Stress-Resilience Impacts Psychological Wellbeing: Evidence from Brain-Gut Microbiome Interactions

Abstract

The brain-gut-microbiome (BGM) system plays an influential role on mental health. We characterized BGM patterns related to resilience using fecal samples and multimodal MRI. Data integration analysis using latent components showed the high resilience phenotype was associated with lower depression and anxiety symptoms, higher frequency of bacterial transcriptomes (related to environmental adaptation, genetic propagation, energy metabolism, anti-inflammation), increased metabolites (N-acetylglutamate; dimethylglycine), and cortical signatures (increased resting state functional connectivity between reward circuits and sensorimotor networks; decreased grey matter volume and white matter tracts within the emotion regulation network). Our findings support a multi-omic signature involving the BGM system suggesting that resilience impacts psychological symptoms, emotion regulation and cognitive function as reflected by unique neural correlates and microbiome function supporting eubiosis and gut barrier integrity. Bacterial transcriptomes provided the highest classification accuracy suggesting that the microbiome is critical in shaping resilience and highlights that microbiome modifications can optimize mental health.

Introduction

Upwards of $300 billion dollars is lost annually due to stress-related health care costs and missed work in the U.S.¹ highlighting the need for greater resilience to stress. Definitions of resilience generally refer to beneficial outcomes in response to threat or stressful events^2,3. Resilience involves positive acceptance of change, tolerance of negative affect, tenacity, and the ability to recover after stressful events⁴. Resilience can predict stress-related depression and anxiety^5,6, traumatic stress⁷, and maladaptive coping mechanisms such as alcohol misuse⁸. Most research investigating resilience has focused its correlations with personality traits⁹, emotional¹⁰ and behavioral¹¹ regulation strategies, and social factors including supportive interpersonal relationships¹². Resilience is also related to autonomic flexibility¹³ and adaptive neurological correlates^11,14.

The composition and function of the human microbiome, an ecosystem consisting of trillions of microorganisms residing in and on the human body¹⁵, have been linked to stress-related disorders and other mental health diagnoses^16–21. The gut microbiome modulates psychological functioning by influencing the brain-gut microbiome (BGM) system²², a bidirectional signaling mechanism between the central nervous system (CNS) and gastrointestinal tract²³. Interestingly, the microbiome has been implicated in conferring stress-resilience^24,25. Evidence reveals that alterations in BGM signatures can influence prosocial behaviors²⁶ related to resilience and stress-related psychopathologies^16–20. Numerous animal studies also reveal the role of the BGM in resilience versus susceptibility traits after stress exposure. For example, gut microbial composition differed in mice susceptible compared to mice resilient to chronic social defeat stress^27,28, as well as to learned helplessness after exposure to inescapable stress^29,30. This opens the intriguing possibility that the endogenous gut microbiome may house stress mitigating therapeutic metabolites supporting neurologically adaptive processes. For instance, bacterial transcriptomes are related to a number of microbiome functions³¹ and serve to maintain a balanced and diverse population of gut microbiota (eubiosis) and gut barrier integrity. Thus, a resilient phenotype involves expression of transcriptomes that ensure appropriate modulation of the BGM system, and therefore, an adaptive CNS²⁴.

Collectively, the gut microbiota produces a number of metabolites, including hormones, neurotransmitters (such as Gamma-Aminobutyric Acid; GABA³², glutamate³³, and serotonin³⁴), and other signaling molecules within the BGM³⁵ implicated in stress-related psychopathology^36,37. Via serotonin, gut microbiota through neural activation of vagal afferents in the gut³⁸, can contribute to a regulated autonomic nervous system and adaptive stress response. The main metabolites produced by gut microbiota are short-chain fatty acids (SCFAs)³⁹ which are known to influence cognitive and emotional processing through effects on the brain via anti-inflammatory properties⁴⁰. In animal studies, low SCFA has been associated with anxiety and depression-related behaviors⁴¹ and also predicted resilience versus susceptibility to traumatic stress⁴².

The gut microbiota can also shape brain structure and function^43–45, as evidenced by probiotic studies. Probiotics have resulted in decreased activity in somatosensory and viscerosensory cortices in response to emotional attentional tasks and reduced grey matter volume and increased resting state functional connectivity (rsFC) within the default mode network (DMN)⁴⁶. A recent neuroscience review of resilience highlighted the mesolimbic reward system, the DMN, and regions involved in fear and stress, namely, the amygdala⁴⁷. The DMN, activated during passive rest, daydreaming, and thinking about the past or others⁴⁸ has been linked to major depressive disorder (i.e., excessive rumination⁴⁹ but also dispositional mindfulness⁵⁰). Resilience-related grey matter volume changes have been observed in cognitive and affective regions (amygdala, subgenual and rostral anterior cingulate cortex)^51,52. Lower activity within the salience network⁵³ but increased connectivity within the sensorimotor network^54,55 and greater responses in reward circuits (basal ganglia)⁵⁶ have been linked to resilience. There are also associations between resilience and white matter tract integrity within structures implicated in social cognition⁵⁷.

Since no study to date has investigated an integrative biological profile of resilience, we aimed to determine how resilience is related to clinical phenomes, microbiome function, and neural characteristics (Figure 1). Considering the microbiome’s role in influencing psychological resilience, we hypothesized that high resilience would be associated with: (1) lower scores on clinical measures of psychological symptoms and higher levels of adaptive coping; (2) microbiome function supporting gut health as evidenced by characteristic bacterial transcriptomes (pathways supporting gut microbial growth and diversity) and metabolome (metabolites supporting anti-inflammation and gut barrier integrity); (3) brain morphometry and connectivity signatures, reflecting increased efficiency in regions important for emotion regulation and cognitive functioning.

Figure 1: Graphical Summary.

A Demographics shows the CD-RISC (Connor-Davidson Resilience Scale) threshold for group label, and the sample size per group. B Clinical and psychological variables were collected from self-report questionnaires. Metabolites and transcriptomics data were collected from stool samples. Multimodal brain images were collected via magnetic resonance imaging. C Each of the six data blocks were cleaned and prepared as individual datasets. D Data integration for biomarker discovery using latent components as used to create signatures for high and low resilience. E Summary of results.

Abbreviations: BMI, body mass index; MRI, magnetic resonance imaging; RNA, Ribonucleic acid; KEGG, Kyoto Encyclopedia of Genes and Genomes; NAG, N-acetylglutamate; DMG, dimethylglycine; Vol, volume; SA, surface area; ERN, emotion regulation network; SMN, sensorimotor network; DMN, default mode network; CAN, central autonomic network; CEN, central executive network; K07335, basic membrane protein A and related proteins; K03315, Na/H antiporter; K01081, 5’-nucleotidase; K07080, uncharacterized protein; K14645, serine protease; K02343, DNA polymerase III subunit gamma/tau; K01596, phosphoenolpyruvate carboxykinase; K01361, lactocepin; K01844, beta-lysine 5,6-aminomutase alpha subunit; K03388, heterodisulfide reductase subunit A2.

Results

Demographics and clinical results

Comparisons of clinical variables between the high resilience group (HR) and low resilience group (LR) are depicted in Table 1. The distribution of BMI that justifies the threshold for resilience group labels is illustrated in Supplementary Figure S1.

Table 1. Demographic, Clinical, Metabolome, Transcriptome, and Multimodal Brain Differences Between the High and Low Resilience Groups.

Demographic, clinical, metabolic, transcriptomic, and neurological characteristics of high and low resilience groups.

