Constructing the human brain metabolic connectome with MR spectroscopic imaging reveals cerebral biochemical organization

Lucchetti, Federico; Céléreau, Edgar; Steullet, Pascal; Alemán-Gómez, Yasser; Hagmann, Patric; Klauser, Antoine; Klauser, Paul

doi:10.1038/s41467-025-66124-w

Download PDF

Article
Open access
Published: 22 December 2025

Constructing the human brain metabolic connectome with MR spectroscopic imaging reveals cerebral biochemical organization

Nature Communications volume 16, Article number: 11344 (2025) Cite this article

5754 Accesses
23 Altmetric
Metrics details

Subjects

Abstract

Network science has mapped brain structural and functional organization, yet capturing metabolic contrast remains a major gap in connectomics. Using fast, high-resolution 3D whole-brain proton magnetic resonance spectroscopic imaging (¹H-MRSI), we derive a within-subject metabolic connectome in 51 healthy subjects, defined as pairwise correlations among five metabolites (tCr, tNAA, Glx, Ins, Cho) across gray-matter parcels. Results show stability, consistency and replicability, including validation in an independent sample (N = 13) scanned at a different site. A dimensionality reduction analysis shows that the leading metabolic similarity mode forms a continuous caudal-to-rostral gradient across gray-matter regions. We show that this progression is reflective of a balance between local metabolic homogeneity and global metabolic diversity, and can be summarized by a principal path through the metabolic network. While the most metabolically active regions overlap with structural hubs, overall metabolic similarity aligns weakly with tractography-based structural connectivity but more closely with cytoarchitectonic similarity and gene co-expression matrices. These findings introduce the metabolic similarity gradient as a signature of the brain’s biochemical organization and position MRSI as a biologically grounded dimension for connectomics in health and disease.

Brain metabolic-functional (de)coupling from health to glioma dysfunction

Article Open access 04 December 2025

Contributions of network structure, chemoarchitecture and diagnostic categories to transitions between cognitive topographies

Article Open access 05 August 2024

Relationships between local metabolic activity and distributed functional connectivity in major depressive disorder

Article Open access 01 December 2025

Introduction

Since their inception in the 1980s, magnetic resonance imaging (MRI) techniques have become central to both clinical and basic neuroscience. MRI offers non-invasive, non-ionizing measurements of the living brain across various ages, with relatively quick acquisition times and moderate costs, making it widely accessible. MRI techniques have not only enhanced our understanding of the brain’s anatomical structure, composition, and function but also enabled the construction of large-scale brain network models. In particular, diffusion MRI maps water diffusion along neuronal axons, revealing white matter tracts and pioneering the structural connectome and the field of connectomics^1,2. BOLD (blood-oxygen-level-dependent) functional MRI introduces a temporal dimension to imaging, enabling functional connectivity to be defined as statistical dependencies–typically correlations–between regional BOLD time series, which index hemodynamic changes used as a proxy for synchronized neural activity³. The field of connectomics has since revealed that, like many other complex systems, both structural and functional brain networks exhibit fundamental organizational principles. The most prominent example is modular architecture, which reflects a balance between local specialization and global integration^4,5,6. This trade-off is a hallmark of healthy brain organization and provides a framework for understanding how alterations in network structure relate to behavior and disease. Individual differences in network organization are linked to cognition and personality⁷, and their disruptions, are increasingly recognized as central to neurological and psychiatric disorders^8,9,10,11,12. Within this framework, network abnormalities often converge on inter-hub connector nodes, which incur high metabolic costs through oxidative glucose metabolism¹³ and are vulnerable to homeostatic disturbances. These hubs thus represent potential points of failure across diverse pathological conditions^6,11. Diffusion MRI and tractography can capture macro-scale changes in the structural connectome, but remain limited in detecting microcircuit-level alterations¹⁴. Functional MRI and dynamic causal modeling provide complementary insights into metabolic activity^15,16,17, yet are constrained by computational demands, model specificity, and interpretative complexity¹⁸. These neuroimaging modalities, along with the network architectures they help reveal, have excelled at characterizing brain connectivity at the macro- to meso-scale. However, they all fundamentally rely on indirect proxies for assessing underlying cellular networks which are deeply shaped and constrained by the brain’s underlying metabolic organization. Therefore, conventional MRI-based techniques have left a significant gap in connectomics research and limited our ability to link observed changes to their neurobiological basis. While ongoing advances in hardware, pulse sequences, and accelerated acquisition hold promise for pushing the spatial resolution of MRI toward finer scales, such improvements alone are unlikely to close this explanatory gap. Ultimately, what remains missing is not simply a matter of spatio-temporal resolution, but of accessing an additional dimension of information which offers more direct insights into the biochemical and physiological processes.

Once the dominant water and lipid signals are suppressed, magnetic resonance spectroscopy (MRS) reveals this hidden dimension—the chemical shift—which enables the separation and quantification of brain metabolites by their unique resonance frequencies. Building on this principle, recent advances in proton magnetic resonance spectroscopic imaging (¹H-MRSI)^{19,20,21,22,23,24,25,26} now allow for whole-brain 3D metabolite mapping at high spatial resolution with substantially reduced acquisition times ( ~ 20 minutes for the 3D compressed-sensing SENSE low-rank ¹H FID-MRSI sequence used in this study). Detailed and scalable analyses of brain metabolism are now facilitated and can be seamlessly integrated into clinical or research MRI protocols. At a magnetic field strength of 3T, the spectrally resolved metabolites include N-acetylaspartate + N-acetylaspartylglutamate (tNAA), creatine + phosphocreatine (tCr), glutamate + glutamine (Glx), phosphocholine + glycerophosphocholine (Cho), and myo-inositol (Ins). Specifically, tNAA may reflect neuronal integrity, tCr is a marker of energy metabolism, Glx is involved in excitatory neurotransmission, Cho is linked to membrane turnover, and Ins is associated with glial activity²⁷.

Given that these metabolites are critical for normal cellular functioning, and in light of previous MRS studies showing how their interplay impacts various neurological and psychiatric disorders^28,29,30,31, it is logical to suppose that these five key compounds must be delicately regulated by homeostatic processes³² to maintain healthy brain function; hence, they participate in a broader neurometabolic network³³. In this work, we propose to leverag high-resolution 3D ¹H-MRSI to construct, a large-scale, within-subject MRI-based metabolic connectome, by mapping the regional similarity among these five metabolites across the gray matter of the entire brain.

We explore the feasibility of constructing an MRSI metabolic connectome via an individual metabolic similarity matrix (MetSiM) using data from a single whole-brain scan ¹H-MRSI in two cohorts of healthy participants. We test five central hypotheses regarding its network structure: (1) its topological characteristics mirror those of naturally occurring networks; it correlates with (2) structural connectivity, (3) cytoarchitectonic similarity, and (4) genetic co-expression patterns; and (5) regions with the highest structural connectivity coincide with those of high metabolic levels.

Results

Metabolic similarity matrices construction

To evaluate the feasibility of a within-subject metabolic similarity matrix (MetSiM), we analyzed ¹H-MRSI data from 51 healthy adolescents in Geneva (i.e., Mindfulteen study³⁴) following the similarity matrix construction pipeline schematically laid out in Fig. 1. After partial-volume and MRI point spread function correction (see Methods MRSI Partial Volume Correction; representative MRSI shown in Supplementary Fig. 9), each subject’s T1-weighted volume was parcellated using the Chimera LFMIHIFS-3 atlas. This yielded 277 cortical, subcortical, and cerebellar GM regions (see Methods Anatomical), which were coregistered to the individual MRSI space and directly applied to the MRSI data, thereby minimizing voxel interpolation and smoothing artifacts. Following an approach similar to that used for extracting morphometric profiles³⁵ (see Methods Estimation of Metabolic Profiles), we derived metabolic profiles from five spectrally resolved ¹H-MRSI compounds (tNAA, tCr, Glx, Cho, Ins) and retained 210 z-score-normalized profiles corresponding to brain regions with reliable ¹H-MRSI coverage (see Supplementary Information Fig. 1).

**Fig. 1: Within-subject metabolic similarity matrix construction.**

The metabolic profiles were augmented via a Monte Carlo-driven uncertainty propagation procedure (see Methods Estimation of Metabolic Profiles), in which each measurement was perturbed K_pert times based on its estimated variance (Cramér-Rao lower bounds issued by the LCModel metabolite quantification maps, shown in Supplementary Fig. 10), which effectively increased sample size per metabolite signal by a factor of K_pert.

Pairwise Spearman correlations between regional metabolic profiles were calculated to create a 210 × 210 MetSiM for each individual. Figure 2a shows an example MetSiM for K_pert = 50, highlighting strong positive correlations clustering along the diagonal, representing super-regional communities (cortex, subcortex, cerebellum). Symmetrical clusters between hemispheres indicate metabolic similarity in corresponding regions, while negative correlations off-diagonally, particularly between distinct super-regions (e.g., cortex and subcortex), reflect metabolic dissimilarity.

**Fig. 2: Individual and Group Average Metabolic Similarity Matrices.**

Stability, consistency and replicability

We evaluated the stability of MetSiM construction by testing two key parameters: the contribution of individual metabolites and the K_pert values. A leave-one-metabolite-out approach assessed the influence of individual metabolites by comparing the original 5-metabolite MetSiM with MetSiM matrices reconstructed after imputing a single metabolite. This was repeated for three K_pert values (1, 50, 100) in all individuals. For each test, pairwise edge strengths from the original and imputed matrices were correlated for each individual’s MetSiM, and these correlations were aggregated across all subjects using Fisher’s z-transformation, with p-values corrected using the Benjamini–Hochberg false discovery rate method. At K_pert = 1, correlations between the original MetSiM and those reconstructed with a single metabolite omitted were low (r = 0.32–0.40, p < 0.05). Increasing K_pert to 50 substantially improved these correlations (tNAA: r = 0.80, Ins: r = 0.96, Cho: r = 0.81, Glx: r = 0.88, tCr: r = 0.96; all p < 0.05), with only marginal gains at K_pert = 100 (r = 0.81–0.97, p < 0.05). To further examine differences in metabolite contributions, an ANOVA on the K_pert = 50 correlations revealed a significant effect (F = 1154, p < 0.001), and post hoc Tukey HSD tests indicated that tNAA, Cho, and Glx contributed significantly more to MetSiM construction than Ins or tCr, whose removal had minimal impact on correlations.

The consistency of MetSiMs was evaluated by comparing individual MetSiMs with a group-level MetSiM, derived by averaging all individual MetSiMs from the Geneva study (Fig. 2b). Variability was quantified by calculating edge-wise correlations between individual- and group-level MetSiMs, Fisher-transformed, and aggregated for overall alignment. Correlations increased from r = 0.51 at K_pert = 1 to r = 0.60 at K_pert = 50, with a minor improvement at K_pert = 100 (r = 0.61), showing that K_pert = 50 sufficiently reduces variability and ensures consistent MetSiM construction. We fixed K_pert = 50 for subsequent analyses. Full results are reported in Supplementary Table 1. Replicability was assessed using a different sample of healthy participants from the Lausanne Psychosis Cohort³⁶ scanned on a different MRI platform. The group-level MetSiMs for the Geneva and Lausanne studies (shown on Fig. 2c) were strongly correlated (r = 0.86, p < 0.001), demonstrating the replicability of the MetSiM construction method across sites (i.e., different sample and scanner).

Metabolic similarity mode and gradients

To show the spatial organization of metabolic similarity and the spatial discrimination between positive and negative MetSiM weights, we analyzed pairwise group-averaged MetSiM weights (N = 51, Geneva study) in relation to inter-regional Euclidean distances, pooling data from both hemispheres (Fig. 3a).

We then binned positive and negative MetSiM values separately in 2 mm intervals, generating density distributions (Fig. 3b). Positive MetSiM values cluster at shorter distance ranges (37 mm), whereas negative MetSiM values peak at longer distance ranges (71 mm) indicating a clear inverse spatial dependence. This pattern suggests that metabolic similarity reflects a spatial gradient of local metabolic coupling and distal metabolic differentiation.

To clearly visualize its spatial embedding, we reduced the 210 × 210 group-averaged MetSiM matrix to a 210 × 1 feature vector using Principal Component Analysis (PCA) for denoising and variance capture, followed by t-distributed Stochastic Neighbor Embedding (t-SNE) to preserve the observed local connectivity patterns in the MetSiM matrix (see Methods Metabolic Similarity Mode; first component accounts for 53% of variance). The resulting 210 × 1 vector–the first Metabolic Similarity Mode (MS mode)–was mapped back onto the anatomical locations of the corresponding brain parcels in the Montreal Neurological Institute (MNI) standard stereotaxic space, enabling 3D visualization of metabolic similarity patterns (Fig. 3c). The MS mode was inverse-mapped onto the five z-score-normalized metabolite level profiles (Fig. 3c; with color scale), allowing each MS mode value to be interpreted as a reduced representation of the five-feature metabolic profile. Common metabolite ratios were mapped onto the MS mode (Supplementary Fig. 11) and are consistent with previously reported regional MRSI ratio patterns. MS mode map generation is robust to cortical parcellation (Supplementary Fig. 3), showing high alignment in terms of normalized mutual information (MI) with alternative schemes: Schaefer197 (MI = 0.65), MIST197 (MI = 0.64), 15 × 15 × 15 mm grid (MI = 0.99), and 10 × 10 × 10 mm grid (MI = 0.99).

