Abstract
Neonatal hypoxic-ischemic (H-I) brain injury, a leading cause of
neurodevelopmental disabilities, severely affects the metabolically
active and neurogenic hippocampus. To investigate its acute effects and
identify drug targets for early therapeutic windows, we applied
single-nucleus RNA sequencing on postnatal day 8 (P8) mouse hippocampi
under sham, hypoxic, and hypoxic-ischemic conditions. We constructed a
comprehensive hippocampal cell atlas and developed a machine-learning
classifier for precise cell type identification. Our analysis reveals
early vulnerabilities in mature neurons and notable resilience in
immature DG, GABAergic, and Cajal-Retzius cells following H-I. Gene
regulatory network analysis identified key transcription factors
associated with neuronal vulnerability, along with upregulated ribosome
biogenesis and dysregulated calcium homeostasis pathways. We observed
rapid activation of astrocytes and microglia, with Runx1 identified as
a potential key transcription factor associated with early microglia
immune responses. Endothelial cells displayed complex transcriptional
changes and predicted intercellular signaling patterns that may
influence vascular repair and recovery. Our study advances the
understanding of immediate cellular and transcriptional responses to
neonatal H-I injury, providing new insights into hippocampal cell
heterogeneity and pathophysiology. The integrated hippocampal atlas,
post-H-I atlas, and machine learning classifier are available at
[40]https://hippo-seq.org.
Supplementary Information
The online version contains supplementary material available at
10.1186/s40478-025-02062-4.
Keywords: Neonatal hypoxia ischemia, Single-cell RNA sequencing,
Hippocampal cell atlas, Machine learning, Neural vulnerability, Gene
regulatory network
Introduction
Hypoxia (oxygen deprivation) and ischemia (restricted blood flow) are
severe conditions that disrupt cellular and molecular functions,
critically affecting brain health [[41]3, [42]29, [43]73]. In the
neonatal period, the brain is particularly susceptible to
hypoxic-ischemic (H-I) injury due to its high metabolic demands,
ongoing developmental processes, and unique vulnerabilities such as
elevated levels of unsaturated fatty acids, a high rate of oxygen
consumption, and low concentrations of antioxidants [[44]19, [45]74].
H-I brain injury remains a leading cause of neonatal mortality and
long-term neurological disabilities, including cerebral palsy,
epilepsy, and cognitive impairments [[46]37, [47]41, [48]54, [49]75].
Among various brain regions, the hippocampus is especially vulnerable
to H-I injury because of its high metabolic activity, extensive
synaptic connectivity, and active neurogenesis during early postnatal
development [[50]15, [51]19, [52]47, [53]90].
Despite advancements in neonatal care, effective therapies to prevent
or mitigate H-I-induced hippocampal damage remain limited [[54]68,
[55]75]. This is partly due to an incomplete understanding of the
immediate cellular and molecular mechanisms underlying neuronal
vulnerability and survival following H-I injury. Most previous studies
have focused on delayed neuronal death and glial activation occurring
days to weeks after the insult [[56]7, [57]20, [58]24, [59]32]. While
informative, these studies may overlook critical early events that
determine the trajectory of injury and recovery. Understanding the
acute responses that occur within hours of H-I injury is crucial, as
this period represents a potential therapeutic window during which
interventions could prevent irreversible damage and improve
neurological outcomes [[60]22, [61]43, [62]77, [63]91]. Elucidating the
cell-type-specific molecular changes during this early phase can reveal
key regulatory mechanisms and potential targets for neuroprotective
strategies.
Various animal models of neonatal brain injury have been developed to
explore cellular and molecular mechanisms and characterize functional
outcomes of H-I, each with its own advantages and limitations [[64]23].
In this context, our lab developed a modified Vannucci procedure to
induce hypoxia–ischemia in postnatal day 8 (P8, DOB = P0) mice,
consistently generating reliable injury outcomes [[65]34, [66]48,
[67]55, [68]60]. The P8 developmental stage, comparable to the human
fetal period of 32–42 weeks of gestation, was selected for its dynamic
balance between neurogenesis and gliogenesis, making it an ideal model
for studying the acute effects of hypoxia–ischemia on these processes.
Single-nucleus RNA sequencing (snRNA-seq) has emerged as a powerful
tool for dissecting cellular diversity in complex tissues, including
brain [[69]85]. Unlike single-cell RNA sequencing (scRNA-seq), which
requires dissociating intact cells—a process particularly challenging
for neurons—snRNA-seq isolates nuclei, making it suitable for profiling
hard-to-dissociate fresh/frozen tissues as well as cryopreserved
clinical samples [[70]4, [71]38, [72]63]. This approach enables the
characterization of cell-type-specific transcriptional responses at
single-cell resolution in both healthy and diseased states [[73]12,
[74]63]. To investigate the immediate cellular and molecular
adaptations in the neonatal hippocampus following H-I brain injury, we
isolated nuclei from the p8 mouse hippocampus under sham, hypoxia-only,
and hypoxia–ischemia conditions, and applied snRNA-seq using split-pool
barcoding technology [[75]56]. This method involves nuclei fixation
followed by three rounds of barcoding, allowing high-resolution gene
expression profiling from individual nuclei across multiple samples in
a single experiment. The combinatory barcoding approach increases
throughput and minimizes cost and batch effects, making it well-suited
for studying complex tissues and comparing multiple conditions.
Accurately annotating cell types in snRNA-seq data is challenging,
especially in the hippocampus (HPC), due to its intricate cellular
diversity [[76]86, [77]89] and the limitations of current computational
methods. Existing annotation methods often struggle to resolve closely
related cell types or identify rare populations even in healthy cell
types [[78]1, [79]16, [80]89], limiting the ability to detect subtle
yet critical alterations in more challenging disease contexts. To
overcome these challenges, we constructed a comprehensive hippocampal
cell atlas by integrating publicly available single-cell transcriptomic
datasets, encompassing over 420,000 cells across 33 distinct cell
types. Leveraging this atlas, we developed a robust machine
learning-based classifier for precise cell-type identification in both
normal and pathological hippocampi. We validated our classifier by
applying it to four recently published hippocampal transcriptomic
datasets not used in its training, demonstrating enhanced cell-type
annotations and correcting misclassifications.
We analyzed our snRNA-seq data and applied this classifier for accurate
cell type annotation across sham, hypoxia-only, and post-H-I
conditions. The analysis revealed specific vulnerabilities in mature
neuronal types within the hippocampal CA1, CA3, and dentate gyrus (DG)
regions as early as three hours post-H-I, while immature DG granule
cells, GABAergic neurons, and Cajal-Retzius cells exhibited remarkable
resilience. Gene regulatory network analysis identified key
transcription factors associated with neuronal vulnerability, such as
Hivep2, Bcl11a, Bcl6, and Thra, which are implicated in regulating
calcium homeostasis and synaptic function. Concurrently, we observed
rapid activation of astrocytes and microglia, marked by the
upregulation of neuroinflammation and tissue repair genes, with Runx1
identified as a potential key regulator in microglia that is associated
with early immune responses following ischemia. Endothelial cells
exhibited a complex transcriptional response and altered intercellular
signaling, potentially influencing vascular repair and modulating
neuroprotection, neuroinflammation, tissue remodeling, and ischemic
recovery.
