Gut Microbiome in Schizophrenia: Altered Functional Pathways Related to Immune Modulation and Atherosclerotic Risk

Abstract

Emerging evidence has linked the gut microbiome changes to schizophrenia. However, there has been limited research into the functional pathways by which the gut microbiota contributes to the phenotype of persons with chronic schizophrenia. We characterized the composition and functional potential of the gut microbiota in 48 individuals with chronic schizophrenia and 48 matched (sequencing plate, age, sex, BMI, and antibiotic use) non-psychiatric comparison subjects (NCs) using 16S rRNA sequencing. Patients with schizophrenia demonstrated significant beta-diversity differences in microbial composition and predicted genetic functional potential compared to NCs. Alpha-diversity of taxa and functional pathways were not different between groups. Random forests analyses revealed that the microbiome predicts differentiation of patients with schizophrenia from NCs using taxa (75% accuracy) and functional profiles (67% accuracy for KEGG orthologs, 70% for MetaCyc pathways). We utilized a new compositionally-aware method incorporating reference frames to identify differentially abundant microbes and pathways, which revealed that Lachnospiraceae is associated with schizophrenia. Functional pathways related to trimethylamine-N-oxide reductase and Kdo₂-lipid A biosynthesis were altered in schizophrenia. These metabolic pathways were associated with inflammatory cytokines and risk for coronary heart disease in schizophrenia. Findings suggest potential mechanisms by which the microbiota may impact the pathophysiology of the disease through modulation of functional pathways related to immune signaling/response and lipid and glucose regulation to be further investigated in future studies.

1. INTRODUCTION

Schizophrenia is a debilitating illness of the brain and body. It is associated with cognitive and functional deficits as well as higher medical comorbidity and shortened life expectancy that limit individuals’ quality and quantity of life.¹ Individuals with schizophrenia are more prone to diseases associated with aging, namely cardiovascular diseases (CVD),² and exhibit age-associated physiological changes, such as inflammation, at earlier ages.^3,4 A growing body of literature strongly supports the hypothesis of accelerated biological aging in persons with schizophrenia.^5–7 Understanding the mechanisms of potential accelerated aging is imperative to improving the quality and quantity of life in schizophrenia. Mounting evidence links the gut microbiome – the modifiable “second genome” consisting of trillions of diverse microbes living in the human gut – to key determinants of human health and disease.⁸ The gut microbiome is critical in maintaining human physiology. It regulates many metabolic processes essential for optimal health that cannot be maintained by human cells, particularly in maintaining homeostasis of host metabolism, stimulating normal immune maturation, and stabilizing the gut barrier.⁹ Imbalance (or dysbiosis) of the microbiota has been associated with most aging-related diseases, including diabetes, obesity, CVD, and neurodegenerative diseases.⁸ These observations have raised a possibility that the human microbiome is a key modulator of host organismal aging, which may contribute to premature morbidity and mortality in schizophrenia. Schizophrenia is also characterized by increased gut permeability.¹⁰ In this way, gut dysbiosis may contribute to increased translocation of enteric microbes into systemic circulation, thus potentially contributing to the pro-inflammatory milieu and other physiological abnormalities implicated in schizophrenia.¹¹

Research on the gut microbiome in psychiatric disorders is in early stages and the role of the gut-brain axis in schizophrenia is not fully understood. Our previous systematic reviews in this area^12,13 found a small number of studies examining differences in the gut microbiota between patients with first-episode psychosis and schizophrenia and non-psychiatric comparison subjects (NCs). Although all investigations reported beta-diversity differences between patients and controls, there was minimal consensus with regards to abundances of microbial taxa. Discrepancies may be explained, at least partly, by misinterpretations of relative abundance data due to the compositionality of microbiome data, which we sought to resolve using novel methods. The present study is an extension of our previous article¹⁴ characterizing the gut microbiome in patients with chronic schizophrenia; it includes participants from the previous study. With an expanded cohort (n = 96), we build upon our previous results (n = 50) and incorporate new analyses on functional pathways, which have not been previously reported in any subjects. The objectives of this study were three-fold. 1) We evaluated differential abundance using Songbird, a newly-developed compositionally-aware method for differential abundance testing.¹⁵ This approach is more robust than previous methods and may help reveal more reproducible results and lead to more stable inferences of compositional change. 2) We also evaluated the predicted functional potential of the microbial community using PICRUSt2.^16,17 Understanding the genetic potential of the intestinal ecosystem may be more important and robust across different studies and cohorts than simply knowing the taxonomic identity of microbes present, particularly in elucidating potential pathways and mechanisms by which the gut microbiome influences downstream clinical outcomes. 3) Finally, using supervised learning (Random Forests models), we evaluated the microbiome as a diagnostic predictive tool by constructing taxonomy-based and function-based classifiers.

Based on previous studies,^14,18,19 we predicted that beta-diversity, but not alpha-diversity, of microbial taxa and predicted functional pathways would be significantly different between schizophrenia and NC groups. Furthermore, consistent with the theoretical framework of accelerated biological aging in schizophrenia, we also hypothesized that schizophrenia would be characterized by alterations in functional pathways involved in immune modulation and cardiovascular risk. We also anticipated that alterations in microbiota function would be associated with indicators of inflammation and CVD risk, in schizophrenia.

