Brain-gut crosstalk associated with brain ageing in young and mid-life adults: a multicohort cross-sectional study

Summary

Background

Brain age deviation is a promising neuroimaging biomarker of brain health, but its relevance in young and mid-life adults and its biological underpinnings remain insufficiently characterised. We aimed to test whether a functional-connectivity-derived brain ageing index (BAI) captures reproducible variability in early brain ageing and whether it is associated with cognitive-affective function and gut-derived biological signatures.

Methods

We analysed resting-state fMRI from a discovery cohort (n = 674) with validation in a replication cohort (n = 444) and an independent cohort (n = 344). Whole-brain functional connectivity was computed using a 100-region Schaefer parcellation, and Bayesian ridge regression was used to predict chronological age; BAI was defined as the age-bias-corrected residual (predicted brain age minus chronological age). We tested associations between BAI and cognitive and affective measures across cohorts. In the independent cohort, we applied multi-view sparse partial least squares to integrate stool metagenomic and metabolomic profiles with BAI, and performed KEGG pathway enrichment analyses on features with non-zero weights.

Findings

Predicted brain age correlated with chronological age across cohorts (r = 0.50–0.59). Higher BAI was consistently associated with connectivity patterns involving posterior cingulate/praecuneus and medial frontal regions, poorer cognitive performance, particularly working memory and executive function, and greater depressive symptoms. Multi-omics integration identified microbial taxa and stool metabolites, including ceramides, 24-hydroxycholesterol, dicarboxylic acids, and inverse associations with estetrol, linked to BAI. Enrichment analyses suggested involvement of neuroimmune, vascular, synaptic, and mitochondrial pathways.

Interpretation

A connectivity-derived BAI captures reproducible variability in early brain ageing and links large-scale brain network organisation to gut-derived biological signatures. These findings suggest that BAI captures individual variability associated with brain health-related phenotypes and support the potential association of peripheral brain-gut biological pathways in early brain ageing.

Funding

National Institutes of Health, National Institute on Ageing.

Research in context.

Evidence before this study

Neuroimaging-derived brain age has emerged as a promising biomarker of brain health and has been associated with cognitive function, neurodegeneration, and psychiatric symptoms, primarily in older adults or clinical populations. We searched PubMed from database inception to March 10, 2026, without language restrictions, using terms related to brain ageing (“brain age”, “brain-predicted age”, “brain ageing index”, “brain-PAD”), neuroimaging (“MRI”, “functional connectivity”, “resting-state fMRI”, “neuroimaging”), cognition and affect (“cognition”, “memory”, “executive function”, “depression”, “psychiatric”), and gut-derived biology (“microbiome”, “metabolomics”, “gut-brain axis”). Existing studies consistently show that advanced brain age relates to poorer cognition, depressive symptoms, and neurodegenerative biomarkers. However, most work has focused on late-life or disease cohorts, leaving uncertainty about whether brain ageing variability is detectable and clinically meaningful in young and mid-life adults. In addition, although gut microbiome composition and metabolite pathways have independently been linked to brain health and ageing biology, few studies have integrated functional-connectivity-based brain age with stool multi-omics within the same individuals, and mechanistic pathways linking central brain dynamics to peripheral biological signatures remain poorly characterised.

Added value of this study

Using large discovery, replication, and independent cohorts of young and mid-life adults, we developed and validated an rsFC-derived brain ageing index that showed consistent associations with poorer cognitive performance, particularly working memory and executive function, and greater depressive symptoms across cohorts. Beyond prior neuroimaging-only studies, we integrated stool metagenomic and metabolomic data using multi-view sparse modelling to identify convergent microbial taxa and metabolites associated with brain ageing variability. These features were enriched for neuroimmune, vascular, synaptic, and mitochondrial pathways and identified coordinated microbe-metabolite-brain connectivity associations. By linking large-scale brain network organisation with peripheral biology in young and mid-life adults, this work extends current understanding of brain-age-related variability beyond late-life and clinical populations.

Implications of all the available evidence

Together with prior literature showing that brain-ageing-related variation can be observed well before the onset of typical neurodegenerative disease, our findings demonstrate that individual differences in rsFC-derived brain ageing are detectable in young and mid-life adults and are associated with cognitive, affective function and gut-derived biological processes. Integrating functional connectivity-based brain ageing measures with microbiome and metabolomic profiles may provide insight into biological processes and pathways associated with variability in rsFC-derived brain ageing. These results support a life-course framework of brain ageing in which central neural dynamics and peripheral metabolic-microbial systems interact, highlighting the potential contribution of brain-gut pathways to individual differences in brain ageing across adulthood.

Introduction

Brain ageing is a central challenge in neuroscience and geroscience as well as in public health as populations age and age-related brain disorders rise.¹ Magnetic resonance imaging (MRI)-based “brain-age” models address this by calculating the Brain Ageing Index (BAI), the discrepancy between predicted and chronological age, which serves as a surrogate for neural health and relates to clinical outcomes like poor cognitive function and depression.2, 3, 4 However, most research has targeted older or clinical populations, despite evidence that ageing processes and cognitive vulnerabilities manifest as early as the second or third decade of life.⁵ This early-to-mid-adulthood window is critical, as biological alterations accumulate long before clinical symptoms emerge.⁶ Consequently, significant gaps remain regarding which neural features define early-life BAI and whether these individual differences correspond to measurable cognitive, affective, and biological dysfunction.

Positioning BAI within the brain-gut-microbiome system provides a mechanistic bridge from peripheral biology to neural ageing.⁷ The gut microbiome and its stool metabolite milieu shape neuroimmune tone, cerebrovascular function, synaptic plasticity, and mitochondrial metabolism—processes central to brain ageing.⁸ These include ceramides (inflammation/apoptosis), asymmetric dimethylarginine (ADMA) (endothelial nitric-oxide inhibition), 24-hydroxycholesterol (oxysterol modulation of synaptic signalling),⁹ and dicarboxylic acids such as nonanedioate/azelate (ω-oxidation/peroxisomal-mitochondrial fatty-acid handling).¹⁰ Integrating metagenomics with stool metabolomics and imaging-derived BAI in younger and mid-life adults may reveal microbe-metabolite-brain interactions underlying individual differences in brain ageing, highlight metabolic pathways (such as bile-acid and short-chain-fatty-acid pathways), and enable the identification of early brain ageing biomarkers for monitoring and prevention.