Data Block	Characteristic	N	High Resilience N = 441	Low Resilience N = 721	p-value2
Demographic/ Clinical	Sex	116			>0.9
	Female		27 / 44 (61%)	44 / 72 (61%)
	Male		17 / 44 (39%)	28 / 72 (39%)
	Age	116	34 (12)	31 (11)	0.14
	Ethnicity	116			0.89
	White		14 / 44 (32%)	18 / 72 (25%)
	Hispanic		4 / 44 (9%)	10 / 72 (14%)
	Asian		9 / 44 (21%)	15 / 72 (21%)
	Black		5 / 44 (11%)	7 / 72 (10%)
	Multiracial		12 / 44 (27%)	22 / 72 (30%)
	BMI	116	28.1 (4.0)	27.6 (4.2)
	Standard American Diet	116
	0		33 / 44 (75%)	49 / 72 (68%)
	1		11 / 44 (25%)	23 / 72 (32%)
	Brief Resilience Score	116	25.6 (3.6)	21.4 (4.3)	<0.001
	CD-RISC Score	116	91 (5)	71 (9)	<0.001
	CD-RISC - Adaptability	116	19.00 (1.43)	15.78 (2.85)	<0.001
	CD-RISC - Control	116	11.05 (1.12)	8.63 (1.63)	<0.001
	CD-RISC - Meaning	114	6.77 (1.27)	4.62 (2.31)	<0.001
	CD-RISC - Persistence	116	29.8 (1.9)	23.8 (3.9)	<0.001
	CD-RISC - Emotional Cognitive Control	116	24.4 (2.5)	18.2 (3.6)	<0.001
	FFM - Describe	91	32 (6)	28 (6)	0.004
	FFM - Nonjudge	89	32 (7)	28 (7)	0.001
	FFM - Total Score	87	149 (20)	132 (15)	<0.001
	HADS - Anxiety	116	3 (3)	5 (3)	0.013
	HADS - Depression	116	1 (2)	2 (2)	0.003
	IPIP - Extraversion	113	38 (6)	33 (7)	<0.001
	IPIP - Neuroticism	113	17 (5)	23 (6)	<0.001
	MASQ - Attention	102	14 (4)	17 (4)	<0.001
	MASQ - Language	102	14 (4)	15 (4)	0.081
	MASQ - Verbal Memory	101	15.4 (3.9)	18.5 (4.6)	<0.001
	MASQ - Visual Perception	102	9.5 (3.2)	11.9 (3.9)	0.001
	PSS Score	116	10 (5)	13 (7)	0.002
	STAI - Anxiety	116	43 (7)	50 (10)	<0.001
	N-acetylglutamate	104	1.37 (0.72)	1.07 (0.57)	0.033
Metabolome	dimethylglycine	98	1.28 (0.67)	1.08 (0.61)	0.13
	creatine	104	5 (16)	9 (31)	0.3
	K07335	116	1.95 (0.57)	1.68 (0.47)	0.012
Transcriptome	K03315	116	0.64 (0.79)	0.17 (0.68)	0.002
	K01081	116	0.51 (0.53)	0.23 (0.50)	0.007
	K07080	116	0.12 (0.94)	−0.33 (1.07)	0.018
	K14645	116	−1.21 (0.95)	−1.82 (1.43)	0.007
	K02343	116	0.42 (0.42)	0.27 (0.44)	0.064
	K01596	116	−0.21 (1.03)	−0.53 (1.16)	0.12
	K01361	116	−0.86 (1.92)	−1.50 (1.76)	0.080
	K01844	116	−0.47 (0.91)	−0.93 (1.20)	0.022
	K03388	116	−0.79 (0.78)	−1.35 (1.09)	0.002
	K03466	116	0.17 (0.30)	0.01 (0.36)	0.013
	K02965	116	1.91 (0.43)	2.18 (0.41)	0.002
	K01119	116	0.07 (0.56)	−0.25 (0.58)	0.004
	K01834	116	1.68 (0.24)	1.89 (0.39)	<0.001
	K05516	116	0.41 (1.50)	0.93 (1.20)	0.053
	K07667	116	−0.39 (0.69)	−0.67 (0.76)	0.045
	R SbCaG Volume	116	−0.32 (0.80)	0.06 (1.08)	0.030
Structural MRI	R SbCaG Surface Area	116	−0.16 (1.01)	0.03 (1.01)	0.3
	L AngG Volume	116	−0.46 (0.85)	0.23 (0.93)	<0.001
	L AngG Surface Area	116	−0.46 (1.01)	0.20 (0.94)	<0.001
	R InfPrCS Surface Area	116	−0.02 (0.93)	0.11 (0.98)	0.5
	R InfPrCS Volume	116	0.00 (0.92)	0.09 (1.04)	0.6
	VTA to R Tha	116	0.12 (0.14)	0.06 (0.13)	0.029
rs Functional MRI	L MRF to R SuMarG	116	0.07 (0.11)	−0.01 (0.12)	<0.001
	R InfCirIns to L Pal	116	0.10 (0.15)	0.04 (0.15)	0.034
	R SbCG S to L SupPL	116	−0.05 (0.16)	−0.08 (0.17)	0.2
	R LORs to L Pu	116	0.06 (0.15)	0.00 (0.15)	0.046
	R SbCaG to L SbCaG	116	0.0006 (0.0002)	0.0007 (0.0003)	0.056
Diffusion MRI	R SupOcG to R Hip	116	0.0001 (0.0001)	0.0000 (0.0001)	0.2
	R SupOcG to R Hip	116	0.0001 (0.0001)	0.0000 (0.0001)	0.2

Abbreviations: BMI, body mass index; CD-RISC, Connor-Davidson Resilience Scale; MASQ, Multiple Ability Self-Report Questionnaire; FFM, Five Facet Mindfulness; HADS, Hospital Anxiety and Depression scale; IPIP, International Personality Item Pool; PSS, Perceived Stress Scale; STAI, Stat-Trait Anxiety Inventory; K07335, basic membrane protein A and related proteins; K03315, Na/H antiporter; K01081, 5’-nucleotidase; K07080, uncharacterized protein; K14645, serine protease; K02343, DNA polymerase III subunit gamma/tau; K01596, phosphoenolpyruvate carboxykinase; K01361, lactocepin; K01844, beta-lysine 5,6-aminomutase alpha subunit; K03388, heterodisulfide reductase subunit A2; K03466, DNA segregation ATPase FtsK/SpoIIIE; K02965, small subunit ribosomal protein S19; K01119, 2’,3’-cyclic-nucleotide 2’-phosphodiesterase / 3’-nucleotidase; K01834, 2,3-bisphosphoglycerate-dependent phosphoglycerate mutase; K05516, curved DNA-binding protein; K07667, two-component system, OmpR family, KDP operon response regulator KdpE. R, right; L, left; SA, surface area; Vol, volume; SbCaG, Subcallosal Gyrus; AngG, Angular Gyrus; InfPrCS, Inferior part of the Precentral Sulcus; VTA, Ventral Tegmental Area; Tha, Thalamus proper; MRF, Mesencephalic Reticular Formation; SuMarG, Supramarginal Gyrus; InfCirIns, Inferior segment of the Circular Sulcus of the Insula; Pal, Pallidum; SbCG S, Subcentral Gyrus and Sulci; SupPL, Superior Parietal Lobule; LORs, Lateral Orbital Sulcus; Pu, Putamen; SupOcG, Superior Occipital Gyrus; Hip, Hippocampus.

¹n / N (%); Mean (standard deviation)

²Pearson’s Chi-squared test; Welch Two Sample t-test

Microbiome analysis on group differences

The alpha and beta diversity did not show significant differences between HR and LR (Supplementary Figure S2 and S3). Performance of on the 16S-identified amplicon sequence variants (ASVs) dataset was not found to be a significant contributor in the DIABLO model (AUC=0.524).

DIABLO identifies a multi-omic signature able to classify high and low resilience

DIABLO Model Performance.

DIABLO identified a highly correlated ‘omics signature capable of discriminating between individuals with high versus low psychological resilience. A receiver operating characteristic (ROC) curve showing the performance of the classification model at all classification thresholds had an area under the ROC Curve (AUC; overall model) of 0.77 and the BER was 0.18. The training set AUC was 68% and the achieved overall testing set (N = 81) AUC was 68%. The training set (N=35) BER was 0.18. Area under the curve by dataset type was 77% for clinical, 63% for metabolome, 77% for transcriptome, 65% for structural MRI, 66% for resting-state functional MRI, and 48% for diffusion MRI. A summary of the final model’s performance is in Supplementary Figure S4.

DIABLO-selected variables.

The standardized values of omics variables selected by the DIABLO model are depicted by resilience group in Figure 2. A total of 45 features, 13 clinical features, including survey sub-scores, 3 metabolome, 16 transcriptome, 6 structural MRI, 5 resting-state functional MRI, and 2 diffusion MRI variables were included in the final model and depicted in the loading plots of Figure 3. Variables from each dataset type are listed below in order of importance based on their loading vectors, which represent the magnitude of contribution to the DIABLO model. Higher loading vectors suggest greater importance in predicting resilience.

Figure 2. Boxplots of DIABLO-selected Variables of Importance.

Standardized values equate to variables of importance for each dataset between High (blue) and Low (orange) resilience. The line inside each box represents the median standardized value. Each box itself represents the interquartile range which captures 50% of standardized values. The lines extending from each box represent the ranges (extending up to the minimum and maximum) of standardized values outside of the interquartile range. Individual data points beyond the lines are considered outliers.