Our study advances the understanding of immediate cellular and
transcriptional responses to neonatal hypoxia–ischemia injury,
providing new insights into hippocampal cell heterogeneity and
pathophysiology, which facilitate the identification of potential drug
targets and the development of neuroprotective strategies during this
critical therapeutic window. All data—including the hippocampal atlas,
post-H-I atlas, and the machine learning classifier—are publicly
accessible via an interactive web application
([81]https://hippo-seq.org).
Materials and methods
Animals
Mice with C57BL6NRj (Janvier labs, France) background were bred and
housed in a 12-h light and dark cycle with food and water ad libitum.
All experiments were approved by the Norwegian Animal Research
Authority and conducted following laws and regulations controlling
experimental procedures in live animals in Norway and the European
Union Directive 86/609/EEC and experiments conducted in Brazil were
approved by the Animal Ethics Committee of the Federal University of
Rio Grande do Norte (CEUA Proj. No. 051/2015). For the snRNA-seq
experiments, both male and female mice were used (Sham 1, Hypoxia 2,
and H-I 2 were females; Sham 2, Hypoxia 1, and H-I 1 were males).
Perinatal hypoxia–ischemia
Cerebral hypoxia and ischemia were produced in P8 (DOB = P0) mice by
permanent occlusion of the left common carotid artery (CCA) followed by
systemic hypoxia, as previously described^1, with some modifications.
In brief, pups were anesthetized with isoflurane (4% induction in
chamber, 2.5% maintenance via mask with a 2:1 mixture of ambient air
and oxygen). A midline incision was made on the ventral side of the
neck, and the common carotid was carefully identified and separated
from the vagus nerve. The artery was electrocoagulated using a
monopolar cauterizer (Hyfrecator 2000; ConMed) set to 4.0 W. The neck
incision was then closed with absorbable sutures (Safil 8–0 DRM6; B.
Braun Melsungen AG). The surgical procedure was completed within 5 min.
After the mice regained consciousness following surgery, they were
returned to their dam in the cage for a 1-h recovery period. After 1-h
recovery, the pups were exposed to a hypoxic (10% oxygen balance
nitrogen; Yara), humidified atmosphere for 45 min at 36.0 °C.
Afterward, the pups were returned to the cage with their dam and
sacrificed 3 h post-hypoxia–ischemia. The ipsilateral brain hemisphere
affected by carotid occlusion experiences hypoxic-ischemic conditions,
while the contralateral brain hemisphere is exposed to hypoxia alone.
Sham-operated animals were subjected to anesthesia and skin incision
but not to occlusion of the CCA or hypoxia. Sham samples were taken
from the left hippocampus, as the left hemisphere was experimentally
manipulated for ischemia. All animals were sacrificed at equivalent
time points (about five hours post-surgery intervention).
Hippocampal nuclei isolation and fixation
Mice were sacrificed by decapitation and brains were excised.
Hippocampi were isolated, snap-frozen in liquid nitrogen, and stored at
− 70 °C freezer for further processing. Hippocampi were homogenized
using a 1 ml glass Dounce Tissue Grinder (Wheaton, VWR, Cat. No.
#62,400–595) in 1 ml of EZ lysis buffer (Sigma Aldrich, Cat. No.
#N3408), with 10 strokes each using both Loose and Tight pestles for
optimal nuclei extraction. The homogenate was filtered through MACS®
SmartStrainers (30 µm) (Miltenyi Biotec, Cat. No. #130-098-458) and
nuclei were pelleted by centrifugation (850xg for 10 min at 4 °C). The
resulting pellet was resuspended in 2 ml of EZ lysis buffer and
centrifuged again. Finally, the nuclei pellet was resuspended in 2 ml
of PBS containing 1% BSA, and pelleted at 450xg for 10 min at 4 °C.
One million nuclei per sample were fixed using Evercode™ Nuclei
Fixation v2 Kit from Parse Biosciences (cat. # ECF2103), following the
manufacturer’s protocol. Briefly, nuclei were resuspended and incubated
in the fixation solution for 10 min on ice, followed by a 3-min
permeabilization on ice. The reaction was quenched with a
Neutralization Buffer. Nuclei were then pelleted by centrifugation
(350xg, 10 min, 4 °C) and resuspended in Nuclei Buffer (Parse
Biosciences cat. #WN101) for a final count. DMSO (Parse Biosciences
cat. #WN105) was added before freezing fixed nuclei at − 80 °C.
Single-nucleus RNA sequencing (snRNA-seq) using split-pool combinatory
barcoding technology
snRNA-seq was performed by Zenit Science
[82]https://zenitscience.com/single-cell-transcriptomics. Briefly,
fixed nuclei samples prepared using the Parse protocol were rapidly
thawed in a 37 °C water bath and immediately placed on ice. Barcoding
of single nuclei, amplification of barcoded cDNA, and preparation of
cDNA libraries for sequencing were performed using Evercode™ WT v2 Kit
from Parse Biosciences (cat. #EC-[83]W02030) following the
manufacturer’s protocol. Briefly, individual nuclear transcriptomes
were uniquely labeled by passing fixed nuclei through four rounds of
barcoding. In the 1 st round of barcoding, cDNA was generated with
in-nucleus reverse transcription (RT) reactions using well-specified
barcoded primers. After RT, nuclei were pooled and distributed in 96
wells for the 2nd round of barcoding by in-nucleus ligation. After
Round 2, nuclei were pooled and redistributed again into 96 wells for
the 3rd round of barcoding by in-nucleus ligation. After Round 3,
nuclei were pooled and split into 8 distinct populations, termed
sublibraries. The final split nuclei were lysed and the barcoded cDNA
underwent template switching and amplification. The cDNA was cleaned
using AMPure XP beads (Beckman Coulter cat. #A63880) and quality
checked using the Qubit dsDNA HS Assay Kit (Thermo cat. #[84]Q33231)
and a Bioanalyzer 2100 High Sensitivity DNA Kit (Agilent cat.
#5067-4626). Each cDNA sublibrary was fragmented and Illumina P5/P7
adapters were ligated during the final amplification (the 4th
sublibrary-specific barcode), followed by size selection and quality
check with the Bioanalyzer and Qubit. Libraries, with 5% PhiX spike-in,
were sequenced on an Illumina NovaSeq 6000 S4 Flow Cell using 150 bp
paired-end reads, achieving an average depth of 50,000 reads per
nucleus.