The overarching goal of this study is to establish a mechanistic framework linking neural network organisation to peripheral biological processes that shape early brain ageing. We focus on resting-state functional connectivity (FC), which reflects coordinated neural communication and may capture ageing-related changes before structural abnormalities emerge.^4,11 We hypothesise that variability in FC reflects inter-individual differences in brain ageing, particularly within default mode and frontoparietal networks associated with cognitive and affective dysfunction. Using large discovery (N = 674), replication (N = 444), and independent (N = 344) cohorts, we derive a robust BAI, examine its neural correlates, and characterise the associated phenotypic differences. We further integrate BAI with gut metagenomic and metabolomic profiles to identify biological pathways related to neuroimmune, vascular, and metabolic processes. Together, this multimodal framework aims to improve understanding of brain ageing-related variability in early and mid-life adults and provide a foundation for future studies of individualised brain health and potential intervention targets.

Methods

In this study, we utilised four datasets—UCLA Consortium for Neuropsychiatric Phenomics (UCLA-CNP), Autism Brain Imaging Data Exchange (ABIDE) I, ABIDE II, and Strategic Research Program for the Promotion of Brain Science (SRPBS)—to construct discovery and independent cohorts. While these datasets were originally developed to study various brain disorders, we focused exclusively on healthy young and midlife adults (ages 18–65) to study normative brain ageing. Individuals with clinical diagnoses or outside the specified age range were excluded. The analysis of this study was implemented on June 2025. Overall study flowchart is summarised in Fig. 1. All analyses were implemented using Python (version 3.10; packages: scipy 1.15.2, numpy 2.2.6, sklearn 1.7.2), except mv-sPLS, which was conducted in R (version 4.4.2) using the mixOmics package (version 6.30.0).

Fig. 1.

Study profile. (A) Functional Connectivity (FC) computation. Resting-state fMRI time signals were extracted from 100 regions of interest (ROIs) defined by the Schaefer parcellation. Pearson correlations between ROI time series produced an FC matrix per participant. Given symmetry, the upper-triangular elements were retained and vectorised for model training. (B) Analysis of Brain Ageing Index (BAI). A Bayesian ridge regression model was trained to predict chronological age based on FCs, yielding BAI. BAI was then correlated with a range of cognitive and affective phenotypic variables and replicated in replication and independent cohorts. Next, a multi-view sparse partial least square (mv-sPLS) model was used to identify microbiome features associated with BAI. Finally, pathway enrichment analysis (EA) was performed to interpret the underlying biological mechanisms. (C) Summary of findings. BAI was associated with cognitive and affective measures, large-scale brain network organisation, gut microbial taxa, and stool metabolites. The identified microbial and metabolic signatures were enriched for pathways related to neuroimmune, vascular, synaptic, and mitochondrial processes. Together, these findings suggest that individual differences in rsFC-derived brain ageing are associated with variation in cognitive function, affective symptoms, brain network organisation, and gut-derived biological features in young and mid-life adults.

Discovery cohort and replication cohort

UCLA-CNP dataset

The UCLA Consortium for Neuropsychiatric Phenomics (UCLA-CNP) dataset, released publicly in 2016, was designed to investigate memory, cognition, and their alterations in healthy individuals and neuropsychiatric disorders.¹² All participants provided written informed consent after a verbal explanation of study procedures, with protocols approved by the UCLA institutional review board and the Los Angeles County Department of Mental Health. Healthy controls were rigorously screened and excluded for any lifetime diagnosis of schizophrenia or other psychotic disorders, bipolar I/II disorder, or substance abuse/dependence (excluding caffeine and nicotine). MRI-specific exclusions included pregnancy or other MRI contraindications, insufficient vision for task procedures, left-handedness, and positive urine toxicology screening on scan day.

ABIDE I/II datasets

The Autism Brain Imaging Data Exchange (ABIDE I) was launched and made publicly available in 2012,¹³ followed by the second phase (ABIDE II) released in 2016.¹⁴ This initiative was originally designed to investigate the heterogeneity of autism spectrum disorder, while also including more than 900 healthy control participants. All contributing sites obtained institutional review board approval, or an explicit waiver for fully anonymised data, prior to data sharing. Common exclusion criteria across sites included history of major neurological disease, significant medical illness affecting brain function, severe head injury, genetic syndromes associated with autism, current substance abuse, and MRI contraindications such as metal implants or excessive motion during scanning.

SRPBS dataset

The Strategic Research Program for the Promotion of Brain Science (SRPBS) was established to generate large-scale, multi-site neuroimaging datasets aimed at classifying and characterising psychiatric disorders in 2013.¹⁵ All participants provided written informed consent, and recruitment procedures as well as experimental protocols were reviewed and approved by the institutional review board at each participating institution in Japan. Individuals with current or past psychiatric disorders, including schizophrenia spectrum disorders, major depressive disorder, bipolar disorder, obsessive-compulsive disorder, autism spectrum disorder, and other major neuropsychiatric conditions, were excluded from the healthy control group. Additional exclusions included significant neurological disease, major medical illness affecting brain function, history of serious head injury, substance abuse, and use of medications that could substantially affect central nervous system function. MRI-specific exclusions included standard contraindications such as metal implants, pregnancy, claustrophobia, or inability to safely undergo MRI scanning, as well as excessive head motion or poor image quality identified during preprocessing.

Independent cohort

UCLA-Omics dataset

The UCLA-Omics dataset, collected by Dr. Arpana Church's lab, investigates brain-gut associations in healthy adults from 2016 to 2025.¹⁶ Participants were recruited from the Los Angeles community, mostly with premenopausal women scanned during the follicular phase. Exclusion criteria included major medical, neurological, psychiatric, or gastrointestinal conditions; substance use; regular CNS-acting medications; MRI contraindications; pregnancy; extreme exercise; obesity; and recent antibiotic/probiotic use. All protocols were IRB-approved, and written informed consent was obtained.

Phenotypic measurements

To comprehensively assess cognitive and affective functions across cohorts, we employed a range of standardised measurements. Among affective domains, we focused primarily on depression, given its high prevalence and consistent association with ageing and cognitive decline.^17,18 In the discovery cohort, cognitive function was assessed using Full-Scale, Verbal, and Performance IQ scores, capturing general intelligence, verbal reasoning, and problem-solving abilities.¹⁹ We also included the Brief Assessment of Cognition in Schizophrenia (BACS),²⁰ which evaluates six cognitive domains: verbal memory, working memory, motor speed, verbal fluency, executive function, and processing speed. Depressive symptoms were measured using the Beck Depression Inventory (BDI).²¹ In the replication cohort, we additionally included the California Verbal Learning Test (CVLT) to capture learning and memory more specifically.²² In the independent cohort, due to differences in data collection protocols, identical measures were not available. Affective function was assessed using the Social Isolation Scale (SI) and Hospital Anxiety and Depression Scale (HADS),^23,24 while executive function was measured using the Trail Making Test (TMT).²⁵ All analyses were conducted using raw scores.

