Foamy microglia link oxylipins to disease progression in multiple sclerosis

Human brain tissue

Human brain samples from 38 donors (28 secondary progressive MS and 10 matched nondemented controls) were obtained from the NBB (brainbank.nl, project 1327). All procedures of the NBB were approved by the Medical Ethical Committee of Amsterdam Academic Medical Centre, and all donors gave written consent to the NBB for the use of their data and tissue for research. Fresh-frozen tissue blocks containing subcortical WM with or without MS lesions were dissected, snap-frozen in liquid nitrogen and stored at −80 °C. Mirrored pieces of this tissue were formalin fixed and paraffin embedded and stained to characterize different MS lesion stages (mixed active/inactive, active, inactive, remyelinated) by the NBB. Selection criteria for donors were a postmortem delay (time interval between the demise of the donor and freezing of the tissue) below 12 h, a pH of the CSF higher than 5.5 and a clinician-confirmed diagnosis of secondary progressive MS for patients with MS and the absence of any neurological disease in controls, nor neurological disease-indicating pathology found in their brains as examined by a neuropathologist. The postmortem delay ranged between 6.5 to 11.5 h, with a median of 8.75 h. The pH of the CSF ranged from 5.8 to 6.8, with a median of 6.4. Donor characteristics were matched as closely as possible (Extended Data Fig. 1a), but there were differences in age between controls and MS donors.

Lesion characterization and definitions

Lesions were classified using our previously defined criteria^1,6 in line with the classification system mentioned in ref. ¹¹. In short, CWM is WM from nondemented control donors, absent any sign of demyelination or aberrant microglial activation. NAWM is macroscopically intact WM from a patient with MS, and within the same tissue block, no lesion or inflammation is visible. PLWM is WM adjacent to a WM lesion. Depending on the lesion type the PLWM is adjacent to, sometimes there is considerable inflammation, or even the presence of foamy microglia. ALs (type 2, also called acute) were defined as (partially) demyelinated areas with increased microglial presence throughout the lesion and no hypocellular core. MLs (type 3, also called mixed, chronic active, smoldering or slowly expanding) were defined as completely demyelinated areas with a hypocellular gliotic core, with an active rim of macrophages at the border of demyelination. Inactive lesions (IL, type 4) were characterized by complete demyelination, the absence of an active HLA⁺ rim and a hypocellular gliotic core. Remyelinated lesions (RL, type 6, also called shadow plaques) were characterized by partial myelination. In addition, residual myelin should not only be myelin debris, but also axonal myelin. Microglial activation was sometimes still present, but not to a larger density of cells than the surrounding PLWM. In addition to these classically defined lesion types, we profile the morphology of the microglia found in the AL and ML⁶⁶. AL and ML with either foamy or ramified and ameboid (together called nonfoamy) microglia were distinguished, in line with our previously published work^6,17.

Analysis of lesion proportions and clinical features

Clinical and pathological data were obtained from the NBB for MS brain donors collected between 1990 and 2021. All lesions identified by macroscopic inspection, or on postmortem magnetic resonance imaging appearance, throughout the brain and in three standard locations in the brain stem and spinal cord were pathologically characterized as described above and are analyzed for the clinical correlations. In total, 8,708 lesions were analyzed in 250 MS donors (average of 34.8 lesions/donor). Proportions of lesions were defined as counts of lesions of interest divided by all WM lesions. Time to EDSS6 and EDSS8 was defined as the year in which EDSS6 or EDSS8 was reached minus the year in which the first symptoms were reported. Disease duration was defined as the number of years between the first reported symptoms and death. Lesion counts were correlated to clinical features using a generalized linear model with a quasibinomial distribution of counts of lesions against other lesions, as previously described in ref. ⁶. P values to test the model were calculated using the likelihood-ratio test and corrected for multiple testing using the Bonferroni method.

Statistics, randomization, blinding and differential expression

No power analysis was performed to predetermine sample sizes, but our sample sizes are similar to those reported in previous publications^24,25,95. Samples were always randomized before analysis. Researchers were not blinded to classifications of samples, because our measurements are quantitative and not sensitive to subjective interpretation. A replicate refers to a unique lesion isolated. Never in this study is the same lesion measured repeatedly. The number of unique lesions isolated per lesion classification is specified in Table 1. Not every lesion was successfully analyzed by all techniques described in this study; missing lesions are specified in Table 1 as well. No datapoints have been excluded after obtaining the data. The only reason a datapoint is not in the analysis is the failure to obtain it.