2. MATERIALS AND METHODS

2.1. Participants

Participants were recruited from the greater San Diego area through the University of California San Diego (UCSD). The research protocol was approved by the UCSD Human Research Protections Program, and all participants provided written informed consent prior to participation. The sample included 48 subjects with schizophrenia and 48 matched NCs between the ages of 27 and 76 years. Participants with schizophrenia were diagnosed based on the Structured Clinical Interview for the DSM-IV-TR (SCID).²⁰ NCs were screened for major neuropsychiatric disorders using the Mini-International Neuropsychiatric Interview (MINI)²¹ and excluded if they had a past or present diagnosis of a major neuropsychiatric illness. Exclusion criteria were: other current major DSM-IV-TR Axis I diagnoses; alcohol or other substance (other than tobacco) abuse or dependence within 3 months prior to enrollment; diagnosis of dementia, intellectual disability disorder, or a major neurological disorder; or any medical disability that interfered with a subject’s ability to complete study procedures. Additional details regarding recruitment and subject selection have been previously reported.¹⁴

To account for potential methodological differences that might bias results and obscure biologically meaningful compositional differences^22,23 and to control for clinical factors and known major drivers of microbiome changes,²⁴ NCs were matched to schizophrenia subjects on the same sequencing plate after initial recruitment. Following a stable matching algorithm, for each SZ subject we found the nearest matching NC neighbor based on age, sex, race, body mass index (BMI), and history of antibiotic use, requiring that the sample must come from the same sequencing plate.

2.2. Sociodemographic and clinical assessment

Sociodemographic characteristics (i.e., age, sex, race/ethnicity, current smoking status) and illness-related factors (i.e., age of onset, duration of illness, antipsychotic medication dosages) were obtained through participant interview and review of the available medical records. Antipsychotic medication dosages were converted to World Health Organization (WHO) defined daily dose.²⁵ BMI was calculated based on height and weight (kg/m²). Participants were interviewed by trained staff and completed the following standardized assessments. Positive and negative psychiatric symptoms were evaluated using interviewer-administered Scales for Assessment of Positive Symptoms and Negative Symptoms (SAPS, SANS).^26,27 Depression was assessed using the Patient Health Questionnaire (PHQ-9).²⁸ Health-related quality of life and functioning was evaluated using the physical and mental health component scores from the Medical Outcomes Study 36-item Short Form (SF-36).²⁹ Medical comorbidity was measured with the total score and severity index from the Cumulative Illness Rating Scale (CIRS).³⁰ The Framingham 10-year Coronary Heart Disease (CHD) Risk Score was calculated according to Wilson et al.³¹

2.3. Inflammatory biomarker assays

A subsample of participants was randomly selected to provide fasting blood samples for testing of specific biomarkers. Of these individuals, a greater proportion was from the schizophrenia group (p < 0.001) and of non-Caucasian race (p < 0.047). Otherwise, the subsample did not significantly differ from the overall sample on age, gender, BMI, medical comorbidity, smoking status, psychiatric symptomatology, or antipsychotic medication use (ps > 0.05). Based on prior empirical literature,^3,32,33 we focused on three pro-inflammatory markers that have been observed to be linked with schizophrenia (high-sensitivity C-reactive protein [hs-CRP], tumor necrosis factor [TNF]α, and interleukin [IL]-6) and explored relationships with an anti-inflammatory cytokine (IL-10). Plasma TNFα, IL-6, and IL-10 were quantified using the Meso Scale Discovery (MSD) MULTI-SPOT® Assay System (MSD, Rockville, MD, USA) and analyzed on a SECTOR Imager 2400 instrument (MSD). Samples were run in duplicates, using V-PLEX Human Biomarker panels (Catalog no. K151A0H-2). Plasma hs-CRP was measured with MSD enzyme-linked immunosorbent assay (ELISA). Intra-assay variability and inter-assay variability was <10% for all assays.

2.4. Fecal sample collection and processing

Participants collected fecal samples using home collection kits (BD SWUBE Dual Swab Collection System; BD Worldwide) and returned them via mail in a self-addressed envelope. Returned samples were immediately frozen and stored at −80°C until aliquoting into 96 well plates for DNA extraction. DNA extraction and 16S rRNA amplicon sequencing were completed using the Earth Microbiome Project standard protocols.^24,34 In brief, DNA was extracted using the Qiagen MagAttract PowerSoil DNA kit,³⁵ followed by PCR amplification on the V4 region of the 16S rRNA gene with unique reverse barcoded primers and sequenced on Illumina MiSeq and HiSeq 4000 platforms, yielding paired-end, 150 base-pair reads, with a minimum of 7,872 (median 47,615) reads per sample.^36,37

2.5. 16S data processing and analysis

The raw sequencing data were processed using QIIME 2 version 2019.7.^38,39 Raw sequences were demultiplexed and processed using Deblur to generate amplicon sequence variants (ASVs)⁴⁰ and previously recognized bloom sequences were removed.⁴¹ Deblur ASVs were inserted into the Greengenes 16S rRNA gene tree using SEPP via q2-fragment-insertion,^42,43 and taxonomy was assigned using UCLUST.⁴⁴ The output feature table was rarefied to 7,000 sequences per sample. PICRUSt2 (phylogenetic investigation of communities by reconstruction of unobserved states, version 2) was performed to predict functional potential of microbial communities, based on metagenomes inferred from 16S data.^16,17 PICRUSt2 predictions were made based on the following gene family databases: Kyoto Encyclopedia of Genes and Genomes (KEGG) orthologs (KO),⁴⁵ Enzyme Classification (EC) numbers, and the MetaCyc Metabolic Pathway Database.⁴⁶