MRI acquisition and preprocessing

In the UCLA-CNP dataset, MRI data were acquired using one of two 3T Siemens Trio scanners at UCLA (syngo MR B15/B17).¹² Functional images were obtained using a T2∗-weighted echo-planar imaging (EPI) sequence with the following parameters: repetition time (TR) = 2000 ms, echo time (TE) = 30 ms, flip angle = 90°, slice thickness = 4 mm, 34 slices, matrix = 64 × 64, field of view (FOV) = 192 mm, voxel resolution = 3.44 × 3.44 × 4 mm³, and oblique slice orientation. The parameters for T1 scanning were: TR = 1.9 s, TE = 2.26 ms, FOV = 250 mm, matrix = 256 × 256, sagittal plane, slice thickness = 1 mm, 176 slices.

In the ABIDE dataset, resting-state fMRI data were acquired across multiple sites without a standardised scanning protocol. For example, at one site, data were collected using a 3T scanner with TR = 2400 ms, TE = 2.91 ms, flip angle = 9°, inversion time = 900 ms, and echo spacing = 6.76 ms. The slice thickness was 1.00 mm, acquisition matrix = 256 × 240, and voxel resolution = 1.00 × 1.00 × 1.20 mm³ across 160 slices. The T1 imaging parameters were the following: voxel resolution = 1.00 × 1.00 × 1.20 mm³, flip angle = 9°, TR = 1800 ms, TE = 3.06 ms. Data were acquired in sagittal orientation with anterior-to-posterior phase encoding. The total acquisition time (TA) was approximately 9 min and 14 s. Full scanning parameters across sites are detailed in the original ABIDE reference.¹⁴

In the SRPBS dataset, resting-state fMRI data were collected using a Siemens ep2d_bold sequence without motion correction or GRAPPA/SENSE acceleration from multiple sites.¹⁵ One representative protocol included TR = 2500 ms; TE = 30 ms; flip angle = 80°; FOV = 212 mm; resolution = 1.00 × 1.00 × 1.00 mm³; and ∼240 volumes per run. Structural T1-weighted scans were obtained using an MPRAGE sequence (1 mm isotropic voxel; TR/TE/TI = 2300/2.98/900 ms; flip angle = 9°) on Siemens Verio systems.

In the UCLA-Omics dataset, a 10-min resting-state fMRI scan was acquired using the following parameters: TR = 2000 ms, TE = 28 ms, flip angle = 77°, acquisition matrix = 64 × 64, slice thickness = 4 mm, voxel resolution = 3.44 × 3.44 × 4 mm³, and a total of 300 volumes.¹⁶ T1-weighted MPRAGE scans were acquired for structural imaging (TR = 2300 ms, TE = 2.98 ms, TI = 900 ms, flip angle = 9°, FOV = 240 × 256 mm, matrix = 240 × 256, slice thickness = 1 mm).

The acquired rs-fMRI data were preprocessed using the reproducible fMRIprep pipeline.²⁶ The T1 weighted image was corrected for intensity nonuniformity and then skull-stripped. Spatial normalisation was done through nonlinear registration, with the T1w reference.²⁷ Using FSL, brain tissue such as cerebrospinal fluid, white matter, and grey matter was segmented from the reference, brain-extracted T1 weighted image.²⁸ The fieldmap information was used to correct distortion in low-frequency and high-frequency components of the fieldmap caused by field inhomogeneity. With less fieldmap distortion, a corrected echo-planar imaging reference was obtained from a more accurate co-registration with the anatomical reference. The BOLD reference was then transformed to the T1-weighted image with a boundary-based registration method, configured with nine degrees of freedom to account for distortion remaining in the BOLD reference.²⁹ Head-motion parameters (rotation and translation parameters of volume-to-reference transform matrices) were estimated with MCFLIRT (FSL). BOLD signals were slice-time corrected and resampled onto the participant's original space with head-motion parameters, susceptibility distortion correction, and then resampled into standard space, generating a preprocessed BOLD run in MNI152NLin2009cAsym space. Automatic removal of motion artefacts using independent component analysis (ICA-AROMA)³⁰ was performed on the preprocessed BOLD time-series in MNI space after removal of non-steady-state volumes and spatial smoothing with an isotropic Gaussian kernel of 6 mm FWHM (full-width half-maximum). Then, we reviewed the imaging parameters of each participant. Ninety-five participants exceeding the quality-control threshold for head motion (framewise displacement > 0.5 mm) were excluded.31, 32, 33

Calculation of resting-state FC

Preprocessed BOLD signals were averaged within 100 regions of interest defined by the Schaefer parcellation atlas grouped into Yeo's seven functional networks.³⁴ Pairwise Pearson correlations between 100 ROI time series generated FC matrices (Fig. 1A). Fisher's r-to-z transformation and subject-level z-score normalisation were applied to improve statistical properties and reduce inter-site variability (statistically, in removing the site effect, this is equal to demeaning). To ensure data quality, 122 subjects with aberrant FC patterns were identified by the isolation forest algorithm and excluded.

Faecal metagenomic collection and preprocessing

A total of 212 participants (only available in the UCLA-Omics dataset) were profiled for gut microbiome composition using shotgun metagenomic sequencing. Metagenomic DNA extraction, sequencing on an Illumina platform and taxonomic identification of gut microbes of individual stool samples were performed by One Codex. Sequencing was performed using the Illumina NextSeq 2000 platform. Raw base call (BCL) files were demultiplexed and converted to FASTQ format using the bcl2fastq software (Illumina Inc.), employing default parameters for quality trimming and adaptor removal. Shotgun metagenomic DNA was processed for microbial community profiling. All further analyses were conducted in the R statistical computing environment (v4.1.3), primarily using packages available through the Bioconductor project.35, 36, 37, 38 To reduce noise and retain taxa with meaningful representation across samples, we applied an abundance filtering step. Taxa were retained if they had read counts greater than 10 in at least 3 samples. This thresholding was performed on the species table generated from One Codex output, resulting in 4614 species used for downstream analyses. Differential abundance analysis was conducted using the limma package,³⁵ which uses linear models and empirical Bayes.

