Lipidome atlas of the adult human brain

The study addresses how lipid composition varies across the adult human brain and how this variation relates to structural (e.g., myelin content, cell-type composition) and functional (e.g., information-processing hierarchy, functional connectivity) organization. Prior molecular atlases have focused largely on gene expression and single-cell profiles, while macroscale brain studies leverage imaging to map structure and function. Lipids constitute a major fraction of brain tissue and influence membrane properties, signaling, metabolism, and disease, yet comprehensive, region-resolved lipidome maps integrating structural and functional brain features have been lacking. The purpose is to systematically map lipid abundance across 75 human brain regions (with macaque validation), classify lipid profiles in relation to myelin and cell types, and test associations with hierarchical processing and functional connectivity, thereby bridging microscale molecular content with macroscale brain organization.

Classical and modern studies have characterized lipid classes in brain tissue, noting differences between gray and white matter, myelin membranes, synaptic membranes, and major cell types. Sphingolipids (e.g., ceramides), cholesterol, and phospholipids are key components, with myelin enriched in sphingolipids and cholesterol. Mass spectrometry has expanded compound-level resolution in selected regions and cell models, and rodent work shows region-, cell type-, projection-, and age-related lipidome differences. Lipids underpin membrane biophysics, metabolic coupling, differentiation, signaling, inflammation control, and protein complex modulation. Altered lipidomes have been implicated in autism, schizophrenia, Alzheimer’s disease, alcohol-related brain damage, and aging. However, prior regional brain lipidomics typically examined fewer than ten regions, limiting integration with structural networks and functional connectivity, motivating a large-scale, multi-region lipid atlas linked to brain architecture.

Samples: Post-mortem adult human brain tissue from four neurotypical individuals (75 anatomically and functionally distinct regions across 11 structures, including 56 neocortical regions) and three adult macaques (38 homologous regions). Tissue was dissected from frozen slices using standardized protocols and stored at −80 °C.

Lipidomics: Two complementary MS approaches were used. (i) Untargeted high-resolution mass spectrometry (HRMS, Bruker Impact II QTOF) in positive and negative modes after UPLC separation on BEH C8, yielding 419 annotated lipids across 21 LIPID MAPS subclasses (with defined exceptions). (ii) Targeted LC-MS/MS multiple reaction monitoring (MRM, Agilent 6495 Triple Quadrupole) measuring 216 lipids, with 169 overlapping HRMS/MRM. Sample prep included MTBE-based extraction spiked with internal standards, QC pools, blanks, and instrument conditioning; stringent filtering excluded features with excessive blank signal, poor QC CV (>25%), or low dilution-series correlation.

Lipid identification: Custom m/z databases per class, adduct-aware matching (10 ppm), retention time grid evaluation, cross-mode RT validation, and MS/MS fragmentation on pooled samples (Q Exactive, DDA) confirmed class and inferred fatty acyl composition where possible (MSI level 2; FA acyl/alkyl level). Ether/plasmalogen annotations leveraged chromatographic behavior and standards.

Data processing: Intensities (peak areas) were log10-transformed, normalized by internal standards and sample wet weight, and then per-individual mean-centered to reduce inter-individual variation. HRMS included XCMS-based feature detection, calibration, retention filters, fillPeaks, and imputation. For MRM, MassHunter integration and QC thresholds were applied.

Region alignment and imaging data: Sample coordinates were mapped to MNI space. Myelin proxy was estimated using T1w/T2w sMRI from Human Connectome Project (humans) and analogous macaque data; mean signal was extracted at region coordinates. Functional connectivity (rs-fMRI) was compiled for 59/75 regions from a connectivity-based atlas.

RNA-seq: For 35 regions (subset of 75), matched-tissue RNA-seq was performed (or reused) with standard trimming, alignment (HISAT2), quantification (StringTie), and normalization to TPM; values were per-individual mean-centered.