All differential expression/abundance analyses on -omics analyses were performed using the limma linear modeling framework⁹⁶. Normality and equal variances were not formally tested.

For the RNA-seq data, we constructed a linear model taking into account covariates, using the formula y = ~0 + lesion type + RQN + sex, with lesion type being the design and RQN (RNA quality number) and sex being covariates. We ran limma’s linear modeling function ‘lmFit’ using the modeling design as above, using ‘voom’ for precision weight analysis⁹⁷. For all other datasets, we ran the lmFit function on log₂-expression values without ‘voom’ and without taking any covariates along in the model (y = ~0 + lesion type).

For calculating differential expression between foamy and nonfoamy microglia, which can appear in several backgrounds (MLs, ALs and also PLWM), we constructed a linear model using lesion type as a covariate, without specifying the morphology (seven groups—CWM, NAWM, PLWM, AL, ML, IL and RL) and specifying the morphology as separate parameters. For the RNA-seq, the linear model was y = ~0 + morphology + lesion type (seven groups) + RQN + sex, and for all the other datasets, the linear model was y = ~0 + morphology + lesion type (seven groups).

These linear models were fed into the limma eBayes algorithm with default settings. Thus, log₂ fold changes and Benjamini–Hochberg (BH) corrected P values were extracted using the TopTable function (two-sided tests). We generally accept a false discovery rate (FDR) of 10%, meaning that we consider variables with a BH-adjusted P value lower than 0.1 statistically significant.

Cryosectioning and isolation of MS lesions

Fresh-frozen tissue was sectioned at 10 μm for stainings or 50 µm for tissue collection in a cryostat (Thermo Fisher Scientific, HM525 NX) between −20 °C and −16 °C with Thermo Fisher Scientific MX35 Ultra microtome blades. Sections for stainings were mounted on Epredia Superfrost slides (J1800AMNZ), dried overnight using silica beads and subsequently sealed and stored at −80 °C until further use. Borders of a demyelinated lesion were identified based on hematoxylin and Sudan Black staining and inspection under a light microscope. A surgical scalpel was used to demarcate the lesioned area, after which the lesion and peri-lesion were separated from 50 µm sections in a cryostat. Continuous-stained sections were prepared to ensure the lesion borders were closely matched by the isolation method. Before and after every sample collected a slide, it was prepared for histological confirmation of lesion pathology. Tissue was collected in 1.5 ml sterile RNase-free SafeLock Eppendorf tubes, weighed and stored at −80 °C. Samples isolated for RNA extraction were stored in Trisure (Bioline, BIO-38032).

Immunohistochemistry (single staining, CD79A, GFAP, CD3, TBXAS1)

Frozen sections were taken from the −80 °C and warmed up to room temperature before opening the seal. The sections were then incubated in 4% PFA in PBS for 30 min, followed by peroxide blocking (1% H₂O₂, 0.5% Triton X-100 in PBS) for 20 min. The sections were subsequently incubated overnight at 4 °C with one of the following primary antibodies: GFAP (1:1,000, D1F4Q, Cell Signaling Technology), CD79A (1:200, M705001-2, DAKO), CD3 (1:100, A0452, DAKO) or TBXAS1 (1:200, 160715, Cayman Chemical) in PBS containing 1% BSA, 0.5% Triton X-100 and 10% horse serum. Subsequently, secondary antibody was applied from the DAKO REAL EnVision Kit (DAKO, K500711-2) for 1 h at room temperature. Horseradish peroxidase immunoreactivity was then visualized with 3,3′-diaminobenzidine (DAB) using the same kit according to the manufacturer’s instructions (substrate was diluted 1:50 in buffer). Then, the sections were washed and stained in hematoxylin solution (50 g l⁻¹ KAl(SO₄), 1 g l⁻¹ hematoxylin, 0.2 g l⁻¹ NaIO₃, 50 g l⁻¹ chloral hydrate, 1 g l⁻¹ citric acid). After washing, the sections were dehydrated in a dehydration series of 50% EtOH (3 min), 70% EtOH (3 min), 96% EtOH (5 min), 100% EtOH (5 min), 100% EtOH (5 min), xylene (10 min) and xylene again (10 min). Then, the sections were dried and mounted in Entellan (Merck Millipore, 1079600500) using 24 × 50 mm coverslips (Corning, CLS2975245). Sections were imaged on a Zeiss Axioscan 7 and further processed using QuPath software (v0.5.0)⁹⁸.