To stabilise variance and approximate Gaussian distributions for multivariate sparse partial least squares (mv-sPLS) analysis, species-level metagenomic data were subjected to a variance-stabilising transformation (VST).^39,40 While centred log-ratio (CLR) transformation is a common compositional alternative, VST was selected to accommodate the sparsity of the metagenomic counts without the prerequisite for zero imputation.

Faecal metabolomics collection and preprocessing

Metabolomic analysis was conducted on 215 human stool samples using the Metabolon® platform.^41,42 Samples were stored at −80 °C until processing. Extraction was performed using the automated MicroLab STAR® system with internal recovery standards added for quality control. Proteins were precipitated with methanol, homogenised, and centrifuged. The resulting extracts were divided across four UPLC-MS/MS acquisition modes—two reverse-phase (positive ion), one reverse-phase (negative ion), and one HILIC (negative ion)—with an additional fraction reserved for archival use.⁴³ Data were acquired using a Waters ACQUITY UPLC system coupled to a Thermo Q-Exactive Orbitrap mass spectrometer operating with a HESI-II source at 35,000 resolution.^44,45 Chromatographic gradients and solvents were optimised to capture hydrophilic, hydrophobic, basic, and acidic analytes.⁴⁶ Compound identification was performed through Metabolon's standardised informatics pipeline, based on retention index, accurate mass (±10 ppm), and MS/MS spectral matching against a reference library containing >3300 authenticated standards. Following curation and run-day normalisation, targeted processing yielded 10 quantifiable metabolites.

We further performed comprehensive shotgun metabolomic profiling on the same 215 human stool samples using the Sapient high-resolution LC-MS platform.⁴⁷ Samples were processed across three 96-well plates and two 384-well plates, with eight pooled stool QC samples inserted as internal brackets to monitor technical variation.^48,49 Sample preparation followed Sapient's standardised extraction workflow, ensuring consistent protein precipitation and pellet formation,⁵⁰ with minimum input requirements met for >90% of samples. Analyses were run in both positive and negative ion modes using rLC02-TTP02 and rLC03-TTP03 systems.⁵¹ Instrument suitability was confirmed prior to acquisition through repeated pooled plasma injections, verifying mass accuracy (<5 ppm), stable pump pressure (±40%) and MCP voltage (<3000 V).^48,50 Multiple internal standards were spiked into every sample set to assess system performance, matrix effects, sample stability, solvent effects, and column integrity.^52,53 QC thresholds were satisfied, with %CV ≤15% for pooled plasma and ≤20% for pooled stool brackets, and median CVs of 3–5% in external plasma injections. The slightly higher variability for an additive-mode internal standard (CV = 42%) was attributed to expected heterogeneity in stool matrices. Raw LC-MS acquisition produced 57,353 metabolite features, of which 991 were annotated to known biochemical compounds. Peak alignment, retention time correction, and signal normalisation were completed using Sapient's proprietary informatics pipeline.⁵⁴

For downstream analysis, compounds with >50% missingness were excluded. For remaining metabolites, missing values (<50%) were imputed using k-nearest neighbours, followed by probabilistic quotient normalisation and log transformation.⁵⁵ After filtering and harmonisation across Metabolon and Sapient platforms, 960 known compounds were retained for mv-sPLS and pathway-level analyses.

Calculation of BAI and its correlation to cognitive variables

BAI was estimated within a multi-cohort framework (Fig. 1B). First, a Bayesian ridge regression model was trained in the discovery cohort to predict chronological age using resting-state FC. Features were selected based on a significant correlation with age (p < 0.05) within the training set. Model performance was evaluated using ten-fold cross-validation. To assess generalisability, models trained in each cross-validation fold were applied to the replication and independent cohorts. Predicted brain age estimates from all folds were then averaged to obtain a robust estimate for each subject in these external datasets. Brain-age-predictive FCs were defined as connections with non-zero weights identified across all Bayesian ridge models.

To assess measurement reliability, we analysed 40 participants from the discovery cohort who underwent two resting-state fMRI runs acquired on the same day. The trained Bayesian ridge model was applied independently to each run, and brain age estimates were compared across runs.

BAI was defined as the bias-corrected residual between predicted brain age and chronological age.³ Positive BAI values indicate a brain age older than expected for an individual's chronological age, whereas negative BAI values indicate a brain age younger than expected relative to age-matched peers. Specifically, we trained the bias-correction model in the discovery cohort and applied it separately to the discovery, replication, and independent cohorts. Brain-age-predictive FCs were subsequently correlated with BAI to identify BAI-related functional signatures (Fig. 1B). To aid interpretation, node strength was computed as the sum of correlation coefficients associated with each brain region in the replication and independent cohorts. Associations between BAI and cognitive or affective phenotypes were then evaluated across cohorts to assess reproducibility.

To examine potential confounding by head motion, we computed correlations between mean framewise displacement (FD) and both predicted brain age and BAI in the discovery cohort. No significant associations were observed (Fig. 2), suggesting that BAI estimates were not substantially influenced by head motion.

Fig. 2.

Pearson correlations between mean framewise displacement and both predicted brain age (A) and BAI (B) in the discovery cohort. No significant associations were observed between mean FD and either predicted brain age (r = 0.06, p = 0.12, n = 674) or BAI (r = 0.01, p = 0.71, n = 674).

Association between BAI and metagenomic, metabolomic profiles

To identify gut-derived biological signatures associated with BAI, multi-view sparse partial least squares (mv-sPLS) was applied to 182 of 344 UCLA-Omics participants with complete BAI, metagenomic, and metabolomic data modalities in the independent cohort.⁵⁶ L1 penalties for the metagenomic and metabolomic views were optimised using five-fold cross-validation. We focused on the first mv-sPLS component because it typically captures the dominant shared covariance structure and exhibits the greatest stability. For each parameter combination, latent scores were generated in held-out folds and concatenated across all validation folds to obtain pooled out-of-fold predictions. Statistical significance of the identified association was further evaluated using a permutation framework.

To elucidate the biological relevance of the identified multi-omics features, pathway enrichment analysis was performed using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database on features with non-zero weights derived from the mv-sPLS model. KEGG is a curated database that links genes and metabolites to known biological pathways, enabling interpretation of their potential physiological functions.⁵⁷ Metabolite features were annotated with KEGG compound IDs, and metagenomic features were mapped to KEGG orthologous groups (KOs).