Primary measures of the gut microbiome included abundances of 1) microbial taxa (e.g., ASVs) and 2) functional pathways. For each, core diversity metrics of alpha-diversity (within-sample) and beta-diversity (between-samples) were calculated. Metrics of alpha-diversity included number of observed features, Shannon diversity index,⁴⁷ and Faith’s Phylogenetic Diversity (PD).⁴⁸ Beta-diversity was calculated using unweighted UniFrac^49,50 and Bray-Curtis dissimilarity⁵¹ for taxa, and Jaccard distance⁵² for predicted functional pathways. Output matrices were ordinated using principal coordinate analysis (PCoA) and visualized using EMPeror.⁵³

2.6. Statistical analysis

Participant sociodemographic and clinical characteristics were summarized and analyzed using paired-samples t tests and McNemar χ² tests for continuous and discrete variables, respectively. Analyses of microbiome 16S sequencing data were performed in QIIME 2. A two-sided alpha level of p < 0.05 was used to determine statistical significance. The Benjamini-Hochberg procedure⁵⁴ was used to control false discovery rate (FDR) at alpha = 0.05 to account for multiple comparisons within each set of analyses.

2.6.1. Diversity analysis.

Alpha-diversity metrics were compared using nonparametric Wilcoxon signed-rank tests. Beta-diversity distances between samples were compared with PERMANOVA⁵⁵ with 999 permutations.

2.6.2. Differential abundance analysis.

Differential abundance testing of microbial taxa and functional pathways were first performed using ANCOM.⁵⁶ We used features identified by ANCOM to seed Songbird to further examine differential abundance.¹⁵ Songbird is a recently developed method that uses reference frames (or balances) to account for the compositional nature of microbiome data. By comparing log-ratios of features, this method circumvents bias introduced by unknown total microbial loads. We employed Songbird to estimate the relative differentials of microbial taxa and functional pathways based on disease status (schizophrenia vs. NC). Relative differentials were then ranked with respect to diagnosis (i.e., most associated with schizophrenia vs. NC groups) and visualized using Qurro.⁵⁷ Balances were created to determine differential abundance, relative to reference frame, and compared using Wilcoxon signed-rank tests. Spearman’s correlations were performed to explore associations of log-ratios of functional pathways with biomarkers of inflammation and CVD risk.

2.6.3. Random forests analysis.

We used supervised learning Random Forests models⁵⁸ to classify samples according to diagnosis (i.e., schizophrenia or NC). For ASV models, data were divided into 80%−20% training-test sets and prior to training the model, the data were imported into Calour,⁵⁹ normalized to 10,000 reads per sample and, in order to limit the number of features in the model, low abundance features (20 reads across all samples or less) were clustered together based on Euclidean distance hierarchical clustering. In PICRUSt2 models, for MetaCyc pathways, we normalized the data to 5,000,000 reads per sample, since there is approximately 100x more pathways than ASVs, and filtered features with <10,000 occurrences across all samples; for KEGG orthologs, we normalized the data to 50,000,000 reads per sample, due to approximately 1,000x more KEGG orthologs than ASVs, and filtered features with <1,000,000 occurrences across all samples. Hyperparameter optimization of the model was performed using the grid search method, as implemented in Python scikit-learn package.^60–62 ROC curves were generated from the stratified 5-fold cross-validation results and the confusion matrices were calculated on the basis of the best model obtained from cross-validation. The best model was selected on the basis of accuracy score.

3. RESULTS

Participant demographic and clinical characteristics are presented in Table 1. As expected, the schizophrenia group had worse psychiatric symptoms, lower levels of physical/mental well-being, higher rates of smoking, and greater medical comorbidity. Although schizophrenia and NC groups did not differ in mean levels of blood-based inflammatory biomarkers, a significantly greater proportion of patients with schizophrenia exhibited hs-CRP levels ≥ 2 mg/L, a level above which is associated with increased risk of major cardiovascular events.⁶³

Table 1.