Immunohistochemistry (double staining, HLA⁺ PLP)

Sections were fixed and peroxide blocked as described above and subsequently incubated in primary antibodies against HLA-DR/DQ/DP (1:1000; DAKO, M0775) in 1% BSA, 0.5% Triton X-100, 10% horse serum in PBS for 1 h at room temperature. Subsequently, the DAKO REAL EnVision Kit (DAKO, K500711-2) was applied for 1 h at room temperature. Horseradish peroxidase immunoreactivity was then visualized using DAB/Ni²⁺ (0.5 mg ml⁻¹ DAB tetrahydrochloride, 2 mg ml⁻¹ nickel ammonium sulfate, 0.01% H₂O₂ in PBS) for ~10 min. Subsequently, a PLP antibody was incubated (1:1,000, clone plpc1; Bio-Rad, MCA839G) in 1% BSA, 0.5% Triton X-100, 10% horse serum in TBS overnight at 4 °C. The next day, DAB staining was developed and the sections were mounted as described for the single staining.

Immunofluorescence stainings and analysis (GPNMB, PLIN2, LAMP1)

For GPNMB and HLA, frozen sections were used; these were taken from the −80 °C and warmed up to room temperature, before fixing in 4% PFA in PBS for 30 min.

Stainings for LAMP1 and PLIN2 were performed in formalin-fixed paraffin-embedded tissue sections (8 µm). Formalin-fixed paraffin-embedded sections were deparaffinized in xylene and subsequently rehydrated through a series of ethanol to demineralized water, followed by antigen retrieval in citrate buffer of pH 6 in the microwave at 700 W for 14 min.

Subsequently, sections were blocked (10% donkey serum, 0.5% Triton X-100 in PBS) for 2 h at room temperature. Sections were subsequently incubated with primary antibodies against GPNMB (1:500; Cell Signaling Technology, E4D7P), HLA-DR/DQ/DP (1:1,000; DAKO, M0775), LAMP1 (1:200; Abcam, ab24170) and PLIN2 (1:200; Progen Biotechnik, GP40), in 0.5% Triton X-100, 5% donkey serum in PBS, overnight at 4 °C. Appropriate secondary antibodies (1:1,000; Jackson Immuno, donkey anti-IgG (H + L)) were incubated subsequently for 2 h at room temperature in PBS with 0.5% Ttriton X-100. Autofluorescence was quenched by incubating the section in 0.5% Sudan Black B in 70% ethanol for 2 min. Subsequently, the sections were briefly rinsed in 50% ethanol and then in PBS. Nuclei were stained with 1 µg ml⁻¹ DAPI (Sigma) in PBS and the sections were mounted under a coverglass (Corning) in mounting medium (100 mg ml⁻¹ Mowiol 4-88, 25% glycerol, 25 mg ml⁻¹ DABCO in 0.1 M Tris–HCl, pH 8.5).

Images were captured using an Axioscan 7 slide scanner (Zeiss), and quantification of positive cells in each lesion was performed using QuPath (v0.5.0). Regions of interest (ROIs) were selected based on demyelination and the border of microglia/macrophages in adjacent HLA/PLP stainings. Then, analysis of HLA and GPNMB was performed using ‘Cell detection’ and ‘Object classifier’ commands. Cell detections were performed within the ROI using DAPI staining to detect cellular nuclei. A Random trees classifier was trained to separate microglia/macrophages from other detections. Objects were filtered for ‘cells’, and all features with the correct channel were selected. Approximately 50 points for each class were annotated by hand. The classifiers were loaded and applied sequentially to the selected image areas. The measurements were exported and further analyzed in R (v4.4.1).

ORO staining

Neutral lipid content in lipid droplets was visualized by an ORO staining¹⁶. ORO powder (Sigma) was dissolved in 100% isopropanol at 0.5 g l⁻¹ (stock). The working solution was prepared by mixing the ORO stock with water in a ratio of 5:4 (55.6% IPA with ORO, 44.4% H₂O) and heating to 60 °C. Sections were taken from −80 °C, warmed up to room temperature, fixed in 4% PFA (15 min, room temperature) and subsequently washed in demineralized water and isopropanol/water (in a ratio of 5:4) before incubating the sections for 15 min in ORO working solution at 60 °C. Then, sections were washed in isopropanol/water (5:4) before further washing in demineralized water. Hematoxylin staining was performed as described above, and the sections were then washed in PBS and mounted in Mowiol 4-88 mounting medium.