Abbreviations: IPIP, International Personality Item Pool; STAI, Stat-Trait Anxiety Inventory; HAD, Hospital Anxiety and Depression scale; FFM, Five Facet Mindfulness; MASQ, Multiple Ability Self-Report Questionnaire; PSS, Perceived Stress Scale; K07335, basic membrane protein A and related proteins; K03315, Na/H antiporter; K01081, 5’-nucleotidase; K07080, uncharacterized protein; K14645, serine protease; K02343, DNA polymerase III subunit gamma/tau; K01596, phosphoenolpyruvate carboxykinase; K01361, lactocepin; K01844, beta-lysine 5,6-aminomutase alpha subunit; K03388, heterodisulfide reductase subunit A2; K03466, DNA segregation ATPase FtsK/SpoIIIE; K02965, small subunit ribosomal protein S19; K01119, 2’,3’-cyclic-nucleotide 2’-phosphodiesterase / 3’-nucleotidase; K01834, 2,3-bisphosphoglycerate-dependent phosphoglycerate mutase; K05516, curved DNA-binding protein; K07667, two-component system, OmpR family, KDP operon response regulator KdpE;

R, right; L, left; SA, surface area; Vol, volume; SbCaG, Subcallosal Gyrus; AngG, Angular Gyrus; InfPrCS, Inferior part of the Precentral Sulcus; VTA, Ventral Tegmental Area; Tha, Thalamus proper; MRF, Mesencephalic Reticular Formation; SuMarG, Supramarginal Gyrus; InfCirIns, Inferior segment of the Circular Sulcus of the Insula; Pal, Pallidum; SbCG S, Subcentral Gyrus and Sulci; SupPL, Superior Parietal Lobule; LORs, Lateral Orbital Sulcus; Pu, Putamen.

Figure 3. Loading Plots from the DIABLO-selected Variables of Importance.

Loading plots of top contributing variables selected by DIABLO separated by dataset type. Blue represents high resilience; orange represents low resilience. Magnitude of loading coefficient represents the level of contribution for prediction. Greater values represent greater contribution.

Abbreviations: MASQ, Multiple Ability Self-Report Questionnaire; FFM, Five Facet Mindfulness; HAD, Hospital Anxiety and Depression scale; IPIP, International Personality Item Pool; PSS, Perceived Stress Scale; STAI, Stat-Trait Anxiety Inventory; K07335, basic membrane protein A and related proteins; K03315, Na/H antiporter; K01081, 5’-nucleotidase; K07080, uncharacterized protein; K14645, serine protease; K02343, DNA polymerase III subunit gamma/tau; K01596, phosphoenolpyruvate carboxykinase; K01361, lactocepin; K01844, beta-lysine 5,6-aminomutase alpha subunit; K03388, heterodisulfide reductase subunit A2; K03466, DNA segregation ATPase FtsK/SpoIIIE; K02965, small subunit ribosomal protein S19; K01119, 2’,3’-cyclic-nucleotide 2’-phosphodiesterase / 3’-nucleotidase; K01834, 2,3-bisphosphoglycerate-dependent phosphoglycerate mutase; K05516, curved DNA-binding protein; K07667, two-component system, OmpR family, KDP operon response regulator KdpE; MRI, magnetic resonance imaging; L, left; R, right; SbCaG, subcallosal gyrus; AngG, angular gyrus; InfPrCS, inferior part of the precentral sulcus; VTA, ventral tegmental area; Tha, thalamus; MRF, mesencephalic reticular formation; SuMarG, supramarginal gyrus; InfCirIns, inferior segment of the circular sulcus of the insula; Pal, pallidum; SupPL, superior parietal lobule; LORs, lateral orbital sulcus; Pu, putamen; SupOcG, superior occipital gyrus; Hip, hippocampus.

Clinical variables in order of importance included IPIP Neuroticism, STAI Anxiety, HAD Anxiety, FFM Total Score and the Describing FFM subscale, MASQ Verbal Memory, PSS Score, IPIP Extraversion, MASQ language, HAD Depression, FFM Non-judgement subscale, MASQ Visual Perception and Attention subscores. HR had higher mean levels of extraversion⁵⁸ and mindfulness⁵⁹ than their LR counterparts. Further, HR had lower mean levels of anxiety (HADS; STAI)^60,61, neuroticism⁵⁸, perceived stress, and difficulties with verbal memory, visual perception, language, and attention⁶² than LR.

Metabolome variables in order of importance included N-acetylglutamate (NAG), dimethylglycine, and creatine. HR had higher mean levels of NAG and dimethylglycine, and similar mean levels of creatine compared to LR. Bacterial transcriptomes in order of importance are depicted in Figure 3 and their functions are listed in Figure 1E. In brief, mean levels of transcriptomes linked to environmental adaptation, genetic propagation, metabolism, and anti-inflammation (i.e., increased frequency of a transcriptome involved in SCFA production) were higher in HR compared to LR.

Structural MRI features in order of importance included right subcallosal gyrus (SbCaG) volume, SbCaG surface area (SA), left angular gyrus (AngG) volume, AngG SA, right inferior part of the precentral sulcus (InfPrCS) SA, and InfPrCS volume. The resting-state functional MRI features in order of importance included connectivity between the ventral tegmental area (VTA) and right thalamus, between the left mesencephalic reticular formation (MRF) and the right supramarginal gyrus(SuMarG), between the right inferior segment of the circular sulcus of the Insula (InfCirIns) and the left pallidum, between the right subcentral gyrus and sulci (SbCG_S) and the left superior parietal lobule (SupPL), and between the right lateral orbital sulcus (LORs) and the left putamen. Diffusion MRI features in order of importance included connections between the right and left SbCaG, and between the right lateral orbital gyrus (SupOcG) and right hippocampus. HR had lower mean levels of all DIABLO-selected structural MRI features but higher mean levels of all resting-state functional MRI features than LR. Regarding diffusion MRI features, HR had lower mean levels in bilateral SbCaG connections but higher connections between the right SupOcG and right hippocampus.

There were also associations between different classes of DIABLO-selected omics variables that were unique to HR. For example, anxiety (HADS) was negatively associated with a transcriptome involved in SCFA production. Although there were positive associations between NAG and transcriptomes involved in anti-inflammatory response and environmental adaptation in both HR and LR, only HR had an additional positive association between NAG and a transcriptome involved in genetic propagation.

The entire model is depicted in the connectogram (Figure 4), where mean levels for each feature are visualized by resilience group in the data blocks forming the circle, and correlations between datasets are depicted inside the circle. Connectograms are built on a similarity matrix⁶³ and represent the correlation between variables from different datasets. A cutoff was chosen as r> 0.55 as this is universally considered a “moderate” correlation.

Figure 4. Connectogram Depicting the Correlations within the Variables of Importance from all the Datasets.

Spearman correlations between features across datasets are shown via the lines inside the circle (cutoff r = 0.55). Mean levels of each feature for each group are represented via the lines within each data block forming the circle. A higher line represents a greater mean value for that group.

Abbreviations: MRI, magnetic resonance imaging; L, left; R, right; SbCaG, subcallosal gyrus; AngG, angular gyrus; InfPrCS, inferior part of the precentral sulcus; VTA, ventral tegmental area; Tha, thalamus; MRF, mesencephalic reticular formation; SuMarG, supramarginal gyrus; InfCirIns, inferior segment of the circular sulcus of the insula; Pal, pallidum; SupPL, superior parietal lobule; LORs, lateral orbital sulcus; Pu, putamen; SupOcG, superior occipital gyrus; Hip, hippocampus; K07335, basic membrane protein A and related proteins; K03315, Na/H antiporter; K01081, 5’-nucleotidase; K07080, uncharacterized protein; K14645, serine protease; K02343, DNA polymerase III subunit gamma/tau; K01596, phosphoenolpyruvate carboxykinase; K01361, lactocepin; K01844, beta-lysine 5,6-aminomutase alpha subunit; K03388, heterodisulfide reductase subunit A2; K03466, DNA segregation ATPase FtsK/SpoIIIE; K02965, small subunit ribosomal protein S19; K01119, 2’,3’-cyclic-nucleotide 2’-phosphodiesterase / 3’-nucleotidase; K01834, 2,3-bisphosphoglycerate-dependent phosphoglycerate mutase; K05516, curved DNA-binding protein; K07667, two-component system, OmpR family, KDP operon response regulator KdpE.