Metabolomic features with positive and negative mv-sPLS weights were analysed separately to identify pathways associated with opposite ends of the BAI-related latent dimension. In contrast, because the majority of non-zero metagenomic weights were positive, only positively weighted metagenomic features were subjected to enrichment analysis. Enrichment significance was evaluated using hypergeometric tests.

Ethics

In this study, all participants in the UCLA-CNP, ABIDE I/II, and SRPBS datasets provided written informed consent under the corresponding Institutional Review Boards (IRBs)-approved protocols,12, 13, 14, 15, 16 in accordance with the 2024 Declaration of Helsinki. Participants in the UCLA-Omics dataset provided written informed consent under the corresponding IRB-approved protocols (numbers: 20-002326, 24-0113, 22-000796, 23-001726), in accordance with the 2024 Declaration of Helsinki.

Statistics

Model performance for brain age prediction was evaluated using Pearson correlation coefficients (r) and coefficients of determination (R²) between predicted and chronological age. Within-subject reliability of rsFC-derived brain age estimates was assessed using Pearson's and Spearman's correlations. Associations among BAI, functional connectivity measures, head motion, and cognitive or affective phenotypes were evaluated using Pearson correlation analyses. For identification of BAI-related functional signatures, correlations between brain-age-predictive FCs and BAI were computed and corrected for multiple comparisons using the false discovery rate (FDR) procedure. For mv-sPLS analyses, model performance was evaluated using cross-validated Pearson correlation coefficients, R², and root mean square error (RMSE), and statistical significance was assessed using permutation testing (1000 iterations). Pathway enrichment analyses were performed using hypergeometric tests. Unless otherwise specified, all tests were two-sided and statistical significance was defined as p < 0.05. For analyses involving multiple comparisons, FDR-corrected q < 0.05 was considered statistically significant.

Role of funders

The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

Results

A total of 1005 healthy participants from the combined ABIDE I/II and SRPBS datasets were randomly divided into a discovery cohort (n = 674; 39% female; median age 30 years [IQR 23–42]) and a primary replication cohort (n = 331) using an approximately 7:3 split. An additional cohort, UCLA-CNP (n = 113; 46% female; median age 29 years [IQR 25–39]), was included for external validation and phenotypic analyses. For analyses involving cognitive and affective measures, the primary replication cohort and UCLA-CNP were combined to form the replication cohort (n = 444; 39% female; median age 30 years [IQR 23–41]). Detailed demographic information is provided in Tables 1 and 2.

Table 1.

Demographic information of the cohorts used in this study.

	Discovery	Replication	Independent
Sex
Male	412	272	32
Female	262	172	312
Handedness
Right	588	396	0
Left	15	15	0
Mixed	6	1	0
Unknown	65	32	344
Age
18–30	337	218	224
31–42	151	121	90
43–54	133	76	27
>54	53	29	3
Median, IQR	30, [23, 42]	30, [23, 41]	26, [21, 32]

We included healthy younger adults aged 18–65 years in each cohort.

Final sample sizes were n = 674 for the discovery cohort, n = 444 for the replication cohort, and n = 344 for the independent cohort.

Table 2.

Demographic information of the dataset used in this study.

	UCLA-CNP	ABIDE I/II	SRPBS	UCLA-Omics
Sex
Male	61	258	365	32
Female	52	57	325	312
Handedness
Right	113	218	653	0
Left	0	8	22	0
Mixed	0	7	0	0
Unknown	0	82	15	344
Age
18–30	59	242	254	224
31–42	33	50	189	90
43–54	21	20	168	27
>54	0	3	79	3
Median, IQR	29, [25, 39]	24, [21, 29]	34, [24, 46]	26, [21, 32]
Race
White	71	0	0	32
Asian	0	0	690	82
Black	0	0	0	12
Hispanic	42	0	0	202
Hawaiian	0	0	0	0
Multiple	0	0	0	15
Unknown	0	315	0	1

We included healthy younger adults aged 18–65 years in each dataset.

In the discovery cohort (N = 674), we calculated resting-state FC using the 100-region Schaefer atlas and trained Bayesian ridge regression models to predict chronological age from whole-brain FC features. Ten-fold cross-validation showed robust performance (Fig. 3A, R² = 0.31, r = 0.59, Mean Absolute Error (MAE) = 7.97). This level of performance is consistent with prior FC-based brain age models.¹⁸ The full and top 50 brain-age-predictive maps, averaged from model weights across the cross-validated folds (Figure S1, Table S1), indicated the strongest predictive connections involved intrinsic connectivity within the default mode network (DMN, left-right middle frontal gyri, weight = 0.40) and within the somatomotor network (SMN, between right precentral and postcentral gyri, weight = −0.44). Node-strength analyses further highlighted the supramarginal gyrus, supplementary motor area, bilateral middle cingulate, praecuneus, and superior medial frontal cortex as major contributors.

Fig. 3.

Brain signature of BAI in the discovery cohort. BAI was defined as the bias-corrected residual (predicted brain age minus chronological age) from connectivity-based age-prediction models. (A) Brain age model performance under ten-fold cross validation (R² = 0.31, Pearson's r = 0.59, p < 0.0001 based on the one-sided test against the alternative hypothesis that r >0, n = 674). (B) Pearson's correlation between BAI and the predictive functional connections (non-zero weights in the Bayesian prediction model) comprising the model is shown, where significance survived after false discovery rate (FDR) correction (q < 0.05). Network abbreviations defined from Yeo's seven canonical function networks: VIS, visual; SMN, somatomotor; DAN, dorsal attention; VAN, ventral attention network; LIM, limbic; FPC, frontoparietal control; DMN, default mode. (C) Node strength mapped onto the cortical surface, computed as the sum of signed correlation coefficients (positive and negative correlations calculated separately) of the FCs incident to each region.

