Microbes on the Mind: Multi-modal Neuroimaging Reveals Gut-Brain Axes in Infants

Technical Summary: Microbes on the Mind: Multi-modal Neuroimaging Reveals Gut-Brain Axes in Infants

Problem Statement
The gut microbiome undergoes rapid maturation during the first three years of life, a period coinciding with critical brain development. While the microbiome-gut-brain axis is known to involve reciprocal signaling via immune, neuroendocrine, and vagal pathways, the specific relationships between early-life microbial composition and concurrent brain development remain poorly defined. Previous studies have been limited by small sample sizes, low taxonomic resolution (e.g., 16S rRNA), restricted metabolic characterization, and incomplete neuroimaging coverage. This study aims to characterize the multivariate covariance between the gut microbiome and multimodal brain imaging phenotypes in a large, typically developing infant cohort.

Methodology
The study utilized an accelerated longitudinal cohort from the UNC/UMN Baby Connectome Project, comprising 223 typically developing children aged 0–3 years. The dataset included 489 paired sessions of fecal shotgun metagenomics and multimodal MRI (structural, diffusion, and resting-state functional MRI) collected within ±30 days of each other.

• Data Processing:

  • Microbiome: Shotgun metagenomics were processed using MetaPhlAn 3.0.2 and HUMAnN to generate taxonomic profiles (species, genera) and functional profiles (554 MetaCyc pathways, 44 gut-brain modules [GBMs], 99 gut metabolic modules [GMMs]).
  • Neuroimaging: Structural MRI (T1w/T2w) was processed with iBEAT V2.0 to derive cortical morphology (volume, surface area, thickness) across 148 regions. Diffusion MRI provided microstructural metrics (FA, MD, MO) across 34 tracts. Resting-state fMRI (acquired during natural sleep) was processed to extract functional connectivity within seven canonical networks.
  • Preprocessing: Age-related trends were removed from neuroimaging phenotypes using generalized additive mixed models (GAMMs) with participant random intercepts. Microbiome features were centered-log-ratio (CLR) transformed.

• Statistical Analysis:

  • Microbiome-Wide Association Study (MWAS): Linear mixed models (MaAsLin2) tested associations between six microbiome feature sets and neuroimaging measures, adjusting for site, age, sex, delivery mode, feeding practice, and sequencing depth.
  • Multivariate Covariance: Ridge-regularized Canonical Correlation Analysis (rCCA) was applied to 107 children with complete data to identify principal covariance patterns between microbial and neuroimaging features.
  • Longitudinal Analysis: Within-participant developmental slopes were estimated for microbial and neuroimaging features. Autoregressive cross-lag models tested temporal ordering between canonical scores over 1–12 month windows.
  • Validation: Sparse canonical Partial Least Squares (PLS) was used to determine if the rCCA patterns were recoverable under a different multivariate objective. Participant-level resampling assessed stability.

Key Results

1. Developmental Trajectories:

   • Microbiome: Alpha diversity increased with age. Beta diversity was most strongly correlated with age ($R^2 = 0.107$), followed by maternal education, feeding practice, and delivery mode. Bifidobacterium species dominated early profiles, shifting toward Bacteroidetes and Firmicutes after the first year.
   • Neuroimaging: Structural measures rose rapidly in the first year. Functional networks showed distinct trajectories: sensorimotor connectivity declined, visual connectivity peaked early then declined, and higher-order networks (default mode, executive control) strengthened gradually.

2. Feature-Level Associations (MWAS):
   Five associations met the strict significance threshold ($q < 0.05$):

   • Total brain volume was positively associated with purine deoxyribonucleoside degradation and purine nucleotide salvage pathways.
   • Actinomyces sp. HMSC035G02 abundance was inversely associated with cortical area and volume PC3 scores.
   • Gordonibacter pamelaeae abundance was positively associated with executive-control functional connectivity.

3. Multivariate Covariance Axes (rCCA):
   Two principal modes were identified:

   • Structural–Metabolic Axis (CCA1): Linked cortical morphology (specifically PC3, representing an auditory–visual regional contrast: larger auditory/perisylvian regions vs. smaller visual regions) with microbial metabolic functions (melibiose and fructan degradation).
   • Functional–Taxonomic Axis (CCA2): Linked species composition (e.g., Bifidobacterium pseudocatenulatum) and GBMs with functional connectivity, particularly in the ventral-attention network.

4. Longitudinal Dynamics:

   • Slope Covariation: Within-participant growth rates of canonical scores were not significantly correlated.
   • Temporal Ordering: Autoregressive cross-lag models revealed that neuroimaging CCA1 scores predicted microbiome CCA1 scores 1–3 months ($\beta = 0.35$) and 3–6 months ($\beta = 0.31$) later. No reverse-direction associations survived correction.

5. Methodological Validation:
   Sparse PLS analysis reproduced the structural contrast of CCA1 and the functional network pattern of CCA2. However, held-out test correlations were modest, indicating that while the patterns are robust within the sample, generalization to external splits is limited.

Significance and Claims
The authors claim that this study provides the first multimodal, longitudinal characterization of gut-brain axes in typically developing infants using shotgun metagenomics and comprehensive neuroimaging. The findings demonstrate:

• Cross-domain covariance: Specific patterns of microbial metabolism and taxonomy covary with distinct aspects of brain structure (auditory-visual contrast) and function (ventral-attention connectivity).
• Temporal precedence: In the measured windows, neuroimaging phenotypes (specifically the structural-metabolic axis) predict subsequent microbiome states, suggesting a temporal ordering where brain development may influence microbial composition, or that shared unmeasured factors drive this sequence.
• Methodological rigor: The study employs rigorous multivariate techniques (rCCA, sparse PLS) and longitudinal modeling to move beyond simple univariate correlations.

The paper explicitly states that these results indicate coordinated variation and temporal ordering but do not establish causality. The authors emphasize that the observational design, lack of time-varying confounders (e.g., illness, antibiotics), and the nature of metagenomic data (genetic potential rather than active flux) preclude mechanistic conclusions. The identified associations serve as a foundation for future studies incorporating metabolomics and experimental models to infer directional mechanisms.