Construction of reference hippocampus cell atlas
To generate a comprehensive hippocampal cell (HCA) atlas, we integrated
publicly available single-cell and single-nucleus RNA sequencing
datasets from mice of various ages and experimental conditions,
totaling 424,225 cells. The datasets included: (1) Drop-seq data from
adult male mice aged P60–70 [[85]58] (n = 151,966; GEO: [86]GSE116470);
(2) nuclear RNA-seq from young (5 months) and aged (24 months) mice
under dietary restriction [[87]46] (n = 90,450; GEO: [88]GSE227515);
(3) snRNA-seq from mice aged 10–16 weeks receiving electroconvulsive
stimulation [[89]85] (n = 13,200; GitHub:
[90]https://github.com/Erik-D-Nelson/ARG_HPC_snRNAseq); (4) scRNA-seq
data from the Allen Brain Cell Atlas [[91]86] (7–10 weeks old;
n = 161,011; GEO: [92]GSE246717); and (5) SMART-seq profiling
of ~ 8-week-old wild-type mice [[93]46] (n = 7598; available at
[94]https://portal.brain-map.org/atlases-and-data/rnaseq/mouse-whole-co
rtex-and-hippocampus-smart-seq). For dataset integration and
annotation, we used scANVI [[95]82], a semi-supervised variational
autoencoder model designed for single-cell data. Each dataset was
loaded into an AnnData object and preprocessed with Scanpy version
1.9.8 [[96]80], including normalization and selection of the top 5,000
highly variable genes per batch. We registered the data with scVI
[[97]45] by specifying the counts layer and batch information, and
initially trained a scVI model with a 50-dimensional latent space to
capture the underlying gene expression patterns. We provided cell type
labels only for cell types that had consistent names across datasets;
otherwise, labels were kept as unlabeled_category ="Unknown". We then
initialized the scANVI model from the pretrained scVI model and trained
it for 20 epochs using both labeled and unlabeled data to improve cell
type annotation. After training, clusters labeled as"Unknown"were
manually examined, and cell type names were unified by majority vote
within each cluster. The latent representations and cell type
predictions were obtained using the get_latent_representation() and
predict() functions, respectively. To visualize the integrated data, we
applied UMAP dimensionality reduction to the latent space embeddings,
enabling us to assess clustering and cell type distributions across
datasets. The interactive version of integrated hippocampal cell atlas
(HCA) atlas is implemented using PHP v7.2 and D3.js version 7 [[98]8],
and is available at [99]https://hippo-seq.org/hca.
Development of classifier for automated hippocampal cell type annotation
To develop a classifier that accurately annotates hippocampal cell
types, we enhanced the ScType [[100]28] cell type classification method
and fine-tuned its accuracy specifically for hippocampal tissue. As a
first step, we identified an optimal set of marker genes that allow for
precise and specific annotation of each cell type within the
hippocampus. This optimization was performed using the integrated
hippocampal cell atlas (HCA) of 424,225 cells. However, there are over
10^83 possible marker gene sets when selecting between 1 and 30 marker
genes per cell type for 33 HCA cell types, rendering exhaustive search
computationally infeasible and prone to overfitting to HCA. To
efficiently navigate this vast search space, we employed advanced
machine learning techniques, specifically a Bayesian optimization
approach in combination with expectation maximization (EM) and a robust
leave-one-cell-type-out (LOCTO) cross-validation strategy.
Our objective was to find the optimal set of marker genes
[MATH:
S∗ :MATH]
that maximizes the classification accuracy
[MATH:
S∗=argmaxSf
(S) :MATH]
, where S represents a candidate set of marker genes for all cell
types. To model the relationship between marker gene sets and
classification accuracy, we constructed a Gaussian Process (GP)
surrogate model
[MATH: GP(μ(S),k(S,S′)) :MATH]
, where
[MATH: μ(S) :MATH]
is the mean prediction of accuracy for gene set S, and
[MATH: k(S,S′) :MATH]
is the kernel function measuring similarity between gene sets S and S′.
At each iteration t, we selected the next candidate marker gene set
[MATH:
St+1
:MATH]
by maximizing the Expected Improvement (EI) acquisition function
[MATH:
St+1=argmaxSEI(S)=∫-<
/mo>∞∞max(0,f-fbest)p(f|S)df :MATH]
, where
[MATH: fbest :MATH]
is the highest accuracy observed up to iteration t and
[MATH: p(f|S) :MATH]
is the posterior distribution of the performance given S. Within each
iteration, we used an Expectation Maximization (EM) algorithm where we
first assigned cell types based on current marker genes, then updated
these markers to maximize the likelihood of the observed cell type
assignments. The Bayesian optimization modelling was implemented using
mlrMBO v1.1.5.1 R package [[101]6].
To evaluate the performance of each candidate marker gene set and
prevent overfitting to HCA atlas, we implemented a
Leave-One-Cell-Type-Out (LOCTO) cross-validation strategy combined with
five-fold cross-validation. For each cell type c ∈ {1,…,C} (where
C = 33), we temporarily excluded its marker genes from S, causing
ScType to label cells of type c as"unknown". ScType was applied with
the modified S to annotate the cells. For the fold k, we calculated the
accuracy
[MATH:
fc,k(S) :MATH]
as the proportion of correctly assigned cells among all cell types
(including cells of type c). We repeated this process across all cell
types and K = 5 data splits (folds), obtaining an overall accuracy:
[MATH: f(S)=1C×K∑c=1C∑k=1K
fc,k(S) :MATH]
.
To enhance computational efficiency, we developed a vectorized version
of ScType. Given the input sc/nRNA-seq data
[MATH: X∈Rm×
n :MATH]
with m genes and n cells, ScType first standardizes each gene
expression profile into z-scores across all cells and multiplies it
with its marker specificity score θ[i]:
[MATH:
X′=(Z(XT))⊆MtT·Θ :MATH]
, where X′ is the transformed scRNA-seq expression matrix of n cells
and ∣M[t]∣ marker genes, within tissue t (in this case, hippocampus).
The marker sensitivity scores θ[i] are calculated based on the current
candidate marker gene set S. Specifically, θ[i] reflects the
specificity of gene i across the cell types in S:
[MATH:
θi=1ni-minj1njmaxj1nj-minj1nj :MATH]
, where n[i] is the number of cell types where gene i appears in S.
Here, the rescaling maps θ[i] to the range [0,1], ensuring that genes
appearing in fewer cell types receive higher scores. Next, the
standardized expressions of marker genes are multiplied by their
specificity scores θ[i]:
[MATH: Z~i,j=Zi,j×θi,fori∈S :MATH]
. Finally, for each cell type c and cell j, the marker enrichment score
E[c,j] is calculated as
[MATH:
Ec,j=1|Sc|∑i∈Sc
Z~i,j :MATH]
, and cell type with the highest enrichment score
[MATH: argmaxcEc,j :MATH]
is set for cell j.
The surrogate model and refined marker genes were iteratively updated
during 1,000,000 iterations and the optimal marker set S^∗ was then
used for cell type annotations with the vectorized ScType version.
We validated the optimized marker-based classifier on four independent
hippocampal transcriptomic datasets. The datasets were downloaded from
Single Cell Portal (SCP,
[102]https://singlecell.broadinstitute.org/single_cell)—SCP2162
[[103]57], SCP1375 [[104]87], SCP2065 [[105]18] and SCP110 [[106]21].
We normalized the raw count values of each dataset using Transcripts
Per Million (TPM). The datasets were further processed using the Seurat
v5.0.2 [[107]25] standard workflow and UMAP visualization was performed
with 20 principal component analysis (PCA) dimensions. Subsequently, we
applied vectorized ScType with a machine-learning-optimized marker gene
set to each of these datasets.