Demographic and Clinical Characteristics of Subjects

	Non-Psychiatric Comparison (n = 48)	Schizophrenia (n = 48)	t or χ2	p
Age (years)	54.1 (12.6)	53.2 (10.3)	N/A1	--
Gender [n (% female)]	19 (40%)	19 (40%)	N/A1	--
Race [n (% Caucasian)	37 (77%)	31 (65%)	N/A1	--
BMI	28.5 (5.9)	31.8 (6.7)	N/A1	--
BMI classification2 [n (%)]
Normal	10 (21%)	9 (19%)	N/A1	--
Overweight	24 (50%)	11 (23%)
Obese	14 (29%)	28 (58%)
Antibiotic use [n (% in past year)]	12 (25%)	12 (25%)	N/A1	--
Current smoking status [n (% smoker)]	2 (4%)	16 (33%)	14.6	0.007
Age of illness onset (years)	--	23.5 (9.1)	--	--
Illness duration (years)	--	30.3 (11.0)	--	--
Current antipsychotic use [n (%)]
On	--	40 (83%)	--	--
Off		8 (17%)
Antipsychotic type [n (%)]
Atypical	--	23 (57%)	--	--
Typical		6 (15%)
Both		11 (28%)
Antipsychotic daily dosage3	--	2.06 (2.20)	--	--
SAPS Total Score	0.50 (0.8)	4.58 (3.5)	−3.72	0.003
SANS Total Score	0.58 (0.9)	4.81 (4.3)	−3.84	0.003
PHQ-9 Depression Score	3.24 (3.9)	7.49 (5.5)	−3.87	<0.001
SF-36 Mental Component	52.2 (8.3)	43.6 (12.8)	2.97	0.005
SF-36 Physical Component	52.3 (8.4)	41.9 (10.8)	4.81	<0.001
CIRS Total Score	3.27 (3.6)	6.74 (3.1)	−2.03	0.07
CIRS Severity Score	0.95 (0.8)	1.63 (0.4)	−2.87	0.02
Framingham 10-year CHD % Risk	0.083 (0.06)	0.010 (0.07)	−1.521	0.19
Medical diagnoses
Diabetes [n (% with)]	0 (0%)	16 (50%)	5.46	<0.001
Heart Disease [n (% with)]	0 (0%)	8 (17%)	2.31	<0.001
Hypertension [n (% with)]	6 (50%)	32 (67%)	1.15	0.052
Inflammatory biomarkers4
hs-CRP (mg/L) [n (%)]	1.38 (2.02)	3.18 (3.80)	−1.63	0.178
hs-CRP clinical high risk5	2 (4%)	16 (33%)	13.4	<0.001
TNFα (pg/mL)6	2.79 (1.17)	2.88 (1.66)	0.09	0.93
IL-6 (pg/mL)6	0.59 (0.12)	1.10 (0.73)	−1.47	0.16
IL-10 (pg/mL)6	0.32 (0.04)	0.42 (0.23)	−0.42	0.68

¹Comparison not applicable as groups were matched on these variables.

²World Health Organization (WHO) classification

³WHO defined daily dose

⁴Values for statistical tests based on log-transformed values

⁵hs-CRP ≥2 mg/L; higher risk of major adverse cardiovascular event⁶³

⁶Independent t-tests were performed, as not all subjects in pairs had inflammatory biomarker data BMI = body mass index; CHD = coronary heart disease; hs-CRP = high sensitivity C-reactive protein; IL = interleukin; PHQ-9 = Patient Health Questionnaire; SANS = Scale for the Assessment of Negative Symptoms; SAPS = Scale for the Assessment of Positive Symptoms; TNFα = tumor necrosis factor alpha; WHO = World Health Organization

3.1. Microbiota analysis

3.1.1. Diversity patterns.

Schizophrenia and NC groups did not differ on any assessed measure of alpha-diversity (Table S1). There were significant beta-diversity differences between schizophrenia and NC groups using unweighted UniFrac and Bray-Curtis (Table 2). PCoA of unweighted UniFrac and Bray-Curtis distances showed that schizophrenia and NC groups formed distinct clusters (Figure 1). Additionally, schizophrenia subjects demonstrated greater variability across the PCoA space, compared to NCs who were more tightly clustered (within-group distance comparison; unweighted UniFrac p < 0.001; Bray-Curtis p < 0.001). Across diagnostic groups, significant beta-diversity differences were also observed for sex, BMI category, and smoking status (Table 2). Within the schizophrenia group, there were no beta-diversity differences between subjects on versus off antipsychotic medication.

Table 2.

Differences in Microbial Beta-Diversity Indicators by Diagnostic Group and Other Demographic and Clinical Variables

	Unweighted UniFrac			Bray-Curtis Dissimilarity
Variable	pseudo-F	p	q†	pseudo-F	p	q
Diagnosis	2.41	0.001	0.010*	2.74	0.002	0.010*
Sex	1.65	0.013	0.022*	1.88	0.007	0.018*
BMI category	1.39	0.011	0.022*	1.40	0.034	0.043*
Smoking status	1.79	0.005	0.167	1.54	0.034	0.043*
Current antipsychotic medication use	0.92	0.60	0.67	0.77	0.826	0.826

^†Adjusted p-value controlling for false discovery rate using Benjamini-Hochberg procedure

^*q < 0.05

BMI = body mass index

Figure 1. Beta-diversity of microbial taxa was significantly different between schizophrenia and NC groups.

PCoA plots are a multivariate reduction method to depict beta-diversity distance matrices of microbial taxa between samples. Each point represents an individual subject. (A) Unweighted UniFrac measures the presence or absence of unique branch lengths in a phylogenetic tree, while (B) Bray-Curtis is a non-phylogenetic method that considers relative abundance. Both unweighted UniFrac (A) and Bray-Curtis (B) distance metrics show a high degree of separation between participants with schizophrenia (red) compared to NC participants (blue) along Axis 3.