Analyses: Global variation visualized by t-SNE and PCA. ANOVA tested regional differences per lipid (BH-adjusted p<0.01). Lipid categories were defined by linear models of lipid intensity vs myelin content: significant positive association (myelin+), negative (myelin−), non-myelin but regionally varying (unexplained), non-varying (housekeeping), or inter-individually variable (variable). Enrichment of lipid classes, unsaturation, chain length, and predicted membrane fluidity (composite of normalized double bonds and chain length) were assessed (hypergeometric and Mann–Whitney tests). Cell-type associations used Pearson correlations between lipid intensity profiles and marker-gene expression profiles for six cell types (OD, MG, OPC, inhibitory, excitatory, astrocytes) across 35 regions (threshold R>0.5 for assignment; enrichment by hypergeometric tests). Cross-species robustness was examined by profile correlations and consistency of category assignments in macaques. Associations with functional hierarchy (four-level neocortical hierarchy) used linear models; FC associations used Mantel tests comparing lipid outer-product matrices to FC matrices. Spatial validation employed MALDI imaging (Orbitrap) on cortical sections and ToF-SIMS on cerebellar sections and standards to visualize lipid class/headgroup ions and WM/GM contrasts. Multiple comparisons were BH-corrected where specified.

• Coverage and variability: The human brain lipidome comprised 419 HRMS-annotated lipids (21 classes); targeted MRM covered 216 lipids (169 overlapping). Macaque data covered 394/419 HRMS lipids and all 216 MRM. Normalized lipid intensities varied significantly among regions for 391/419 (93%) human lipids and 279/394 (71%) macaque lipids (ANOVA, BH p<0.01). Cross-technique and cross-species lipid profiles correlated significantly beyond chance (Mann–Whitney U, p<0.0001).
• Myelin association: T1w/T2w myelin proxy strongly correlated with lipidome PC1 (humans R=0.78; macaques R=0.77; p<0.0001). Lipid categorization yielded: myelin+ 65% (N=274), myelin− 17% (N=71), unexplained 11% (N=46), housekeeping 5% (N=20), variable 2% (N=8). Category assignments were consistent in macaques for 84% of overlapping HRMS and 88% of MRM lipids. MALDI imaging in human cortex and comparison to published WM/GM datasets validated myelin+/myelin− assignments (Fisher p<0.05), and GO analysis linked myelin+ to axon/myelin terms and myelin− to synaptic terms.
• Biochemical distinctions: Myelin-associated lipids showed class enrichments (e.g., sphingolipids, DG, sulfatides) and characteristic FA properties (e.g., PUFAs in PE/PE-P; class-specific chain lengths). Housekeeping lipids were enriched in free fatty acids (hypergeometric p<0.001) with unsaturated, short-chain FAs. Predicted membrane fluidity differed by category, with myelin+ exhibiting highest and myelin− lowest fluidity (Mann–Whitney, BH p<0.001). ToF-SIMS imaging of headgroup ions across 12 classes supported spatial distributions.
• Cell-type associations: 353/419 (84%) lipids correlated with at least one cell-type marker profile across 35 regions (R>0.5, p<0.01). Most cell-type-associated myelin+ lipids aligned with oligodendrocyte markers (~82%), while myelin− lipids aligned with inhibitory and excitatory neuronal markers. Unexplained lipids were enriched for astrocytes, OPCs, and inhibitory neurons (BH p<0.05). Cross-species comparison with mouse data showed significant profile similarity (Pearson R=0.66, p=0.00012), and lipid-class-specific cell-type assignments (HexCer, PE) were conserved (BH p<0.001). Sorted mouse pyramidal neurons showed enrichment of human neuron-assigned lipids (Wilcoxon p=0.0077).
• Functional architecture links: The lipidome PC2 correlated with neocortical processing hierarchy (R=0.58, p=3.4×10⁻⁶). Sixty lipids associated with hierarchy (HR lipids): 26 positively (HR+) and 34 negatively (HR−). HR+ lipids were overrepresented in myelin category and phosphatidylcholines and enriched for omega-3 DHA-containing PUFAs; HR− lipids were enriched in sphingomyelins with saturated/oligo-unsaturated residues and showed omega-6 prevalence (BH p<0.05). Cell-type association mapped HR+ to astrocytes/OPCs/microglia and HR− to oligodendrocytes/inhibitory neurons. The lipidome also related to functional connectivity: FC PC1 inversely correlated with lipidome PC2 (R=0.35, p=0.006). Seventy-six lipids showed significant lipid–FC profile correlations (Mantel p<0.005), with myelin+ lipids correlating best (Mann–Whitney p=1.5×10⁻⁴). Sulfatides, hexosylceramides, and diacylglycerols exhibited stronger FC associations than the bulk (BH p<0.05).
• Regional gradients and class/unsaturation patterns: Cholesterol enriched in subcortical/WM; PUFA-containing lipids (e.g., DHA 22:6) decreased in WM but varied across neocortex (lower in prefrontal; higher in motor/visual/parietal areas).

