ABCA7 variants impact phosphatidylcholine and mitochondria in neurons

Isolation of nuclei from post-mortem brain tissue

Batch 1 nuclei (BA10 region, frozen tissue) were isolated according to a protocol adapted from a previous study¹⁸, performed entirely at 4 °C or on ice. In brief, tissue was homogenized (700 µl homogenization buffer: 320 mM sucrose, 5 mM CaCl₂, 3 mM Mg(CH₃COO)₂, 10 mM Tris-HCl pH 7.8, 0.1 mM EDTA pH 8.0, 0.1% IGEPAL CA-630, 1 mM β-mercaptoethanol, 0.4 U µl⁻¹ recombinant RNase inhibitor (Clontech)) using a Wheaton Dounce tissue grinder (15 strokes, loose pestle), filtered (40 µm cell strainer), then mixed 1:1 with working solution (diluent (30 mM CaCl₂, 18 mM Mg(CH₃COO)₂, 60 mM Tris pH 7.8, 0.6 mM EDTA, 6 mM β-mercaptoethanol) and OptiPrep density-gradient solution (Sigma-Aldrich, D1556-250ML), 1:5). The sample was layered onto an OptiPrep density gradient consisting of 750 µl of 30% OptiPrep (1.5:1 ratio of working solution:homogenization buffer) above 300 µl of 40% OptiPrep (4:1 ratio of working solution:homogenization buffer), centrifuged (10,000g, 5 min, 4 °C) and nuclei were collected from the 30/40% interface (100 µl). Nuclei were washed twice (1 ml PBS, 0.04% BSA, 300 g, 3 min), resuspended (100 µl PBS, 0.04% BSA), counted (C-Chip hemocytometer) and diluted to 1,000 nuclei per µl (PBS 0.04% BSA).

Batch 2 nuclei (fresh post-mortem PFC BA10 tissue) were prepared as part of a previous study¹⁸.

Informed consent and Anatomical Gift Act consent were obtained, including repository consent to allow sharing of data and biospecimens. Rush University Medical Center IRB approved protocols (Religious Orders Study, Rush Memory and Aging Project).

Droplet-based snRNA-seq

Batch 1 libraries were prepared using Chromium Single Cell 3′ Reagent Kits v3 (10x Genomics) and sequenced on the NovaSeq 6000 S2 (paired-end, 28 + 91 bp, 8-nt index) system. Each sample was sequenced twice across two lanes to increase the depth. Batch 2 libraries were prepared using the Chromium Single Cell 3′ Reagent Kits v2 and sequenced using the NextSeq 500/550 High Output v2 kits (150 cycles), as previously described¹⁸. All raw reads were processed together for alignment and gene counting.

Variant calling and ROSMAP participant selection

We selected 36 individuals from the ROSMAP cohort, a longitudinal study of ageing and dementia⁵⁴. Whole-genome sequencing (WGS) variant calls (n = 1,249 available ROSMAP samples) were downloaded from Synapse (syn11724057) for genes with rare damaging variants linked to AD: SORL1, TREM2, ABCA7, ATP8B4, ABCA1 and ADAM10². For participants with multiple WGS samples, the highest-quality sample was chosen (Genomic Quality Score). Samples with sex mismatches or genotype inconsistencies were excluded (see Synapse accession syn12178037). Only variants passing quality control (FILTER_PASS) were used.

PTC variants flagged as splice, frameshift, nonsense, missense or premature stop variants annotated as loss-of-function (LOF) were identified. For ABCA7, known LoF variants from the literature were captured, except for c.5570+5G>C, which was manually added. Additional WGS details (library preparation, quality control, annotations, impact predictions) can be viewed at Synapse (syn10901595).

We selected 12 individuals (LoF samples) who carried ABCA7 PTC variants, had no PTC variants in the other candidate genes listed above and had fresh-frozen tissue available from Rush University. Moreover, we chose 24 matched controls without any PTC variants in ABCA7 or the other listed genes. Controls were matched by age, sex and pathology.

Read counting and alignment

Libraries were demultiplexed using the MIT BioMicroCenter BMC/BCC 1.8 pipeline (updated 9 December 2020; https://openwetware.org/wiki/BioMicroCenter:Software#BMC-BCC_Pipeline). Fastq reads were aligned to the human reference genome (GRCh38) and counted using Cell Ranger (v.6.1.2; 10x Genomics) with intron counting enabled and an expected cell count of 5,000 per sample. The default parameters were otherwise used. Counts from all samples were aggregated using a custom script, yielding a total of 150,456 cells.