3.1.2. Differential abundance.

Differential abundance testing using ANCOM identified two taxa from family Lachnospiraceae to be significantly different between schizophrenia and NC groups (W = 647, 688). The W value is used to evaluate the null hypothesis that the two groups have no significant difference; the higher the W value, the more significantly the taxon differs across groups. Songbird was then employed to identify taxa that were most positively and negatively associated with diagnosis, and Qurro was used to display the sorted differential rankings associated with schizophrenia vs. NC groups, as determined by Songbird (Figure 2A). The differential microbial ranks are listed in Table S2. A log-ratio was computed with the sequences associated with the family Lachnospiraceae (as identified by ANCOM) in the numerator, compared to the top-20 ranked taxa (i.e., taxa most associated with NCs) in the denominator as a reference frame. The difference in this log-ratio was compared between groups, and across this balance, Lachnospiraceae was significantly higher in the schizophrenia group compared to NC group (p = 0.002) (Figure 2B).

Figure 2. Differential rankings of taxa associated with diagnosis.

(A) Qurro visualization displaying the sorted differential rankings of taxa associated with schizophrenia vs. NC groups, as determined by Songbird. Left on the x-axis indicates relative over-expression in NCs, right indicates relative over-expression in schizophrenia. Sequences associated with family Lachnospiraceae are highlighted in blue; the top-20 ranked taxa are highlighted in yellow. (B) The log-ratio of Lachnospiraceae to the top-20 ranked taxa is significantly increased in schizophrenia, compared to NCs (Z = −3.072, p = 0.002).

3.2. Functional analysis

To identify potential genetic functional differences between microbial communities present, PICRUSt2 was performed to predict microbial metagenomes, which can reveal the specific metabolic and/or biological functions of the gut microbiome.

3.2.1. Diversity patterns.

There were no group differences in the alpha-diversity of KEGG orthologs, EC numbers, or MetaCyc pathways (ps > 0.08). Beta-diversity was significantly different between schizophrenia and NC groups for KEGG orthologs (Jaccard pseudo-F = 2.41, p = 0.008, q = 0.012) and EC numbers (Jaccard pseudo-F = 2.49, p = 0.007, q = 0.012), but not MetaCyc pathways (Jaccard pseudo-F = 1.28, p = 0.252, q = 0.252).

3.2.2. Differential abundance.

ANCOM revealed three predicted functional pathways to be differentially abundant between groups: trimethylamine-N-oxide (TMAO) reductase (KO W = 3725; EC W = 1243), glycerol degradation to 1,3-propanediol (W = 109), and Kdo₂-lipid A biosynthesis (W = 76). The differential ranking of functional pathways associated with diagnosis were generated using Songbird. The differential ranks for KEGG orthologs, EC numbers, and MetaCyc pathways are listed in Table S3. The log ratios of the above-identified pathways to either the top-20 or bottom-20 ranked pathways were computed and compared between groups. Log-ratios with TMAO reductase as the numerator, relative to bottom-20 pathways, were significantly lower in the schizophrenia group compared to the NC group (q < 0.001) (Figure 3A). The log-ratio of glycerol degradation to 1,3-propanediol pathway, relative to the top-20 pathways, was significantly increased in the schizophrenia group, whereas the log-ratio of Kdo₂-lipid A biosynthesis pathway, relative to the bottom-20, was significantly decreased in the schizophrenia group, compared to NCs (q < 0.001) (Figure 3B).

Figure 3. Three predicted functional pathways were differentially abundant between schizophrenia and NC groups: trimethylamine-N-oxide (TMAO) reductase, glycerol degradation to 1,3-propanediol, and Kdo₂-lipid A biosynthesis.

(A) The log-ratios of TMAO reductase (TMAOr), relative to the bottom-20 pathways (i.e., pathways most associated with schizophrenia), were significantly lower in schizophrenia, compared to NCs. K07821: Z = −3.80, p < 0.001, q < 0.001; EC 1.7.2.3: Z = −4.19, p < 0.001, q < 0.001. (B) The log-ratio of glycerol degradation to 1,3-propanediol pathway (GlycProp), relative to the top-20 pathways (i.e., most associated with NCs), was significantly increased in schizophrenia, whereas the log-ratio of Kdo₂-lipid A biosynthesis pathway, relative to the bottom-20 pathways, was significantly decreased in schizophrenia group, compared to NCs. GOLPDLCAT-PWY: Z = −4.36, p < 0.001, q < 0.001; KDO-NAGLIPASYN-PWY: Z = −3.33, p = 0.001, q < 0.001 (C) Greater log-ratios of TMAOr/bottom-20 were associated with greater Framingham Risk Scores in schizophrenia, and greater log-ratios of Kdo₂-lipid A/bottom-20 was correlated with decreased levels of IL-10.

The extant literature suggest links between TMAO and Kdo₂-lipid A with CVD outcomes and immune responses, respectively,^64,65 which are both related to the hypothesis of accelerated aging in schizophrenia. Spearman’s correlations were performed to examine the association of these functional pathways with age, Framingham Risk Score, and inflammatory biomarkers (Figure 3C). Log-ratios of TMAO reductase/bottom-20 were positively correlated with Framingham Risk Score. The log-ratio of Kdo₂-lipid A/bottom-20 was negatively correlated with IL-10. No significant correlations were observed with age, hs-CRP, IL-6, or TNFα in schizophrenia. No relationships between pathways and clinical variables were observed in the NC group (ps > 0.13).