Next, we computed the corrected BAI, where positive BAI values indicate a brain age older than expected for an individual's chronological age, whereas negative BAI values indicate a brain age younger than expected relative to chronological-age-matched peers. We observed that corrected BAI is not correlated with chronological age (Figure S2A). In addition, a within-subject stability analysis using repeated rsFC runs from 40 participants in the discovery cohort demonstrated strong consistency of brain age predictions across runs (Pearson's r = 0.82, Spearman's ρ = 0.81; Figure S3), supporting the reliability of the rsFC-derived brain age estimates. Next, correlation analysis between predictive FCs and BAI (FDR q < 0.05, Fig. 3B, top 40 correlated FCs in Table S2, and node strength in Fig. 3C) showed that higher BAI was linked to stronger connectivity in bilateral praecuneus, mid/posterior cingulate cortices, and inferior parietal lobule within DMN and frontoparietal control network (FPC). Negative associations were concentrated in the supplementary motor area, posterior insula, temporal pole, and superior medial frontal cortex spanning ventral attention network (VAN), limbic (LIM), SMN, and DMN networks. The network-level feature importance, defined as the average absolute feature importance within a network and between the network and all other networks from the BAI-related brain signatures, further indicated that BAI association was mainly driven by between-network connections involving the DMN, VIS, FPC, and VAN (Figure S4).

Higher BAI was significantly associated with poorer cognition and greater depressive symptoms. In the discovery cohort (Table 3), higher BAI correlated with greater depressive symptoms (BDI, r = 0.12), lower Performance IQ (r = −0.22), and worse working memory (BACS, r = −0.27). These effects remained after controlling for chronological age (Figure S3B–D). BAI-related brain signatures weighted network-level FCs involving DMN, FPC, and VIS were positively associated with depression and negatively associated with IQ and BACS performance (Figure S5), supporting a link between large-scale network dysregulation and cognitive-affective vulnerability. To further evaluate the added value of BAI, conventional within- and between-network (Yeo's seven canonical networks) functional connectivity summaries showed no significant associations with these phenotypes after FDR correction, whereas BAI consistently explained additional variance in depressive symptoms, IQ, and working memory beyond conventional network measures (ΔR² = 0.01–0.09; Figure S6).

Table 3.

Correlation between cognitive, affective function and BAI measured by Pearson correlation in discovery cohort.

Variables	Median	[Q25, Q75]	N	R	q
BDI	6	[3, 10]	304	0.12	0.038
FIQ	113	[106, 122]	215	−0.13	0.078
VIQ	115	[108, 123]	131	−0.12	0.179
PIQ	112	[105, 121]	150	−0.22	0.018
BACS (verbal memory)	52	[45, 58]	99	−0.16	0.238
BACS (working memory)	23	[20, 25]	99	−0.27	0.037
BACS (motor speed)	80	[71, 90]	99	−0.06	0.577
BACS (verbal fluency)	51	[44, 56]	99	−0.13	0.312
BACS (executive function)	72	[65, 79]	99	−0.06	0.578
BACS (processing speed)	19	[18, 21]	99	−0.23	0.069

P-values were FDR-adjusted across all items from each measure (e.g., correction of six multi-comparisons for BACS).

Significance q values were bolded.

Q25: first quartile, Q75: third quartile.

These findings replicated in both the replication cohort (N = 444) and the independent cohort (N = 344). Predicted brain age remained strongly associated with chronological age (Fig. 4A and D: replication: r = 0.56, R² = 0.28, MAE = 9.03; independent: r = 0.50, R² = 0.23, MAE = 6.34). Using the same correlation analysis, the spatial organisation of FC-BAI associations reproduced the discovery findings. Positive associations clustered in connectivity involving the bilateral praecuneus, posterior cingulate, and inferior parietal lobule, whereas negative associations were mainly localised to connectivity involving the bilateral supplementary motor area, posterior insula, temporal pole, and superior medial frontal cortex (Fig. 4B and C, Fig. 3E and F). Finally, we quantified signature similarity by Spearman correlation between the discovery FC pattern (Fig. 3B) and those from the replication (Fig. 4B) and independent (Fig. 4E) cohorts, yielding ρ = 0.89 and ρ = 0.57, respectively. These spatial patterns closely recapitulate the discovery results, supporting the reproducibility of the BAI-related brain signatures.

Fig. 4.

Reproducibility of BAI-related brain signatures. (A, D) Predicted brain age versus chronological age in the replication and independent cohorts (n = 444, n = 344). (B, E) Pearson's correlations between BAI and predictive functional connections (non-zero weights in the Bayesian prediction model) in the replication and independent cohorts. Only FCs surviving FDR correction (q < 0.05) are displayed. (C, F) Regional node strength in the replication and independent cohorts, computed as the sum of signed correlation coefficients (positive and negative correlations calculated separately) for connections incident to each region, mapping onto the cortical surface.

We next tested whether the cognitive and affective associations with BAI replicated across the replication and independent cohorts. Higher BAI was again related to poorer memory and executive function (Table 4, e.g., CVLT recall, r = −0.28, BACS working memory, r = −0.34) and greater depressive symptoms (BDI, r = 0.20, HADS depression, r = 0.15), overall reproducing the discovery-cohort pattern.

Table 4.

Correlation between cognitive, affective function and BAI measured by Pearson correlation in replication and independent cohorts.

Cohort	Variables	Median	[Q25, Q75]	N	R	q
Replication	BDI	6	[3, 9]	150	0.20	0.016
Replication	FIQ	114	[106, 120]	103	−0.24	0.016
Replication	VIQ	113	[106, 120]	59	−0.18	0.172
Replication	PIQ	112	[106, 119]	74	−0.27	0.019
Replication	BACS (verbal memory)	52	[49, 59]	43	−0.01	0.974
Replication	BACS (working memory)	24	[22, 25]	43	−0.34	0.048
Replication	BACS (motor speed)	82	[72, 90]	43	−0.38	0.048
Replication	BACS (verbal fluency)	53	[46, 60]	43	−0.09	0.647
Replication	BACS (executive function)	72	[63, 80]	43	−0.36	0.048
Replication	BACS (processing speed)	19	[17, 20]	43	−0.16	0.440
Replication	CVLT (total correct)	57	[51, 62]	113	−0.21	0.073
Replication	CVLT (long-delay recall)	0.5	[0, 1]	113	−0.28	0.016
Replication	CVLT (free-recall intrusions)	0	[−0.5, 0.75]	107	0.17	0.145
Replication	CVLT (learning slope)	0.5	[−0.5, 1]	107	−0.08	0.402
Replication	CVLT (recognition discriminability)	0.5	[0, 1]	113	−0.11	0.320
Independent	HADS (anxiety)	8	[4, 11]	336	0.09	0.119
Independent	HADS (depression)	3	[2, 6]	336	0.15	0.013
Independent	TMT-A	19	[16, 24]	72	0.19	0.107
Independent	TMT-B	51	[40, 63]	72	0.28	0.032
Independent	SI (social disconnectedness)	20	[15, 25]	225	0.17	0.040
Independent	SI (lack of social support)	7	[5, 9]	225	0.04	0.758
Independent	SI (perceived loneliness)	5	[4, 6]	225	−0.07	0.758

p-values were FDR-adjusted across all items from each measure.

q ≤ 0.05 (corrected 15, 7 times for replication and independent cohorts).