Processing snRNAseq data
For single-nucleus RNA sequencing analysis, we utilized the Seurat
v5.0.2 [[108]25] R package. Following quality control filtering
(minimum 200 genes per nucleus, maximum 20,000 nuclear UMIs/nUMIs),
data was normalized using SCTransform and principal component analysis
(PCA) was performed on the top 3000 variable genes. The first 30
principal components were used for uniform manifold approximation and
projection (UMAP) dimensionality reduction and graph-based clustering
with a resolution of 0.5. We utilized canonical correlation analysis
(CCA) on the top 30 principal components with Seurat’s IntegrateLayers
function, resulting in a unified dataset. Cell types were annotated
using vectorized ScType with a machine-learning-optimized marker gene
set. For differential expression analysis, we aggregated single-cell
data into pseudobulk samples by averaging expression within each
biological replicate (two mice per condition) and cell type using
Seurat's AverageExpression function, thereby enhancing statistical
power and mitigating false discovery rates [[109]65]. Differential
expression analysis was performed using the limma-voom pipeline from
the limma version 3.56.2 package in R [[110]53], utilizing pseudobulk
samples, with sex included as a covariate in the design matrix to
account for potential sex-specific effects. SCENIC (Single-Cell
rEgulatory Network Inference and Clustering) analysis was performed
using pySCENIC version 0.12.1 [[111]2, [112]70] command line interface
(CLI) within a Docker container downloaded from
[113]https://hub.docker.com/r/aertslab/pyscenic. The expression data
were normalized using TMM normalization and a log1p transformation,
followed by generating co-expression modules with the GRNBoost2
algorithm using ‘pyscenic grn’, separately for each condition. We then
refined the networks using ‘pyscenic ctx’ with RcisTarget version 10 to
retain only direct target genes with TF binding motifs, and finally
applied AUCell to quantify regulator activity in each cell with
‘pyscenic aucell’. The pathway enrichment analysis employed
clusterProfiler R package version 4.8.3 [[114]81] to perform KEGG
pathway and GO enrichment analyses on differentially expressed genes,
with significance thresholds of adjusted p value less than 0.05,
minimum gene count of 3, and gene ratio ≥ 5%. For network
visualizations, igraph 1.5.1 R package [[115]11] and D3.js version 7
were utilized. The interactive version of integrated
post-hypoxia–ischemia cell atlas is implemented using PHP v7.2 and
D3.js version 7, and is available at [116]https://hippo-seq.org/hi.
Paraffin brain sectioning and staining
Mice were deeply anesthetized with a single subcutaneous dose of
fentanyl/fluanisone plus midazolam (ZRF) followed by transcardial
perfusion with PBS. The mouse brains were excised and immersion-fixed
in 4% paraformaldehyde for at least 24 h. The fixed brains were then
dehydrated and hemisected along the midline, and both hemispheres were
embedded side-by-side in the paraffin block. Sagittal brain
Sects. (4 μm thick) of both hemispheres were sectioned using the
microtome Leica RM2255 (Leica Biosystems) and mounted on SuperFrost
Plus™ Adhesion slides (Epredia™, Thermo Fisher Scientific cat. #
J1800AMNZ). The sections were allowed to dry in the oven at 37 °C
overnight and stored at 4 °C for subsequent processing.
Sagittal brain Sects. (4 µm) were subjected to antigen retrieval in a
buffer containing 40 mM trisodium citrate (pH 6.0) at 99 °C for 3 min
and then washed with PBS. After blocking the sections in a blocking
buffer (PBS with 5% normal goat serum, 1% BSA, and 0.1% Triton X-100)
at room temperature for 1 h, the sections were incubated with primary
antibodies in a dilution buffer (PBS with 1% normal goat serum, 1% BSA,
and 0.1% Triton X-100) overnight at 4 °C. The next day, sections were
washed three times with 1 × PBST and incubated with secondary
antibodies in the same dilution buffer at room temperature for 1 h.
After another three washes in 1 × PBST, sections were mounted on glass
slides and left to dry overnight. Finally, sections were stained with
DAPI (1 μg/mL in PBS), washed, and cover-slipped using ProLong™ Gold
Antifade Mountant with DAPI (Thermo Fisher). The first antibodies used
are NeuN (mouse IgG1 1:500, Merck Millipore, Cat. No. MAB377, RRID:
AB_2298772), Calbindin1 (rabbit IgGs, 1:1000, Swant, Cat. No. CB 38a,
RRID: AB_10000340), Calretinin (mouse IgG1, 1:1000, Swant, 7699/3H,
RRID: AB_10000321). Secondary antibodies are from Thermo Fisher: Alexa
Fluor 488 anti-rabbit (1:1000, Cat. No. A32731, RRID: AB_2633280) and
Alexa Fluor 555 anti-mouse IgG1 (1:1000, Cat. No. A-21127,
RRIDAB_2535769).
Tunel assay was performed using the In Situ Cell Death Detection Kit,
Fluorescein following the manufacturer’s protocol (Roche, Cat. No.
#11,684,795,910).
3D image analysis and statistics
Microscopy was carried out using a Zeiss LSM 880 confocal laser
scanning microscope with a Plan-Apochromat 40x/1.4 Oil DIC M27
objective (Carl Zeiss, Jena, Germany). Z-stack images (1 µm intervals)
of the entire hippocampus were captured for analysis. 3D nucleus
segmentation and fluorescence intensity measurements were carried out
within the nuclei. Briefly, Nuclei were segmented using the Stardist
segmentation algorithm [[117]78], with a custom model trained on
DAPI-stained, 4 and 30-um thick mouse hippocampal images. The mean
fluorescence intensity (MFI) of each labeled nucleus was measured using
the scikit-image [[118]71]. NeuN-positive cells were classified based
on the mean NeuN fluorescence intensity within the nucleus. K-means
clustering (three clusters) was applied to categorize nuclei into three
groups: NeuN-, weak NeuN^+, and strong NeuN^+. The clustering algorithm
determined threshold values by minimizing variance within each group,
thereby assigning nuclei to one of the three categories according to
NeuN expression levels. The hippocampal subregions (CA1, CA3, and DG)
were manually annotated and cell densities for each subregion were
calculated. The number of NeuN^−, weak NeuN^+, and strong NeuN^+ cells
was quantified in the hippocampus under sham, hypoxia-only, and
hypoxia–ischemia conditions. The cell density was calculated by
manually annotating the analysis area and determining its volume. For
calbindin and calretinin-positive cells, clustering was performed using
K-means clustering with two clusters, and classification was limited to
nuclei in the dentate gyrus (DG). Apoptotic cells were analyzed using
the TUNEL assay and the TUNEL-positive cells were identified based on
an MFI threshold of 1.6 × that of all nuclei. For quantifying
cytoplasmic markers, such as GFAP, CD86, the cytoplasmic signal was
approximated by dilating the nuclear masks approximately 1 μm and
assigning each voxel to the closest nucleus. The mean fluorescence
intensity (MFI) of each labeled area was then measured. GFAP and
CD86-positive cells were classified using k-means clustering with two
clusters. Only cells within the hippocampal region were selected for
analysis.