Spearman correlations.

Spearman correlations revealed two CD-RISC factors most robustly associated with DIABLO selected variables: control (sense of control and purpose in life, knowledge of where to turn for help) and persistence (personal competence, high standards, and tenacity; see Table 2). Additionally, Spearman’s rank correlations were performed to assess relations between all ASV’s on the genus level and DIABLO selected metabolites. Results showed that NAG was negatively associated with the genus Bacteroides rs (114) =−0.32, p=.004, q=.02.

Table 2. Correlations of Participant Characteristics and Omics Variables Selected by the DIABLO Model with CD-RISC Sub-scales.

Spearman’s correlations of participant demographics and DIABLO selected variables with Total CD-RISC score, and Adaptability, Control, Meaning, Persistence, and Emotional Cognitive control sub-scales. Adaptability represents ability to bounce back and relates to the positive acceptance of change and secure relationships. Control meaning refers to individual’s perception of control over their life circumstances. Higher scores represent having a sense of control and a tendency to view challenges as manageable and belief that one has some influence over outcomes which enhance resilience by promoting problem-solving and adaptive coping strategies. Meaning relates to spiritual influences. Persistence refers to a sense of self-efficacy and reflects a sense of personal competence and high standards. Emotional Cognitive control scores control under pressure, trust in one’s instincts and tolerance of negative affects. P-value < 0.05 (Significant)

Data Block	Variable	Total Score		Adaptability		Control		Meaning		Persistence		Emotional Cognitive Control
		R	p	R	p	R	p	R	p	R	p	R	p
Demographics	Sex	−0.0424	0.6511	−0.0618	0.5097	0.0896	0.3387	0.1287	0.1724	−0.0851	0.3636	−0.1027	0.2728
	Age	0.1281	0.1704	−0.0312	0.7393	0.2072	0.0256	0.1756	0.0616	0.0871	0.3524	0.1342	0.1509
	BMI	0.0808	0.3883	0.0428	0.6486	0.0586	0.5321	0.0676	0.4750	0.0562	0.5493	0.0932	0.3199
	Standard American Diet	0.0611	0.5148	−0.0993	0.2891	0.1172	0.2101	0.1152	0.2223	0.0607	0.5174	0.0722	0.4413
Clinical	Brief Resilience Score	0.6715	0.0000	0.6152	0.0000	0.5149	0.0000	0.1736	0.0647	0.6143	0.0000	0.6138	0.0000
	FFM - Describe	0.3389	0.0010	0.3059	0.0032	0.4181	0.0000	0.2044	0.0547	0.2681	0.0102	0.2371	0.0237
	FFM - Nonjudge	0.4687	0.0000	0.3911	0.0002	0.4519	0.0000	0.1582	0.1434	0.4769	0.0000	0.3518	0.0007
	FFM - Total Score	0.5345	0.0000	0.4454	0.0000	0.5229	0.0000	0.1938	0.0755	0.4819	0.0000	0.4531	0.0000
	HADS - Anxiety	−0.4190	0.0000	−0.2951	0.0013	−0.4311	0.0000	−0.0876	0.3540	−0.4274	0.0000	−0.3599	0.0001
	HADS - Depression	−0.5104	0.0000	−0.5093	0.0000	−0.5065	0.0000	−0.1163	0.2177	−0.4418	0.0000	−0.4217	0.0000
	IPIP - Extraversion	0.4690	0.0000	0.3881	0.0000	0.4782	0.0000	0.2304	0.0150	0.4321	0.0000	0.3522	0.0001
	IPIP - Neuroticism	−0.5935	0.0000	−0.5298	0.0000	−0.4870	0.0000	−0.0645	0.5013	−0.6341	0.0000	−0.4980	0.0000
	MASQ - Attention	−0.5460	0.0000	−0.3877	0.0001	−0.4351	0.0000	−0.1388	0.1686	−0.5352	0.0000	−0.5286	0.0000
	MASQ - Language	−0.3905	0.0000	−0.4462	0.0000	−0.3115	0.0014	0.0530	0.6004	−0.4040	0.0000	−0.3333	0.0006
	MASQ – Verbal Memory	−0.4836	0.0000	−0.3035	0.0020	−0.4830	0.0000	−0.1775	0.0788	−0.4305	0.0000	−0.4660	0.0000
	MASQ - Visual Perception	−0.3730	0.0001	−0.3467	0.0004	−0.3190	0.0011	−0.0225	0.8238	−0.3515	0.0003	−0.3511	0.0003
	PSS - Score	−0.5098	0.0000	−0.4116	0.0000	−0.5298	0.0000	−0.1402	0.1368	−0.4940	0.0000	−0.4109	0.0000
	STAI - Anxiety	−0.6287	0.0000	−0.6153	0.0000	−0.5603	0.0000	−0.1436	0.1276	−0.5816	0.0000	−0.5170	0.0000
Metabolome	N-acetylglutamate	0.1623	0.0997	0.0933	0.3464	0.2240	0.0223	0.0634	0.5267	0.1032	0.2972	0.1710	0.0827
	Dimethylglycine	−0.0139	0.8917	−0.0863	0.3983	0.0537	0.5996	0.0552	0.5932	−0.0376	0.7134	−0.0060	0.9533
	Creatine	−0.2416	0.0135	−0.1297	0.1894	−0.2656	0.0064	−0.1605	0.1071	−0.2333	0.0172	−0.1782	0.0703
Transcriptome	K07335	0.2408	0.0092	0.1396	0.1351	0.2866	0.0018	0.0766	0.4182	0.2982	0.0011	0.1494	0.1095
	K03315	0.2063	0.0263	0.1328	0.1553	0.1873	0.0441	0.0529	0.5761	0.2226	0.0163	0.1829	0.0494
	K01081	0.1358	0.1461	0.0572	0.5420	0.1898	0.0412	0.1293	0.1705	0.1318	0.1584	0.0771	0.4105
	K07080	0.1183	0.2058	0.1346	0.1496	0.1241	0.1845	−0.0048	0.9594	0.1386	0.1380	0.0686	0.4643
	K14645	0.1442	0.1224	0.0720	0.4426	0.1819	0.0507	0.1832	0.0510	0.1119	0.2319	0.0942	0.3146
	K02343	0.1749	0.0604	0.0996	0.2872	0.1842	0.0478	0.0163	0.8635	0.2127	0.0219	0.1427	0.1265
	K01596	0.2705	0.0033	0.1767	0.0578	0.2848	0.0019	0.1656	0.0783	0.2645	0.0041	0.1973	0.0338
	K01361	0.1916	0.0394	0.0697	0.4570	0.2007	0.0307	0.1531	0.1039	0.1887	0.0424	0.1607	0.0849
	K01844	0.2080	0.0251	0.0775	0.4083	0.2676	0.0037	0.1762	0.0607	0.2084	0.0248	0.1427	0.1264
	K03388	0.2522	0.0063	0.2002	0.0312	0.2725	0.0031	0.0502	0.5962	0.2720	0.0031	0.1832	0.0490
	K03466	0.1844	0.0475	0.1019	0.2764	0.2238	0.0158	0.0105	0.9119	0.2408	0.0092	0.1240	0.1848
	K02965	−0.1880	0.0433	−0.1876	0.0438	−0.1580	0.0902	−0.0887	0.3482	−0.1604	0.0854	−0.1528	0.1015
	K01119	0.1712	0.0662	0.0963	0.3039	0.1672	0.0729	0.2378	0.0108	0.1086	0.2457	0.1326	0.1558
	K01834	−0.2441	0.0083	−0.1453	0.1195	−0.2214	0.0169	−0.2148	0.0217	−0.2492	0.0070	−0.1580	0.0902
	K05516	−0.1948	0.0361	−0.1940	0.0369	−0.1877	0.0436	−0.0055	0.9534	−0.1875	0.0438	−0.1703	0.0677
	K07667	0.0165	0.8602	−0.0300	0.7490	0.1004	0.2833	0.0553	0.5588	0.0315	0.7371	−0.0306	0.7440
Structural MRI	R SbCaG Vol	−0.2085	0.0247	−0.2589	0.0050	−0.2150	0.0204	−0.0037	0.9691	−0.1790	0.0545	−0.1582	0.0900
	R SbCaG SA	−0.1138	0.2240	−0.1785	0.0552	−0.1488	0.1108	0.0305	0.7473	−0.0939	0.3158	−0.0712	0.4474
	L AngG Vol	−0.2401	0.0094	−0.2194	0.0180	−0.2163	0.0197	0.0302	0.7496	−0.2220	0.0166	−0.2553	0.0057
	L AngG SA	−0.2190	0.0182	−0.2461	0.0078	−0.1763	0.0583	0.0218	0.8182	−0.1922	0.0388	−0.2162	0.0198
	R InfPrCS SA	−0.1191	0.2028	−0.1170	0.2108	−0.1306	0.1625	−0.1423	0.1311	−0.0694	0.4592	−0.0711	0.4483
	R InfPrCS Vol	−0.0930	0.3206	−0.0736	0.4324	−0.0992	0.2896	−0.1036	0.2725	−0.0508	0.5885	−0.0775	0.4083
rs Functional MRI	VTA to R Tha	0.1827	0.0496	0.1555	0.0956	0.1567	0.0930	0.1320	0.1615	0.1529	0.1013	0.1401	0.1335
	L MRF to R SuMarG	0.2659	0.0039	0.1848	0.0470	0.2537	0.0060	0.0651	0.4915	0.2836	0.0020	0.2231	0.0161
	R InfCirIns to L Pal	0.1801	0.0531	0.1135	0.2252	0.1842	0.0478	0.1468	0.1190	0.1500	0.1079	0.1467	0.1162
	R SbCG S to L SupPL	0.1432	0.1251	0.1144	0.2216	0.1889	0.0423	0.1305	0.1662	0.1350	0.1485	0.0576	0.5388
	R LORs to L Pu	0.1956	0.0354	0.1127	0.2284	0.1751	0.0601	0.1891	0.0439	0.2058	0.0266	0.1163	0.2137
Diffusion MRI	R SbCaG to L SbCaG	−0.2406	0.0093	−0.1567	0.0930	−0.3306	0.0003	−0.1463	0.1204	−0.2416	0.0090	−0.1350	0.1485
	R SupOcG to R Hip	0.1912	0.0398	0.1712	0.0662	0.1685	0.0706	0.0663	0.4835	0.1332	0.1542	0.2027	0.0291