3.3. Random Forests Classifier

To evaluate the ability of the microbiome to classify diagnostic groups, we used Random Forests models to classify groups using microbial composition (ASVs) and functional profiles (KEGG orthologs and MetaCyc pathways) as key input features (Figure 4). The model built using microbial taxonomy differentiated participants with schizophrenia from NCs with 75% accuracy. Feature importance scores were all <2%, indicating that no single (or group of) microbial feature(s) contributed to model performance in a significant way. This suggests that many different features and their unique combination are jointly responsible for the overall model performance. Functional classification models were built using PICRUSt2 data. The accuracy of supervised classification was 67% for KEGG orthologs (0.65 AUC) and 70% for MetaCyc pathways (0.70 AUC). Similar to taxonomic classification models, feature importance scores from functional models were <5%.

Figure 4. Accuracy of supervised Random Forests classification models.

(A) Microbial classification predicted schizophrenia vs. NCs with 75% accuracy. Area under the ROC curve was 0.82, indicating good classification accuracy. Functional classification predicted diagnosis with (B) 67% accuracy for KEGG orthologs and (C) 70% for MetaCyc pathways. All ROC curves were produced from stratified 5-fold cross-validation results, while the confusion matrices were generated from the best model obtained in cross-validation.

4. DISCUSSION

Consistent with our hypotheses, we found significant beta-diversity differences in both microbial composition and functional potential between individuals with schizophrenia and NCs. On the other hand, alpha-diversity of microbial taxa and functional pathways were not different between groups. Using a novel compositionally-sound method to conceptualize and interpret differential abundance, we found that the family Lachnospiraceae was associated with schizophrenia. One of the most notable findings of this study is that pathways related to TMAO reductase and Kdo₂-lipid A biosynthesis were altered in schizophrenia, indicating that these enzymes and pathways may have clinical significance and, if verified in independent studies, may prove to be useful markers of disease. Also as hypothesized, we observed that these functional profiles were associated with decreased anti-inflammatory cytokine IL-10 and increased risk for CVD. This study extends our previous report¹⁴ and presents new findings on microbial functional potential and pathways in this cohort of patients with chronic schizophrenia.

Our findings regarding community-level differences in microbiota composition is consistent with previous studies in schizophrenia.^14,18,19,66 Beta-diversity findings are further reinforced by random forests analysis, which revealed that microbial taxa can predict patients with schizophrenia from NCs with 0.82 AUC. Feature importance scores from our model did not reveal any microbial features that strongly differentiated groups, despite the strong overall accuracy of the model. This suggests that no single microbial feature was responsible for model performance, but rather the unique distribution of all taxa in relation to one another are most associated with schizophrenia. Despite differences in beta-diversity of microbes, individuals with schizophrenia and NCs did not differ in terms of within-sample alpha-diversity. It typically observed that low alpha-diversity is a hallmark of dysbiosis and represents “worse” health.⁶⁷ However, the literature on alpha-diversity in psychiatric populations has been mixed (e.g., no differences,^{14,18,68–72} decreased,^19,73,74 and increased^66,75). A recent meta-analysis of the gut microbiota in major depressive disorder revealed no difference in alpha-diversity between patients and NCs.⁷⁶

The transdisciplinary nature of microbiome research brings challenges to reproducibility and replicability,⁷⁷ particularly in the field of clinical psychiatric research. The microbiome revolution has opened new frontiers for examining host-microbe associations in the context of understanding psychiatric disorders, leading to an increase of studies investigating microbial abnormalities in schizophrenia and related psychotic disorders. However, such rapid expansion has pitfalls, as various research groups seek to incorporate microbiome collection and analysis into ongoing clinical studies, employing varied experimental techniques at hand. Technical differences (e.g., DNA extraction methods, PCR primers, sequencing platforms, bioinformatic pipelines, taxonomy databases) can produce systematic biases that can obscure biologically meaningful results.^22,23,78 Systematic variation is especially problematic when the effects of a biological parameter are expected to be subtle and makes replication and integration of findings difficult, if not impossible.⁷⁹ Small sample sizes and lack of determination of within-individual sample variation are among other pitfalls of the current literature. Our systematic review in this area¹³ found that, among a handful of studies examining microbiota differences between patients with first-episode psychosis and schizophrenia and healthy controls, there was minimal consensus with regards to alpha-diversity patterns, relative abundance, or directionality of differences in taxa. Nowhere is this problem more apparent than in differential abundance testing. Across these studies,^14,18,19 over 130 taxa were observed to be different with little consistency across investigations.