The n-value for any analysis is reported as both the number of brain
sections analyzed and the number of independent animals from which
these sections were derived. Specific n-values for different markers
and experimental conditions are provided in the accompanying Table. For
clarity, ‘Contra’ refers to the hypoxia-only (contralateral) condition,
while ‘Ipsi’ denotes the hypoxic-ischemic (ipsilateral) condition.
Images with substantial damage to the analyzed area, such as rifts and
tears, were excluded before analysis. After quantification, outliers
were detected and removed using the interquartile range method.
Statistical significance between groups was determined using ANOVA,
with Tukey’s HSD test applied for post-hoc comparisons.
Marker Sham Contra Ipsi Total
Animals Sections Animals Sections Animals Sections Animals Sections
TUNEL 3 4 3 4 3 6 6 14
CD86 2 5 3 4 3 6 5 15
GFAP 2 7 3 6 3 6 5 19
Calb1 3 11 5 12 4 8 8 31
Calr 2 9 5 9 5 10 7 28
NeuN 3 6 3 4 3 6 6 16
[119]Open in a new tab
Results
Generating a comprehensive hippocampal atlas for accurate cell-type
annotation via machine learning
To explore early cellular and transcriptional adaptations in the
neonatal hippocampus following H-I, we induced cerebral ischemia in
wild-type mice at postnatal day 8 (P8) by permanently occluding the
common carotid artery in the ipsilateral brain hemisphere, followed by
systemic hypoxia (Fig. [120]1a). Histological examinations of samples
demonstrated significant injury in the cortex, hippocampus, striatum,
and thalamus of the affected hemisphere, whereas the contralateral
hemisphere remained largely undamaged, showing no visible harm when
compared to the brains of sham-treated mice [[121]60].
Fig. 1.
[122]Fig. 1
[123]Open in a new tab
snRNA-seq analysis of early transcriptional changes in developing mouse
hippocampus. a A schematic illustration of the experimental and
analytical workflow. Micro-dissected hippocampal samples from postnatal
day 8 (P8, DOB = P0) mice were subjected to three different conditions:
sham, 45-min hypoxia, and 3-h post-H-I. Single-nucleus RNA sequencing
(snRNA-seq) was performed on these samples, and the resulting data were
analyzed to identify molecular and cellular changes. The processed data
are available for exploration at the interactive web portal:
[124]https://hippo-seq.org. b Upper panels: Integration of publicly
available scRNA-seq mouse hippocampus datasets. The right panel
presents a UMAP visualization of the integrated reference hippocampus
atlas (HCA) utilized in the machine learning model, which can be
interactively explored at [125]https://hippo-seq.org/hca. Lower panels:
The left panel illustrates the machine-learning framework, where
Gaussian process optimization was used to refine gene sets that define
specific cell types (see methods). The optimized model was applied to
annotate four independent public single-cell transcriptomic datasets
(right panel), where it automatically assigned cell types as reported
in the original studies, while reannotating five unspecifically
classified and four incorrectly assigned cell types. c Integrated UMAP
visualizations of all three hypoxia–ischemia conditions in P8 mice,
with major cell types automatically annotated using the
machine-learning-based model. An interactive version of the data is
available at [126]https://hippo-seq.org/hi. d Violin plots showing the
expression levels of high-specificity canonical marker genes for each
cell type. e UMAP visualizations for selected marker genes
To further dissect the cellular alterations induced by H-I, we
performed single-nuclei RNA sequencing (snRNA-seq) using the split-pool
combinatorial barcoding technology (Parse Biosciences, see Methods) on
micro-dissected hippocampal samples from perinatal mice subjected to
different conditions: sham treatment, 45-min hypoxia, and 3-h
post-hypoxia–ischemia (Fig. [127]1a, Suppl. Fig. 1). Given the
complexity and heterogeneity of the hippocampal cellular landscape,
accurately annotating cell types in snRNA-seq data is critical,
particularly under conditions of hypoxia and ischemia, which
significantly alter transcriptional profiles.
To address the challenge of cell type annotation in the hippocampus, we
first established a robust foundation by constructing a comprehensive
reference hippocampal atlas. This atlas was built using well-annotated,
publicly available transcriptomic datasets from the hippocampus,
encompassing over 420,000 cells across 33 distinct cell types
(Fig. [128]1b upper panel, Suppl. Fig. 2,
[129]https://hippo-seq.org/hca) (see Methods). The integration of these
datasets was facilitated by variational autoencoders, which effectively
harmonized diverse datasets, and Gaussian Process (GP)-based
optimization, which identified a robust set of marker genes with high
discriminatory power for different cell types. The resulting model
automatically assigns cell types to query cells based on their gene
expression profiles (Fig. [130]1b left bottom panel) (see methods).
To validate the robustness and accuracy of our model, we applied it to
four recently published hippocampal transcriptomic studies, including a
spatial transcriptomics dataset [[131]18, [132]21, [133]57, [134]87].
Notably, these test datasets were not part of the data used for model
training. Our analysis revealed significant discrepancies in cell type
annotations across the published datasets when compared to the model's
predictions (Suppl. Fig. 3). For example, the model precisely
classified CA3 neurons into dorsal (Iyd ^high) and ventral (Nkd2^high)
subtypes, astrocytes into reactive (Gfap^high) and resident types, and
reclassified incorrectly annotated neuroblast granule cells into
immature (Prox1^+Calb1^low) and mature (Prox1^+Calb1^high) granule
cells in Farsi et al. study (Suppl. Fig. 3a). Similarly, in the spatial
transcriptomics dataset of Russel et al.[[135]57], the model accurately
distinguished mature and immature granule cells and reannotated
oligodendrocytes into oligodendrocytes (Mog^+) and oligodendrocyte
precursor cells (Pdgfra^+) (Suppl. Fig. 3d, e). Altogether, among the
four datasets, our model reannotated five unspecifically assigned and
four incorrectly identified cell types (Fig. 1b, right bottom panel).
These findings underscore the limitations of current annotation
methodologies in accurately characterizing the complex cellular
architecture of the hippocampus and highlight the superior accuracy and
specificity of our approach. Similarly, the use of supervised learning
techniques has proven to be highly effective in distinguishing
transcriptionally ambiguous cell subtypes across various single-cell
datasets from multiple systems and disease contexts [[136]64, [137]72,
[138]89]. The implemented machine learning-based model for annotating
user-provided single-cell hippocampus data, along with the integrated
hippocampal cell atlas, is publicly accessible at
[139]https://hippo-seq.org.
Cell census of the neonatal hippocampus under hypoxia and ischemia
Next, we applied the developed machine learning model to our snRNA-seq
dataset, which consisted of 27,380 nuclei from six hippocampal
samples—2 × sham-control, 2 × hypoxic, and 2 × hypoxic-ischemic.
Unsupervised clustering based on global gene expression profiles,
followed by UMAP visualization for dimensionality reduction, revealed
fourteen distinct cell populations (Fig. [140]1c). The pre-trained
model was then employed to assign cell types to each of these clusters.