Sample-swap analysis

Sample-swap analysis was performed using an established pipeline (MBV; QTLtools v1.1)⁵⁵, comparing allelic concordance between genomic (VCF) and transcriptomic (BAM, generated by Cell Ranger) data. We specifically analysed chromosome 19 variants (location of ABCA7). Each single-cell sample matched the expected WGS sample clearly, showing higher concordance (fewer mismatches) compared with all other ROSMAP WGS samples (examples are shown in Extended Data Fig. 1e).

Cell filtering metrics

Aggregated counts underwent quality control before cell annotation. Cells with fewer than 500 or more than 10,000 detected genes (count > 0) were removed. Next, we filtered cells by mitochondrial fraction (total mitochondrial counts divided by total gene counts), a measure of nuclear integrity. We log-transformed mitochondrial fractions and fitted a Gaussian mixture model (GMM, sklearn GaussianMixture) to identify and remove cells assigned to the GMM component with the highest mean mitochondrial fraction. This step removed approximately 20,000 low-quality cells.

We next considered cells in marker-gene expression space defined by known major cell-type markers for human PFC: astrocytes (159 markers), excitatory neurons (113), inhibitory neurons (83), microglia (97), oligodendrocytes (179), OPCs (143) and vascular cells (124)¹ (Supplementary Table 3). Marker counts were normalized to total library size, mean-centred and scaled to unit variance. Incremental principal component analysis (sklearn IncrementalPCA) reduced dimensionality (top 50 principal components). Visually, cells projected onto the first two principal components formed distinct Gaussian-like clusters. Assuming each Gaussian cluster corresponded to a distinct brain cell type, we fitted another GMM to the projected data. The resulting ten clusters aligned clearly with known brain cell types.

Cells poorly modelled by this GMM (log-probability < −100) and two clearly outlying clusters were removed. These excluded cells had lower total counts and higher mitochondrial fractions, suggesting low quality. This step removed approximately 12,000 cells, leaving a final dataset of 118,668 cells.

Gene filtering metrics

Downstream analyses included only nuclear-encoded, protein-coding genes (total 19,384) based on Ensembl GRCh38p12 annotations.

Cell type annotations

We first corrected variance due to sequencing batch and individual-of-origin by applying Harmony⁵⁶ to the top 50 principal components from the quality-controlled data. Using Harmony-corrected principal components, we computed a neighbourhood graph (default Scanpy parameters)⁵⁷ and clustered cells with the Leiden algorithm (Scanpy implementation)⁵⁸.

Major cell types (excitatory neurons, inhibitory neurons, astrocytes, microglia, oligodendrocytes, OPCs, vascular cells) were assigned to Leiden clusters by computing cell-type-specific marker gene enrichment. Specifically, we calculated enrichment scores as the average log-ratio of expression for marker genes inside versus outside each cluster and assigned labels based on the highest enrichment.

We then subclustered each major cell type using the Leiden algorithm and removed subclusters with excessively high mitochondrial fraction or extreme total counts. Thresholds were set at two s.d. above the mean for these metrics within each major cell type. Removed clusters were small, poorly represented across individuals and weakly connected on manual inspection.

Individual-level filtering

After all of the previous quality-control steps, six individuals with fewer than 500 cells were excluded from further analyses, leaving 24 control individuals and 12 individuals with ABCA7 LoF. None of these excluded individuals carried ABCA7 PTC variants, and their removal did not substantially affect clinical variable distributions across genotypes.

Differential gene expression

Pseudo-bulk gene expression values were generated by summing cell-level counts per gene per individual (matrix multiplication). For each major cell type, we considered genes detected in >10% of cells. Counts were normalized by TMM (edgeR), and residual mean-variance trends were removed using Limma-Voom. Unknown variance was modelled through surrogate variable analysis (SVA). Differential expression analysis (Limma: lmFit, eBayes, topTable) was performed separately for each major cell type using the following linear model for each gene (G_i):

where n is the number of surrogate variables determined by num.sv() per cell type and ABCA7 LoF indicates individuals carrying ABCA7 LoF variants. Additional covariates (defined in Supplementary Note 1) included sex, NFT, amyloid burden, age at death, PMI, sequencing batch and APOE4 status.

Gene perturbation projections across cell types

We computed cell-type-specific gene perturbation scores summarizing differential expression significance and direction associated with ABCA7 LoF as S = sign(log₂[FC]) × −log₁₀(P), where positive log₂[FC] indicates upregulation in ABCA7 LoF. Scores for genes not detected in >10% of cells per cell type were set to zero. Genes with \(| S| > 1.3\) in at least one of six major cell types (excitatory neurons, inhibitory neurons, astrocytes, microglia, oligodendrocytes and OPCs) were projected from 6D perturbation-score space into 2D using UMAP (Python umap).