Significance q values were bolded.

Q25: first quartile, Q75: third quartile.

To examine gut-derived contributors to brain ageing, we applied multi-view sparse partial least squares (mv-sPLS) to metagenomic, stool metabolomic, and BAI data from the UCLA-Omics cohort. The optimal model (L1 for metagenomic = 0.03, L1 for metabolomic = 0.08) showed a modest but significant cross-validated association between the first latent component and BAI (r = 0.144, R² = 0.021, p = 0.04; RMSE = 3.05; Figure S7), which exceeded chance expectations in permutation testing (p_permute = 0.02). Model performance was stable across folds (r = 0.122 ± 0.024, R² = 0.03 ± 0.01, RMSE = 3.07 ± 0.06).^58,59

Given that the majority of top-weighted microbial features were positively associated with BAI, our interpretation focused on these taxa and their related biological pathways. Microbial taxa positively associated with higher BAI included Clostridium sp. CAG:508 (weight = 0.185), Clostridiales bacterium UBA7200 (weight = 0.167), Collinsella vaginalis, Christensenella minuta, and Oscillibacter species (top 40 weights in Fig. 5A). KEGG enrichment showed strong involvement in biosynthesis of secondary metabolites (q < 0.0001, hypergeometric test), amino acid biosynthesis (q < 0.0001, hypergeometric test), general metabolism, histidine metabolism, and peptidoglycan biosynthesis, suggesting that microbial features associated with higher BAI are enriched for pathways related to neuroimmune and metabolic processes (Fig. 5B).

Fig. 5.

Microbial taxa, metabolomic molecules, and related pathways associated with BAI. Mv-sPLS was applied to integrate microbial metagenomic sequences and stool metabolite concentrations in relation to BAI (n = 182). Microbial features with non-zero mv-sPLS weights were further analysed through KEGG pathway enrichment to identify biologically relevant pathways. (A) Top 40 microbial taxa with the highest positive mv-sPLS weights associated with BAI. (B) Top 15 KEGG pathways enriched among microbial features contributing to BAI. (C) The top 20 metabolites with the highest positive (red) and negative (blue) mv-sPLS weights associated with BAI. (D) Top 15 KEGG pathways enriched among negatively associated metabolites. (E) Top 15 KEGG pathways enriched among positively associated metabolites. The p-values were FDR-adjusted across the pathways.

Metabolomic analyses showed that negatively weighted metabolites such as estetrol (weight = −0.240), histidine (weight = −0.186), norvaline, and asymmetric/symmetric dimethylarginine were associated with lower BAI-related latent scores, whereas positively weighted metabolites including Cer(36:1; O3) (weight = 0.235), PC(33:1) (weight = 0.186), nonanedioate, and 24-hydroxycholesterol were associated with higher BAI-related latent scores (Fig. 5C). Negatively associated metabolites were enriched in amino acid metabolism (q < 0.01, hypergeometric test), neurotransmitter pathways, GABAergic synapse, and metabolic homoeostasis (Fig. 5D), whereas metabolites with positive mv-sPLS weights were enriched in fatty acid biosynthesis (Fig. 5E, q = 0.02, hypergeometric test), cutin, suberine and wax biosynthesis (q = 0.02, hypergeometric test).

Next, we performed Spearman correlation analysis between BAI-associated FCs (Fig. 3B) and BAI-associated microbial metagenomic and metabolomic features identified by mv-sPLS to examine direct brain-gut interactions in brain ageing. We also incorporated cognitive and affective phenotypic measures to assess their clinical relevance. The top cross-modal correlations (Figure S8, p < 0.00005, Spearman's correlation) revealed that Ruminococcaceae bacterium UBA5891 and Clostridiales bacterium UBA7200 were positively associated with 12-deoxywithastramonolide, which in turn was positively associated with FCs involving DMN, FPC, and VIS. Although the phenotypic measures were not directly linked to metagenomic or metabolomic features, they showed associations with FCs in the VAN and FPC networks, suggesting that clinical symptoms may be primarily connected to the brain network component of the brain-gut-microbiome system. Further network-level summarisation of significant FC associations (Table S3) showed that cross-modal effects were mainly concentrated in networks including VIS, DMN, FPC, and VAN.

Discussion

Our findings demonstrate that a connectivity-derived BAI captures meaningful variation in early brain ageing and links it to both cognition and gut-derived biology. Using a large discovery cohort, we built and validated an fMRI-based age-prediction model and derived BAI. BAI was anchored in networks centred on the mid/posterior cingulate, praecuneus, and superior medial frontal cortex, and higher BAI related to poorer cognitive performance, especially lower performance IQ and working memory/executive scores, and greater depressive symptoms. These associations replicated across two additional cohorts. Extending beyond the brain, multi-view integration of metagenomics and stool metabolomics identified convergent microbial and molecular correlates of BAI, including taxa such as Clostridium, Collinsella, and Butyricimonas paravirosa, and metabolites spanning ceramides, estetrol, 24-hydroxycholesterol, and sphingomyelin. Together, BAI links large-scale brain networks to gut-metabolic pathways in young and mid-life adults, highlighting potential biological pathways that may be relevant for future investigation.

At the brain-network level, elevated BAI was characterised by stronger coupling among the praecuneus, mid/posterior cingulate, and inferior parietal lobule within the default mode and frontoparietal control networks (DMN-FPC), together with weaker connectivity in the VAN, particularly involving the supplementary motor area and posterior insula. This configuration was associated with lower IQ, weaker working memory and executive performance, and greater depressive burden. Reduced superior medial frontal connectivity suggested impaired top-down control,⁶⁰ while weaker supplementary motor and posterior insula coupling pointed to psychomotor inefficiency and disrupted interoceptive-affective processing.⁶¹ These patterns support models in which early brain ageing shifts network balance towards internally focused processing and away from flexible cognitive control, highlighting DMN-FPC-VAN interactions as a potential biomarker linked with cognitive and affective vulnerability.