The identified cell populations include a majority of neuronal cells
(83.4% of total sequenced cells) alongside a smaller fraction of
non-neuronal cells (16.6%). Specifically, the neuronal clusters
comprise Mki67^+ intermediate neural progenitors (IPCs), immature
Prox1^+ Calb1^low and mature Prox1^+Calb1^high Dentate Gyrus granule
cells (imGCs and mGCs), Fibcd1^+/Htr2c^+/Prss35^+ Cornu Ammonis (CA)
neurons, Gad1^+ inhibitory (GABA) neurons, and Trp73^+Reln^+
Cajal-Retzius (CR) cells. The non-neuronal populations include
Aldh1l1^+ astrocytes, Mbp^+ oligodendrocytes, Pdgfra^+ oligodendrocyte
progenitor cells (OPCs), vascular Flt1^+ endothelial cells,
Col3a1^+Ptgds^+ fibroblast-like cells, immune Cd86^+ microglia cells,
and Foxj1^+ ependymal cells (Fig. [141]1d, e, Suppl. Fig. 4). An
interactive version of the data is available at
[142]https://hippo-seq.org/hi, featuring unintegrated datasets for each
condition (sham, hypoxia-only, post-H-I) and an integrated dataset
combining all conditions (shown in Fig. [143]1c). This platform allows
for detailed exploration of hippocampal cell types under each condition
and enables investigation of gene expression and co-expression patterns
for user-defined genes and gene sets.
Selective vulnerability of neonatal hippocampal neurons to hypoxia–ischemia
To gain insights into the neuronal response to hypoxia-only and
hypoxia–ischemia in the neonatal hippocampus, we performed a focused
re-clustering of neuronal populations identified by our machine
learning model. Subtypes of neurons were isolated and analyzed,
including those from the subiculum, CA1, CA2, and CA3 areas, immature
and mature granule cells (imGC and mGC) in the dentate gyrus (DG),
Cajal-Retzius cells, and various GABAergic neurons characterized by
specific markers (Htr3a^+, Pvalb^+, Hapln1^+, Npy^+Sst^+, Tmem a/b^+)
(Fig. [144]2a). Post-hypoxia–ischemia, we observed a significant
proportional cell type reduction in mature neurons, particularly in the
CA and DG regions, with CA1 cells decreasing by 65.88%, mGCs by 54.31%,
CA3 cells by 26.41%, and pro-subiculum neurons by 15.55% (Fig. [145]2b,
c, Suppl. Fig. 5), consistent with prior findings of CA1's highest
vulnerability to ischemia [[146]35, [147]39, [148]59]. In contrast,
imGCs, GABAergic neurons, and Cajal-Retzius cells remained
statistically unaffected, indicating notable resilience. The reduction
of hippocampal CA1/CA3 neurons and DG-mGCs was further validated by
immunohistochemistry (Fig. [149]2d-i). The neuronal marker NeuN was
used to detect both mature and immature hippocampal neurons, based on
its expression levels reflected by strong and weak fluorescent
intensity [[150]9, [151]36] (Suppl. Fig. 6). A marked decrease in NeuN
fluorescent intensity was observed following hypoxia–ischemia
(Fig. [152]2d), along with a significant reduction in NeuN^+ cells in
the CA3 and DG (Fig. [153]2e, Suppl. Fig. 6c). Notably, the number of
strongly NeuN-expressing CA1/CA3 neurons and mGCs was significantly
reduced post-H-I (Fig. [154]2f, Suppl. Fig. 6c). This reduction in mGCs
was further validated by the mGC-specific marker Calbindin (Calb1^+)
(Fig. [155]2g), while no significant decrease in Calretinin-positive
(Calr^+) imGCs was detected (Fig. [156]2h). Additionally, the TUNEL
assay revealed a significant increase of apoptotic cells across all
hippocampal subregions, particularly in the DG-mGC layer
(Fig. [157]2i). Taken together, these results underscore the
vulnerability of mature hippocampal neurons to acute stress following
hypoxic-ischemic brain injury.
Fig. 2.
[158]Fig. 2
[159]Open in a new tab
Early neuronal responses in the neonatal hippocampus to hypoxia and
hypoxia–ischemia. a UMAP visualization of reclustered scRNA-seq data
highlighting neuronal subtypes in the hippocampus including granule
cells, Subiculum, CA1, CA2, CA3, Cajal-Retzius cells, and GABAergic
neurons. b UMAP plots of neuronal subtypes visualized under three
distinct conditions: sham (control), hypoxia alone, and H-I. c
Quantitative analysis shows a significant proportional reduction in
mature neuronal populations, especially in CA1 neurons and mature
dentate gyrus granule cells, with immature granule cells and GABAergic
neurons showing resilience. Statistical significance was determined
using permutation tests with alternative hypothesis that values in
hypoxia or hypoxia–ischemia groups are less than in sham group
(one-sided test), *p = 0.05, **p < 0.05. d Representative images show
NeuN + cells (green) and Dapi + nuclei (blue) in the hippocampus of P8
mice under three conditions. e Quantification of NeuN + cells per 100
um^3 tissue volume across hippocampal subregions. Data are presented as
mean ± SEM (n = 4–6 brain sections of 3 animals per condition),
*p < 0.05, **p < 0.01 (two-sided Wilcoxon test). f Quantification of
high-intensity NeuN + mature neurons per 100 um^3 DG tissue volume.
Data are presented as mean ± SEM (n = 4–6 brain sections of 3 animals
per condition), **p < 0.01 (two-sided Wilcoxon test). g Representative
images show NeuN + (green) and Calb + (red) mature granular cells
(mGCs) in DG of P8 mice under each condition. Quantification of
Calb + mGCs per 100 um^3 DG tissue volume is presented alongside. Data
are presented as mean ± SEM (n = 8–11 brain sections of 3–5 animals per
condition), *p < 0.05. h Representative images show NeuN + (green) and
Calr + (red) immature granular cells (imGCs) under sham, hypoxia-only,
and H-I. Quantification of Calr + imGCs per 100 um^3 DG tissue volume
is presented alongside. Data are shown as mean ± SEM (n = 9–10 brain
sections of 2–5 animals per condition). i Representative images show
increased apoptotic cells (green) detected by the TUNEL assay in the
mature DG granular cell layer after hypoxia–ischemia. Quantification of
apoptotic cells 100 um^3 tissue volume across hippocampal subregions is
presented alongside. Data are presented as mean ± SEM (n = 4–6 brain
sections of 3 animals per condition), *p < 0.05 (two-sided Wilcoxon
test). j Alignment of distinct conditions: sham (control),
hypoxia-only, and H-I with Slide-seq spatial transcriptomics data
To better understand the spatial and molecular context of these
changes, we aligned our scRNA-seq data with Slide-seq spatial
transcriptomics of the mouse hippocampus [[160]67]. This alignment
revealed that hypoxia alone did not significantly alter gene expression
profiles, allowing consistent matching with Slide-seq data
(Fig. [161]2j). In contrast, post-hypoxia–ischemia, neurons in
hippocampal subregions became poorly aligned, while non-neuronal cells
such as endothelial cells, microglia, and astrocytes remained mostly
unaffected. Notably, even the relatively resilient immature granule
cell (imGC) population showed similar misalignment, indicating that
transcriptional profiles and RNA integrity were compromised post-injury
in various neuronal populations, reflecting complex stress-induced
alterations.