This representation reveals two distinct yet co-occurring features: first, a local property characterized by a high degree of local similarity in MS mode values; and second, a global organizational pattern defined by a high degree of diversity in MS mode values across regions.

The first property could result from spatial blurring inherent in MRI acquisition or from the underlying biological organization. Regardless of its origin, we tested to what extent the observed MetSiM structure could be accounted for by spatial autocorrelation alone, using a distance-based random geometric null model (RandGeom) (see Methods Spatial Null Models) which reproduces both the distance dependence and the heterogeneity of variance observed in the empirical MetSiM distribution (Fig. 3b).

We generated 1000 RandGeom models and correlated each with the empirical MetSiM, producing a distribution of correlation coefficients (Supplementary Fig. 4 and Supplementary Table 2). The best RandGeom model (top, Fig. 3d) achieved only r = 0.082 and failed to reproduce key features such as bilateral homotopy or separation of nearby but non-adjacent regions (e.g., cerebellum-occipital cortex), as shown by the MS mode map of the simulated MetSiM (right, Fig. 3d). Enforcing hemispheric symmetry - by fitting to one hemisphere and mirroring across the midline - improved correlations (LH: r = 0.114, RH: r = 0.131). Further constraining connections to gray matter-adjacent regions (RandGeom ⋅ GMAdj) produced the highest correlations (LH+RH: r = 0.133, LH: r = 0.208, RH: r = 0.223), capturing local adjacency patterns but still lacking full bilateral symmetry in the whole brain (LH+RH) model (bottom, Fig. 3d). These results show that spatial proximity alone fails to reproduce key features of the MetSiM structure, such as bilateral homotopy and the separation of nearby but distinct regions. Adding biologically grounded constraints (gray matter adjacency and hemispheric symmetry) improves alignment with the empirical data, indicating that metabolic similarity cannot be reduced to spatial autocorrelation from geometric distance or imaging-induced smoothness, but instead reflects underlying biological organization at the systems level.

The second property corresponds to a global organizational pattern, with the MS mode defining the principal axis of variation in the MetSiM and producing a continuous, widely distributed range of values across the brain. This axis separates cortical from subcortical regions, differentiates within the cortex along a rostro-caudal trajectory, and distinguishes neighboring but functionally distinct areas such as the primary motor cortex (precentral gyrus) and primary somatosensory cortex (postcentral gyrus). The limbic (e.g., cingulate and parahippocampal cortex, thalamus, hippocampus) and paralimbic (e.g., insula) systems, along with related subcortical nuclei (e.g., striatum), exhibit metabolic profiles clearly distinct from the isocortex. Notably, rostro-caudal progression in the MS mode variation is also evident within individual structures such as the cingulate cortex and insula. The cerebellum is the only exception, metabolically aligning with the subcortical-prefrontal axis.

These systems, long recognized as functionally and structurally distinct, are here shown to be differentiated also in neurometabolic terms. This distinction is reflected in the spatial variation of the five metabolic profiles, which drive a monotonically increasing caudal-to-rostral MS mode and give rise to a recognizable gradient that we term the metabolic similarity gradient (MS gradient; schematized in Fig. 4c and formally defined in Methods Metabolic Principal Path Construction, Eq. (3)). Specifically, segregation is largely explained by a reversal in the tNAA-to-Cho ratio, tNAA dominating in the cortical regions and Cho in the subcortical regions, which underpins the observed inversion from predominantly positive MetSiM values (cortical-to-cortical correlations) to negative values (cortical-to-subcortical correlations).

Topology, metabolic hubs and principal paths

We examined metabolic similarity network topology by thresholding individual MetSiMs (see Methods Weighted Matrix Binarization) to generate binary graphs and calculating rich-club coefficients across edge densities (2%, 10%, 18%, and 40%). At all densities, metabolic networks exhibited with rich-club coefficients being statistically significant (p < 0.001) when tested against both a degree-preserving null model and a spatially embedded random geometric model (see Methods Rich-Club and Supplementary Fig. 5). These findings were replicated on the group-averaged MetSiM (binarized MetSiM shown in Fig. 4a), and the resulting rich-club network mapped to the MNI coordinates is shown in Fig. 4b. Rich-club nodes were most concentrated in the occipital lobe and less so in the frontal lobe. Coloring nodes and edges by MS mode-derived values, revealed metabolically distinct and spatially segregated clusters–each containing at least one rich-club node, except the cerebellum.

To illustrate the integrative nature of these seemingly disparate high-similarity nodes, we leveraged the previously observed caudal-to-rostral MS gradient, which gives rise to a unique topological structure, and distilled it through the construction of a discrete analog of a principal curve^37,38. Inspired by the original definition of principal curves as smooth, self-consistent trajectories through high-dimensional data, our approach identifies an optimal path within the metabolic network–constrained by gray-matter adjacency (GM constrained random geometric model aligns more closely with empirical MetSiMs), limited to individual hemispheres (which are metabolically homotopic), spatially aligning with local MS gradient progression and capturing the maximal variation of the MS mode distribution (see Methods Metabolic Principal Path Construction) between the lowest and highest MS mode nodes (see Fig. 4c). The resulting network paths were fitted with a piecewise cubic spline for visualization, 3D-rendered on Fig. 4d and schematically visualized in Fig. 4g.

The principal path, originating at the isthmus cingulate, first extends into the visual pericalcarine region, then traverses the lateral occipital cortex (LOC), continues through the lateral frontal lobe, and terminates in the rostral anterior cingulate cortex (rACC). A subcortical path, structurally disjoint yet metabolically connected via an edge between the rACC and caudate nucleus (interrupted path), links the thalamus, brainstem hubs, and the cerebellum. MS mode values vary smoothly and monotonically along both cortical and subcortical trajectories in each hemisphere (Fig. 4e). To test whether these paths reflect the joint influence of local spatial effects and global metabolic diversity, we compared them to 10,000 random MetSiM matrices (RandGeom), evaluating the normalized MS gradient (${\widetilde{{{\mathscr{G}}}}}_{\mu }$, the sum of L² norms of local MS mode gradients along the path, normalized by the maximum attainable value for the same endpoints) and the normalized entropy ($\widetilde{H}$, path entropy relative to a maximal-entropy Erdős-Rényi graph) (see Methods Metabolic Principal Path Construction). The empirical paths jointly optimized ${\widetilde{{{\mathscr{G}}}}}_{\mu }$ and $\widetilde{H}$, with both occurring more extremely than expected under the spatial null model (joint one-tailed permutation test: LH-CTX ${\widetilde{{{\mathscr{G}}}}}_{\mu }=0.08$, $\widetilde{H}=0.79$, p = 0.0358; RH-CTX 0.09, 0.81, p = 0.0110; LH-SUBC 0.11, 0.70, p = 0.0030; RH-SUBC 0.13, 0.76, p = 0.0005; Fig. 4f). This pattern indicates convergence towards high local metabolic similarity and high global metabolic diversity.

The schematic in Fig. 4h summarizes the spatial organization of metabolic similarity in the brain, as revealed by the principal path analysis. In our data, nearby adjacent gray matter regions tend to be metabolically similar, whereas distant regions are more metabolically diverse. This co-occurrence produces a coherent spatial structure that organizes into a traversing MS gradient, positioning the observed network topology between that of a highly structured lattice (low local variation, multiple gradients) and a random network (high global variability, no gradients). Importantly, these principal paths constitute measurable features of the metabolic network, allowing their properties to be quantitatively assessed and compared in subsequent analyses. The broader implications of this organization are considered in the Discussion.

Metabolic similarity and structural connectivity

To test the hypothesis that MetSiM edges reflect axonal connectivity, we analyzed diffusion spectrum imaging (DSI) data from the Geneva study, reconstructing tractograms and deriving weighted structural connectivity matrices (SC) based on the same Chimera gray matter parcellation LFMIHIFS-3 (see Methods DSI and Structural Connectivity) to allow direct MetSiM and SC edge-wise comparisons (see Fig. 5a). After excluding individuals with poor quality tractograms, 42 SC matrices were retained, with only the nodes covered by both matrices used to produce 42 190 × 190 matrices.

A direct edge-wise Pearson correlation between structural connectivity and metabolic similarity yielded low scores (0.01 to 0.05, p < 0.001). We then deliberately extended the analysis to probe whether metabolic similarity between nodes could depend on indirect structural connections via intermediate nodes. By computing Sⁿ (the n-th power of the SC matrix, see Methods Higher Order Connectivity), we found a peak correlation of 0.1 (p < 0.001) at second-order connections, which declined for higher orders (see Fig. 5b). Thus, despite the weak overall correspondence, metabolically similar nodes exhibit a very modest tendency to converge on the same structural targets.

We also tested whether perturbations of metabolic nodes could be influenced by distant structural nodes using communicability analysis³⁹. Figure 5c shows correlations between the predicted communicability model and the observed metabolic network were consistently weak (r < − 0.022, p < 0.001) in binary and weighted matrices. These findings indicate that the metabolic similarity network is effectively independent of structural connectivity under the tested communicability model.

Lastly, we applied the same principal path extraction procedure to the averaged structural connectivity matrix by first computing its first Uniform Manifold Approximation and Projection (UMAP) component, commonly referred to as the connectopy⁴⁰ (i.e., the spatial map of structural connectivity profiles; see Fig. 5d), and then deriving a path that aligns with its spatial gradient while maximizing the diversity of connectopic values along the trajectory. Despite the presence of smooth connectopic variation across brain lobes (LH-CTX ${\widetilde{{{\mathscr{G}}}}}_{\mu }=0.16$ and RH-CTX ${\widetilde{{{\mathscr{G}}}}}_{\mu }=0.09$), only disjoint paths could be identified, and no trajectory traversing all lobes was found–likely due to the limited diversity of structural connectivity profiles (LH-CTXH = 0.63 and RH-CTX H = 0.64). Cost function optimization results are shown in Supplementary Fig. 7.

Lastly, we tested the hypothesis that structurally central nodes exhibit higher metabolic activity. Spearman correlations between structural degree centrality and metabolic profiles (z-score normalized) (sample correlation shown in Fig. 6a) yielded values ranging from 0.19 to 0.25 across connection densities (1–20%) (see Fig. 6b). Taken together, these results confirm the poor resemblance between structural connectivity and metabolic similarity but establish a link between structural centrality and high ¹H-MRSI metabolite levels.

**Fig. 6: Metabolic levels and structural connectivity.**

Metabolic similarity and cytoarchitecture

To test the hypothesis that metabolic networks align with cytoarchitectonic similarity networks, we analyzed three datasets: the Cognitive-Consilience dataset, the Von Economo dataset, and the BigBrain dataset (see Methods Cytoarchitecture).

The Cognitive-Consilience dataset, based on cortical laminar patterns⁴¹, was digitized by Seidlitz et al.³⁵ and extended in this study to include subcortical and cerebellar regions (see Methods Cognitive Consilience Classification), comprising nine bilaterally symmetric cytoarchitectonic classes (see Fig. 7a). We applied Gaussian Mixture clustering with nine clusters to the Geneva study group’s averaged MetSiM MS mode map and compared the resulting clusters to the cytoarchitectonic classes. Alignment between the clusters and cytoarchitectonic classes was assessed using overlap metrics and spatial permutation tests (see Methods Inter-Class Overlap Score). Individual MetSiMs yielded a significant overlap score of 0.30 ± 0.02 (AdjPerm p = 0.008), and the group average MetSiM showed a higher score of 0.36 (AdjPerm p < 0.001). Cluster correspondence analysis revealed strong alignment in the primary somatosensory cortex (class 1 and metabolic class 7, overlap: 0.632), subcortical regions (class 8 and metabolic class 1, overlap: 0.577) and in the premotor cortex (class 2 and metabolic class 6, overlap: 0.730). Full results are reported in Supplementary Table 4.

**Fig. 7: Metabolic Similarity and Cytoarchitectural Similarity.**

The Von Economo dataset included 42 bilaterally symmetric cortical cytoarchitectonic profiles mapped onto the 210-node LFMIHIFS-3 parcellation (left Fig. 7b). A direct edge-wise comparison between the group-average MetSiM and the Von Economo similarity matrix (see Methods Cognitive Consilience Classification) yielded a Spearman correlation of 0.09 (AdjPerm p < 0.001) which dropped to 0.05 (p < 0.001) after controlling for spatial autocorrelation. This global association was primarily driven by the occipital lobe, medial superior prefrontal regions, precentral gyrus, superior temporal gyrus, and posterior insula (r = 0.45 − 0.51) as shown in the absolute nodal correlation metric on Fig. 7b (see Methods Absolute Nodal Correlation). In contrast, the anterior cingulate cortices showed minimal correlation.