MGLL and CD68 ISH

Duplex in ISH was performed using the Ventana Discovery Ultra automated staining platform (Roche Diagnostics), following the manufacturer’s standard protocol for RNA detection. Frozen sections were postfixed with 4% PFA and treated with Cell Conditioning 1 (Roche, 950-124) for target retrieval. RNAscope probes targeting MGLL and CD68 mRNA (Advanced Cell Diagnostics, Bio-Techne) were hybridized at 43 °C for 2 h using the Discovery Ultra’s onboard protocols. Signal amplification and detection were performed using duplex chromogenic labeling, with MGLL visualized in green and CD68 in red using proprietary chromogenic substrates provided in the RNAscope Duplex Detection Kit. All reagents and amplification steps were integrated into the Ventana automated workflow. Slides were counterstained with hematoxylin, dehydrated in graded alcohols, cleared in xylene and mounted using VectaMount (Vector Laboratories). Whole-slide brightfield images were acquired at ×40 magnification using a Hamamatsu NanoZoomer digital slide scanner. Quantitative analysis of ISH signals was performed using the HALO image analysis platform (Indica Labs), applying the RNAscope ISH algorithm module. ROIs were manually annotated to define specific areas based on adjacent myelin and microglia stainings (HLA/PLP). The algorithm was configured to detect and quantify red (CD68) and green (MGLL) chromogenic puncta per cell. Positive cells were defined based on a minimum threshold of signal dots per nucleus, with settings optimized for signal-to-noise ratio and validated across control and experimental samples. Data were exported as the number of positive cells per mm².

MAGL activity-based histology using LEI-463-Cy5

The activity-based histology procedure was performed as detailed in ref. ⁶⁴. In short, tissue was freshly sectioned at 10 µm and fixed in 4% PFA in PBS for 10 min at room temperature. Then, the sections were incubated with 1 µM LEI-463-Cy5 in 1% BSA, 20 mM HEPES buffer (pH 7.2). Afterward, the tissue was serially washed in PBS, water, 50% THF/H₂O, water and finally in PBS. Subsequently, the sections were stained with antibodies as described for immunofluorescence stainings above, without Sudan Black quenching of autofluorescence. Autofluorescence was captured in an empty channel and used to control the other channels, as described in ref. ⁶⁴.

Proteomics sample preparation

Lysates were prepared from tissues by glass-bead mechanical lysis in 20 mM HEPES, 1 mM MgCl₂, 2 U ml⁻¹ benzonase. After protein quantification using the Bradford assay (Bio-Rad), lysates were snap-frozen in LN₂ and stored at −80 °C until further use. Three micrograms of protein were resuspended in 20 µl of sodium deoxycholate (SDC) buffer (2% SDC, 10 mM Tris(2-carboxyethyl) phosphine, 10 mM Tris (pH 8.5), 40 mM chloroacetamide) supplemented with complete mini ethylenediaminetetraacetic acid-free protease inhibitor cocktail (Roche). The samples were denatured at 95 °C for 5 min. Cooled, reduced and alkylated samples were mixed with 162 µl of 50 mM ammonium bicarbonate, followed by the addition of trypsin (Promega) and LysC (Wako) in 1:50 and 1:75 ratios, respectively. The digestion took place overnight at 37 °C. The digestion was quenched with formic acid (FA) at the final concentration of 2% and centrifuged at 20,000g for 20 min to remove the SDC. The acidified supernatant containing 800 ng of peptides was loaded onto the EvoSep StageTips (EV2018) according to the manufacturer’s protocol.

LC–MS/MS analysis proteomics

The samples were analyzed on EvoSep One liquid chromatography system (EvoSep) coupled to Exploris 480 mass spectrometer (Thermo Fisher Scientific). The peptides were eluted from the EvoSep StageTips and separated on an EvoSep analytical column (15 cm × 150 µm, 1.9 µm; EV-1106) using a 44-min gradient (30 SPD program) and the data were acquired in data-independent mode. The following mass spectrometric parameters were used for the full MS scan: scan mass range set to 375–1,100 m/z, resolution of 60,000 at 200 m/z, AGC target set to standard, maximum injection time set to auto. For the MS/MS spectra, the following parameters were used: quadrupole isolation window of 15 Da with 1 Da overlap between the windows (total of 40 windows), the precursor mass range was set to 400–1,000 m/z, resolution of 15,000 at 200 m/z, normalized AGC target was set to 1000%, injection time set to auto, normalized collisional energy was set to 27% and isotope exclusion set to on (or activated).