Abbreviations: BMI, body mass index; CD-RISC, Connor-Davidson Resilience Scale; FFM, Five Facet Mindfulness; HADS, Hospital Anxiety and Depression Scale; IPIP, International Personality Pool; MASQ, Multiple Ability Self-Report Questionnaire; PSS, Perceived Stress Scale; STAI, State-Trait Anxiety Inventory; K07335, basic membrane protein A and related proteins; K03315, Na/H antiporter; K01081, 5’-nucleotidase; K07080, uncharacterized protein; K14645, serine protease; K02343, DNA polymerase III subunit gamma/tau; K01596, phosphoenolpyruvate carboxykinase; K01361, lactocepin; K01844, beta-lysine 5,6-aminomutase alpha subunit; K03388, heterodisulfide reductase subunit A2; K03466, DNA segregation ATPase FtsK/SpoIIIE; K02965, small subunit ribosomal protein S19; K01119, 2’,3’-cyclic-nucleotide 2’-phosphodiesterase / 3’-nucleotidase; K01834, 2,3-bisphosphoglycerate-dependent phosphoglycerate mutase; K05516, curved DNA-binding protein; K07667, two-component system, OmpR family, KDP operon response regulator KdpE; MRI, magnetic resonance imaging; L, left; R, right; SbCaG, subcallosal gyrus; AngG, angular gyrus; InfPrCS, inferior part of the precentral sulcus; VTA, ventral tegmental area; Tha, thalamus; MRF, mesencephalic reticular formation; SuMarG, supramarginal gyrus; InfCirIns, inferior segment of the circular sulcus of the insula; Pal, pallidum; SupPL, superior parietal lobule; LORs, lateral orbital sulcus; Pu, putamen; SupOcG, superior occipital gyrus; Hip, hippocampus

Discussion

Our study shows that several key BGM markers distinguish HR from LR. HR exhibits adaptive psychological features, microbiome function facilitating gut health, and neurological signatures supporting emotion regulation and cognitive-emotional connections. Notably, among all data blocks, bacterial transcriptomes most strongly differentiated high from low resilience phenotypes. Furthermore, our findings suggest resilient individuals, particularly those demonstrating tenacity and perceived ability to control life outcomes, possess a microbiome that supports gut barrier integrity and eubiosis and a cortical signature that reflects adaptive emotional and cognitive regulation.

Association of resilience with clinical symptoms

In addition to higher levels of extraversion, HR was associated with lower scores of depression, anxiety, perceived stress, and neuroticism, which is consistent with previous literature^5,6,9,64,. HR was also higher in trait-like mindfulness, particularly in abilities to express emotions in words, be non-judgmental, and express empathy, which are characteristics linked to increased resilience to stress⁶⁵. Additionally, HR was related to better self-reported cognitive abilities including verbal fluency, verbal memory, comprehension, visual memory, and sustained attention. Such cognitive skills have been associated with coping ability^11,66,67. Collectively, these findings suggest low resilience individuals may deplete psychological and cognitive resources when confronted with stressful events while resilient individuals may instead reappraise stressful events to promote advantageous psychological outcomes and coping.

Association of resilience with gut transcriptomics and metabolites

The functional categories of transcriptomes associated with HR included adaptive response to environmental changes, genetic propagation, energy metabolism, and anti-inflammatory response. HR exhibited increased frequency of pathways that improve gut bacteria adaptation in unfavorable conditions (e.g., uptake of essential nutrients or extrusion of toxic substances; adjustment to fluctuations in pH and osmotic changes; communication within cellular community to adjust gene expression accordingly and cooperatively behave in virulence regulation, resource utilization, or even antibiotic resistance⁶⁸). In support of bacterial adaptation, HR also exhibited increased frequency of pathways facilitating genetic proliferation and pathways providing bacterial energy sources (e.g., carbohydrate metabolism). In addition, HR exhibited increased frequency of pathways related to anti-inflammatory response; therefore, maintaining gut barrier integrity (e.g., metabolic degradation of lysine into SCFAs such as acetate and butyrate⁴⁰). See Supplemental Table S1 for other transcriptomes associated with high resilience. Our findings were consistent with animal studies showing that high resilience phenotypes are associated with increased frequency of pathways related to SCFA production^41,42, as well as carbohydrate metabolism and genetic information processing⁴². Overall, HR possessed a microbiome that functions to maintain eubiosis and gut barrier integrity, thereby promoting intact communication between the gut and brain to optimize psychological functioning.

In the metabolome, HR was associated with higher levels of N-acetylglutamate (NAG) and dimethylglycine (DMG). Bacteria use NAG, derived from the amino acid glutamate, to synthesize arginine⁶⁹. NAG’s role in the BGM system is unclear. Increased levels of NAG may be secondary to endogenous metabolism or altered dietary intake, at least in healthy individuals⁷⁰. In the present study, increased levels of NAG in HR were associated with higher frequency of metabolic pathways related to environmental adaptation and anti-inflammation. Increased levels of NAG may be associated with HR secondary to stimulating arginine synthesis which has anti-inflammatory effects in the gut⁷¹. On the other hand, DMG is derived from dietary amino acid glycine and is used in methylation reactions crucial for various metabolic pathways, including energy metabolism, anti-oxidative activity, and DNA synthesis and metabolism⁷². DMG has also been shown to increase gut microbiota strains involved in anti-inflammatory response (e.g., SCFA production)⁷³. See Supplemental Table S1 for other metabolites associated with HR. Our findings show the HR group may host a gut microbiome that can withstand perturbations as evidenced by an increase in metabolites that quell inflammation which, in turn, may support optimal neurological processes.

Association of resilience with multimodal brain signatures

Regarding brain morphometry, HR was associated with reduced SbCG grey matter volume and surface area. The SbCG, an important node in a network including the limbic system and thalamus, is strongly associated with cognitive-emotional processing⁷⁴ and inhibiting fear responses⁷⁵. Similarly, reduced anterior cingulate cortex volume (subcomponents of which include the SbCG) was associated with PTSD remission after treatment completion, showing that volume reduction may indicate neuroplastic changes evidencing increased emotional regulation and extinction of maladaptive cognitive-emotional connections⁵². Reduced SbCG volume was also positively associated with greater resilience subscale scores⁵¹. These morphological signatures suggest dampened but adaptive reactivity towards acute emotional challenges and greater brain efficiency⁷⁶, highlighting that resilient individuals may require less cognitive effort in emotion and fear modulation^11,76. Gut microbiota alterations can also influence grey matter changes^44,77. Taken together, a resilient individual demonstrates microbiome functions that exert characteristic CNS changes, which then provide the physiological means for adaptive coping.