Genes were clustered in the resulting 2D embedding using Gaussian mixture modelling (Python sklearn). Clusters were annotated by hypergeometric enrichment (Python gseapy) for Gene Ontology Biological Process pathways (Supplementary Table 3), using all genes in the embedding as background. Pathways with enrichment P < 0.01 were selected for naming each cluster. Per-cell-type perturbation scores for each cluster were calculated as the mean gene score within clusters. Statistical significance was assessed by permuting cluster assignments (100,000 permutations).

Gene-set enrichment and Kernighan–Lin pathway clustering

Genes were ranked by perturbation scores S (see the ‘Gene perturbation projections across cell types’ section). Fast GSEA (fGSEA; R implementation⁵⁹) with 10,000 permutations tested enrichment of WikiPathways gene sets (Supplementary Table 3) among differentially expressed genes. Only gene sets with 5–1,000 genes were considered.

To simplify gene–pathway associations, we constructed a bipartite graph using genes from the fGSEA leading-edge (LE) subset (268 genes, enriched at P < 0.05 in ABCA7 LoF excitatory neurons) and WikiPathways associated with ≥4 LE genes. We treated gene–pathway grouping as a graph partitioning problem (Supplementary Note 2). Among three graph-partitioning algorithms tested (Supplementary Note 2), the METIS and the Kernighan–Lin algorithms showed the lowest loss and highly comparable performance (within 1.8% loss; Rand index = 0.98 after 5.0 × 10⁴). We selected the Kernighan–Lin algorithm because it consistently outperformed the METIS algorithm across a wider range of graph sizes. The Kernighan–Lin algorithm was implemented in Python as described previously⁶⁰ with the parameters C=0, KL_modified=True, random_labels=True, unweighted=True, and K=50 to partition the graph into eight groups. We performed 5.0 × 10⁴ random initiations and selected the lowest-loss solution.

Graph layouts were computed using the spring layout algorithm (networkx, 10,000 iterations) and visualized using matplotlib. Representative pathways for each cluster were identified by averaging ABCA7 LoF perturbation scores (S) of genes in the cluster connected directly to each pathway. Pathways with ≥5 intracluster gene connections are highlighted in the figures.

Excitatory neuronal layer annotation

Excitatory neurons were annotated by cortical layer using published marker gene sets⁶¹ (Supplementary Table 3) according to the procedures described in the ‘Cell type annotations’ section. In brief, the normalized expression matrix was filtered to include only layer-specific marker genes and cells expressing ≥15% of these genes. Dimensionality was reduced using iterative principal component analysis, followed by batch-effect correction using Harmony. A neighbourhood graph was constructed, and cells were clustered using the Leiden algorithm. Clusters enriched for layer-specific markers (average log-transformed FC > 0.1) were labelled accordingly, while ambiguous clusters were excluded. Layers 5 and 6 were combined into a single ‘L5/6’ category. Annotations were validated using independent marker genes⁶² (Supplementary Table 3). Layer-specific differential expression analysis was performed as described in the ‘Differential gene expression’ section, followed by gene-set enrichment analysis (fGSEA, as described in the ‘Gene-set enrichment and Kernighan–Lin pathway clustering’ section) testing enrichment of ABCA7 LoF-associated gene clusters identified by Kernighan–Lin clustering (as described in the ‘Gene-set enrichment and Kernighan–Lin pathway clustering’ section).

ABCA7 p.Ala1527Gly variant calling and gene–pathway clustering comparisons

Participants carrying the ABCA7 p.Ala1527Gly variant with available PFC snRNA-seq data from a previous study (Supplementary Table 3) were identified using methods described in the ‘Variant calling and ROSMAP participant selection’ section. Differential expression was computed as described in the ‘Differential gene expression’ section, followed by fGSEA to test enrichment of ABCA7 LoF-associated gene clusters identified by Kernighan–Lin clustering (see the ‘Gene-set enrichment and Kernighan–Lin pathway clustering’ section).

Culture and generation of human isogenic iPS cells

A control parental iPS cell line (AG09173; 75-year-old female individual, APOE3/3 genotype) was generated previously by the Picower Institute iPSC Facility⁶³. Two ABCA7 LoF isogenic lines were derived from AG09173: ABCA7 p.Glu50fs*3, containing a novel premature stop codon in exon 3 (generated by Synthego), and ABCA7 p.Tyr622*, containing a patient-derived mutation (Y622*)⁶⁴ generated in-house by CRISPR–Cas9 editing.