Raw data processing proteomics

All raw files were processed with DIA-NN software (v1.8.1) with the deep learning in silico spectral library generation option. The digestion was set to trypsin with one missed cleavage allowed. Cysteine carbamidomethylation was set as a fixed modification and methionine oxidation was set as a variable modification. The threshold FDR for the protein identification was set to 1%. The full-scan mass accuracy was 6 ppm, and the optimized MS/MS mass accuracy was set to 22 ppm. All other settings were set to default. The UniProt human database, with 20,398 entries, was used for the search (released in April 2023). The quantification was based on the unique peptides for the downstream analysis. Two samples were excluded from the downstream analysis as they contained a very high percentage of missing values.

All the downstream analyses were performed in R (v4.4.1). Proteins that had a label-free quantification (LFQ) value in at least 70% (found, total ≥ 0.7) of samples in one group (lesion type) were included, and immunoglobulin variable regions were excluded. This led to 3,237 proteins included in downstream analysis. Missing values were imputed by sampling from a normal distribution around the lowest value found for a given protein, with an s.d. of one-third of the original s.d. The imputed dataset was used only for principal component analysis (PCA), while all other analyses were performed on the original dataset without imputation.

Cytokine profiling

Cytokine concentrations were measured using a custom-made Human Luminex Discovery Assay (LXSAHM-26) according to the manufacturer’s protocol. Lysates from tissue samples were prepared as described for proteomics. Twenty-five microliters of 1 mg ml⁻¹ lysate were thawed and diluted 1:2 in calibrator diluent buffer (RD6-52). The standard cocktails were reconstituted in RD6-52 buffer and further diluted 1:3 following the manufacturer’s instructions. In total, 50 µl of sample or standard was mixed with 50 µl of diluted microparticles cocktail and incubated for 2 h at room temperature on a horizontal orbital microplate shaker operating at 800 rpm. After the incubation, the plate was washed thrice with a magnetic plate washer, using 150 µl of wash buffer and allowing 1 min before removing the liquid. Next, 50 µl of diluted Biotin-Antibody Cocktail was added to each well and the plate was incubated for 1 h at room temperature while shaking at 800 rpm, followed by another washing step and the addition of 50 µl of diluted Streptavidin-PE. After 30-min incubation and a washing step, the microparticles were resuspended in 100 µl of wash buffer. The plate was analyzed with FLEXMAP 3D using Standard PTM settings and doublet discriminator gates set at 8,000 and 16,500. After assigning the microparticle region for each measured analyte, 50 µl of the samples was acquired. Cytokine concentrations were quantified by the instrument software (BioPlex Manager) using a five-parameter logistic curve.

RNA extraction

Isolated lesions were dissolved in 500 µl Trisure (Bioline, BIO-38032). Subsequently, 100 µl CHCl₃ was added to the tube and the solution was centrifuged (15 min, 11,000 rcf, 4 °C). The aqueous layer was carefully removed and 500 µl were added to the same tube for an additional extraction. After combining the two aqueous layers, 1 equal volume 70% EtOH was added and the solution was added to an RNAeasy column (QIAgen). The RNA was washed with RW1 buffer, after which any DNA contamination was removed by incubation with DNase (15 min, room temperature). After this, the column was washed in subsequent steps with RW1 buffer, then with RPE buffer and finally with 80% EtOH. The column was air dried for 5 min before eluting the RNA in 16-µl RNAse-free water. One microliter of the extracted RNA was used for analysis on a Denovix DS-11 spectrofotometer to analyze the purity and concentration.

RNA-seq and analysis

Isolated RNA was sent for total RNA-seq at Genomescan (Leiden). RNA integrity was assessed, delivering an RNA quality (RQN/RIN) value, which is used in downstream analysis as a covariate. No minimum RQN value was set as a requirement. After library prep, a library quality control (QC) was performed, which gave satisfactory results. Subsequently, the libraries were sequenced on a NovaSeq 6000 sequencer for a target library size of 30 million reads. Alignment was performed with HISAT (v2.2.0) against the human Ensembl GRCh38 reference genome. Ensembl identifiers were converted into gene names using AnnotationDbi and the human reference genome database Org.Hs.eg.db (v3.16.0) from Bioconductor.

The raw count data were subsequently loaded in R (v4.4.1) and a data analysis pipeline was conducted using the edgeR, voom and limma packages in R^96,97,99. Detailed filtering criteria and analysis choices are specified in Supplementary Methods. Methods for weighted gene-coexpression network analysis (WGCNA)³⁹, single-cell deconvolution and gene set enrichment can be found in Supplementary Methods.

