mechanistic studies.

A notable strength of this study is our analytic approach: DiffCoEx provides a more nuanced bioinformatic approach relative to other techniques used to analyze microbiota data. Differential abundance analysis of metagenomics data that employ standard analytics such as ANOVA and Kruskal-Wallis, and subsequent feature selection and machine learning methods including LefSe, RandomForest, Support Vector Machines, are extensively employed in the current literature (e.g., Fung et al., 2021; Estaki et al., 2020) - including our previous reports (Saulnier et al., 2011; Riehle et al., 2012; Aagaard et al., 2012; Madan et al., 2020; Thompson et al., 2021; El Saie et al., 2022). The focus of this study, however, was to explore the network-based analysis of gut microbiome in conjunction with key clinical responses to psychiatric care. The aforementioned approaches do not allow for network-based analysis. A seminal network analysis methodology is weighted correlation network analysis (WGCNA); it has been used extensively in microbiome analysis, in disease models spanning cancer (e.g., Park et al., 2019; Vernocchi et al., 2020) to dietary interventions in humans and farm animals (e.g., Murga-Garrido et al., 2021; Xie et al., 2021). An alternative differential co-expression method to DiffCoEx is Differential Gene Correlation Analysis (DGCA), which identifies and reports differential co-expression at the level of individual pairs of features. DGCA has been successfully employed in analysis of skin (Meisel et al., 2018) and gut (Zou et al., 2022) microbiomes. Unlike WGCNA, which finds modules of correlated features, and DGCA, which identifies differential correlation at the level of individual pairs, DiffCoEx uniquely identifies modules of microbiome species differentially co-expressed; DiffCoEx also provides a robust statistical framework for the final modules reported, using permutation testing.

The relatively large number of donors and the systematic assessment of treatment response in a controlled physical environment are strengths of the present study. Nonetheless, several limitations must be acknowledged and should be addressed in future efforts. The study sample includes psychiatric inpatients with considerable comorbidity, including personality disorders that may be associated with health behaviors (e.g., substance misuse) that influence the structure and function of the gut microbiota. A more diagnostically homogenous patient population may have discordant findings. Though the study sample had a restricted number of dietary options during their hospitalization, they were free to choose from a menu of options for meals during their hospitalization. This variability in dietary intake was not systematically tracked and may have confounded the observed data. Future research should better account for diet, given that there is increasing evidence of the influence of diet on depression and anxiety severity and response to treatment (Adan et al., 2019). The longitudinal measures of treatment outcome were exclusively patient reported with significant potential for under/overreporting of distress. Additionally, results from this study are based on a single fecal sample collected early in treatment. Future research should include serial collection of additionally performance-based outcome measures and fecal samples and to allow for an examination of potential changes (or lack thereof) in the gut microbiota in response to treatment.

5. Conclusion

In conclusion, a novel, data-reduction analytic strategy was used to examine microbial features for the identification of unique bacterial co-expression patterns associated with treatment outcomes for depression, treatment outcomes for anxiety, and common and concordant features associated with both depression and anxiety treatment resistance. This study only begins to examine the co-abundant expression patterns of gut bacteria and pathways in subjects with depression and anxiety. Further research is needed to advance our understanding of the mechanistic role of the gut microbiome in the development, maintenance, and treatment response of psychiatric disorders. This study advances the nascent literature of the complex interdependence between the gut and neural functioning, wherein greater interdisciplinary study of psychiatric microbiology integrates fields of psychiatry and microbiology as well as advanced analytics. Psychiatric microbiology promises a more nuanced understanding of interdependent human functions with the possibility of directly and indirectly manipulating bacterial structure and function to treat psychiatric symptomatology in the future.