HR was associated with decreased anatomical connectivity involving the right and left SbCG, regions linked to dysphoric mood⁷⁸. Decreased anatomical connectivity between the bilateral SbCG regions may imply less activation of distressing emotions⁷⁹. In addition, HR was associated with decreased anatomical connectivity involving the hippocampus, a region together with the amygdala that is involved in fear⁸⁰ and anxiety⁷⁹. Based on these findings, HR demonstrates a stress-resilient neurological signature.

Several functional connections were significant in differentiating HR vs LR, including the reward circuit (e.g., basal ganglia), sensorimotor network (SMN), DMN, and brainstem. HR had increased rsFC between the reward circuit (ventral tegmental area, pallidum) and SMN (thalamus, insula). Connections between these regions are related to fear-related motor neurocircuitry and thought to be involved in maladaptive fear emotions^81,82. Abnormal connectivity between these areas may be a vulnerability factor. For example, restricted range of emotion is a common symptom in stress-related mental health conditions like PTSD and depression⁸³, which highlights a possible link to reward circuitry alterations. Other brain networks associated with HR involved the DMN (supramarginal gyrus) and brainstem (MRF). Normal connectivity between the DMN and MRF may be a resilient characteristic given that self-monitoring (DMN) of threatening stimuli (MRF) facilitates readiness to react to stressors appropriately. The DMN may also play a role in resilience given its involvement in treatment recovery and “bouncing back” from trauma^55,84,85,. DMN connectivity has also been linked to an abundance of certain gut microbiota⁴⁵; therefore, it is plausible the microbiome optimizes stress response by modulating relevant areas of brain connectivity. Our findings highlight that a resilient individual demonstrates an intact BGM system that shapes connectivity within the CNS to allow for adaptive coping.

Association of resilience with integrated brain-gut interactions

Although SCFAs were nonsignificant predictors of resilience within our DIABLO model, lower anxiety scores in HR were associated with higher frequency of a transcriptome involved in SCFA production. Thus, an individual’s level of resilience may modulate microbiome function to produce SCFAs and consequently mitigate stress-related conditions such as anxiety. Importantly, SCFAs are involved in maintaining intestinal barrier integrity⁸⁶ and anti-inflammation⁸⁷. Lower anxiety scores in HR were also associated with a lower degree of neuroticism and perceived stress in life. In addition, increased frequency of certain bacterial transcriptomes in HR was associated with increased production of the bacterial metabolite, NAG. Such transcriptomic pathways may facilitate overall gut health by supporting increased production of NAG, a metabolite implicated in having anti-inflammatory effects in the gut. Furthermore, resilient individuals may be able to mitigate development of stress-related psychopathology due to bacterial transcriptomes aiding in maintaining overall gut health^69,88.

Conclusion and Clinical Implications

This study is the first to identify a HR BGM phenotype and reveals promising pathways by which the onset and severity of stress-related psychiatric conditions might be prevented or mitigated. This cross-sectional study provides theoretical support for the development of longitudinal studies needed to establish causality. While the current study focuses on healthy samples to facilitate identification of potential factors that enhance and maintain health, future studies should consider comparing microbial profiles of healthy controls to those presenting with psychiatric diagnoses such as depression or PTSD. Future work should also consider metabolomics analysis of plasma samples and, importantly, vagal tone as it may mediate the relation between stress-resistance and microbial composition. Interestingly, the afferent vagus nerve can differentiate between pathogenic and non-pathogenic bacteria and can transmit signals that either exacerbate or mitigate stress responses depending on the bacterial stimulus⁸⁹ making it a promising avenue for future inquiry. Some clinical implications to explore are whether dietary modifications, prebiotics, probiotics, or other clinical interventions (e.g., fecal transplantation) may improve coping and resilience to stress. Collectively, our findings support numerous avenues for novel inquiries and suggest that features of the brain and the gut microbiome work together to build stress-resilience.

Methods and Materials

Participants

A cohort of 116 healthy individuals (71 females) were recruited from the Los Angeles community through advertisements. Premenopausal women were included as determined by self-report of the last day of the previous menstrual cycle, and enrolled women were scanned during the follicular phase of the menstrual cycle. Participants were excluded if they had any major medical/neurological conditions, current or past psychiatric illnesses, gastroenterological issues, abdominal surgeries, substance use, tobacco dependence (half a pack or more daily), or metal implants (due to MRI contraindications); regularly used medications that interfere with the CNS; regularly used analgesics; were pregnant or breastfeeding; performed extreme strenuous exercise (> 8 hours of continuous exercise per week); weighed over 400 pounds; or used antibiotics or probiotics in the past 3 months.

All procedures complied with institutional guidelines and were approved by the Institutional Review Board at UCLA’s Office of Protection for Research Subjects. All participants provided written informed consent.

Study design

In this cross-sectional study, all participants underwent multimodal MRI brain imaging, provided a stool sample within 2-3 days prior to scans, and answered questionnaires including detailed diet information during the week before the MRI scan (consistent with published studies^90–95). This study was a secondary data analysis on existing data and was pooled from two studies: (IRB#s 16-000281, 16-000187).

Questionnaires

Questionnaire data included the following:

Resilience was measured using the Connor-Davidson Resilience Scale (CD-RISC), which is a self-reported scale that consists of 25 items, evaluated on a five-point Likert scale ranging from 0-4; not true at all (0), rarely true (1), sometimes true (2), often true (3), and true nearly all the time (4) resulting in a number between 0-100 with higher scores indicating higher resilience. The resilience total score is comprised of 5 factors: 1) personal competence, high standards, and tenacity; 2) trust in one’s instincts, tolerance of negative affect, and strengthening effects of stress; 3) positive acceptance of change and secure relationships; 4) control; 5) spiritual influences⁴. The mean resilience score of the general US population is 80.7. However, given our study’s participants having a median body mass index (BMI) of 28.06, we based the threshold for high and low resilience using the mean score (83.1) of a study involving mostly overweight patients (normally distributed data with Cronbach’s alpha of 0.92)⁹⁶. In this study, a CD-RISC score ≥ 83.1 was labeled as “High” resilience (N = 50) and < 83.1 was “Low” Resilience.

Other measures included BMI, socioeconomic status (SES), Early Trauma Inventory (ETI)⁹⁷, Adverse Childhood Experiences (ACE) questionnaire⁹⁸, Hospital Anxiety and Depression Scale (HADS)⁶⁰, Coping Strategies Questionnaire (CSQ)⁹⁹, Perceived Stress Scale (PSS)¹⁰⁰, State-Trait Anxiety Inventory (STAI)⁶¹, Positive and Negative Affect Schedule (PANAS)¹⁰¹, 12-item Short Form (SF12) Survey¹⁰², International Physical Activity Questionnaires (IPAQ)^103,104, Behavioral Inhibition System, Behavioral Approach System (BISBAS)¹⁰⁵, Everyday Discrimination Score (EDS)^106,107, Brief COPE (BCope)¹⁰⁸, Patient Health Questionnaire (PHQ)¹⁰⁹, Multiple Ability Self-Report Questionnaire (MASQ)⁶², Mindful Attention Awareness Scale (MAAS)¹¹⁰, Five Facet Mindfulness (FFM)⁵⁹, Patient-Reported Outcomes Measurement Information System Sleep Scale (PROMIS_Sleep9)¹¹¹, Visceral Sensitivity Index (VSI)^112,113, Pain Vigilance and Awareness Questionnaire (PVAQ)¹¹⁴, Pain Catastrophizing Scale (PCS)^115,116, Normal Personality Assessment (NEO)^117,118, International Personality Pool (IPIP)⁵⁸, Diet Questionnaire (used in our previous studies)⁹³.

Gut Microbiome

The methods used for sample collection, processing and analysis are described in detail in published papers^119,120.

Collection and storage

Participants were given “at-home kits” with specified instructions for when to collect and how to store their stool sample. Stool was collected 2-3 days before the MRI scan. 2-3 consecutive diet diaries were collected from the time of enrollment to the time of the MRI scan and stool collection (1-2 weekdays and 1 weekend). Participants were asked to collect the stool before the first meal of the day and immediately freeze the fresh stool after collection. If participants were on antidiarrheal or laxatives, they were asked to refrain from use for 2-3 days before the sample collection. Any deviations from the stool sample collection were documented to account for in the analysis. Fecal samples were stored at −80°C, then ground to coarse powder by mortar and pestle under liquid nitrogen and aliquoted for nucleic acid extraction and metabolomic profiling.