PCA

PCA was performed using the ‘prcomp’ function in base R. ‘Scale’ was set to FALSE, to preserve highly variant variables to have more weight in the PCA. PCA was performed on the full proteomics, lipidomics and ABPP datasets. For RNA-seq, PCA was performed on the 1,000 most variable genes. The obtained principal components were rotated to focus loadings of variables onto a single component using the ‘varimax’ function. Correlation of principal components to metadata was performed using the PCAtools package, using the function ‘eigencorplot’. The number of components was determined using the elbow criterion, in PCAtools automated in the function ‘getelbow’.

MOFA

MOFA was performed based on the workflow discussed in ref. ⁵³, using MOFA2 (v1.8.0). Data input for MOFA was the full lipidomics dataset (712 lipids) and a reduced RNA-seq and proteomics dataset. Only highly variable transcripts were selected by filtering for genes with an s.d. higher than 2× mean(s.d.) of the full dataset, which yielded 1,032 transcripts. Proteins were selected by filtering for proteins with an s.d. higher than 1× mean(s.d.) of the full dataset, which yielded 1,218 proteins. A total of 92 samples had data for all three data modalities. A total of 11 samples had data for two data modalities and 7 samples had data from only one data modality (Supplementary Fig. 10a).

We ran the MOFA model with largely default settings. Scale_views was set to ‘TRUE’, convergence mode was ‘medium’ and maximum iterations was 2,000. We obtained seven factors that explained a minimum of 5% variance in at least one data modality. The total variance explained by seven factors were 53.5% for the RNA-seq input, 57.7% for the proteomics input and 72.8% for the lipidomics dataset. Factor values were tested for differential expression across groups using two-sided Wilcoxon rank-sum test using the rstatix package (v.0.7.2), with correction for multiple testing using BH FDR.

Dimensionality was further reduced using uniform manifold approximation and projection using the package uwot (v.0.1.16) using a n_neighbours parameter of 20. Pseudotime trajectories were calculated using Slingshot (v2.6.0) using clusters generated by MClust (v6.1) using a G parameter of 5. The starting cluster was defined as the cluster containing CWM samples. Generated trajectories were extracted using the ‘slingCurves’ function and plotted onto the uniform manifold approximation and projection using ‘geom_path’. Trajectories were plotted against the original input data (factor values) and a smooth fit was plotted using the ‘loess’ fitting function of ggplot2.

MOFA factor 3 was associated with the clinical features in a generalized estimating equation model, accounting for donor identity, because our samples were nested within donors (100 samples from 28 MS donors). To this end, we used the function ‘geeglm’ from the package geepack (v1.3.10), with a Gaussian distribution and an exchangeable correlation structure.

Lipidomics sample preparation

Samples (50-µm sections) were lysed, and protein concentrations (Bradford method) were quantified. Two study QC pools (normal WM pool and lesion pool) were prepared using 10–100 μl of more than 50% of all the samples. A total of 11 QC aliquots were made separately from the normal WM pool and lesion pool.

Purchased or synthesized standards, internal standards (IS) were dissolved in methanol, ethanol, chloroform or ACN. These stock solutions were further diluted and mixed to make the standard stock solutions and IS stock solutions. The eCB synthesis IS mix contained 18:1, 18:1-PE-N-17:0, p18:1, 18:1-PE-N-17:0, 18:1-OH-PE-N-17:0, p18:1-OH-PE-N-17:0, OH-OH-PE-N-17:0. The lipids IS mix contained deuterated version of ceramides, (lyso)phospholipids, diacylglycerols, triacylglycerols and CEs. The signaling IS mix contained deuterated version of oxylipins, eCBs, free fatty acids and bile acids. The eCB synthesis calibration standards contained standards of precursors of eCBs; the signaling calibration standards contained oxylipins, eCBs, free fatty acids and bile acids; and the oxidized lipids calibration standards contained oxidized versions of phosphatidylcholines. The IS mixes and calibration standards were stored at −80 °C. The mixed IS working solutions were prepared and stored at −20 °C till further use.

Aliquots of sample lysates (~1 mg protein per ml, 50 μl) were thawed on ice. To each sample, 10-μl IS work solution was added. Calibration samples were prepared by spiking 10 μl of each calibration standard into 50 μl of water. Extraction was performed by 100-μl extraction buffer (0.2 mM ammonium formate) and 1,000-μl extractant (MTBE:BuOH, 50:50, vol/vol). Samples were then mixed in a Next Advance Bullet Blender (5 min, 90% speed, room temperature), followed by centrifugation (16,000g, 10 min, 4 °C). In total, 950 μl of the organic layer was transferred into clean, precooled tubes and concentrated in a SpeedVac vacuum concentrator (Thermo Fisher Scientific), followed by adding 50 µl of reconstitution solution (MeOH:ACN, 30:70, vol/vol) and agitating for 15 min. The reconstituted samples were centrifuged (16,000g, 10 min, 4 °C) and 40 µl were transferred into autosampler vials with inserts. Samples were kept at −80 °C till LC–MS analysis.