For the ABCA7 p.Tyr622* line, an sgRNA targeting ABCA7 (oligos: forward, 5′-CACCGCCCCTACAGCCACCCGGGCG-3′; reverse, 5′-AAACCGCCCGGGTGGCTGTAGGGGC-3′; designed at http://crispr.mit.edu) was cloned into pSpCas9-2A-GFP (PX458, Addgene, 48138) as previously described⁶⁵. The plasmid was confirmed by Sanger sequencing, then nucleofected (Amaxa, Lonza Human Stem Cell Nucleofector Kit I, program A-23) along with 15 μg of a single-stranded oligodeoxynucleotide template into dissociated AG09173 iPS cells (Accutase, Thermo Fisher Scientific; 10 μM ROCK inhibitor, Tocris). Cells (around 5 × 10⁶) were sorted (BD FACS Aria IIU, Whitehead Institute), plated at single-cell density in medium supplemented with penicillin–streptomycin (Gemini Bio-products) and ROCK inhibitor. Colonies were expanded, screened by genomic DNA extraction (DNeasy Blood & Tissue Kit, Qiagen, 69504) and Sanger sequencing to confirm the Y622* mutation (Supplementary Table 16).

All iPS cell lines were regularly tested for karyotypic normality (Cell Line Genetics) and cultured at 37 °C, 5% CO₂, in feeder-free conditions using mTeSR-1 medium (StemCell Technologies, 85850) on Matrigel-coated plates (Corning; hES-cell-qualified, 354277). Cells were passaged at 60–80% confluence using ReLeSR (StemCell Technologies, 05872) onto Matrigel-coated plates at a 1:6 to 1:24 split ratio.

rTTA and NGN2 virus production

HEK293T cells were seeded at 5 × 10⁶ cells per 10 cm plate and transfected using a third-generation lentiviral system. Per plate, transfection mixtures contained 10 µg plasmid DNA (EF1a-rtTA-Hygro, Addgene 66810, or pLV-TetO-hNGN2-eGFP-Puro, Addgene, 79823), 5 µg pMDLg/pRRE, 2.5 µg pRSV-Rev, 2.5 µg MD2.G and 48 µl polyethyleneimine (1 mg ml⁻¹) diluted in 600 µl OptiMEM (Thermo Fisher Scientific, 51-985-034). Mixtures were incubated 20 min at room temperature, added dropwise to cells and replaced with fresh medium after 16 h. Virus-containing supernatant collected 72 h after transfection was clarified (3,000g, 5 min, 4 °C) and the supernatant was ultracentrifuged (Beckman Optima L-90K Ultracentrifuge, SW32Ti rotor, 25,000 rpm, 2 h), resuspended in 1 ml PBS per 10 cm plate and stored at −80 °C.

Lentivirus-mediated NGN2 induction in iPS cells and drug treatments

iPS cells were dissociated into single-cell suspensions (Cell Dissociation Buffer, Life Technologies, 13151-014), resuspended in mTeSR1 medium with ROCK inhibitor (Rockout; Abcam, ab285418), and plated onto Matrigel-coated six-well plates at 50–60% confluence after 24 h. After 1 day, cells were co-transduced overnight with 80 µl each of pLV-TetO-hNGN2-eGFP-Puro and EF1a-rtTA-Hygro lentivirus per well. NGN2 expression was induced 24 h later with doxycycline (1 µg ml⁻¹) and ROCK inhibitor. Puromycin selection was performed 24 h after viral transduction. Immature neurons were replated on PDL/laminin-coated plates (1 × 10⁶ cells per well in six-well plates, or 5 × 10⁴ cells per well in 96-well plates), and maintained in BrainPhys neuronal medium (StemCell Technologies, 05793) with Neurocult SM1 neuronal supplement (StemCell Technologies, 05711), (N2-supplement-A StemCell Technologies, 07152), laminin (1 µg ml⁻¹) and doxycycline (1 µg ml⁻¹). Half-medium changes were performed every 3–4 days, and cultures were matured for 28 days before experiments.

Neurons were treated with cytidine 5′-diphosphocholine (CDP-choline, Millipore Sigma-Aldrich, 30290) at a final concentration of 100 µM starting at day 14, continuing with each medium change until day 28. Choice of treatment concentration and duration was based on a previous study by our laboratory³².

Cortical organoid generation

Dorsal cortical organoids were generated as previously described⁶⁶. In brief, iPS cells at 80–90% confluence were dissociated into single-cell suspensions (1 × 10⁵ cells per ml) in mTeSR with 10 µM ROCK inhibitor, seeded at 100 µl per well in PrimeSurface 96 Slit-well plates (S-Bio, MS9096SZ) and induced to differentiate using neural induction medium consisting of DMEM/F12 (Life Technologies, 11330-032), 100 mM GlutaMAX (Life Technologies, 35050-061), 0.1 mM 2-mercaptoethanol (Sigma-Aldrich, M3148), 1% penicillin–streptomycin (Life Technologies, 15070-063) and 10 µM SB-431542 (R&D Systems, 1614), and 2.5 µM dorsomorphin (Sigma-Aldrich, P5499-CONF) with daily medium changes (days 0–5). The medium was then switched (days 6–16) to neural differentiation medium (Neurobasal A, B27 supplement, GlutaMAX, penicillin–streptomycin, human recombinant EGF and FGF2, 20 ng ml⁻¹ each), with daily changes until day 16, then every other day until day 25. From day 25 onwards, EGF and FGF2 were replaced with 20 ng ml⁻¹ each of BDNF and NT3, with medium changes twice weekly after day 45.