Fecal microbial profiling

DNA extraction with bead beating was performed using the QIAGEN Powersoil DNA Isolation Kit (MO BIO Laboratories, Carlsbad, CA), following the manufacturer’s protocol. The V4 hypervariable region of the 16S rRNA gene was then amplified using 515F and 806R primers to generate a sequencing library according to a published protocol¹²¹. The PCR products were purified with a commercial kit. The library underwent 2x250 sequencing on an Illumina HiSeq 2500 to a mean depth of 250,000 merged sequences per sample. The DADA2 pipeline was used for quality filtering, merging paired-end reads, removing chimera, and assigning taxonomy to each amplicon sequence variant (ASV) using the SILVA 138 reference database^122–125.

Alpha and Beta Diversity.

Bacterial counts were derived using the DADA2 package v1. 29.0¹²² and analyzed using R statistical software. To preserve statistical power, only bacterial genera that were present in at least 10% of the samples were included in analyses resulting in the exclusion of 146 genera and inclusion of 125 genera for analyses. Microbial alpha diversity was assessed on datasets rarefied to 10000 using the Shannon index to assess richness and evenness. Permutational analysis of variance analyses (PERMANOVA) of Bray-Curtis dissimilarities was used to quantify variation in genus between groups. Specifically, beta diversity was calculated using the adonis2 function in Vegan package 2.6-4 to determine Bray-Curtis distance matrices and conduct PERMANOVA significance testing for compositional data with 999 permutations.

Fecal metabolomics processing

Aliquots of fecal samples were shipped to and processed by Metabolon, Inc and run as a single batch through their global HD4 Metabolomics platform¹²⁶, which involves running methanol-extracted samples through ultrahigh performance liquid chromatography-tandem mass spectroscopy under four separate chromatography and electrospray ionization conditions, separating the compounds by their chemical properties. The amount of missing data was low (<3%). However, missing values in the raw data were median imputed using median values, and ineffective peaks were dropped through interquartile range denoising, and an internal standardization normalization method was employed. The metabolites data block was compiled from the metabolite profiling results, and a 3-dimensional matrix with metabolite numbers, sample names, and normalized peak intensities processed with the MetaboAnlyst web software 3.0 (http://www.metaboanalyst.ca).

Fecal Transcriptomics Processing

Fecal samples were submitted to Viome Life Sciences, Inc., where RNA extraction, metatranscriptomics sequencing and annotation were conducted. A detailed description of these procedures is provided in previously published work¹²⁷. In sum, RNA extraction by beading was performed, DNA degraded by DNase, and 16S/23S ribosomal RNA depleted by subtractive hybridization. Sequencing libraries were prepared from the resulting RNA and underwent 150x2 paired-end sequencing on Illumina NovaSeq. Taxonomy was assigned by aligning sequencing reads to a precomputed database of unique k-mers; 898 taxa were identified, including bacteria, fungi, viruses, and bacteriophages. Functional annotation was performed by aligning sequencing reads to the integrated gene catalog from the MetaHIT consortium then mapping these genes to the KEGG database; 5,896 distinct transcripts annotated as KOs were identified¹²⁸.

Multimodal neuroimaging

Magnetic resonance imaging acquisition.

Each subject underwent imaging in a 3.0T Prisma MRI Scanner with a 20-channel head coil (Siemens Healthcare, Erlangen, Germany) for a high resolution T1 structural scan, a resting state functional scan, and a diffusion weighted scan. Participants were asked to fast for an average of 6 hours prior to scanning. The acquisition parameters are as follows: T1 weighted MP-RAGE scans acquired to assess brain structure (TR: 2300ms, TE: 2.98ms, TI: 900ms, flip angle: 9°, field of view: 240 x 256 mm, acquisition matrix: 240 x 256, slice thickness: 1 mm, voxel resolution: 1 x 1 x 1 mm). A 10-minute resting-state fMRI scan was acquired to assess resting-state functional connectivity (TR: 2000ms, TE: 28ms, flip angle: 77°, acquisition matrix: 64 x 64, slice thickness: 4 mm, voxel resolution: 3.44 x 3.44 x 4 mm, 300 volumes). A diffusion weighted image was acquired to assess white matter anatomical connectivity (64 noncollinear directions, b = 1000 s/mm2, 9 b = 0 s/mm2 images, TR: 9500ms, TE: 88ms, field of view: 2304 x 2304, acquisition matrix: 128 x 128, slice thickness: 2mm, spacing between slices: 2mm).

MRI preprocessing.

Scans from each neuroimaging modality was considered as separate datasets and processed separately with the appropriate respective modality specific pipelines. Structural scans were processed and passed quality control using Statistical Parametric Mapping 12.¹²⁹ Processing included motion correction, skull stripping, segmentation into gray matter, white matter, and cerebral spinal fluid (CSF), and normalization onto a NMI153 T1 template.

Structural image processing.

Cortical reconstruction and volumetric segmentation was done using the FreeSurfer6 analysis suite¹³⁰. All participants’ T1 structural data was first parcellated using the Destrieux cortical atlas¹³¹ and the Harvard-Oxford subcortical atlas and the Harvard Ascending Arousal Network (AAN)¹³². FreeSurfer was used to compute values of cortical thickness, surface area, mean curvature and volume for cortical ROIs and volume for subcortical ROIs. Multimodal scans were processed similar to previously published studies^95,133–135.

Functional image processing.

Functional scans were preprocessed using the volume-based rs-FC analysis pipeline in the functional connectivity toolbox (CONN)¹³⁶. All scans underwent realignment and unwarping, slice-timing correction, and outlier identification (advanced retrospective technique-based identification of outlier scans, ART) for scrubbing. Functional and structural data (T1 scans) were normalized and segmented into grey matter, white matter and CSF tissue¹³⁷. Denoising was done using ordinary least squares regression of potential confounding effects and temporal band-pass filtering. The default anatomical component-based noise correction procedure (aCompCor) includes noise components from white matter¹³⁸, estimated subject-motion parameters¹³⁹, outlier scans or scrubbing based on frame-wise displacement¹⁴⁰, and effect of rest repressing potential ramping effects (at the start of the session)¹⁴¹. The influence of physiological, head-motion and other noise sources were minimized using a 0.008 Hz to 0.09 Hz after regression temporal band-pass filter¹⁴². Fisher transformed correlations (Z) were computed between the functional time series of all the FreeSurfer parcellated regions to derive a 165x165 matrix for each participant. A single vector representing the correlation strength between each ROI pair was concatenated from the bottom half of the undirected matrix for each subject.

Diffusion image processing.

Diffusion-weighted images, corrected for eddy current-induced distortions and movement with FSL’s eddy_correct tool¹⁴³, along with the associated b-vectors and v-values were converted into Camino data formats with Camino’s fsl2scheme and image2voxel¹⁴⁴. Weighted linear least squares regression was used to fit a diffusion tensor on the voxel order data in Camino (wdtfit)¹⁴⁴. Deterministic and probabilistic tensor-based approaches have shown to have similar performance¹⁴⁵. The track command in Camino Euler algorithm performed whole-brain deterministic tractography with a step size of 0.5 and curve threshold of 76. Connectivity matrices were constructed using the conmat command in Camino and produced a 165x165 matrix that represents the number of streamlines connecting each ROI-to-ROI pair. Every subject’s matrix went through within-subject normalization by taking a sum of all counts between each ROI, then dividing each pair’s count by that total. The bottom half of the undirected matrix was then concatenated into one vector for each subject representing every ROI pair.

Data Integration Analysis for Biomarker discover using Latent cOmponents (DIABLO)

Approach Overview.

Integrating multiple omics approaches is essential for a thorough understanding of stress resilience and its associated phenotypes. Studying multimodal brain imaging, microbiome, transcriptome, metabolome, and clinical/behavioral variables in isolation cannot provide a comprehensive biological insight. DIABLO¹⁴⁶ facilitates employing integrative multi-omics analysis, examining the interplay between various datasets to unravel the intricacies of stress-resilient phenotypes.