Differential abundance analysis is controversial throughout microbiome research, and there is widespread misconception about how to interpret microbial abundance.¹⁵ Data from next-generation high-throughput sequencing methods are compositional, meaning that abundance information is relative to the biological specimen itself, and contains information about the relationship between parts (e.g., proportions or probabilities).^80,81 Thus, the reporting and interpretation of relative abundance data as absolute differences or changes can lead to misinterpretations of microbial community structures, as the increase of one taxon inevitably leads to the concurrent decrease of others. Since the changes of components are mutually dependent, high false discovery rates occur when compositional data are analyzed using traditional statistical methods − a fact that is frequently swept under the rug in microbiome studies.⁸² Instead, log ratios are a preferable way to examine differences within compositional datasets. Ratio transformations of data capture relationships between the features, and, importantly, these ratios are the same whether the data are represented as counts or proportions. This method allows for a more robust biological understanding of microbial contributions to schizophrenia. We performed ANCOM, a compositional approach that performs statistical test based transformed by an additive log ratio,⁵⁶ followed by Songbird, which uses reference frames to establish microbial composition measurement standards.¹⁵ Using this approach, we found that the log-ratio with Lachnospiraceae in the numerator and the top-20 ranked taxa associated with NCs in the denominator was higher in individuals with schizophrenia compared to NCs, indicating that the ratio of this taxa may have clinical significance and, if verified in independent studies, may prove to be a useful marker of the disorder. Lachnospiraceae is one of the most abundant families from the phylum Firmicutes found in the gut and has been associated with the production of butyrate,⁸³ a short chain fatty acid associated with anti-inflammatory effects.⁸⁴ Results relying on relative abundance information for this family have been mixed. Prior studies of affective disorders have reported increased levels of Lachnospiraceae in patients with MDD, compared to healthy controls,⁷⁰ and in patients with bipolar disorder on atypical antipsychotic medications, compared to patients not on atypical antipsychotics.⁸⁵ A majority of participants on antipsychotic medication in this sample were taking atypical antipsychotics (either alone or with typical antipsychotics), which may have contributed to the finding of increased Lachnospiraceae. On the other hand, Lachnospiraceae has also been found to be depleted in schizophrenia and positively correlated with psychosis symptom severity.¹⁹ A recent systematic review⁸⁶ noted that the most consistent finding when pooling results from affective and psychotic disorders is lower levels of Lachnospiraceae. It is possible that this taxon may be a transdiagnostic marker of psychiatric disease, perhaps related to common mood symptoms or medications across disorders. Discrepant results, with regards to directionality, may be due to differences in differential abundance methods. Thus, the method currently presented holds much promise for resolving outstanding inconsistencies across previous reports by re-analyzing those datasets using reference from to make more stable inferences of compositional differences. It is critical to use appropriate statistical tools for data analyses and to continuously benchmark analytic methods in order to better understand their strengths and limitations. It is beyond the scope of the current paper to benchmark our methods against others. Instead, we refer the reader to other published studies, including benchmarking, that highlight that compositionally-aware methods are more accurate and robust with regards to microbiome data.^15,87

Differences between differential abundance and random forests findings, which did not identify any microbial features that strongly differentiated groups, may be explained by several reasons. Generally speaking, important features in a random forests model (i.e., taxa highly contributing to a model’s performance) should be differentially abundant, but differentially abundant features may not necessarily be important in the random forests model. First, differential abundance testing examines ratios of microbes, whereas the random forests model considers normalized counts. Second, rare features may be differentially abundant, but ubiquitous features are important to the model. Third, the overall microbiome composition drives diagnotic prediction in random forests models, not any single of several taxonomic features; on the other hand, in differential abundance testing, we are looking for ratios of taxa that are different between the groups.

PICRUSt2 analysis revealed broad community functional differences between individuals with schizophrenia and NCs, and functional classification using random forests predicted diagnosis with good accuracy. Similar to taxonomic data, no differences in alpha-diversity of KEGG orthologs, EC numbers, or MetaCyc pathways were observed between groups. However, functional pathways related to TMAO reductase and Kdo₂-lipid A biosynthesis were associated with NCs. In humans, TMAO is formed from trimethylamine, which is a byproduct of the bacterial metabolism of dietary choline in the intestine.⁶⁴ TMAO modulates lipid and glucose homeostasis⁸⁸ and has been associated with a number of chronic diseases, most notably CVD and diabetes.⁶⁴ In mouse models, TMAO represents the end of a complex dietary phosphatidylcholine–choline metabolic pathway involving the gut microbiome that contributes to the pathogenesis of atherosclerotic coronary artery disease.⁸⁹ Similarly, TMAO induces glucose intolerance and insulin resistance in mice fed a high fat diet, and effects were concurrent with increased mRNA levels of pro-inflammatory cytokines.⁹⁰ Among individuals at risk for CVD, high plasma levels of TMAO is correlated to increased atherosclerotic plaque burden⁸⁹ and risk of major adverse cardiovascular events,⁹¹ and people at risk for diabetes have higher plasma TMAO concentrations.^91–93 TMAO reductase is an enzyme that catalyzes the reduction of TMAO to trimethylamine. Our findings suggest the hypothesis that individuals with schizophrenia may have less ability to clear TMAO due to lower levels of TMAO reductase, leading to increased levels of the metabolite. In patients with schizophrenia, higher ratios of TMAO reductase were correlated with higher Framingham Risk Scores. Further blood-based metabolomics are needed to examine this hypothesis and determine levels of compounds specific to TMAO pathways present in circulation, which might provide greater insights into the role of this potential mechanism in cardiovascular disease in schizophrenia.