Acute transcriptional signatures of neuronal resilience and vulnerability to
hypoxia–ischemia in the neonatal hippocampus
To further dissect transcriptional changes in distinct neuronal cell
clusters, we performed differential gene expression (DE) analysis under
both post-hypoxia (hypoxia-only vs sham) and post-hypoxia–ischemia (H-I
vs sham) conditions. A significant increase in differentially expressed
genes (DEGs) was observed across all neuronal subtypes following
hypoxia–ischemia, with the majority being upregulated (Fig. [162]3a,
Suppl. File 1), highlighting a pronounced ischemia-induced
transcriptional shift.
Fig. 3.
[163]Fig. 3
[164]Open in a new tab
Cell type-specific transcriptional responses and key transcription
factors underlying neuronal vulnerability and resilience post-H-I. a
Differential expression analysis of neuronal subtypes under the
post-hypoxia and hypoxia–ischemia conditions compared to sham
(FDR < 0.1, |log[2]FC|> 1). b Gene set enrichment analysis (GSEA) of
neuronal subtypes post-hypoxia–ischemia. c Scaled average expression of
ribosome biogenesis-related genes across neuronal subtypes under sham,
hypoxia-only, and H-I conditions (left), and their expression across
various neuronal cell types following H-I (right). d UMAP plots of
cells re-clustered based on regulon activity profiles (AUCell scores)
for sham and post-H-I conditions. e Volcano plot displaying
differential regulon activity between H-I and sham conditions
(FDR < 0.05, |AUCell difference|> 0.1). f Left panel: Regulons specific
to imGCs ranked by regulon specificity score (RSS), see methods. Higher
RSS indicates greater specificity for imGCs compared to other cell
types. Top 5 regulons are shown in red. Right panel: AUCell score
histogram for the Neurod1 regulon in imGCs and UMAP plot under H-I
conditions, highlighting cells with high (blue) and low (grey) regulon
activity (right), demonstrating strong specificity to imGCs. g Regulons
specific to vulnerable or resilient neurons post-hypoxia–ischemia.
Regulons with significantly altered activity (as shown in panel e) and
high specificity to vulnerable or resilient groups are highlighted in
bold. h Gene regulatory network (GRN) depicting regulons Hivep2,
Bcl11a, Bcl6, and Thra and their top downstream target genes with the
most altered regulatory importance after hypoxia–ischemia. i Pairwise
correlation matrix of transcription factors in neuronal subtypes,
revealing distinct co-expression modules: M1 (common across cell
types), M2 (GABAergic neurons), M3 (CA1 pyramidal neurons), and M4
(Cajal-Retzius cells). j Interaction network of transcription factors
within each co-expression module. Nodes represent TFs, colored by the
cell type with the highest AUCell score. Edges represent correlations
between TFs (using a Pearson correlation threshold of ρ > 0.45)
Gene set enrichment analysis revealed distinct molecular adaptations in
affected neuronal populations (Fig. [165]3b). Specifically, upregulated
ribosomal biogenesis was identified in neuronal populations exhibiting
increased apoptosis (CA1-3 and mGCs) following hypoxia–ischemia
(Fig. [166]3b). Genes related to ribosome biogenesis (e.g., Nol6, Imp4,
Surf6, Rrp9, Rpl6, Rpl26, Rpl41, Rps20) were predominantly upregulated
in these neurons (Fig. [167]3c). This response was absent in GABAergic
neurons and Cajal-Retzius cells, as well as non-neuronal cells (Suppl.
Fig. 7a), suggesting that affected neurons initiate an acute
transcriptional response prioritizing protein synthesis for repair,
despite the energy crisis induced by hypoxia–ischemia. Immature granule
cells (imGCs) showed upregulation of pathways related to autophagy
(e.g., Atg12, Becn1, Gabarapl2, Sesn2) and proteolysis (e.g., Furin,
Psmc3, Ctsl). Similar patterns, though not statistically significant
across all populations, were observed in all neuronal types (Suppl.
Fig. 7b), implying conserved stress-response mechanisms in neurons
following hypoxia–ischemia [[168]10, [169]40]. Additionally, imGCs
exhibit specific downregulation of pathways associated with
neurogenesis (e.g., Prox1, Ntf3, Sema3c) and axon extension (e.g.,
Sema5a, Sema3c, Myo5b), suggesting that imGCs prioritize survival over
growth by conserving resources and stabilizing existing neuronal
structures rather than promoting new development. Interestingly,
Cajal-Retzius cells show upregulated genes (Arl13b, Traf3ip1, Hspg2,
Dcdc2a) associated with Smoothened signaling, a key component of the
neuroprotective Hedgehog pathway, known for its role in promoting cell
survival and neuroprotection [[170]42] (Fig. [171]3b, Suppl. File 1).
Taken together, these transcriptional changes underlie the observed
increased vulnerability of excitatory neurons in CA1-3 and mGCs in DG
post-H-I, while also supporting the greater resilience observed in
GABAergic neurons, Cajal-Retzius cells, and imGCs.
Next, we utilized SCENIC (Single-Cell rEgulatory Network Inference and
Clustering) analysis to reconstruct cell type-specific gene regulatory
networks (GRNs) and identify key transcription factors (TFs) underlying
neuronal vulnerability and resilience after H-I. Cells were reclustered
based on their regulon activity profiles using AUCell (see methods),
with neuronal cell types remaining largely separated post-ischemia,
suggesting preserved core regulatory identities despite stress-induced
transcriptional changes (Fig. [172]3d). Differential regulon activity
analysis between H-I and sham conditions revealed TFs with
significantly altered regulatory influence, including downregulated
Hivep2, Bcl6, Bcl11a, Thra, Zbtb20, Bach2, Klf7, Emx1, Nr3c1, Rarb,
Jun, and Elk4, and upregulated Pdlim5, Zfp148, Zfp14, Klf12, Pax6, Maf,
Nfib (Fig. [173]3e). To characterize cell-type-specific regulatory
roles of these TFs, we performed regulon specificity analysis,
identifying key regulons for each cell type (Suppl. File 2), such as
Neurod1 for imGCs (Fig. [174]3f). Strikingly, common regulons were
identified in vulnerable and resilient hippocampal cell populations
post-H-I (Fig. [175]3g), highlighting the distinct molecular responses
that underly cell-type-specific reactions to hypoxia–ischemia.
Integration of differential regulon activity and regulon specificity
analyses revealed that certain regulons, including Hivep2, Bcl6,
Bcl11a, and Thra, exhibited both decreased activity (Fig. [176]3e) and
increased specificity (Fig. [177]3g) in vulnerable neurons
post-ischemia, suggesting their critical role in neuronal
susceptibility to damage. Conversely, Klf12 and Pdlim5 regulons
demonstrated upregulated activity and higher specificity in resilient
neurons, indicating their potential importance in neuronal survival
under ischemic conditions (Fig. [178]3g, marked in bold). Delving
deeper into the transcription factors linked to neuronal vulnerability,
we found that Hivep2, Bcl11a, Bcl6, and Thra significantly impacted
genes involved in calcium homeostasis and signaling (Fig. [179]3h).