The BigBrain similarity matrix (left Fig. 7c)⁴², derived from ultra-high-resolution cytoarchitectural data of the cortex⁴³ (see Methods Big Brain Similarity Matrix), shows a node-wise correlation of 0.29 (AdjPerm p < 0.001) with the group-average MetSiM and dropped to 0.26 (p < 0.001) after correcting for spatial autocorrelation. The highest absolute nodal similarities were observed in the occipital lobe, postcentral gyrus, superior frontal gyrus (medial part), caudal-medial superior frontal cortex, and posterior cingulate cortex (r = 0.45–0.51). Similar to the Von Economo similarity matrix, the anterior cingulate cortex showed minimal correlation.

Lastly, we applied the principal path extraction procedure to the Von Economo and BigBrain similarity matrices by computing their first PCA-t-SNE component to obtain a cytoarchitectonic similarity index (CSI) (bottom panels of Fig. 6b,c) and attempting to derive a path minimizing the normalized MS gradient and maximizing normalized entropy. No valid path was identified, likely due to a weak continuous similarity gradient (LH/RH-CTX Von Economo ${\widetilde{{{\mathscr{G}}}}}_{\mu }=0.12$; BigBrain 0.07) visible in Fig. 7b,c, in the presence of a high CSI diversity (LH/RH-CTX Von Economo $\widetilde{H}=0.84$; BigBrain LH-CTX 0.86, RH-CTX 0.76). Cost function optimization results are provided in Supplementary Fig. 7.

These results suggest that MetSiM topology partly reflects the spatial organization of cytoarchitectonic patterns, and is primarily driven by the high diversity of distinct metabolic similarity profiles across the cortex.

Metabolic similarity and genetic co-expression

We next tested our second biological hypothesis, that MetSiM edges are characterized by high levels of gene co-expression, by leveraging the nearly complete human genome dataset (20,737 genes) from the Allen Brain Institute. Using data from six adult human post-mortem brains⁴⁴, we mapped whole-genome transcriptional profiles onto the same LFMIHIFS-3 parcellation used to define the MetSiM’s 277 nodes (see Methods Genetic Co-Expression). This approach allowed us to estimate inter-regional co-expression patterns (see Fig. 8) for every possible pair of nodes within the same anatomical reference frame as the MetSiM. A significant positive correlation was observed between MetSiM edge weights and whole-genome co-expression values (PermAdj r = 0.34, p < 0.001) and r = 0.32 (p < 0.001) when accounting for spatial autocorrelation, indicating that metabolically similar regions also share similar gene expression profiles. Regions in the occipital lobe and superior frontal cortex (medial) showed the strongest absolute nodal correlation (r = 0.60 − 0.74), while the anterior cingulate cortex, brainstem and cerebellar regions contributed the least. This strong alignment between metabolic and genetic similarity is reflected in the principal path of genetic similarity from a PCA-t-SNE reduction (Fig. 8c and optimization results show in Supplementary Fig. 7), which follows a similar caudal-to-rostral trajectory as the MetSiM path (Fig. 8e; LH/RH-CTX ${\widetilde{{{\mathscr{G}}}}}_{\mu }=0.07$; LH-CTX $\widetilde{H}=0.88$, RH-CTX 0.83). This concordance, and the topological prerequisites for such paths to emerge, suggest that the local and global effects characterizing metabolic organization may, in part, be shaped by underlying transcriptional architecture. Furthermore, after a gene ontology (GO) analysis of the 20,737 genes, we extracted 250 GO categories with brain-enriched contributions (FDR-corrected enrichment ratio > 1). A leave-one-GO-class-out analysis, identified, via the elbow method, 17 parent GOs that contributed the most (see Supplementary Fig. 8). Most of these contributing categories are related to developmental processes (e.g., cell division and polarity, cell migration, neuron projection development), transport (e.g., inorganic cation and anion, amino acids, organophosphate ester), and cell-cell signaling including synaptic signaling and transmission. Multicellular organismal-level homeostasis and ensheathment of neurons by glia cells are also among the most contributing GO categories. By contrast, the least contributing GO categories include embryo development, development of organs not directly related to the brain, organelle assembly, RNA metabolic and catabolic processes. Full results are reported in Supplementary Data 2 and 3. These results indicate that MetSiM topology aligns with spatial expression patterns of brain-expressed genes that play a central role in supporting brain development and function as well as homeostasis.

Discussion

We have demonstrated how MetSiMs, derived from 3D high resolution whole-brain ¹H-MRSI, can be used to estimate the metabolic similarity between brain regions in a sample of healthy participants at both the individual and group levels. We developed a reliable and consistent method for deriving MetSiMs, which we have shown to be also replicable in an independent sample scanned using a different MRI scanner. Furthermore, we recurred to a brain metabolic network analysis, which enabled us to first investigate the biological basis of the brain metabolic organization and test the resemblance of the MetSiMs-derived network with other known brain networks.

We observed that MetSiM distribution follows an inverse relationship with distance: spatially proximate nodes are metabolically similar, whereas distant nodes are metabolically dissimilar. This distance-dependent organization parallels patterns reported in other brain networks, where short-range anatomical connections tend to be stronger and more prevalent, while long-range connections are sparse and weak⁴⁵; similarly, positive functional correlations typically occur at short distances, whereas anti-correlations are more common between distant regions and antagonistic networks⁴⁶. Such spatial dependencies are characteristic of the complex topological organization observed in naturally embedded networks⁴⁷, where modular architectures balance local clustering with global integration. Beyond this spatial tendency, MetSiMs also align along functional dimensions, showing clustering of functionally similar regions (e.g., the limbic system) and segregation of functionally distinct regions (e.g., primary somatosensory vs. primary motor cortices). Notably, this pattern is primarily driven by a decrease in the tNAA/Cho ratio (relative to colinear levels of tCr, Glx, and Ins) from the occipital lobe (high ratio) through the parietal and frontal cortices, extending into subcortical structures, the brainstem, and the cerebellum (low ratio).

This decrease is consistent with existing MRS literature, which demonstrates higher tNAA (a neuronal density marker) in gray matter-rich cortical regions, and lower tNAA with relatively higher Cho (a membrane turnover marker) in subcortical structures such as the basal ganglia and brainstem^48,49,50. Even after partial volume correction, regional variation persists: cortical GM–particularly in the occipital lobe–has a higher density of mature neurons⁵¹ and lower glial content than subcortical GM⁵², which is richer in glial cells, including a higher density of oligodendrocytes⁵³, and exhibits greater myelin-related membrane dynamics. The observed decrease in tNAA/Cho ratio from the occipital lobe to subcortical structures thus reflects a shift from neuron-rich to glia-enriched regions with elevated membrane turnover. Of note, one GO categorie that contributes to the correlation between MetSiM and gene co-expression is related to ensheathment of neurons by glia cells.

Topographic organization of network properties is common in biology, with one of the most striking examples being the tonotopy along the basilar membrane of the cochlea. There, a monotonic graded change in membrane properties (stiffness, width, and mass distribution) maps low- to high-frequency responses of outer hair cells along the cochlear spiral⁵⁴. Analogously, the MS gradient revealed by the metabolic principal path shows a spatially ordered, monotonically increasing progression of MS mode values along the caudo-rostral axis, mapping distinct functional brain specializations into a continuous biochemical space. Albeit the occurennce of such a gradient in other biological system, we do not claim that the resulting metabolic principal paths have a direct physical or biological instantiation; rather, they represent measurable signatures of a network topology balancing local clustering and global diversification. Similar to structural brain networks, where nearby gray matter regions form dense connections to minimize wiring cost and sparse long-range links preserve global efficiency⁶, metabolic connectomics appears to balance local metabolic variability with global metabolic diversity. Minimizing the former creates a coherent spatial structure on an otherwise random topology, while maximizing the latter organizes this structure into a traversing MS gradient with a monotonic MS mode. This economy of metabolic distribution is reminiscent of the small-world topology in structural networks⁶, where such balance enables efficient communication routes and, beyond its biophysical interpretation, has served as a marker of network integrity in health and disease^7,8. By analogy, we propose that the occurrence of metabolic principal paths reflects a direct measure of metabolic coherence in brain networks.

Large-scale spatial gradients as continuous patterns of variation in brain organization have been found across structural, functional, microstructural, transcriptomic, and neurotransmitter receptor data⁵⁵. Constant, whole-brain-traversing rostro-caudal gradients (resulting from monotonically increasing modes) are rarely reported in the literature, likely because most brain organization follows more complex, multidirectional patterns and is typically confined to specific brain systems, such as the principal morphometric similarity mode separating motor from sensory cortices⁵⁶, the developmental and evolutionary cortical expansion where association cortices enlarge most relative to primary areas⁵⁷, or the microstructural “sensory-fugal” gradient that extends from sensory anchors toward transmodal cortex⁵⁸. The study most closely aligned with our results is that of Oldham et al.⁵⁹, who combined in vivo MRI-based structural connectivity with post-mortem gene expression data to identify a medial-lateral thalamic gradient mapping onto an anterior-posterior (rostro-caudal) cortical pattern. We observed a congruent anterior-posterior gradient in our MS gradient, obtained purely from non-invasive in vivo ¹H-MRSI data. In this context, the thalamus–identified by Oldham et al. as the driver of the thalamo-cortical gradient and recognized as a central structural and functional hub, may likewise act as the orchestrator of metabolite diversity, assembling locally coherent metabolic similarity patterns into the complex topology observed at the macroscale. We found that MetSiMs show only weak correspondence with structural connectivity–both directly, where the presence of a structural edge does not imply metabolic similarity, and indirectly, where shared neighbors mediate similarity only to a limited extent. Instead, their organizational structure is better captured by gradients: locally, MS modes align with connectopies, reflecting high proximal similarity, but globally they diverge, as metabolic similarity rarely extends over long distances. By contrast, connectopies display redundancy both proximally (neighboring nodes sharing similar profiles) and distally (distant nodes converging on the same set of hubs). As hypothesized, cortical areas similar in MetSiM networks also exhibit moderate cytoarchitectural similarity and high genetic co-expression, consistent with the established link between shared cellular machinery and parallel metabolic demand^33,42,60,61. Notably, the genetic co-expression gradient not only follows the same rostro-caudal trajectory as the MS gradient but also suggests that both the local and global properties of MS are shaped by underlying genetic organization. Taken together, these results indicate that metabolic similarity networks provide a biologically grounded, node-centric organizational framework that is locally consistent with but globally distinct from the white matter edge-centric connectome.

Building on these previous observations, we propose that the metabolic principal path extends beyond a usefull mathematical construction, and it is a remnant of the neurodevelopmental processes governing the maturation of the neural tube and its subsequent unfolding into subcortical and cortical structures. During early neurodevelopment, the neural tube differentiates into three primary divisions from rostral to caudal; the prosencephalon (forebrain), mesencephalon (midbrain), and rhombencephalon (hindbrain), preserving a sequential arrangement such that initially adjacent regions remain proximal in the mature central nervous system. While each region adopts a distinct gene expression profile that guides its development, function, and metabolism, the extent of its spatial variation remains constrained by proximity effect: adjacent areas share similar microenvironments⁶², exhibit local diffusion of signaling molecules (morphogens)⁶³, and display aligned epigenetic modifications (e.g., DNA methylation patterns)⁶⁴. Spatially assimilated regions share similar gene expression, hence produce comparable metabolic enzymes and display analogous metabolic activity⁶¹. Consequently, proximity in the neural tube translates into adjacency in the mature metabolic principal path, yielding its unbroken continuity observed from the forebrain through the midbrain to the hindbrain. Consistent with this interpretation, the GO categories that most strongly contribute to the MetSiM-gene co-expression association are enriched for neurodevelopmental processes.

Our findings show that variation in ¹H-MRSI-measured metabolite levels across high-centrality nodes predicts the structural backbone of the human connectome, suggesting that metabolic organization may precede and support structural connectivity. Whereas PET–particularly FDG-PET–has demonstrated associations between glucose metabolism and structural centrality^13,65,66, ¹H-MRSI captures a broader spectrum of cellular processes, including neuronal integrity (tNAA), energy buffering and storage (tCr), excitatory neurotransmission (Glx), membrane turnover (Cho), and glial activity (Ins)⁶⁷, providing a richer biochemical basis for network inference. Moreover, compared to PET, the proposed ¹H-MRSI-based metabolic connectome offers key advantages: it is non-invasive, avoids ionizing radiation, can simultaneously quantify multiple metabolites in a single acquisition (8-20 minutes depending on protocol, field strength and acceleration²⁶), and provides direct insight into neurochemical pathways beyond energy metabolism. In contrast, PET and hybrid PET/MRI approaches, while powerful for mapping glucose or receptor binding^{68,69,70,71,72}, remain limited by long scan times, higher motion sensitivity, cost, neurovascular coupling effects, and the use of radioactive tracers^73,74.