Kdo₂-lipid A is an essential structural component for the survival of most Gram-negative bacteria and is the active component of lipopolysaccharide (LPS), which stimulates host immune responses through a protein complex of Toll-like-receptor 4 (TLR4) and myeloid differentiation protein 2 (MD-2).^65,94 The TLR4/MD-2 complex activates a cascade of signal transductions that orchestrate an inflammatory response. Many bacteria can modify the structure of their Kdo₂-lipid A, as a way to adapt to different environments and modulate their virulence or infectivity. Our results showing alterations in the Kdo₂-lipid A biosynthesis pathway in schizophrenia and its negative relationship with anti-inflammatory cytokine IL-10 is consistent with evidence that schizophrenia is associated with a lasting pro-inflammatory state.⁴ To our knowledge, the clinical impact of Kdo₂-lipid A has not been established in humans; as such, more research is needed to determine how the up- or down-regulation of this pathway may impact downstream systemic inflammation. Kdo₂-lipid A has potential clinical relevance, particularly its capabilities to elicit host innate immune responses. The enzymes and receptors involved in Kdo₂-lipid A biosynthesis is an emerging target for immunopharmocological exploitation.⁶⁵

4.1. Limitations

The most important limitation of this study is its relatively small sample size, given the complex nature of schizophrenia and heterogeneity of the disorder. Additionally, only one microbiome sample was evaluated per individual. The composition of the gut microbiome is known to fluctuate across time and as a result of many factors,^24,67 and intra-individual variability may be greater than between-group variability.⁹⁵ The cross-sectional design limits our ability to make causal inferences. Future prospective longitudinal studies are needed to characterize and control for temporal variations within and between groups in order to understand causal relationships between the microbiome and health in schizophrenia. This study was not designed to assess the impact of antipsychotic medications on the gut microbiome. Our clinical population of interest is individuals with long-time, chronic disease. As such, a majority of our sample were taking antipsychotic medications. Relatedly, we did not exclude individuals with schizophrenia with cardiometabolic disorders (e.g., diabetes, hypertension, heart disease). Although these conditions may influence the stability of the gut microbiota, patients with schizophrenia tend to have higher rates of these conditions;² as such, these are important to consider in understanding the phenotype of accelerated aging in schizophrenia, and excluding them would create a sample that is unrepresentative of the general schizophrenia population. It is possible that differences in the gut microbiome between groups may have been driven by these comorbid medical illnesses; however, we did not find any differences in the beta-diversity of microbial taxa (ps > 0.38) or functional potential (ps > 0.20) between persons with schizophrenia with and without these conditions, suggesting that these medical conditions were unlikely to be driving the reported findings. Likewise, we did not match samples on or control for smoking prevalence in analyses. Patients with chronic schizophrenia are markedly prone to smoke tobacco and it has been suggested that biological factors may underlie the association between this disorder and tobacco use. To disentangle the effects of smoking and diagnostic group, we compared smokers to non-smokers within the schizophrenia group did not find differences in beta-diversity of taxa (ps > 0.068) or functional pathways (ps > 0.105), suggesting that differences in microbial community composition between study groups were driven by disease rather than cigarette smoking. Functional data from PICRUSt was predicted based on 16S rRNA marker gene sequences, which does not provide direct information about the functional composition of sampled communities. Nevertheless, PICRUSt predictions have overall high accuracy¹⁷ and can yield new biological insights, particularly in a novel field of research. Future studies should incorporate shotgun metagenomic sequencing, which directly measures genetic functional potential, to validate predictions of functional potential. Finally, as this was an exploratory study, we did not have circulating levels of metabolites to further validate TMAO reductase findings. Performing blood-based metabolomics will be an important next step to discover whether compounds specific to TMAO pathways are observed in circulation, which could impact host metabolism and health.

4.2. Conclusions and next steps

Strengths of the present study include a matching of NC participants to control for other non-disease related factors known to have major influences on the gut microbiome. We build upon our previous article by utilizing a new compositionally-aware method for differential abundance testing that can be used in future studies with larger sample sizes and/or meta-analysis of existing datasets to help resolve inconsistencies in the extant literature. Our study is one of the few to have examined the functional potential of the gut microbiome in patients with chronic schizophrenia. Previous studies have found altered functional potentials related to metabolism of amino acids, lipids, and carbohydrates and degradation of xenobiotics.¹⁸ To our knowledge, this study is the first to show that individuals with schizophrenia have altered metabolic pathways related to TMAO reductase and Kdo₂-lipid A biosynthesis. The microbiota may impact that pathophysiology of the disease through modulation of functional pathways related to immune signaling/response and lipid and glucose regulation, which might have implications for accelerated biological aging in schizophrenia. For example, therapies that target TMAO pathways are being explored,⁶⁴ including dietary interventions targeted towards reducing TMAO levels.^96,97

Research is increasingly moving toward characterizing the functional capacity of the community by quantifying the abundances of genes or pathways, which can better elucidate potential downstream effects of taxonomic shifts (i.e., significant differences in abundance observed between case and control samples). Greater effort should be made to link these two facets of the microbiome,⁹⁸ and future studies should focus on identifying taxonomic drivers of disease-associated functional imbalances. Deep functional profiling of the microbiome, along with integration of other omics data such as metabolomics and proteomics, offers greater promise for therapeutic discovery and finding microbiome-related interventions.⁹⁹

HIGHLIGHTS.

• SZ is associated with altered gut microbial composition and functional potential.

• Models based on the gut microbiome differentiated SZ from NCs with high accuracy.

• Lachnospiraceae was associated with SZ.

• TMAO reductase and Kdo₂-lipid A biosynthesis functional pathways were altered in SZ.

• Pathways were associated with inflammatory cytokines and risk for coronary heart disease in SZ.