We used DIABLO to accomplish our Integrative Multi-omics Analyses goal to elucidate the interactions among central (brain), peripheral (microbiome, metabolome), and clinical/behavioral markers linked to resilience phenotypes. This advanced multi-block integration strategy simultaneously models outcomes based on multiple data matrices, identifies key predictive variables, and unveils relationships between different dataset types. DIABLO calculates linear combinations (multi-omic signatures) maximally correlated with a specified outcome while performing variable selection, controlling for relevant covariates (e.g., gender, age, BMI), and correcting for multiple comparisons. The result is a yield of a minimal subset of variables associated with resilience outcomes. Additionally, we identify a limited number of multi-omics signatures distinguishing high resilience from low resilience in both training and test sets.

DIABLO Analysis.

DIABLO¹⁴⁶ with a supervised learning framework was conducted to determine a sparse subset of correlated phenotypic and behavioral features from the 6 high-dimensional input data blocks (Q) that predict resilience group given the 83.1 score threshold⁹⁶. Data was split into a 70% training (N = 81) and 30% testing set (N = 35). Training data is used to calculate the design matrix and train the model. DIABLO extends sparse generalized canonical correlation analysis, a generalization of partial least squares (PLS)^146,147, to a supervised machine learning framework with scarcity constraints for variable selection. Pairwise sparse PLS (sPLS) models were run (e.g. clinical versus resting-state functional MRI; metabolome versus diffusion MRI) prior to the DIABLO analysis to gauge the overall correlation structure between the dataset types and guide the data integration in DIABLO¹⁴⁶. A weighted Q x Q design matrix, a DIABLO input that informs the integration process, was calculated taking the correlation of the first principal component of individual sPLS models between pairwise data blocks. The data-driven design matrix contains values from 0 to 1 and represents if and by how much each data block Q should be correlated to one another. The final model consists of a limited number of features across datatypes that show high correlation with one another, which gives insight to both which ‘omic types’ are relevant to the discriminatory process and how different ‘omic types,’ or datasets interact with one another.

Data Preparation.

Following the independent processing of the six data blocks (clinical, microbiome, metabolite, morphological MRI, rs-FC MRI, DTI MRI), involved the examination of each block to eliminate variables not suitable for analysis. Due to the nature of the data integration algorithm, variables exhibiting near zero variance (NZV) were detected and excluded from further consideration. Regarding the clinical dataset, categorical variables of interest were replaced with dummy variables before identifying and 16 variables with NZV were eliminated. For the fecal metabolome dataset, 38 NZV variables were dropped before performing median normalization. All clinical and metabolome variables with 50% or more NA’s were removed.

Neuroimaging data preparation for DIABLO.

Measures of brain morphometry, resting state functional connectivity and anatomical connectivity were derived for each individual and pairwise ROI similar to our previous published studies^{131,148–151}. Data from the three neuroimaging modalities was considered as separate datasets and went through their own specific preprocessing methods. To construct the structural MRI dataset, the observations that used the ascending arousal network (AAN) and Destrieux Harvard-Oxford atlas parcellations were merged and measures of surface area and volume were residualized by the estimated total intracranial volume (eTIV) to control for effects driven by brain size^152–154. The NZV features in the resting-state functional MRI and diffusion MRI dataset, which can consist of many variables with zeros (i.e., regions that do not share anatomic connections), were dropped. Data blocks for all three neuroimaging datasets were resilience group median imputed.

The final clinical dataset had 99 features, metabolome dataset had 714 features, transcriptome dataset had 1,424 features, structural MRI dataset had 626 features, resting-state functional MRI dataset had 15,753 features, and diffusion MRI dataset had 3,408 features. All six datasets were then scaled and centered separately by calculating mean and standard deviation of each vector, then “scaling” each element by subtracting the mean and dividing by the standard deviation within the DIABLO function call. These datasets were used in subsequent analyses.

Variable selection.

An initial DIABLO model with 10 components and all features from each dataset was fit on the training subset and the global performance was assessed with 5-fold cross validation to identify the number of components that produces the lowest Balanced Error Rate (BER) with mahalanobis distance. One component produced the lowest BER. Mahalanobis distance was set as the distance metric parameter as it is robust at handling imbalanced groups¹⁵⁵. Manual tuning was then performed to determine the optimal number of variables for each data block for the selected number of components to obtain the lowest BER on the unseen test subset. The testing set’s performance was used to determine the final model. DIABLO outputs a set of components (i.e., latent variables), a set of loading vectors (i.e., coefficients assigned to each variable) and a subset of selected variables in each dataset associated with each component. The outputs are obtained by maximizing the covariance between a linear combination of X variables and Y labels and the magnitude of the loading vectors represents the importance of that variable in the model with higher values suggesting greater importance^146,147. Loading plots show the importance of each variable within each component. Connectograms represent the correlation between variables from each datatype selected by the final DIABLO model. Performance metrics, including a confusion matrix, BER, and area under the receiver operating characteristic curve (AUC) curve were calculated to determine how well the model performed for each data block and overall. AUC is a value ranging from 0 to 1 that summarizes the overall diagnostic accuracy of a model. 1 represents a perfectly accurate classification model, 0.5 represents a poor model with no discriminatory ability, and values between 0.7 and 0.8 are considered acceptable¹⁵⁶.

Spearman Correlations.

Spearman correlations were conducted to calculate associations between the five factors that compose the total resilience score and DIABLO selected variables.

Supplementary Material

Supplementary Material

Supplemental Table S1. Loading Values of DIABLO-selected Variables. Magnitude of the loading vectors represent the relative importance of each variable in the final model. Variables are grouped by dataset type and ranked from most to least important.

Abbreviations: FFM, Five Facet Mindfulness; HADS, Hospital Anxiety and Depression Scale; IPIP, International Personality Pool; MASQ, Multiple Ability Self-Report Questionnaire; PSS, Perceived Stress Scale; STAI, State-Trait Anxiety Inventory; MRI, magnetic resonance imaging; L, left; R, right; SbCaG, subcallosal gyrus; AngG, angular gyrus; InfPrCS, inferior part of the precentral sulcus; VTA, ventral tegmental area; Tha, thalamus; MRF, mesencephalic reticular formation; SuMarG, supramarginal gyrus; InfCirIns, inferior segment of the circular sulcus of the insula; Pal, pallidum; SupPL, superior parietal lobule; LORs, lateral orbital sulcus; Pu, putamen; SupOcG, superior occipital gyrus; Hip, hippocampus.

Figure S1

Supplemental Figure S1. Distribution of Group Labels. A Histogram of CD-RISC (Connor-Davidson Resilience Scale) Scores for 116 heathy subjects. 83.1 threshold is based on previous literature for overweight and obese subjects.

Figure S2

Supplemental Figure S2. Shannon Diversity of those in the Low and High Resilience Groups. Box plots showing Shannon diversity across low (pink) and high (blue) resilience groups. All samples were sequenced using 16S sequencing. The center line within each box defines the mean, boxes define the upper and lower quartiles, and whiskers define the interquartile range.

Figure S3

Supplemental Figure S3. Differences in Beta Diversity between Low and High Resilience Groups. Nonmetric multidimensional scaling (MDS) ordination using Bray-Curtis dissimilarity to calculate distances on the genus level between resilience groups: low (pink) and high (blue). Each point represents one sample at one time point. The x and y-axis only provide a reference in which to gauge Bray-Curtis distances in relation to the other samples but are not a meaningful value independently.

Figure S4

Supplemental Figure S4. DIABLO-derived Testing Set Performance and Confusion Matrix of Testing Set. Testing set includes a total N = 35 subjects and training set includes a total N = 81 subjects. A Receiver operating characteristic curve (ROC) for each data block. Area under the curve (AUC) is the measure the ability of a binary classifier to distinguish between the two classes and ranges from 0 to 1; Higher AUC suggests it is a better classifier. The Clinical AUC is 0.7692; Metabolites AUC is 0.6346; Transcriptome AUC is 0.7727; Structural AUC is 0.6469; Functional AUC is 0.6573; Diffusion AUC is 0.4825. The Transcriptome and Clinical data blocks contributed the most to the classification. The Metabolome and Diffusion data blocks contributed the least. B Confusion matrix of testing set. C. The Overall Model AUC is 0.7692 (BER=0.1836).

Abbreviations: MRI, magnetic resonance imaging.

Data Availability:

Deidentified individual participant data (brain) can be shared upon request and will be made available through the Center’s pain repository portal (https://www.painrepository.org/). To access the data, participants will fill out a user agreement, upon which access to the data will be made available through a secure password protected portal. The raw microbiome sequences can be accessed NIH NCBI BioProject (BioProject ID: PRJNA946906).