This study has several limitations to consider. MRSI’s limited coverage excludes orbitofrontal and basotemporal regions due to susceptibility artifacts and signal distortion. Additionally, the 5 mm isotropic spatial resolution challenges the accurate assessment of smaller subcortical regions. As a result, the MetSiMs constructed here may lack critical information, potentially omitting key network structures like the default mode network located in these areas. Furthermore, the constructed MS mode map in this study was based on group-averaged data, which overlooks inter-individual variability. Factors such as participants’ resting or task states could influence metabolic similarity and were not addressed here. Future work with larger multi-institutional cohorts will allow a more systematic assessment of generalizability and the influence of neurological and psychiatric conditions on MetSiMs. Although we have not yet evaluated the alignment between MetSiMs and functional connectivity, our findings reveal that these networks possess a functional modular structure, which suggests that a strong correlation may exist.

In summary, our work provides a topological analysis of whole-brain gray matter metabolic similarity networks, made possible by recent advances in ¹H-MRSI. Beyond replicating canonical principles of brain network organization, we show that the observed topological properties can be directly linked to regional metabolite variation and their interplay, shaped by genetic and cytoarchitectural constraints, thereby establishing metabolic similarity as a biologically meaningful proxy for underlying cellular and molecular processes. Through their association with neurodevelopmental processes, such networks may play a fundamental role in shaping both structural and functional brain systems. This highlights their potential as biomarkers, capable of revealing metabolic disruptions linked to disease even before structural or functional changes occur or symptoms manifest. Comparing these networks between controls and individuals at risk for disorders such as schizophrenia could provide deep insights into underlying pathological mechanisms and identify metabolic alterations affecting central brain regions prior to the onset of the first psychotic episode. Thus, the development of MRSI-based metabolic networks should be viewed not merely as a technical advance but as a practical tool poised for integration into a comprehensive suite of clinical applications.

Methods

This research complies with all relevant ethical regulations. This study is a secondary analysis of de-identified MRI/MRSI data from two independent cohorts that were conducted previously and approved by their respective ethics committees: the Geneva Cantonal Research Ethics Committee (CCER; protocol 2018-01731) and the Cantonal Research Ethics Commission on Human Research, Canton of Vaud, Switzerland (CER-VD; protocol PB 2017-00675 82/14). Data handling complied with applicable data-protection laws and institutional policies; only de-identified data were used and no re-identification was attempted.

Groups

Geneva Mindfulteen study

Participants were recruited and scanned in the context of the Mindfulteen study, a randomized controlled trial to assess the effects of a mindfulness-based intervention on adolescent well-being. Details on recruitment, inclusion, and exclusion criteria can be found in the published protocol³⁴. Briefly, the study recruited adolescents (n = 69,sex assigned at birth: 39 female, 30 male, gender not assessed, age range = 13–15 years) from the general population. Exclusion criteria included chronic somatic diseases or significant medical conditions, recent psychotherapy (< 6 months), recent use of psychotropic medications (<6 months), and any history of psychiatric disorders, except for current anxiety disorders or a past episode of major depressive disorder resolved at least 6 months earlier. The study protocol was approved by the Geneva Regional Ethical Committee (CCER 2018-01731). All participants (assent) and their legal guardians (consent) provided written informed consent/assent in accordance with the Geneva Regional Ethics Committee. No a priori sex- or gender-based analyses were planned given the study’s scope and sample size.

Lausanne psychosis cohort

For the replication of our results in an independent sample scanned at a different site, we included 13 healthy controls (sex assigned at birth: 4 female, 9 male; gender not assessed, age range = 15-35 years), from the Lausanne Psychosis Cohort³⁶. They were assessed by the Diagnostic Interview for Genetic Studies⁷⁵ to exclude a major mood, psychotic, or substance use disorder or had a first-degree relative with a psychotic disorder. Written informed consent was obtained from adult participants; for minors, written parental/guardian consent and participant assent were obtained. The study protocol was approved by the Cantonal Ethics Committee for Research on Human Beings (PB 2017-00675 82/14).

MRI image acquisition

For the Geneva study, magnetic resonance (MR) data were acquired using a 3-Tesla scanner (Magnetom TrioTim, Siemens Healthineers, Forchheim, Germany) equipped with a 32-channel head coil at the Brain and Behavior Laboratory in Geneva. Each scanning session included a magnetization-prepared rapid acquisition gradient echo (MPRAGE) T1-weighted sequence with an in-plane resolution of 1 mm and a slice thickness of 1.2 mm, covering a volume of 240 × 257 × 160 voxels. The repetition time (TR), echo time (TE), and inversion time (TI) were set to 2300 ms, 2.98 ms, and 900 ms, respectively. Additionally, a diffusion spectrum imaging (DSI) sequence was performed, acquiring 128 diffusion-weighted images with a maximum b-value of 8000 s/mm² and one b₀ reference image. The acquisition volume for DSI consisted of 96 × 96 × 34 voxels with a resolution of 2.2 × 2.2 × 2.2 mm. The TR and TE for the DSI sequence were 6800 ms and 144 ms, respectively.

The 3D ¹H-FID-MRSI sequence accelerated by compressed-sensing²⁴ was acquired with 1.50 ms TE, 372 ms TR and 35 deg flip angle. The Field-of-View (FoV) size was 210 × 160 × 105 mm (anterior-posterior, right-left, head-foot directions) with a 95 mm-thick slab selection, with a spatial resolution of 5 × 5 × 5.3 mm and the spectral bandwidth was 2 kHz acquired with 512 points. The reference water acquisition had the following parameters: same TE, TR of 25 ms, lower flip angle of 3^∘, same FOV size, lower resolution of 6.6 x 6.7 x 6.6 mm, same bandwidth and free induction decay (FID) size of 16 points.

For the Lausanne Psychosis Cohort, MRI sessions were conducted on a 3-Tesla scanner (MAGNETOM Prisma fit, Siemens Healthineers, Forchheim, Germany) at the Lausanne University Hospital for 13 healthy controls. While the MPRAGE and DSI sequences parameters were identical to those used in the Mindfulteen study, the the 3D ¹H-FID-MRSI sequence had different parameters than the one implemented for the Geneva study (TE = 1.0 ms, TR = 353 ms and flip angle of 40 deg. The FoV was the same as in Lausanne, with the same spatial resolution of 5 × 5 × 5.3 mm, as the spectral bandwidth and FID size. The water acquisition had the following parameters: same TE of 1.07 ms, TR of 25 ms, and a flip angle of 5 deg. FOV size, resolution, bandwidth and vector size were the same as for the Geneva study.

Complete details on MRSI acquisition, reconstruction, quantification, and validation are given in the original methodological reference²⁴; MRSI parameters are summarized in Supplementary Table 5 following the MRSinMRS checklist⁷⁶.

Preprocessing

Anatomical

For each subject, the anatomical T1-weighted image was used to parcellate the brain into different regions using the FreeSurfer package (version 7.2.0, http://surfer.nmr.mgh.harvard.edu) and subsequently identify nine distinct supra-regions: cortex, basal ganglia, thalamus, amygdala, hippocampus, hypothalamus, cerebellum, brainstem, and white matter.

We relied on the Chimera parcellation software tool⁷⁷ to create a combined volumetric parcellation for each input subject anatomical image based on the following atlases (as listed in the original Chimera repository) and referred to as the LFMIHIFS-3 parcellation scheme:

L: Lausanne cortical parcellation scale 3⁷⁸,
F: Basal ganglia parcellation⁷⁹,
M: Thalamus parcellation⁸⁰,
I: Amygdala parcellation⁸¹
H: Hippocampus parcellation⁸²
I: Hypothalamus parcellation⁸³,
F: Brainstem parcellation⁸²,
S: Cerebellum parcellation⁸⁴,

DSI and structural connectivity

The DSI volumes were visually inspected for signal loss, which could indicate motion artifacts and potential exclusion from the analysis⁸⁵. The DSI data were reconstructed following the methodology described by Wedeen et al.⁸⁶. Preprocessing, anatomically constrained tractogram construction using deterministic tractography, and structural connectivity computation were performed using the MRtrix3 software tool⁸⁷.

Particularly, the peaks of the local fiber orientation distribution functions (fODFs) (maximum of three peaks per fODF) were computed and served as input for the deterministic tractography⁸⁸. The tractography parameters included a maximum track length of 300 mm, 32 seeds per white matter voxel per diffusion direction, and termination criteria of a 60-degree angle between subsequent propagation steps or reaching the white-gray matter boundary. To ensure reproducibility, 10⁶ streamlines were selected⁸⁹.

A weighted, undirected brain connectivity matrix was reconstructed for each subject by filtering the whole-brain tractogram using Spherical-deconvolution Informed Filtering of Tractograms (SIFT)⁹⁰, which refined streamline contributions based on their alignment with underlying diffusion data and termination coordinates. We defined the sub-regions of the Chimera parcellation (see Section Anatomical) as the nodes of the structural connectivity. The strength of the connections S_ij between any two node i and j were computed as the number of streamlines ns_ij connecting these two regions weighted by the inverse of their average length $\frac{1}{{l}_{ij}}$ and the inverse of the node’s volumes $\frac{1}{{V}_{i}+{V}_{j}}$ i.e., ${S}_{ij}=\frac{1}{{l}_{ij}}\cdot \frac{1}{{V}_{i}+{V}_{j}}\cdot n{s}_{ij}$. The length and volume terms were used to eliminate the bias towards longer fibers introduced by the tractography algorithm and the bias for the variable size of cortical ROIs, respectively⁴⁵. For consistency and to limit possible false- positive connections, the connections that present in less than 50% of participants were discarded^91,92.

MRSI reconstruction

The MRSI data was reconstructed using a low-rank model constrained by total-generalized variation, with prior removal of subcutaneous lipid contamination and residual water signals²⁵. Following reconstruction, the spatio-spectral data was analyzed using LCModel⁹³ to quantify metabolite signal in each voxel, using the water signal from the additional acquisition as a reference. Because of the ultra-short echo time (1.5 ms), no T₂ relaxation corrections were required²⁵. Although T₁ weighting is present because of the reduced TR and would differentially affect GM and WM, only GM-GM correlations were considered in this study. Thus, explicit T₁ correction would merely introduce a global scaling factor without altering the reported results. However, the 5 mm isotropic resolution may introduce partial signal leakage across GM-WM boundaries; therefore, potential GM-WM composition effects were further mitigated using voxel-wise partial volume correction in combination with a MRI point-spread-function correction (see Methods Section MRSI Partial Volume Correction).

The basis set for LCModel fitting included the following metabolites: N-acetylaspartate (NAA), N-acetylaspartylglutamate (NAAG), creatine (Cr), phosphocreatine (PCr), glycerophosphocholine (GPC), phosphocholine (PCh), myo-inositol (mI), scyllo-inositol (sI), glutamate (Glu), glutamine (Gln), lactate (Lac), gamma-aminobutyric acid (GABA), glutathione (GSH), taurine (Tau), aspartate (Asp), and alanine (Ala). Only a subset of these metabolites could be reliably resolved and due to overlapping spectral peaks, certain metabolites were combined: NAA and NAAG (denoted tNAA), Cr and PCr (denoted tCr), GPC and PCh (denoted Cho), mI (denoted Ins), and Glu and Gln (denoted Glx), resulting in five distinct metabolite volumes. LCModel also provided spectral quality metrics, such as the signal-to-noise ratio (SNR), Cramer-Rao Lower Bound (CRLB) for each metabolite estimation.

MRSI registration

The MRSI tCr volume was co-registered to the corresponding T1-weighted anatomical image using Advanced Normalization Tools (ANTs)⁹⁴, with mutual information as the cost function for optimization. In addition, each individual T1-weighted image was nonlinearly registered to the MNI152 standard space. Both registration steps yielded a combination of affine and symmetric image normalization (SyN) transforms, which were subsequently applied or inverted, as appropriate, in later preprocessing and parcellation steps.

MRSI partial volume correction

To account for partial volume effects (PVE) arising from the limited spatial resolution of MRSI (5 mm isotropic), we applied a correction combining region-based voxel-wise (RBV) modeling⁹⁵ with point spread function (PSF) deconvolution to account for signal blurring. High-resolution T1-weighted structural images (1 mm isotropic) were first segmented into probabilistic tissue maps (gray matter, white matter, and CSF) using the Computational Anatomy Toolbox (CAT12)⁹⁶ for SPM12⁹⁷. These tissue maps were then transformed into MRSI space using the inverse of the composite affine and symmetric image normalization transforms previously estimated during MRSI registration (see Section MRSI Registration). Partial volume correction was performed using the petpvc command-line tool⁹⁸, applying the RBV method with an isotropic Gaussian PSF of 5 mm full-width at half-maximum in all directions. Sample MRSI volume from the Geneva study, shown before and after partial volume correction, is presented in Supplementary Fig. 9.