Confocal imaging experiments

All confocal images were acquired on a Zeiss LSM900 microscope using ZEN software.

For mitochondrial health staining, live cells were incubated with MitoHealth dye (Thermo Fisher Scientific, H10295) according to the manufacturer’s protocols for 30 min at 37 °C, fixed (4% paraformaldehyde/4% sucrose, 15 min, room temperature), permeabilized (0.1% Triton-X, 5 min), blocked (2% BSA, Fisher Bioreagents, BP9703) and incubated overnight at 4 °C with NeuN antibody (1:500), followed by incubation with secondary antibodies (1:1,000) for 2 h and Hoechst (1:2,000, Invitrogen, H3570) for 10 min (Supplementary Table 17). Images were captured as z stacks (1 µm intervals).

Live imaging of mitochondrial membrane potential used TMRM (0.1 µM, 30 min at 37 °C; Thermo Fisher Scientific, I34361), followed by imaging before and immediately after adding the mitochondrial uncoupler FCCP (1 µM; Cayman Chemical, 15218). ROS were assessed by live staining with CellROX Orange (5 µM, 30 min at 37 °C; Thermo Fisher Scientific, C10443). TMRM and CellROX images were acquired as single optical sections.

For immunostaining, iNs cultured on coverslips and cortical organoid cryosections (20 µm) were fixed (4% formaldehyde, 10 min), permeabilized (0.2% Triton X-100) and blocked (10% BSA, 1 h), and incubated overnight at 4 °C with primary antibodies (MAP2 and NeuN, both 1:1,000). Alexa-Fluor-conjugated secondary antibodies (1:500) and Hoechst (1:1,000) were used for visualization. Coverslips were mounted with Fluoromount-G, and images were captured as single optical sections.

For visualization, confocal images were pseudocoloured to enhance the signal contrast; representative unprocessed images are provided in Supplementary Fig. 6.

Confocal image quantification

Confocal images (.czi format; 8 or 16 bits; voxel size: 1 × 0.62 × 0.62 µm) were loaded into Python (aicsimageio) and normalized to floating-point format [0,1]. Acquisition settings were consistent within each imaging batch.

For fixed z-stack images, NeuN-positive cell bodies were segmented in 3D using the pre-trained cyto2 model (Cellpose⁶⁷). The segmentation quality was manually verified (blinded), and low-quality images were excluded. Cell-level fluorescence intensities were computed as probability-weighted sums of voxel intensities, using segmentation-derived voxel probabilities. Measurements from multiple differentiation batches (independent staining and imaging experiments) were combined by uniformly sampling cells per condition per batch, batch-wise z-scaling fluorescence values, and including batch and well-of-origin indicator variables in downstream analyses. Clipping was minimal (<0.1%), and the confocal microscope response was assumed linear. A linear mixed-effects model (mixedlm() from statsmodels) tested cell-level fluorescence intensities, modelling genotype or treatment as a fixed effect and well of origin as a random effect.

For single-plane live imaging (TMRM, CellROX), images were binarized at the 75th percentile intensity threshold per channel to identify regions occupied by neuronal soma or processes, according to established methodology⁶⁸. Mean fluorescence intensities were quantified within these masked areas. For time-course imaging, images were spatially aligned by Fourier-based registration (phase cross-correlation), with alignment accuracy confirmed manually. A mask from the baseline (pre-FCCP) TMRM image (75th percentile threshold) was consistently applied across timepoints. For all live-imaging experiments, masked regions (wells) were treated as individual observations in statistical tests. Batch-wise z-scaling was not required here, as data were not combined across batches for these experiments.

One outlier (p.Tyr622*+H20; value 0.34) was identified and removed in the TMRM p.Tyr622* (with or without CDP-choline) experiment using the interquartile range (IQR) method (values outside Q1−2 × IQR or Q3 + 2 × IQR) and removed for plotting convenience. This did not affect the statistical significance of the results.

Aβ ELISA assays

Culture media were collected and analysed for Aβ40 and Aβ42 levels using enzyme-linked immunosorbent assay (ELISA) kits (Thermo Fisher Scientific, KHB3481 and KHB3441, respectively) according to the manufacturer’s protocols. For 4-week-old iNs, media were flash-frozen before analysis. For cortical organoids (aged 5–6 months; days 176–182), media were analysed immediately after collection following 3–4 weeks of treatment with 500 µM or 1 mM CDP-choline.