Extending beyond the brain, multi-view integration of metagenomics and stool metabolomics identified convergent microbial and molecular correlates of BAI. Microbiota profiling revealed that higher BAI is associated with an ecological configuration featuring Clostridium, Butyricimonas, Oscillibacter, and Collinsella. These taxa are central to short-chain fatty acid (SCFA) synthesis and bile acid metabolism—pathways with established immunomodulatory and neuroactive properties. While Collinsella is frequently linked to pro-inflammatory signalling, its co-occurrence with core homoeostatic commensals like Christensenella and Faecalibacterium suggests that in healthy young and mid-life adults, these relationships represent metabolically contingent ecological profiles.^62,63 These taxonomic signatures converge with functional enrichment in peptidoglycan remodelling and protein-export processes, mirroring systemic biological hallmarks of ageing such as disrupted proteostasis and heightened neuroimmune activity.⁶⁴

Metabolomic profiling revealed complementary signatures. Lower BAI was associated with metabolites enriched in amino acid, nitrogen, and neurotransmitter pathways, including GABAergic signalling and glycine/serine/threonine metabolism, which support synaptic plasticity and neuroimmune homoeostasis. Estetrol showed one of the strongest inverse associations, suggesting a possible brain-ageing-protective role of reproductive-neuroendocrine pathways.⁶⁵ In contrast, higher BAI was linked to sphingolipids, ceramides, and sterol derivatives such as 24-hydroxycholesterol, which are associated with lipid dysregulation, oxidative stress, inflammation, and neuronal cholesterol turnover. Together, these findings suggest that lower and higher BAI are associated with distinct metabolic profiles, with the former enriched for neurotransmitter and amino-acid-related pathways and the latter enriched for lipid-related pathways implicated in neuroimmune and metabolic processes.⁶⁶

Our integrated brain-gut-microbiome analysis revealed a potential microbe-metabolite-brain network axis. Taxa within the Ruminococcaceae and Clostridiales families positively correlated with 12-deoxywithastramonolide, a compound structurally similar to the neuroprotective withanolides found in Withania somnifera (Ashwagandha).⁶⁷ These findings align with emerging evidence that microbiome-derived metabolites mediate links between gut composition, systemic inflammation, and neural decline.⁶⁸ For example, a recent AD-continuum study identified a microbiota → metabolites → neuroimaging → cognition mediation pathway,⁶⁹ while a multi-omics study in major depressive disorder showed that Alistipes-related metabolites were associated with multimodal brain imaging measures, including spontaneous brain activity, grey matter morphology, and white matter integrity.⁷⁰ Then, increases in 12-deoxywithastramonolide were linked to increases in FCs spanning DMN, FPC, and visual networks, large-scale systems known to undergo early connectivity changes in ageing.⁷¹ This tripartite association suggests a potential microbe-metabolite-FC axis, whereby gut-derived biochemical pathways may interface with and potentially modulate large-scale brain network communication.

Several limitations should be considered when interpreting the present findings. First, the predominantly female microbiome subset and its relatively small size necessitate replication in larger, sex-balanced cohorts. Second, the mv-sPLS analysis was performed in a relatively small subset of the independent cohort. Larger cohorts will be needed in future studies to validate these identified microbial and metabolic signatures. Third, our cross-sectional design precludes causal or temporal inferences. Longitudinal studies will be necessary to determine whether individual differences in rsFC-derived BAI reflect stable trait-like characteristics or transient brain states, to establish the directionality of brain-gut-metabolite relationships, and to assess whether cross-sectional deviations in BAI correspond to differences in brain ageing trajectories over time. Fourth, although shotgun metagenomics enables species-level classification, it does not resolve strain-level heterogeneity within taxa such as Clostridium and Collinsella; strain-resolved approaches will be necessary to identify specific variants underlying these associations. Fifth, although the associations between BAI and cognitive and affective phenotypes were consistently replicated, their effect sizes were modest, suggesting that BAI represents one of multiple factors contributing to individual differences in these complex phenotypes. Finally, extending analyses to older adults will clarify whether elevated BAI consistently signals advanced biological ageing across the lifespan.

This study identifies a reproducible rsFC-derived brain ageing index (BAI) associated with cognitive, affective, and large-scale brain network variation in young and mid-life adults. By integrating gut-derived metagenomic and metabolomic features, we reveal a biologically meaningful brain-gut-microbiome signature linked to individual differences in brain ageing. These findings suggest that BAI captures variation relevant to brain health and highlight the potential role of gut-metabolome interactions in brain-age-related variability during young and mid-life adulthood.

Contributors

K.Z. conceptualised and designed the work, wrote the code, analysed and interpreted the data, and drafted and revised the manuscript. G.A.V. provided guidance on the microbiome analyses, contributed to drafting the relevant sections, and revised the manuscript. J.S.L. refined the design of the work, provided consultation on statistical analyses and methodology, interpreted the data, and revised the manuscript. A.V. and M.D. assisted with the acquisition of the data. K.V., E.A.M., C.G., D.D.R., C.S., R.P.R., J.F.C., and R.K.D. interpreted the data and revised the manuscript. P.V. and M.W. preprocessed the fMRI data for the independent (UCLA-Omics) cohort. Y.Z. preprocessed the fMRI data for the discovery and replication cohorts and provided revision of the results. A.C. conceptualised and designed the work, acquired data, sought funding, oversaw the analysis and interpretation of the data, and revised the manuscript. All authors read and approved the final version of the manuscript. All authors accessed and verified the data, and K.Z. and A.C. were responsible for the decision to submit the manuscript.

Data sharing statement

All raw image data are openly available. The fMRI sources that are publicly available for direct download are the following: UCLA-CNP (USA) (https://openneuro.org/datasets/ds000030/versions/1.0.0), ABIDE (USA) (https://fcon_1000.projects.nitrc.org/indi/abide/), and SRPBS (Japan) (https://bicr-resource.atr.jp/srpbsopen/). The UCLA-Omics dataset requires IRB approval and a formal data-sharing agreement. Interested researchers should contact Arpana Church at ArpanaChurch@mednet.ucla.edu. The main code for running analysis can be found at https://github.com/alienx-zero/Brain-Gut-Crosstalk-Shapes-Early-Brain-Aging-in-Young-and-Mid-Life-Adults.

Declaration of interests

A.C. is a scientific advisor and consultant to Yamaha. R.K.D. is an inventor on a series of patents on the application of metabolomics for the diagnosis and treatment of central nervous system diseases and holds equity in Metabolon Inc., Chymia LLC, and Metabosensor. J.F.C. had contracts from the Kerry group and the Reckitt company. All other authors declare no competing interests.