MRSI parcellation

The multi-scale cortical parcellation proposed by Cammoun et al.⁷⁸ was selected for its hierarchical organization, which allows for flexible regional granularity. This parcellation provides coarse-scale segmentation with fewer but heterogeneously sized regions (scale 1), as well as finer-scale subdivisions with a greater number of parcels of relatively uniform size (scale 5). Due to the limited spatial resolution of ¹H-MRSI data (5 mm isotropic) compared to the higher resolution of T1-weighted anatomical images (1 mm), scale 3 was chosen as it offers a balance between anatomical specificity and voxel-wise compatibility. This scale was integrated into the Chimera parcellation framework(see Anatomical), resulting in the LFMIHIFS-3 schema.

To project the Cammoun parcellation to MRSI space, the parcellation image in individual T1 space was transformed using the inverse of the composite affine and symmetric normalization (SyN) transforms previously estimated during MRSI registration (see Section MRSI Registration). Nearest-neighbor interpolation was used to preserve discrete parcel labels. This procedure was repeated for each subject, resulting in personalized MRSI-space parcellations and yielding five MRSI volumes (tNAA, tCr, Cho, Ins, and Glx) distributed across 277 brain regions.

In addition to the Cammoun atlas, we evaluated two functionally derived cortical parcellations: the Schaefer-200 atlas⁹⁹ and MIST197¹⁰⁰. Both were integrated into the Chimera LFMIHIFS-3 schema by replacing the Lausanne (Cammoun) parcellation, resulting in the SFMIHIFS-200 and MFMIHIFS-197 schemes, respectively. These atlases were selected for their comparable number of cortical parcels and their derivation from functional connectivity analyses, offering a meaningful contrast to the anatomically driven Cammoun atlas. To assess the uniformity of these schemes, we computed the Gini coefficient of parcel sizes. All three atlases exhibited similarly low Gini values (Lausanne: 0.20, Schaefer-200: 0.21, MIST197: 0.22). To provide a structurally and functionally agnostic reference, we also implemented two geometric parcellation approaches based on regular cubic volumes. A standard MNI152 template (1 mm isotropic) was divided into cubes of 10 mm and 15 mm edge lengths, corresponding to approximately 4 and 9 MRSI voxels per parcel, respectively. These parcellations were then transformed into each subject’s MRSI space and used to assess the influence of different parcellation atlases on the results presented in this study.

Estimation of metabolic profiles

After parcellation, regions smaller than 8 MRSI voxels (<1 cm³) or those with poor MRSI coverage in more than 70% of subjects were excluded from the 277 regions, resulting in N = 210 retained regions. The discarded regions were mainly located in orbitofrontal and baso-temporal regions where ¹H-MRSI has limited coverage due to susceptibility artifacts. These factors reduce signal-to-noise ratio and introduce spectral contamination, hindering reliable metabolite quantification in these areas. Additionally, the ¹H-MRSI sequence covered only the superior portion of the cerebellum. The complete ¹H-MRSI mask coverage is shown in the Supplementary Information Fig. 1. Median metabolite values were calculated inside a given parcel and z-score normalized across regions, yielding a 5-feature metabolic profile per region. To address variability in the metabolites quantification, we followed the Monte Carlo uncertainty propagation methodology described by Instrella et. al¹⁰¹, repeating the process K_pert = 50 times to generate perturbed versions of the original metabolic profile. Specifically, for each voxel (i, j, k), the metabolite signal was resampled from a normal distribution with mean equal to the original MRSI signal Y_ijk and standard deviation equal to $3{\sigma }_{ijk}^{2}$, where ${\sigma }_{ijk}^{2}$ is variance corresponding to the Cramér-Rao lower bound (CRLB; see Methods MRSI reconstruction) derived from the LCModel. Mathematically,

$${Y}_{ijk}^{{\prime} } \sim {{\mathscr{N}}}\left({Y}_{ijk},\,3{\sigma }_{ijk}^{2}\right)$$

(1)

where Y_ijk is the original MRSI signal value. This augmentation increased the sample size of the metabolic profiles by a factor of K_pert, providing K_pert − 1 additional perturbed profiles for each brain region, in addition to the original profile.

Metabolic similarity mode

To visualize the spatial embedding of the 210 × 210 dimensional group-averaged MetSiM (Geneva sample, N = 51), we reduced each of its 210 high-dimensional feature vectors x_i–corresponding to the parcel connectivity profiles–into a lower-dimensional representation. First, Principal Component Analysis (PCA) was applied for denoising and variance capture, resulting in a 50-dimensional embedding. Subsequently, t-distributed Stochastic Neighbor Embedding (t-SNE) was employed to further reduce the dimensionality to obtain the metabolic similarity mode (MS mode) μ_i at every nodes i. As prescribed by ref.¹⁰², PCA initialization was used to stabilize and improve the convergence of t-SNE, which in turn was chosen to preserve the observed local connectivity patterns of the MetSiM.

t-SNE converts high-dimensional distances between data points into conditional probabilities (or similarities) and then maps them into a low-dimensional space. A crucial parameter in t-SNE is perplexity, which can be interpreted as a measure of how many close neighbors each point has. It also balances local and global connectivity patterns, where lower perplexity emphasizes local structures at the expense of global structure, and higher perplexity does the opposite. The method adjusts the low-dimensional output features μ_i by minimizing the Kullback-Leibler (KL) divergence between the probability distributions in the high-dimensional and low-dimensional spaces, as computed via a t-Student-based metric.

We explored perplexity values from 20 to 60 and observed minimal changes in the KL divergence (ranging approximately from 0.55 to 0.42). Consequently, following¹⁰², we chose a perplexity of 30 to strike a balance between preserving both local and global structures. Finally, an early exaggeration factor of 12, as suggested by¹⁰³, was applied to enhance cluster separation during the initial stages of optimization. From these parameters, we reconstructed a filtered version of the original MetSiM, whose elements q_ij were computed using the t-Student-based metric:

$${q}_{ij}=\frac{{\left(1+\parallel {{{\boldsymbol{\mu }}}}_{i}-{{{\boldsymbol{\mu }}}}_{j}{\parallel }^{2}\right)}^{-1}}{{\sum }_{k\ne l}{\left(1+\parallel {{{\boldsymbol{\mu }}}}_{k}-{{{\boldsymbol{\mu }}}}_{l}{\parallel }^{2}\right)}^{-1}}$$

(2)

where μ_i is the dimension-reduced representation of the i-th node’s connectivity profile x_i. We then rescaled these values to 2q_ij − 1 to restore the original MetSiM value range [ − 1, 1] and verify the overall pattern conservation of the original MetSiM shown in Supplementary Fig. 2.

An intrinsic dimensionality analysis based on ISOMAP, which estimates geodesic distances on a neighborhood graph¹⁰⁴, revealed that the first component accounted for 53% of the variance, the second for 20%, and the third for 4%, consistently across a wide range of ISOMAP neighborhood sizes (3-20).

Metabolic principal path construction

In diffusion MRI tractography white matter streamlines are traced by aligning their tangent vector to the principal eigenvectors field of the diffusion tensor field, weighted by regions of strong fractional anisotropy and thereby structurally connecting distant gray matter nodes. Here, we define metabolic paths by aligning their tangent vector with the MS gradient vector field and privileging those that globally capture the overall variability of metabolic profiles.

Instead of identifying a continuous curve in space, the path we aim to construct is a sequence of edges in a discrete lattice (finite set of nodes) within the metabolic network. We proceed to introduce the definitions for a path, the MS gradient, and normalized entropy.

Let G = (V, E, w) be a weighted network where V is a finite set of nodes, E ⊆ V × V is a set of edges and w: E ↦ [ − 1, 1] is the weight function which assigns a MetSiM value w_ij to nodes i and j. We define a path γ of length L as an ordered sequence of nodes

$$\gamma=({l}_{0},\,{l}_{1},\,\ldots,\,{l}_{L}),$$

such that for each integer t(0≤t < L), the consecutive pair (l_t, l_t+1) forms an edge in the network. We designate l₀ as the start node and l_L as the end node. We define the MS gradient between node i and j as the finite-difference directional derivative of the MS mode scalar field along the edge connecting them:

$${({{\boldsymbol{\nabla }}}{{\boldsymbol{\mu }}})}_{ij}=\frac{{\mu }_{j}-{\mu }_{i}}{\parallel {{{\bf{r}}}}_{j}-{{{\bf{r}}}}_{i}\parallel }\,{{{\bf{e}}}}_{ij},$$

(3)

where μ_i is the MS mode value at node i, and d_ij = ∥r_j − r_i∥ is the Euclidean distance between nodes i and j and e_ij the unit vetor between node i and j. To fully capture the overall smoothness by accumulating local fluctuations of MS mode values, we summed the L² norm of each sequential gradient ((3)) along the path to yield the total MS gradient generated along the path γ as

$${{{\mathscr{G}}}}_{\mu }(\gamma )=\frac{1}{N-1}{\sum}_{i\in E(\gamma )}\parallel {{\boldsymbol{\nabla }}}{{\boldsymbol{\mu }}}{\parallel }_{i,i+1}$$

(4)

then we write the normalized MS gradient as

$${\widetilde{{{\mathscr{G}}}}}_{\mu }(\gamma )=\frac{{{{\mathscr{G}}}}_{\mu }(\gamma )}{{{{\mathscr{G}}}}_{\mu }({\gamma }_{{{\rm{rand}}}})}$$

(5)

where the denominator is the maximum MS gradient generated by a uniformly random MS mode sequence γ_rand of length N with the same start and end nodes as γ, resulting in the normalized final quantity.

We define the normalized entropy of the MS mode values along the path γ = (l₀, …, l_L) by

$$\widetilde{H}(\gamma )=-\frac{1}{{H}_{max}}{\sum }_{i=1}^{{N}_{{{\rm{bins}}}}}{p}_{i}\,\log ({p}_{i}),$$

(6)

where is ${H}_{max}=\log ({N}_{bins})$ is the maximum entropy (entropy of a random uniform distribution). We used the Rice rule for determining the number of bins N_bins = ⌈2 (L+1)^1/3⌉ of the MS mode histogram $p=({p}_{1},\ldots,{p}_{{N}_{{{\rm{bins}}}}})\{{\mu }_{{l}_{0}},\ldots,{\mu }_{{l}_{L}}\}$ along γ.

The optimal path should simultaneously traverse nodes which

Condition 1 are highly metabolically similar to their proximal neighbors (i.e., minimize MS gradient)
Condition 2 globally capture the highest diversity of metabolic profiles (i.e., maximize the Shannon entropy of the MS mode distribution they generate).

The first condition is analogous to optimal alignment of the tangent vector of a smooth curve in continuous space with a vector field (MS gradient), while the second condition prioritizes nodes that globally capture the principal variability of the network. Then, for a given path γ = (l₀, …, l_L), its mean total cost is

$${{\mathscr{L}}}(\gamma )=\frac{{\widetilde{{{\mathscr{G}}}}}_{\mu }(\gamma )}{\widetilde{H}(\gamma )}$$

(7)

where the optimal path γ^* minimizes:

$${\gamma }^{*}=\arg {\min }_{\gamma }{{\mathscr{L}}}(\gamma )$$

(8)

Due to the non-convexity of the cost function, we employed a brute-force algorithm via Depth-First Search backtracking to:

1.
identify all possible paths connecting the start and end nodes,
2.
calculate the cost C(γ) induced by each path,
3.
select the paths which yielded the minimal cost value.

Given the hemispheric symmetry of the MS mode map, pathfinding was performed independently in each hemisphere. In both cases, the isthmus cingulate; lowest index) served as the starting node, and the rostral anterior cingulate cortex (rACC; highest index) as the ending node. Because subcortical regions form a separate gray matter network, we applied the same pathfinding approach to these regions by splitting each hemisphere while keeping the brainstem as a shared structure.

Results for both hemispheres–showing the total costs of the cortical and subcortical-region paths–are displayed in Supplementary Fig. 6. The optimal path γ^* was selected from 180304 and 229670 possible pathways in the left and right neocortical hemispheres, respectively, and from 20 and 220 pathways in the left and right subcortical hemispheres. Paths with implausible trajectories, such as loops or abrupt detours, were excluded from the selection. We applied the same procedure to the other datasets, including the Geneva Study average connectivity matrix, the cytoarchitectonic similarity matrices (Von Economo and BigBrain), and the genetic co-expression matrix, with results shown in Supplementary Fig. 7. Full results are reported in Supplementary Table 3.