Electrophysiology recordings

Electrophysiology recordings were performed using the Axon Multiclamp 700B amplifier and Clampex 11.2 software (Molecular Devices). Cells were visualized using infrared differential interference contrast imaging (Olympus BX-50WI microscope), placed in a recording chamber and perfused continuously at 2 ml min⁻¹ (32 °C) with oxygenated artificial cerebrospinal fluid (containing 125 mM NaCl, 2.5 mM KCl, 1.2 mM NaH₂PO₄·H₂O, 2.4 mM CaCl₂·2H₂O, 1.2 mM MgCl₂·6H₂O, 26 mM NaHCO₃, and 11 mM d-glucose).

Action potentials were elicited by injecting current steps in current-clamp mode. Whole-cell currents were recorded from a holding potential of −80 mV by stepping to various voltages in voltage-clamp mode. Spontaneous firing was recorded in cell-attached configuration. Recordings were filtered at 1 kHz (four-pole Bessel filter), digitized at 10 kHz with a Digidata 1550B interface (Molecular Devices). Pipette solution contained 120 mM K-gluconate, 5 mM KCl, 2 mM MgCl₂·6H₂O, 10 mM HEPES, 4 mM ATP and 0.2 mM GTP. Data were analysed using pClamp 11.2 and GraphPad Prism 10.

For electrophysiology recordings from cortical organoids, day 150 organoids were dissociated using Accutase (StemCell Technologies, 07920, 40 min, 37 °C), plated onto #1 glass coverslips (Thermo Fisher Scientific, 50-194-4702) coated with PDL, laminin and Matrigel, and maintained in 2D culture with or without 100 µM CDP-choline for 2 weeks before recordings.

Spontaneous action potential outliers were identified using the IQR method (values outside Q1–Q2 × IQR or Q3 +Q2 × IQR) and removed, resulting in the exclusion of two datapoints (9.38 in p.Tyr622*; 6.15 in p.Tyr622* + CDP-choline). Cells recording zero spontaneous potentials (likely glial) were also excluded.

Seahorse metabolic assays and OCR analysis

iPS-cell-derived neurons were differentiated directly in Seahorse XFe96/XF Pro microplates for 28 days before metabolic assays on a Seahorse XFe96 Analyzer. Seahorse XF cell mito stress and oxidation stress tests were conducted according to manufacturer’s protocols using final drug concentrations of 2.5 µM oligomycin, 1 µM FCCP and 0.5 µM rotenone/antimycin. Data were analysed using XFe Assay v.2.6.3.5 software.

OCRs were monitored over time, with curves visually inspected (blinded) to exclude wells not responsive to drug injections. The following OCR metrics were computed from integrals of OCR curves between specific experimental intervals: (1) basal respiration (before oligomycin injection); (2) proton leak (after oligomycin, before FCCP); (3) maximal respiration (after FCCP, before rotenone/antimycin); (4) relative uncoupling (proton leak divided by basal respiration); and (5) spare respiratory capacity (maximal respiration divided by basal respiration).

mRNA-seq and analysis of iNs

Total RNA was extracted from iNs using the RNeasy Mini Kit (Qiagen). RNA quality was assessed (Fragment Analyzer, Agilent), and only samples with RNA quality number > 9.5 were selected. Full-length cDNA libraries were generated (SMART-seq v4 kit, Takara Bio), and sequencing libraries prepared (Nextera XT DNA Library Preparation Kit, Illumina) for sequencing on the Element AVITI platform (Element Biosciences; 75 bp paired-end reads with dual 8-nucleotide indexes) at the MIT BioMicro Center.

Sequencing data were processed through the MIT BioMicro Center BMC/BCC pipeline v1.8 (updated 6 June 2023; https://openwetware.org/wiki/BioMicroCenter:Software#BMC-BCC_Pipeline). Reads were adapter-trimmed (Trim Galore, Nextera-specific settings, minimum overlap 3 bases), aligned to the human reference genome (GRCh38.p14, GENCODE release 47; STAR aligner), and counted (featureCounts, paired-end settings). Read counts were summarized at the exon level and aggregated by gene identifier.

Differential expression analysis (edgeR, limma-voom) retained protein-coding genes expressed at ≥1 CPM in ≥1 sample, normalized counts and used linear modelling with empirical Bayes moderation with contrasts based on experimental conditions (treatment/genotype). fGSEA (10,000 permutations) of WikiPathways gene sets (Supplementary Table 3) was performed using ranked differentially expressed genes (score: sign(log[FC]) × −log₁₀[P]), as described above (see the ‘Gene-set enrichment and Kernighan–Lin pathway clustering’ section). Significant pathways (adjusted P < 0.05) were identified, and leading-edge genes underwent gene–pathway clustering (Kernighan–Lin heuristic, described above). Gene–pathway cluster similarity was assessed by computing Jaccard indices based on pathways and genes assigned to each Kernighan–Lin cluster. The significance of observed overlaps was determined empirically through comparison to 1,000 random permutations, with P values adjusted using the Benjamini–Hochberg method to control the FDR.