This work demonstrates that the adult human brain lipidome varies substantially across regions and that these variations align with both structural features (myelin content, cell-type composition) and functional architecture (processing hierarchy, functional connectivity). The strong coupling between myelin proxies and lipid profiles accounts for most lipid variation, with conserved patterns across macaques, consistent with myelin’s distinct biochemical signature. Nonetheless, nearly one-fifth of lipids follow myelin-independent regional patterns; these include an “unexplained” category enriched for astrocyte/OPC associations and show lower human–macaque conservation, consistent with faster astrocyte evolution. Lipids tracking hierarchical information processing segregate into biochemically distinct groups: HR+ lipids are enriched for PCs and omega-3 DHA and map to glial populations (astrocytes/OPCs), whereas HR− lipids are enriched for sphingomyelins, omega-6 residues, and map to oligodendrocytes and inhibitory neurons. Thus, lipid composition gradients may reflect varying neuronal and glial densities or subtypes along cortical hierarchies. In contrast, correlations with resting-state FC are strongest for myelin+ and myelin-enriched classes (sulfatides, HexCer, DG), supporting a link between axonal myelination properties and network synchrony. Together, the findings bridge molecular lipid composition with macroscale brain structure and function, providing a reference for future studies of regionally distributed neurological and psychiatric disorders.

The study provides a high-resolution lipidome atlas across 75 adult human brain regions, integrating molecular lipid data with myelin content, cell-type composition, hierarchical information processing, and functional connectivity. It identifies robust, conserved myelin-associated lipid categories, reveals biochemically and cell-type-distinct lipid sets linked to processing hierarchy, and shows that myelin-enriched lipids relate most strongly to functional connectivity. Validation across macaques, targeted MRM, MALDI and ToF-SIMS imaging, and sorted neuronal lipidomes supports robustness. The atlas establishes a baseline for interpreting lipid alterations in brain disorders. Future work should expand sample sizes and demographics, include additional lipid classes (e.g., cardiolipins, gangliosides), refine cell-subtype resolution and spatial mapping, and extend coverage of subcortical and white matter tracts to deepen structure–function–lipidome integration.

• Sample size: Only four human brains (and three macaques) may limit generalizability and power to detect region-specific effects; some lipids could be misassigned to housekeeping/variable categories.
• Class coverage: Important classes like cardiolipins and gangliosides were not included in the LC-MS pipeline, potentially missing mitochondrial and glycosphingolipid biology.
• Cell-type assignment: Lipid–cell-type links relied on correlation with tissue-level mRNA markers across regions, limiting subtype resolution and susceptible to correlated marker profiles.
• Regional sampling bias: Emphasis on neocortex; less detailed coverage of subcortical structures, axonal tracts, and infracortical white matter.
• Spatial resolution: LC-MS lacks intraregional spatial context; partially addressed via MALDI and ToF-SIMS in selected regions.
• Normalization: Per-individual mean-centering improves regional comparability but precludes absolute abundance comparisons between different lipids.