Datasets

Cytoarchitecture

Cognitive consilience classification

We acquired three distinct cytoarchitectonic datasets, referred to as the cortical-laminar, Von Economo, and BigBrain cytoarchitectures. The first cytoarchitecture is based on an independent modular decomposition representing the five classic cortical laminar patterns described by Von Economo and Koskinas¹⁰⁵. Following the approach of Seidlitz et al.³⁵, we manually assigned each node from our 308-region parcellation to one of these five cortical classes, guided by the methodology outlined by Solari and Stoner⁴¹. The cingulate and insular cortices were assigned respectively to a sixth and a seventh class. In addition, we categorized subcortical regions–including the thalamus, hypothalamus, amygdala, hippocampus, and brainstem–into a distinct eighth class to account for their unique cytoarchitectonic characteristics. Finally, the cerebellum parcels were assigned to a ninth class, as they are composed of distinctive cell types (e.g., Purkinje cells) found exclusively in this region of the brain.

Von Economo similarity matrix

The Von Economo cytoarchitecture¹⁰⁶ represents a digitized version of the original cytoarchitectonic brain profiles described by Von Economo and Koskinas¹⁰⁵, forming a comprehensive brain atlas. This atlas comprises 48 ‘most important` distinct cortical areas and provides detailed layer-specific histological information, including neuronal count, neuron size, and cortical thickness. The Von Economo cytoarchitectonic similarity matrix was constructed as a pairwise Pearson correlation matrix of all 48 profiles.

Big brain similarity matrix

The BigBrain cytoarchitectonic profiles⁴² were derived from the ultrahigh-resolution BigBrain dataset⁴³ (see Methods Big Brain Similarity Matrix), where cortical profiles capturing the laminar cell number and density of the cortex were extracted. The resulting BigBrain similarity matrix, including pairwise similarity scores between N brain regions (excluding subcortical and cerebellar regions), was directly shared with us by the authors of Wei et al.⁴².

Genetic co-expression

Gene expression data were obtained from the publicly available dataset provided by the Allen Institute for Brain Sciences (AIBS). To reduce redundancy in complementary RNA (cRNA) hybridization probes measuring overlapping gene expression, expression values were averaged across probes targeting the same gene, while probes without matched genes were excluded. This process resulted in a dataset comprising 20,152 genes across 3,702 samples.

Given the symmetry in gene expression observed between hemispheres¹⁰⁷, AIBS data include samples from both hemispheres for only two subjects. However, due to the under-sampling of the right hemisphere, all analyses were restricted to the left hemisphere (n = 152 regions). Gene expression data for each subject were mapped onto the fsaverage (MNI152) volumetric template. This was achieved by assigning samples to the nearest centroid within the left hemisphere regions using T1-weighted MRI scans from each AIBS subject. For the two subjects with right hemisphere data, coordinates of the right hemisphere samples were reflected before mapping to ensure consistency.

Median gene expression values were calculated for each region across participants (N = 6) and standardized using z-scoring, resulting in a 152 × 20,152 matrix representing genome-wide expression across 152 left hemisphere regions. A 152 × 152 gene co-expression matrix was then generated by computing pairwise Pearson correlations of gene expression between each region. This co-expression matrix, along with regional expression values, was used for comparisons with the left hemisphere data of the group MetSiM. Gene ontology enrichment analysis was performed using GOATools¹⁰⁸.

Statistical analysis

MetSiM variablity

The variability of a set of metabolic similarity matrices was assessed by calculating the mean and standard deviation of variances across the set, and then performs a t-test to determine if the observed dispersion is significantly different from zero.

The variability of a set of metabolic similarity matrices with respect to the group average was determined by calculating the Pearson correlation coefficients between each matrix and the group average matrix. The resulting p-values were adjusted for multiple comparisons using the Benjamini–Hochberg method, yielding the average and standard deviation of the Fisher Z-transformed correlations, along with an overall p-value.

Spatial null models

To assess the extent to which the observed MetSiM structure could be explained by spatial autocorrelation alone, we implemented a family of distance-based random geometric null models (RandGeom). These models generate synthetic similarity matrices in which pairwise weights depend solely on the Euclidean distance between brain regions, while reproducing key statistical features of the empirical data.

RandGeom (full brain)

In the baseline RandGeom model, each pairwise MetSiM weight was sampled from a Gaussian distribution whose mean linearly decreased with inter-regional Euclidean distance, based on a linear regression fit to the empirical data. Formally, for any pair of brain regions i and j, the expected weight μ_ij was defined as: μ_ij = a ⋅ d_ij + b where d_ij is the Euclidean distance between the centroids of regions i and j, and a and b are regression coefficients estimated from the empirical MetSiM. To account for heteroscedasticity, we stratified distances into 2 mm bins and computed the residual variance within each bin. The variance ${\sigma }_{ij}^{2}$ of each pairwise connection was then sampled from the distribution of residuals corresponding to its distance bin ${M}_{ij}^{{{\rm{rand}}}} \sim {{\mathscr{N}}}({\mu }_{ij},{\sigma }_{ij}^{2}).$

RandGeom (within hemisphere)

To test whether hemispheric separation influenced model fit, we constrained the RandGeom model to within-hemisphere connections only. Separate synthetic matrices were generated for left and right hemispheres using the same procedure as described above, applied independently within each hemisphere.

RandGeom ⋅ GMAdj

To further impose a biologically plausible spatial constraint, we constructed a variant of the RandGeom model in which only region pairs that were adjacent in the cortical gray matter were eligible for sampling. Cortical adjacency was defined based on voxelwise Euclidean proximity of gray matter masks in MNI space. Pairwise weights were then sampled from the same distance-based distributions as in the full RandGeom model, but applied only to gray matter-adjacent region pairs.

Adjacent parcel permutation

To assess statistical significance when comparing spatially embedded datasets (e.g., similarity matrices, brain region classification scores) while preserving local spatial properties, we employed a customized spatial permutation method termed AdjPerm. This approach maintains local topology by restricting permutations to occur only between adjacent regions. Specifically, we defined adjacency based on anatomical proximity, considering each brain parcel to be adjacent to its immediate neighboring parcels. During each permutation, data values were randomly shuffled among these adjacent parcels, preserving local spatial dependencies. We fixed the number of unique permutations to 1000. By maintaining local topology, this method effectively accounts for inherent spatial autocorrelation without introducing anatomically implausible configurations.

Normalized mutual information score

We evaluated the agreement between two 3D scalar maps using the Normalized Mutual Information score. If both datasets were spatially embedded, we supplemented the significance of the overlap score with the AdjPerm test (see Methods Adjacent Parcel Permutation).

Inter-class overlap score

We evaluated the agreement between two labeled datasets using the Hungarian algorithm (Kuhn-Munkres algorithm). This approach identifies an optimal one-to-one mapping between the sets of labels, maximizing the number of samples that are consistently classified under this mapping. We supplemented the significance of the overlap score with the AdjPerm test (see Methods Adjacent Parcel Permutation).

Network based statistics

Weighted matrix binarization

Network graphs were generated by binarizing weighted matrices (similarity or connectivity) based on a specified edge density ρ. For matrices containing negative values, negative weights were rectified to their absolute values before applying the binarization threshold. This threshold was chosen to achieve the desired proportion of binarized connections ρ. Graph analyses were conducted across a wide range of edge densities (1 to 30%) for each weighted matrix.

Rich-Club

We computed the rich-club coefficient ϕ(k) for a given binarized network as

$$\phi (k)=\frac{2{E}_{\ > \ k}}{{N}_{\ > \ k}({N}_{\ > \ k}-1)}$$

(9)

where E_>k is the number of edges among nodes with a degree greater than k, and N_>k is the number of such nodes.

Degree centrality

We calculated the centrality of a node using degree centrality, which measures the number of direct connections a node has to other nodes in the network. For a node i, the degree centrality C_D(i) is defined as:

$${C}_{D}(i)={{\rm{deg}}}(i)$$

(10)

where deg(i) represents the number of edges connected to i. Degree centrality highlights the immediate influence of a node based on its local connectivity.

Random networks

For each binarized graph, we assessed the statistical significance of the observed network metrics by generating an ensemble of 1000 randomized networks using the Maslov-Sneppen rewiring procedure, which preserves the number of nodes, total number of edges, and the degree distribution of the original graph while randomizing the specific wiring of edges. We then computed empirical p-values for each topological metric by comparing the observed value to the distribution of corresponding values across the 1000 random graphs. Specifically, for each metric, the p-value was calculated as the fraction of randomized networks in which the metric was greater than or equal to the observed value.

Higher order connectivity

The higher-order connectivity of a weighted or binarized network matrix A was computed as the k-th order cosine similarity of A with itself:

$${({A}^{m})}_{ij}={\sum}_{{k}_{1},\ldots,{k}_{m-1}}{A}_{i,{k}_{1}}\cdot {A}_{{k}_{1},{k}_{2}}\cdot \ldots \cdot {A}_{{k}_{m-2},{k}_{m-1}}\cdot {A}_{{k}_{m-1},j}$$

(11)

where ∑_l,mA_k,l ⋅ A_m,n denotes the cosine product of the l-th node feature with the m-th node feature of A.

Network communicability model

To compute the influence of distant nodes from a structural connectivity matrix S on the state of the MetSiM M, we relied on a communicabilty model¹⁰⁹ and estimated the Green’s function of the network heat equation, which corresponds to the communicability expression:

$${{\bf{M}}}=\exp (g\cdot {{\bf{S}}}),$$

(12)

The decay parameter g controls the influence of distant nodes structural edges on the metabolic connectivity and was estimated by minimizing the mean squared loss between the model’s output (M) and the observed metabolic matrix (M_obs). Predictive performance was then quantified as the mean Spearman correlation between the predicted MetSiM (M) and the observed MetSiM (M_obs) given the optimal g.

Absolute nodal correlation

When comparing the nodal similarity between two matrices A and B of N nodes, we define the absolute nodal correlation of node i as the absolute value of the Pearson correlation between its connectivity profiles in A and B. Here, the connectivity profile of a matrix X is given by

$${{{\bf{x}}}}_{i}={\{{x}_{ij}\}}_{j\in \{1,\ldots,N\}},$$

which is the row vector containing all similarity measures of node i with the other nodes in the network.

Visualization software

Brain visualizations were generated programmatically in Python using Nilearn¹¹⁰ and FURY/DIPY¹¹¹; backgrounds were the MNI ICBM152 2009a template¹¹² and FreeSurfer fsaverage surfaces¹¹³. All software/templates permit inclusion of rendered images in commercial publications with proper attribution.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

The processed derivative data (parcellation masks, within subject and group-level metabolic similarity matrices, figure source data, and analysis scripts) are available at https://github.com/MRSI-Psychosis-UP/MRSI-Metabolic-Connectome. Source data underlying the main figures are provided with this paper in the accompanying Source Data files. The full tabulations underlying the statistical analyses are provided as Supplementary Data.

Code availability

The analysis code used in this study is available at GitHub: https://github.com/MRSI-Psychosis-UP/MRSI-Metabolic-Connectome. The exact version used for the results reported here is archived on Zenodo (https://doi.org/10.5281/zenodo.17293201) and should be cited as¹¹⁴.