Samples were randomized and run in one batch. Each batch included QC samples and blank samples. QC samples are used to assess data quality. Method blanks (proc blanks) are used to check for background signal. These samples are composed of analyte-free matrix and have undergone all steps of the sample preparation procedure using only reagents. The detailed LC/MS–MS procedures for all lipidomics platforms are specified in the Supplementary Methods.

Chemical proteomics (ABPP) sample preparation

The chemical proteomics workflow was based on the previously reported procedures discussed in ref. ⁵⁷. Lysates were prepared by glass-bead mechanical lysis of isolated lesions in 20 mM HEPES, 1 mM MgCl₂, 2 U ml⁻¹ benzonase. After protein quantification using the Bradford method (Bio-Rad), lysates were snap-frozen in LN₂ and stored at −80 °C until further use.

Lysates (100 µl, 1 mg ml⁻¹ protein) were thawed on ice and subsequently treated with a cocktail of 10 µM FP-biotin and 10 µM THL-biotin as described previously⁵⁷. As controls, ten heat and SDS-inactivated (1% SDS, 5 min, 95 °C) lysates were taken along in the procedure. Probe-treated lysates were precipitated using the MeOH/CHCl₃ precipitation. The protein pellet was washed in MeOH and subsequently redissolved in PBS containing 0.5% SDS and 5 mM DTT. Proteins were dissolved by sonication and the solution was incubated at 65 °C for 15 min to reduce cysteines. Afterward, cysteines were alkylated using iodoacetamide and excess iodoacetamide was quenched using DTT. Subsequently, the samples were added to a suspension of 5-µl high-capacity streptavidin agarose beads (Thermo Fisher Scientific, 20361) and 15-µl control agarose beads in PBS (Thermo Fisher Scientific, 26150) containing 0.25% SDS. The suspension was incubated for 2 h at room temperature while rotating head-overhead to let probe proteins bind to the beads. After incubation, the beads were washed with PBS containing 0.5% SDS (4×) and subsequently with PBS (5×). On-bead digestion was performed with 0.25 µg sequence-grade trypsin (Promega) in 100 mM Tris, 100 mM NaCl, 1 mM CaCl₂ and 2% (vol/vol) acetonitrile overnight at 37 °C while shaking at 1,000 rpm. Trypsin was quenched with 10 µl FA. Peptides were then desalted on Oasis C18 plates (Waters) and eluted in 60% acetonitrile with 0.5% FA. The solvent was evaporated in a SpeedVac (45 °C, Eppendorf), and samples were stored at −80 °C until further use.

LC–MS/MS for chemical proteomics (ABPP)

Samples were randomized and assigned to six blocks, each containing 23 samples. Each block also contained four reference samples that were measured in every measurement block as described in ref. ¹⁰⁰. Per block of samples, peptides were then reconstituted in 3% acetonitrile, 0.1% FA in H₂O and analyzed on a QExactive HF LC–MS/MS (Thermo Fisher Scientific). All gradients and solutions were prepared according to previously published procedures⁵⁷. Before and after each block, the LC/MS machine was cleaned, recalibrated and its performance was assessed.

Data analysis ABPP with chemical proteomics

Raw spectral data from the chemical proteomics experiments were analyzed using MaxQuant (v2.0.1.0) using largely default settings¹⁰¹. Match across runs was enabled. The ‘proteingroups.txt’ output file from MaxQuant was imported in R. Identified proteins were filtered for potential contaminants identified by MaxQuant, and additional criteria for the proteins were that (1) a protein was identified with two unique peptides, (2) the ratio of protein raw intensity of native, heat inactivated lysate was at least 2.0 in at least five of ten QC pairs, (3) the protein is annotated as a serine hydrolase, or has a annotation in UniProt as ‘charge-relay system’ or ‘nucleophile’ as a catalytic residue and (4) the protein has an LFQ value for more than 60% samples in at least one biological group (lesion type). A total of 97 enzymes met these criteria and were further processed.

The LFQ values were then corrected for LC/MS performance by block design following the procedure described in ref. ¹⁰⁰. In short, the median log₂ LFQ for a protein in the reference samples was compared to the same samples in other blocks. The difference in log₂ median was subtracted from the same protein in the experimental samples. Analysis of the reference samples themselves showed that this correction improved the Pearson’s correlations of the reference samples to each other (data not shown). All subsequent analyses were then performed using the corrected log₂-transformed LFQ values.