LC–MS lipidomics on iNs

iPS-cell-derived neurons were washed in cold PBS, scraped, centrifuged (2,000g, 5 min), counted, pelleted to equal number and resuspended in cold methanol (2 ml). Biphasic extraction was performed by sequentially adding cold chloroform (4 ml) and cold water (2 ml), vortexing after each addition, then centrifuging (3,000 rcf, 10 min) for phase separation. Samples prepared at the Harvard Center for Mass Spectrometry were similarly processed from provided pellets (in 500 µl methanol), supplemented with additional methanol (1.5 ml) and chloroform (4 ml), sonicated (10 min), mixed with water (2 ml) and centrifuged (800 rcf, 10 min, 4 °C). Upper aqueous phases were collected for metabolomics, while chloroform phases were reserved for lipidomics. At least one blank control (no cells) was included in each extraction run. All LC–MS analyses were performed by the Harvard Center for Mass Spectrometry.

Extracted samples were dried under nitrogen, fully evaporated, resuspended in chloroform (scaled by biomass (cell count); ≥60 µl), and split equally for positive and negative ionization analyses (or unsplit if only positive mode). After centrifugation (18,000 rcf, 20 min, 4 °C), the supernatants were transferred into microinserts for LC–MS.

LC–MS analyses were performed on an Orbitrap Exactive plus MS (Thermo Fisher Scientific) consistent with an Ultimate 3000 LC (Thermo Fisher Scientific) in positive- and negative-ionization modes (in WT versus p.Tyr622* only in positive mode), in top five automatic data-dependent MS/MS mode. Chromatography separation was performed on the Biobond C4 column (4.6 × 50 mm, 5 µm particle size; Dikma Technologies). The flow rate began at 100 µl min⁻¹ with 0% mobile phase B (MB) for the initial 5 min, followed by an increase to 400 µl min⁻¹ over the next 50 min with a linear gradient of MB from 20% to 100%. The column was subsequently washed at 500 µl min⁻¹ for 8 min with 100% MB, then re-equilibrated for 7 min at 500 µl min⁻¹ using 0% MB. For positive-ion mode, mobile phases consisted of buffer A (MA: 5 mM ammonium formate, 0.1% formic acid and 5% methanol in water) and buffer B (MB: 5 mM ammonium formate, 0.1% formic acid, 5% water and 35% methanol in isopropanol). For negative-ion mode, buffer A (MA) contained 0.03% ammonium hydroxide and 5% methanol in water, and buffer B (MB) contained 0.03% ammonium hydroxide, 5% water and 35% methanol in isopropanol.

Lipids were identified, and their signals integrated using the Lipidsearch software (v.4.2.27, Mitsui Knowledge Industry, University of Tokyo). Integrations and peak quality were curated manually. Peak areas were background-corrected (subtracting 3× median blank peak areas; negative values set to zero). Statistical analyses were performed using Welch’s t-tests (unequal variance) to compare different cell lines, and Student’s t-tests (equal variance) for treatment comparisons within identical genetic backgrounds.

LC–MS metabolomics on iNs

Samples were dried under nitrogen, evaporated completely and resuspended in biomass-scaled volumes (≥20 µl) of 50% acetonitrile in water. After centrifugation (maximum speed, 10 min), consistent volumes (12 or 15 µl, depending on batch) of supernatants were transferred to microinserts. The remainder of the sample volumes was combined to create a pool sample used for MS2/MS3 data acquisition.

LC–MS metabolomics analyses were performed at the Harvard Center for Mass Spectrometry using a Vanquish LC system coupled with an ID-X mass spectrometer (Thermo Fisher Scientific). Samples (5 µl injection) were analysed on a ZIC-pHILIC peek-coated column (150 mm × 2.1 mm, 5 µm particle size; Sigma-Aldrich) held at 40 °C. Mobile phases comprised buffer A (20 mM ammonium carbonate and 0.1% ammonium hydroxide in water) and buffer B (97% acetonitrile in water). The gradient initiated at 93% B, decreasing linearly to 40% B over 19 min, further decreasing to 0% B over the subsequent 9 min, held at 0% B for 5 min, returned to 93% B within 3 min and finally was re-equilibrated at 93% B for 9 min. The flow rate was held constant at 0.15 ml min⁻¹, except for an initial 30 s ramp from 0.05 to 0.15 ml min⁻¹. MS data were acquired in polarity-switching mode at 120,000 resolution, with an AGC target of 1 × 10⁵, covering an m/z range from 65 to 1,000. MS1 acquisition used polarity switching for all samples. MS2 and MS3 analyses were performed on pooled samples using the AcquireX DeepScan method, with five reinjections each in positive- and negative-ion modes separately. A mixture containing standards of targeted metabolites was prepared and analysed immediately after the sample runs for targeted metabolite analysis.