References

Hagmann, P. From diffusion MRI to brain connectomics. Ph.d. thesis, EPFL, Lausanne, (2005).
Sporns, O., Tononi, G. & Kötter, R. The human connectome: a structural description of the human brain. PLoS Comput. Biol. 1, e42 (2005).
Article ADS PubMed PubMed Central Google Scholar
Biswal, B., Yetkin, F. Z., Haughton, V. M. & Hyde, J. S. Functional connectivity in the motor cortex of resting human brain using echo-planar mri. Magn. Reson. Med. 34, 537–541 (1995).
Article PubMed CAS Google Scholar
Chen, Y., Wang, S., Hilgetag, C. C. & Zhou, C. Features of spatial and functional segregation and integration of the primate connectome revealed by trade-off between wiring cost and efficiency. PLoS Comput. Biol. 13, e1005776 (2017).
Article ADS PubMed PubMed Central Google Scholar
Ma, J. et al. Trade-offs among cost, integration, and segregation in the human connectome. Netw. Neurosci. 7, 604–631 (2023).
Article PubMed PubMed Central Google Scholar
Bullmore, E. D. & Sporns, O. The economy of brain network organization. Nat. Rev. Neurosci. 13, 336–349 (2012).
Article PubMed CAS Google Scholar
van den Heuvel, M. P. & Sporns, O. A cross-disorder connectome landscape of brain dysconnectivity. Nat. Rev. Neurosci. 20, 435–446 (2019).
Article PubMed PubMed Central Google Scholar
Martin, G. Network analysis and the connectopathies: current research and future approaches. Nonlinear Dyn. Psychol. Life Sci. 16, 79–90 (2012).
MathSciNet Google Scholar
Stam, C. J. Modern network science of neurological disorders. Nat. Rev. Neurosci. 15, 683–695 (2014).
Article PubMed CAS Google Scholar
Bullmore, E. D. & Sporns, O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nat. Rev. Neurosci. 10, 186–198 (2009).
Article PubMed CAS Google Scholar
de Lange, S. C. et al. Shared vulnerability for connectome alterations across psychiatric and neurological brain disorders. Nat. Hum. Behav. 3, 988–998 (2019).
Article PubMed Google Scholar
Xia, M., Sun, X., Bu, X., Li, Q. & He, Y. Big connectome imaging data in psychiatric disorders. Med. Plus 1, 100038 (2024).
Vaishnavi, S. N. et al. Regional aerobic glycolysis in the human brain. Proc. Natl. Acad. Sci. USA 107, 17757–17762 (2010).
Article ADS PubMed PubMed Central CAS Google Scholar
Thomas, C. et al. Anatomical accuracy of brain connections derived from diffusion MRI tractography is inherently limited. Proc. Natl. Acad. Sci. USA 111, 16574–16579 (2014).
Article ADS PubMed PubMed Central CAS Google Scholar
Friston, K. Beyond phrenology: what can neuroimaging tell us about distributed circuitry? Annu. Rev. Neurosci. 25, 221–250 (2002).
Article PubMed CAS Google Scholar
Stephan, KlaasEnno et al. Nonlinear dynamic causal models for fMRI. Neuroimage 42, 649–662 (2008).
Article PubMed Google Scholar
Seghier, M. L., Zeidman, P., Neufeld, N. H., Leff, A. P. & Price, C. J. Identifying abnormal connectivity in patients using dynamic causal modeling of fMRI responses. Front. Syst. Neurosci. 4, 142 (2010).
PubMed PubMed Central Google Scholar
Heinzle, J. & Stephan, K. E. Dynamic causal modeling and its application to psychiatric disorders. In Computational Psychiatry 117–144 (Elsevier, 2018).
Klauser, A. et al. Fast high-resolution brain metabolite mapping on a clinical 3t mri by accelerated h-fid-mrsi and low-rank constrained reconstruction. Magn. Reson. Med. 81, 2841–2857 (2019).
Article PubMed CAS Google Scholar
Lecocq, Angèle et al. Whole-brain quantitative mapping of metabolites using short echo three-dimensional proton mrsi. J. Magn. Reson. Imaging 42, 280–289 (2015).
Article PubMed Google Scholar
Maudsley, A. A. et al. Mapping of brain metabolite distributions by volumetric proton MR spectroscopic imaging (MRSI). Magn. Reson. Med. Off. J. Int. Soc. Magn. Reson. Med. 61, 548–559 (2009).
Article CAS Google Scholar
Steel, A. et al. Metabolite-cycled density-weighted concentric rings k-space trajectory (dw-CRT) enables high-resolution 1 h magnetic resonance spectroscopic imaging at 3-Tesla. Sci. Rep. 8, 7792 (2018).
Article ADS PubMed PubMed Central Google Scholar
Moser, P. et al. Intra-session and inter-subject variability of 3d-fid-mrsi using single-echo volumetric epi navigators at 3t. Magn. Reson. Med. 83, 1920–1929 (2020).
Article PubMed Google Scholar
Klauser, A., Klauser, P., Grouiller, Frédéric, Courvoisier, S. & Lazeyras, F. Whole-brain high-resolution metabolite mapping with 3d compressed-sensing sense low-rank 1h fid-mrsi. NMR Biomed. 35, e4615 (2022).
Article PubMed Google Scholar
Klauser, A., Strasser, B., Thapa, B., Lazeyras, F. & Andronesi, O. Achieving high-resolution 1h-mrsi of the human brain with compressed-sensing and low-rank reconstruction at 7 Tesla. J. Magn. Reson. 331, 107048 (2021).
Article PubMed PubMed Central CAS Google Scholar
Klauser, A. et al. Eccentric: a fast and unrestrained approach for high-resolution in vivo metabolic imaging at ultra-high field MR. Imaging Neurosci. 2, 1–20 (2024).
Article Google Scholar
van der Graaf, M. In vivo magnetic resonance spectroscopy: basic methodology and clinical applications. Eur. Biophys. J. 39, 527–540 (2010).
Article PubMed Google Scholar
Li, X. et al. The distribution of major brain metabolites in normal adults: Short echo time whole-brain MR spectroscopic imaging findings. Metabolites 12, 543 (2022).
Article PubMed PubMed Central CAS Google Scholar
Du, Y. et al. Neurometabolite levels in the brains of patients with autism spectrum disorders: A meta-analysis of proton magnetic resonance spectroscopy studies (n= 1501). Mol. Psychiatry 28, 3092–3103 (2023).
Article PubMed CAS Google Scholar
Griffith, H. R. et al. Brain metabolism differs in alzheimer’s disease and parkinson’s disease dementia. Alzheimer’s. Dement. 4, 421–427 (2008).
Article CAS Google Scholar
Wang, A. M. et al. Assessing brain metabolism with 7-t proton magnetic resonance spectroscopy in patients with first-episode psychosis. JAMA Psychiatry 76, 314–323 (2019).
Article PubMed PubMed Central Google Scholar
Marder, E. & Goaillard, Jean-Marc Variability, compensation and homeostasis in neuron and network function. Nat. Rev. Neurosci. 7, 563–574 (2006).
Article PubMed CAS Google Scholar
Choi, W. T. et al. Metabolomics of mammalian brain reveals regional differences. BMC Syst. Biol. 12, 77–88 (2018).
Article Google Scholar
Piguet, C. et al. Randomized controlled trial of a mindfulness-based intervention in adolescents from the general population: The mindfulteen neuroimaging study protocol. Early Intervent. Psychiatry 16, 891–901 (2022).
Article Google Scholar
Seidlitz, J. et al. Morphometric similarity networks detect microscale cortical organization and predict inter-individual cognitive variation. Neuron 97, 231–247 (2018).
Article PubMed CAS Google Scholar
Baumann, PhilippSebastian et al. Treatment and early intervention in psychosis program (TIPP-L Ausanne): implementation of an early intervention programme for psychosis in Switzerland. Early Intervent. Psychiatry 7, 322–328 (2013).
Article Google Scholar
Hastie, T. & Stuetzle, W. Principal curves. J. Am. Stat. Assoc. 84, 502–516 (1989).
Article MathSciNet Google Scholar
Einbeck, J., Evers, L. & Powell, B. Data compression and regression through local principal curves and surfaces. Int. J. Neural Syst. 20, 177–192 (2010).
Article PubMed Google Scholar
Estrada, E. & Hatano, N. Communicability in complex networks. Phys. Rev. E–Stat. Nonlinear Soft Matter Phys. 77, 036111 (2008).
Article ADS MathSciNet Google Scholar
Haak, K. V., Marquand, A. F. & Beckmann, C. F. Connectopic mapping with resting-state fMRI. Neuroimage 170, 83–94 (2018).
Article PubMed Google Scholar
Van Hout Solari, S. & Stoner, RichMartin Cognitive consilience: primate non-primary neuroanatomical circuits underlying cognition. Front. Neuroanat. 5, 13956 (2011).
Google Scholar
Wei, Y., Scholtens, L. H., Turk, E. & Van Den Heuvel, M. P. Multiscale examination of cytoarchitectonic similarity and human brain connectivity. Netw. Neurosci. 3, 124–137 (2018).
Article PubMed PubMed Central Google Scholar
Amunts, K. et al. Bigbrain: an ultrahigh-resolution 3d human brain model. Science 340, 1472–1475 (2013).
Article ADS PubMed CAS Google Scholar
Hawrylycz, M. J. et al. An anatomically comprehensive atlas of the adult human brain transcriptome. Nature 489, 391–399 (2012).
Article ADS PubMed PubMed Central CAS Google Scholar
Hagmann, P. et al. Mapping the structural core of human cerebral cortex. PLoS Biol. 6, e159 (2008).
Article PubMed PubMed Central Google Scholar
Van Den Heuvel, M. P. & Fornito, A. Brain networks in schizophrenia. Neuropsychol. Rev. 24, 32–48 (2014).
Article PubMed Google Scholar
Barabási, Albert-L. ászló Network science. Philos. Trans. R. Soc. A Math. Phys. Eng. Sci. 371, 20120375 (2013).
Article ADS Google Scholar
Ratai, E. et al. Integrated imaging of [11C]-pbr28 PET, MR diffusion and magnetic resonance spectroscopy 1H-MRS in amyotrophic lateral sclerosis. Neuroimage Clin. 20, 357–364 (2018).
Article PubMed PubMed Central Google Scholar
Guan, J. et al. Detection and application of neurochemical profile by multiple regional 1H-MRS in parkinson’s disease. Brain Behav. 7, e00792 (2017).
Zimny, A. et al. In vivo evaluation of brain damage in the course of systemic lupus erythematosus using magnetic resonance spectroscopy, perfusion-weighted and diffusion-tensor imaging. Lupus 23, 10–19 (2013).
Article PubMed Google Scholar
Azevedo, FredericoA. C. et al. Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. J. Comp. Neurol. 513, 532–541 (2009).
Article PubMed Google Scholar
Von Bartheld, C. S., Bahney, J. & Herculano-Houzel, S. The search for true numbers of neurons and glial cells in the human brain: a review of 150 years of cell counting. J. Comp. Neurol. 524, 3865–3895 (2016).
Article Google Scholar
Mosharov, E. V. et al. A human brain map of mitochondrial respiratory capacity and diversity. Nature 1–10 (Nature Publishing Group, UK, London, 2025).
von Békésy, G. E. O. R. G. Travelling waves as frequency analysers in the cochlea. Nature 225, 1207–1209 (1970).
Article ADS Google Scholar
Bernhardt, B. C., Smallwood, J., Keilholz, S. & Margulies, D. S. Gradients in brain organization. NeuroImage 251, 118987 (2022).
Article PubMed Google Scholar
Yang, S. et al. Cortical patterning of morphometric similarity gradient reveals diverged hierarchical organization in sensory-motor cortices. Cell Rep. 36, 109582–109582 (2021).
Hill, J. et al. Similar patterns of cortical expansion during human development and evolution. Proc. Natl. Acad. Sci. USA 107, 13135–13140 (2010).
Article ADS PubMed PubMed Central CAS Google Scholar
Paquola, C. et al. Microstructural and functional gradients are increasingly dissociated in transmodal cortices. PLoS Biol. 17, e3000284 (2019).
Article PubMed PubMed Central CAS Google Scholar
Oldham, S. & Ball, G. A phylogenetically-conserved axis of thalamocortical connectivity in the human brain. Nat. Commun. 14, 6032 (2023).
Article ADS PubMed PubMed Central CAS Google Scholar
Hansen, J. Y. et al. Integrating multimodal and multiscale connectivity blueprints of the human cerebral cortex in health and disease. PLoS Biol. 21, e3002314 (2023).
Article PubMed PubMed Central CAS Google Scholar
Oldham, M. C., Horvath, S. & Geschwind, D. H. Conservation and evolution of gene coexpression networks in human and chimpanzee brains. Proc. Natl. Acad. Sci. USA 103, 17973–17978 (2006).
Article ADS PubMed PubMed Central CAS Google Scholar
Shin, J., Ming, G. L. & Song, H. Dna modifications in the mammalian brain. Philos. Trans. R. Soc. B Biol. Sci. 369, 20130512 (2014).
Article Google Scholar
Goulas, A., Betzel, R. F. & Hilgetag, C. C. Spatiotemporal ontogeny of brain wiring. Sci. Adv. 5, eaav9694 (2019).
Guo, J. U. et al. Neuronal activity modifies the DNA methylation landscape in the adult brain. Nat. Neurosci. 14, 1345–1351 (2011).
Article PubMed PubMed Central CAS Google Scholar
Scholtens, L. H., Schmidt, R., de Reus, M. A. & van den Heuvel, M. P. Linking macroscale graph analytical organization to microscale neuroarchitectonics in the macaque connectome. J. Neurosci. 34, 12192–12205 (2014).
Article PubMed PubMed Central CAS Google Scholar
Horwitz, B., Duara, R. & Rapoport, S. I. Intercorrelations of glucose metabolic rates between brain regions: application to healthy males in a state of reduced sensory input. J. Cereb. Blood Flow. Metab. 4, 484–499 (1984).
Article PubMed CAS Google Scholar
Horská, A. & Barker, P. B. Imaging of brain tumors: MR spectroscopy and metabolic imaging. Neuroimaging Clin. 20, 293–310 (2010).
Article Google Scholar
Wang, M., Yan, Z., Xiao, Shu-yun, Zuo, C. & Jiang, J. A novel metabolic connectome method to predict progression to mild cognitive impairment. Behav. Neurol. 2020, 2825037 (2020).
Article PubMed PubMed Central Google Scholar
Ceccarini, J., Liu, H., Van Laere, K., Morris, E. D. & Sander, C. Y. Methods for quantifying neurotransmitter dynamics in the living brain with PET imaging. Front. Physiol. 11, 792 (2020).
Article PubMed PubMed Central Google Scholar
Finnema, S. J. et al. Application of cross-species pet imaging to assess neurotransmitter release in brain. Psychopharmacology 232, 4129–4157 (2015).
Article PubMed PubMed Central CAS Google Scholar
Stiernman, LarsJonasson et al. Dissociations between glucose metabolism and blood oxygenation in the human default mode network revealed by simultaneous pet-fMRI. Proc. Natl. Acad. Sci. USA 118, e2021913118 (2021).
Article PubMed PubMed Central CAS Google Scholar
Shan, Y. et al. Integrated positron emission tomography/magnetic resonance imaging for resting-state functional and metabolic imaging in human brain: what is correlated and what is impacted. Front. Neurosci. 16, 824152 (2022).
Article PubMed PubMed Central Google Scholar
Pichler, B. J., Wehrl, H. F., Kolb, A. & Judenhofer, M. S. Positron emission tomography/magnetic resonance imaging: the next generation of multimodality imaging? Semin. Nucl. Med. 38, 199–208 (2008).
Article PubMed PubMed Central Google Scholar
Sander, C. Y., Hansen, H. D. & Wey, Hsiao-Ying Advances in simultaneous PET/MR for imaging neuroreceptor function. J. Cereb. Blood flow. Metab. 40, 1148–1166 (2020).
Article PubMed PubMed Central Google Scholar
Preisig, M., Fenton, B. T., Matthey, Marie-Louise, Berney, A. & Ferrero, François Diagnostic interview for genetic studies (DIGS): inter-rater and test-retest reliability of the French version. Eur. Arch. Psychiatry Clin. Neurosci. 249, 174–179 (1999).
Article PubMed CAS Google Scholar
Lin, A. et al. Minimum reporting standards for in vivo magnetic resonance spectroscopy (mrsinmrs): experts’ consensus recommendations. NMR Biomed 34, e4484 (2021).
Article PubMed PubMed Central Google Scholar
Aleman-Gomez, Y. Chimera: An open source framework for combining multiple parcellations. https://github.com/connectomicslab/chimera, (2024). Available on GitHub.
Cammoun, L. et al. Mapping the human connectome at multiple scales with diffusion spectrum MRI. J. Neurosci. methods 203, 386–397 (2012).
Article PubMed Google Scholar
Fischl, B. et al. Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron 33, 341–355 (2002).
Article PubMed CAS Google Scholar
Najdenovska, E. et al. In-vivo probabilistic atlas of human thalamic nuclei based on diffusion-weighted magnetic resonance imaging. Sci. Data 5, 1–11 (2018).
Article Google Scholar
Saygin, Z. M. et al. High-resolution magnetic resonance imaging reveals nuclei of the human amygdala: manual segmentation to automatic atlas. Neuroimage 155, 370–382 (2017).
Article PubMed CAS Google Scholar
Iglesias, JuanEugenio et al. A computational atlas of the hippocampal formation using ex vivo, ultra-high resolution MRI: application to adaptive segmentation of in vivo MRI. Neuroimage 115, 117–137 (2015).
Article PubMed Google Scholar
Billot, B. et al. Automated segmentation of the hypothalamus and associated subunits in brain MRI. Neuroimage 223, 117287 (2020).
Article PubMed Google Scholar
Diedrichsen, J. örn, Balsters, J. H., Flavell, J., Cussans, E. & Ramnani, N. A probabilistic MR atlas of the human cerebellum. Neuroimage 46, 39–46 (2009).
Article PubMed Google Scholar
Yendiki, A., Koldewyn, K., Kakunoori, S., Kanwisher, N. & Fischl, B. Spurious group differences due to head motion in a diffusion MRI study. Neuroimage 88, 79–90 (2014).
Article PubMed Google Scholar
Wedeen, V. J., Hagmann, P., Tseng, Wen-YihIsaac, Reese, T. G. & Weisskoff, R. M. Mapping complex tissue architecture with diffusion spectrum magnetic resonance imaging. Magn. Reson. Med. 54, 1377–1386 (2005).
Article PubMed Google Scholar
Tournier, J.-Donald et al. Mrtrix3: a fast, flexible and open software framework for medical image processing and visualisation. Neuroimage 202, 116137 (2019).
Article PubMed Google Scholar
Tournier, J.-Donald, Calamante, F. & Connelly, A. Mrtrix: diffusion tractography in crossing fiber regions. Int. J. Imaging Syst. Technol. 22, 53–66 (2012).
Article Google Scholar
Newlin, N. R., Rheault, François, Schilling, K. G. & Landman, B. A. Characterizing streamline count invariant graph measures of structural connectomes. J. Magn. Reson. Imaging 58, 1211–1220 (2023).
Article PubMed PubMed Central Google Scholar
Smith, R. E., Tournier, Jacques-Donald, Calamante, F. & Connelly, A. Sift: Spherical-deconvolution informed filtering of tractograms. Neuroimage 67, 298–312 (2013).
Article PubMed Google Scholar
Zalesky, A. et al. Connectome sensitivity or specificity: which is more important? Neuroimage 142, 407–420 (2016).
Article PubMed Google Scholar
de Reus, M. A. & van den Heuvel, M. P. Estimating false positives and negatives in brain networks. Neuroimage 70, 402–409 (2013).
Article PubMed Google Scholar
Provencher, S. W. Automatic quantitation of localized in vivo 1h spectra with LCModel. NMR Biomedicine: Int. J. Devoted Dev. Appl. Magn. Reson. Vivo 14, 260–264 (2001).
Article CAS Google Scholar
Tustison, N. J. et al. The ANTsX ecosystem for quantitative biological and medical imaging. Sci. Rep. 11, 9068 (2021).
Article ADS PubMed PubMed Central CAS Google Scholar
Thomas, B. A. et al. The importance of appropriate partial volume correction for PET quantification in alzheimer’s disease. Eur. J. Nucl. Med. Mol. imaging 38, 1104–1119 (2011).
Article PubMed Google Scholar
Gaser, C. et al. Cat: a computational anatomy toolbox for the analysis of structural MRI data. Gigascience 13, giae049 (2024).
Article PubMed PubMed Central Google Scholar
Ashburner, J. & Friston, K. J. Unified segmentation. neuroimage 26, 839–851 (2005).
Article PubMed Google Scholar
Thomas, B. A. et al. Petpvc: a toolbox for performing partial volume correction techniques in positron emission tomography. Phys. Med. Biol. 61, 7975 (2016).
Article PubMed Google Scholar
Schaefer, A. et al. Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Cereb. Cortex 28, 3095–3114 (2018).
Article PubMed Google Scholar
Urchs, S. et al. Mist: a multi-resolution parcellation of functional brain networks. MNI Open Res. 1, 3 (2019).
Article Google Scholar
Instrella, R. & Juchem, C. Uncertainty propagation in absolute metabolite quantification for in vivo mrs of the human brain. Magn. Reson. Med. 91, 1284–1300 (2024).
Article PubMed CAS Google Scholar
Kobak, D. & Berens, P. The art of using t-SNE for single-cell transcriptomics. Nat. Commun. 10, 5416 (2019).
Article ADS PubMed PubMed Central Google Scholar
Linderman, G. C. & Steinerberger, S. Clustering with t-SNE, provably. SIAM J. Math. Data Sci. 1, 313–332 (2019).
Article MathSciNet PubMed PubMed Central Google Scholar
Camastra, F. & Staiano, A. Intrinsic dimension estimation: advances and open problems. Inf. Sci. 328, 26–41 (2016).
Article Google Scholar
von Economo, Constantin Freiherr & Koskinas, G. N. Die Cytoarchitektonik der Hirnrinde des Erwachsenen Menschen. J. (Springer, 1925).
van den Heuvel, M. P., Scholtens, L. H., Barrett, LisaFeldman, Hilgetag, C. C. & de Reus, M. A. Bridging cytoarchitectonics and connectomics in human cerebral cortex. J. Neurosci. 35, 13943–13948 (2015).
Article PubMed PubMed Central Google Scholar
Pletikos, M. et al. Temporal specification and bilaterality of human neocortical topographic gene expression. Neuron 81, 321–332 (2014).
Article PubMed CAS Google Scholar
Klopfenstein, D. V. et al. Goatools: a Python library for gene ontology analyses. Sci. Rep. 8, 10872 (2018).
Article ADS PubMed PubMed Central CAS Google Scholar
Kondor, Risi Imre & Lafferty, J. Diffusion kernels on graphs and other discrete structures. In Proceedings of the 19th International Conference on Machine Learning, 2002, 315–322, (2002).
Abraham, A. et al. Machine learning for neuroimaging with scikit-learn. Front. Neuroinform. 8, 14 (2014).
Article PubMed PubMed Central Google Scholar
Garyfallidis, E. et al. Dipy, a library for the analysis of diffusion MRI data. Front. Neuroinform. 8, 8 (2014).
Article PubMed PubMed Central Google Scholar
Fonov, V. et al. Unbiased average age-appropriate atlases for pediatric studies. Neuroimage 54, 313–327 (2011).
Article PubMed Google Scholar
Fischl, B. Freesurfer. Neuroimage 62, 774–781 (2012).
Article PubMed Google Scholar
Lucchetti, F., Céléreau, L. & Klauser, P. Mrsi-metabolic-connectome (v1.0.0), (2025). Source code for the manuscript: “Constructing the Human Brain Metabolic Connectome with MR Spectroscopic Imaging for Insights into Biochemical Organization".