Missing values were imputed using the random Gaussian distribution centered around the minimum LFQ value found for that protein in the dataset, with an s.d. equal to one-third of the s.d. of that protein in the dataset. This is based on the assumption that, if a protein was not found, it was most likely below the detection threshold, and therefore very low.

Enzyme activity values were associated with the MOFA factors derived from the other datasets using limma. The ABPP dataset was log₂-transformed, centered around 0 and then fed into a linear model using the design y ~ f1 + f2 + f3 + f4 + f5 + f6 + f7. P values for the derived coefficients were calculated using the eBayes function, and subsequently adjusted for multiple testing using the BH method with an FDR of 10%.

CSF lipidomics

Three hundred microliters of fresh-frozen CSF were spiked with 10-µl IS calibration standard, and subsequently extracted using the same method as described for the lesions. For the CSF lipids, only the signaling lipids platform was measured, which includes the oxylipins and prostaglandins.

Integrated MS/MS peaks (response ratio) were log₂-transformed, centered around 0, and missing values were not imputed. These response ratio values were correlated to the proportion of foamy lesions in the WM of the originating donor. Given the non-normal distribution of the proportions (between 0 and 1), the Spearman correlation was used. Then, to calculate if there were significant associations, the CSF oxylipin levels were modeled in a quasibinomial distributed generalized linear model using the counts of lesions with foamy microglia against other lesion types, as described for the clinical correlations presented in Fig. 1g,h. P values were determined using the likelihood-ratio test, and, because this was an exploratory analysis of oxylipins in the CSF, we did not correct for multiple testing.

LPC spinal cord injection in mice

Animals

Adult male and female C57BL/6JRj mice (14 weeks old; 20–30 g) were used in this study and obtained from Janvier Labs (France). Mice were housed under a 12-h light/12-h dark cycle with ad libitum access to food and water. All animal experiments were conducted in accordance with the Swiss Federal Act on Animal Protection and were approved by the Cantonal Veterinary Office Basel-Stadt (license 3095).

Anesthesia

Animals received a subcutaneous injection of buprenorphine (0.2 mg kg⁻¹) for 45 min to 2 h before surgery for pre-emptive analgesia. Anesthesia was induced with ~3.5% isoflurane and maintained at ~2% isoflurane through a face mask. Throughout the procedure, oxygen saturation, pulse and body temperature were continuously monitored.

Surgical procedure

Mice were placed on a heated surgical platform and the dorsal fur was shaved and disinfected with Betadine. Local anesthesia (0.1 ml of 0.1% lidocaine and 0.025% bupivacaine) was administered subcutaneously at the incision site. A midline skin incision (~2 to 3 cm) was made above the thoracic spine. Paraspinal muscles were carefully retracted to expose the vertebral column. The intervertebral space between T10 and T12 was identified and cleaned of overlying muscle. The animal was stabilized by gently clamping the tissue anterior to the injection site with a hemostat. The meninges were pierced using a 30-G needle adjacent to the central dorsal vessel. A Hamilton syringe fitted with a custom-pulled glass capillary was inserted at a 65 ° angle, targeting the ventral WM. A total of 1 μl of 1% LPC (L-α-lysophosphatidylcholine; Sigma, L4129) dissolved in 0.9% NaCl was injected in two steps—0.5 μl, followed by a 0.5 mm retraction of the syringe and injection of an additional 0.5 μl. After a 1-min delay, the syringe was slowly withdrawn. The paraspinal muscles were sutured using 6-0 Prolene (C-1), and the skin was closed with sutures.

Postoperative care

Animals received meloxicam (5 mg kg⁻¹, subcutaneous) immediately after surgery and daily for the following 2 days. They were housed in a warmed recovery cabinet until fully awake. Postoperative monitoring included daily scoring for 3 days (and weekly thereafter), with close attention to body weight, wound healing and hindlimb motor function. Weight loss was limited to ≤10% and typically recovered within 2–3 days.

RO7232432 treatment

Mice received daily intraperitoneal injections of RO7232432 (MAGLi-432) at 5 mg kg⁻¹ (10 ml kg⁻¹ dosing volume) starting from the second day after LPC injection. The treatment continued once daily until the designated experimental endpoint. Doses were prepared freshly in vehicle solution and administered at the same time each day to minimize variability. Control groups received vehicle only.

Reporting summary

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