Data were analysed using Compound Discoverer 3.2 (Thermo Fisher Scientific). Metabolite identification was based either on MS2/MS3 spectral matching against a local mzVault library and corresponding retention times from pure standards (level 1), or spectral matching using mzCloud (level 2). Each metabolite identification was manually inspected. Blank samples were used to exclude background compounds (compounds for which the area in at least one sample was not higher than three times the area in the blanks). Median-centred peak areas were scaled to zero-mean and unit variance before principal component analysis. The Harvard Center for Mass Spectrometry identified three samples with notably low overall metabolite intensities, which were subsequently excluded from downstream analyses.

LC–MS metabolomics on medium

Medium samples (100 µl each) were transferred into microcentrifuge tubes containing 1 ml of methanol and incubated at −20 °C for 2 h. After incubation, the samples were centrifuged at 18,000 rcf for 20 min at −9 °C, and the supernatants were transferred into new tubes and evaporated to dryness under nitrogen flow. The dried samples were resuspended in 50 µl of 30% acetonitrile in water containing 2 mM medronic acid, centrifuged again at 18,000 rcf for 20 min at 4 °C and the resulting supernatants were transferred into glass microinserts for LC–MS analysis.

Peak areas from targeted metabolite analysis of media samples were compared for CDP, CDP-choline and choline. To ensure accurate detection, solvent blanks were analysed: CDP and CDP-choline were not detected in these blanks, while choline was detected at levels several orders of magnitude lower than in medium samples.

Molecular dynamics simulations

ABCA7 structures (unbound-open and bound-closed conformations; Protein Data Bank (PDB): 8EE6, 8EOP) containing the G1527 variant were retrieved from the PDB. The A1527 variant was generated by mutation (Gly to Ala) using PyMOL v.2.0. ABCA7 residues 1517–1756 were embedded in a DPPC membrane (CHARMM-GUI) and oriented according to the OPM database. Four simulations were performed (GROMACS 2022.3; CHARMM36M force field; Supplementary Table 15).

The protein–membrane system was solvated in a cubic box with a minimum distance of 1.0 nm between the protein and the box edge, using the TIP3P water model. Energy minimization was performed using the steepest descent algorithm with a maximum force threshold of 1,000 kJ mol⁻¹ nm⁻¹ to relieve any steric clashes or bad contacts. The system was equilibrated in six phases, each 125 ps long, to equilibrate volume (NVT) and pressure (NPT). The production run, 300 ns long, was performed in the NPT ensemble at 323 K using a V-rescale thermostat and 1 bar using the Parrinello–Rahman barostat. A 2 fs time step with H-bond constraints was used with periodic boundary conditions applied in all directions. Long-range electrostatics were handled using the particle mesh Ewald method with a cut-off of 1.0 nm for non-bonded interactions.

The r.m.s.d. was calculated to monitor the conformational stability of a given structure over the course of the simulation by comparing the position of C_α at time t under simulation to its reference position (in 8EOP or 8EE6). The ϕ and ψ dihedral angles were calculated using the gmx rama tool, followed by post-processing. Secondary structure analysis was performed using gmx dssp -hmode dssp, with subsequent post-processing using custom Python scripts. Visualization of the trajectories was carried out using VMD v.1.94 software. Principal component analysis was conducted on C_α atom positional fluctuations to identify the major conformational changes during the simulation.

Eukaryotic cell lines

Human iPS cell lines used in this study were generated by the Picower Institute for Learning and Memory iPSC core. The initial parental cell line (AG09173) was obtained from the Coriell Institute. HEK293T cells (ATCC, CRL-3216) were sourced from ATCC. iPS cell lines were confirmed by cell marker staining, RNA-seq and karyotyping. No further authentication of HEK293T cells was performed. All cell lines used here tested negative for mycoplasma contamination.

Use of large-language models

ChatGPT (OpenAI) was used to edit portions of the manuscript text for brevity and clarity, and to assist in generating selected plotting code.

Ethics statement

The study protocol involving the use of human stem cells was approved by the Coriell Institutional Review Board (Coriell IRB) in compliance with DHHS regulations (45 CFR Part 46). The initial cell lines were obtained from the Coriell Institute, which ensured that informed consent was received from all donors. Donors were informed that their tissue donations would be used for the creation of cell lines intended for educational and research purposes, and that all biological materials would be anonymized. For post-mortem human brain samples, informed consent was obtained from each participant, and the Religious Orders Study and Rush Memory and Aging Project were approved by an Institutional Review Board (IRB) of Rush University Medical Center.

Reporting summary

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