Download references

Acknowledgements

This project was supported by the Swiss National Science Foundation (grant number 215728); P.K. was supported by a fellowship from the Adrian & Simone Frutiger foundation. We gratefully acknowledge Jacob Seidliz for providing the digitized version of cytoarchitectonic classes based on the Cognitive-Consilience dataset, and Martijn van den Heuvel and Yongbin Wei for providing the Big Brain cytoarchitectonic similarity matrix. Moreover, we gratefully acknowledge Jean-Baptiste Ledoux for the acquisition of MRI data from the Lausanne study, and Arnaud Merglen and Camille Marie Piguet for sharing the MRI dataset of the Geneva study.

Author information

Authors and Affiliations

Center for Psychiatric Neuroscience, Department of Psychiatry, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland
Federico Lucchetti, Edgar Céléreau, Pascal Steullet & Paul Klauser
Connectomics Lab, Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland
Yasser Alemán-Gómez & Patric Hagmann
Swiss Innovation Hub, Siemens Healthineers International AG, Lausanne, Switzerland
Antoine Klauser
Division of Child and Adolescent Psychiatry, Department of Psychiatry, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland
Paul Klauser

Authors

Federico Lucchetti
View author publications
Search author on:PubMed Google Scholar
Edgar Céléreau
View author publications
Search author on:PubMed Google Scholar
Pascal Steullet
View author publications
Search author on:PubMed Google Scholar
Yasser Alemán-Gómez
View author publications
Search author on:PubMed Google Scholar
Patric Hagmann
View author publications
Search author on:PubMed Google Scholar
Antoine Klauser
View author publications
Search author on:PubMed Google Scholar
Paul Klauser
View author publications
Search author on:PubMed Google Scholar

Contributions

F.L. developed the methodology, performed the data analysis, conducted the experiments, contributed to data interpretation, and drafted the initial manuscript. P.S., and P.H. provided critical revisions. A.K. developed the MRSI technique and provided critical revisions. E.C., Y.A.-G. performed part of the data preprocessing and provided critical revisions. P.K. conceptualized the study, supervised the project, and provided critical revisions.

Corresponding author

Correspondence to Paul Klauser.

Ethics declarations

Competing interests

Antoine Klauser is employed by Siemens Healthcare as a research scientist. He contributes to on-site research and development initiatives. He does not hold fiduciary responsibilities for Siemens Healthcare. All the other authors declare no competing interests.

Peer review

Peer review information

Nature Communications thanks Paul Mullins and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information (download PDF )

Description of Additional Supplementary Files (download PDF )

Supplementary Data 1 (download CSV )

Supplementary Data 2 (download CSV )

Supplementary Data 3 (download CSV )

Reporting Summary (download PDF )

Transparent Peer Review file (download PDF )

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Lucchetti, F., Céléreau, E., Steullet, P. et al. Constructing the human brain metabolic connectome with MR spectroscopic imaging reveals cerebral biochemical organization. Nat Commun 16, 11344 (2025). https://doi.org/10.1038/s41467-025-66124-w

Download citation

Received: 26 March 2025
Accepted: 30 October 2025
Published: 22 December 2025
Version of record: 23 December 2025
DOI: https://doi.org/10.1038/s41467-025-66124-w

This article is cited by

Constructing the human brain metabolic connectome with MR spectroscopic imaging reveals cerebral biochemical organization
- Federico Lucchetti
- Edgar Céléreau
- Paul Klauser
Nature Communications (2025)