Hivep2 showed increased regulatory importance for Cacnb2 (voltage-gated
calcium channel subunit) and decreased influence on Atp1b1
(Na⁺/K⁺-ATPase subunit) as well as on Prkcg, Cpne6, and Grm7
(calcium-dependent neuronal signaling and function). Bcl11a exhibited
increased regulatory importance for Camk2a
(calcium/calmodulin-dependent protein kinase II alpha) and decreased
importance for Slc8a1 (sodium/calcium exchanger). Bcl6 demonstrated
increased regulatory influence on Atp2b1 (plasma membrane calcium
ATPase 1), Chrm3 (muscarinic acetylcholine receptor M3), and Grin2a
(NMDA receptor subunit 2 A). Conversely, Thra exhibited decreased
regulatory influence on calcium-related genes, including Camk2a,
Camk2b, and Calm2 (calmodulin 2). The altered regulatory influence of
key transcription factors on calcium-related genes underscores the
critical role of calcium dysregulation in promoting neuronal death,
aligning with the well-established mechanisms of calcium-dependent
excitotoxicity in neurons. In addition, Hivep2 decreased regulation of
Atp1b1; Bcl11a altered Kalrn and Rbfox1; Bcl6 modified Bcl11b, Negr1,
and Lsamp; and Thra reduced influence on Itgb1, Matn2, Dync1i2, Mapt,
Stmn2, Sptbn2, and Shank1. These changes in genes related to ion
homeostasis, synaptic function, neuronal differentiation, adhesion, and
cytoskeletal organization may contribute to neuronal vulnerability
post-injury.
To further elucidate the regulatory landscape, we performed regulon
co-expression analysis by calculating pairwise correlations of
transcription factors (TFs) in neuronal cell types and identified four
distinct modules, specific to different neuronal cell types
(Fig. [180]3i). A correlation matrix was used to visualize interactions
between TFs within each module, with each TF colored according to the
cell type with the highest AUCell score (Fig. [181]3j). This analysis
revealed that the transcription factors Hivep2, Bcl6, Bcl11a, and Thra
clustered together, further suggesting a coordinated regulatory program
with potential synergistic action of these TFs in regulating
calcium-related genes in the vulnerable neuronal cell types.
Rapid glial activation in the neonatal hippocampus following hypoxia–ischemia
Next, we expanded our analysis to non-neuronal cell populations by
reclustering and annotating them using our machine learning-based
approach, identifying seven distinct cell types: astrocytes, microglia,
oligodendrocytes, oligodendrocyte precursor cells (OPCs), endothelial
cells, fibroblast-like cells, and ependymal cells (Fig. [182]4a).
Remarkably, just three hours after the H-I event, we observed a
significant proportional increase in activated microglia and astrocytes
(Fig. [183]4b). This rapid increase suggests cellular activation rather
than proliferation, as evidenced by elevated expression of Cd86 in
microglia and Gfap in astrocytes—both markers associated with reactive
gliosis [[184]17, [185]30]—implying an immediate neuroinflammatory
response following H-I injury (Fig. [186]4c). Interestingly, when
examining the effects of hypoxia alone, without concurrent ischemia,
only astrocytes exhibited increased activation, with no significant
changes observed in microglia (Fig. [187]4b,c). The upregulation of
activated microglia (Cd86 +) and astrocytes (Gfap +)
post-hypoxia–ischemia was confirmed by immunohistochemistry
(Fig. [188]4d, e). Statistical analysis using a two-sided Wilcoxon test
supported these findings, reinforcing the notion of a rapid glial
response to H-I injury.
Fig. 4.
[189]Fig. 4
[190]Open in a new tab
Early non-neuronal responses in the neonatal hippocampus to hypoxia and
hypoxia–ischemia. a UMAP visualization of reclustered scRNA-seq data
highlighting non-neuronal cell subtypes in the hippocampus. b
Proportional changes in the populations of activated microglia and
astrocytes across sham (control), hypoxia-only, and hypoxia–ischemia
conditions. c Expression levels of Cd86 in microglia and Gfap in
astrocytes under different conditions. d. Representative
immunohistochemistry images of hippocampal tissue stained for Cd86
(microglia) and Gfap (astrocytes) in control, post-hypoxia, and
post-hypoxia–ischemia conditions. e Quantification of
immunohistochemistry staining intensity for Cd86 and Gfap. Data are
presented as mean ± SEM (n = 4–7 brain sections of 2–3 animals per
condition), *p < 0.05 (two-sided Wilcoxon test)
Multifaceted early transcriptional responses in hippocampal non-neuronal
cells to hypoxia–ischemia
Similarly, we performed differential gene expression analysis in
non-neuronal cell populations, comparing hypoxia-only versus sham and
hypoxia–ischemia versus sham conditions to identify early
transcriptional changes in these cell types. Astrocytes exhibited
hundreds of DEGs in both hypoxia-only and hypoxia–ischemia conditions,
indicating a transcriptional shift in astrocytes in response to both
stressors (Fig. [191]5a). Gene set enrichment analysis revealed that
astrocytes, activated under both hypoxia-only and H-I, exhibited a
multifaceted response (Fig. [192]5b). Under both conditions, the
cholesterol biosynthesis pathway was upregulated, indicated by
increased expression of genes such as Hmgcr, Hsd17b7, Pmvk, Fdft1,
Msmo1, and Lbr, which are essential for membrane repair, synaptic
maintenance, and neuroprotection. Additionally, genes involved in
positive regulation of catabolic processes (Lpl), wound healing (Cd44,
Pdgfa, Hbegf, Tgfb1), and response to external stimuli (Atf3, Map3k14,
Nrros) were upregulated, suggesting that astrocytes enhance metabolic
activity, facilitate tissue repair, and adapt to stress signals
following hypoxic insult. Following ischemia, astrocytes uniquely
upregulated genes involved in the regulation of apoptotic processes,
including Dapk2, Pawr, Faim, Bcl2l11, and Sesn2, indicating a response
to increased cellular damage by modulating apoptosis to balance
survival and programmed cell death. Genes associated with the
stress-activated MAPK pathway—such as Dusp3, Dusp10, Dusp14, Junb, and
Fos—were also specifically upregulated after ischemia, which may enable
astrocytes to manage heightened stress signals, regulate inflammatory
responses, and initiate repair mechanisms. Furthermore, upregulation of
genes related to cytoskeletal remodeling and cell migration, including
Rnd1, Amotl2, Itga1, Thbs1, and Ror2, suggests that astrocytes begin
preparing for migration to H-I injury sites even at this early time
point.
Fig. 5.
[193]Fig. 5
[194]Open in a new tab
Early transcriptional responses in non-neuronal populations. a
Differential expression analysis of non-neuronal cells under
hypoxia-only versus sham and H-I versus sham conditions. Bars represent
the number of DE genes (false discovery rate [FDR] < 0.1, |log[2]fold
change|> 1). b Gene set enrichment analysis (GSEA) of non-neuronal cell
populations under hypoxia-only and H-I conditions. c Volcano plot
illustrating differential regulon activity between H-I and sham
conditions. Significantly up- and down-regulated regulons are labeled
(FDR < 0.05, |AUCell difference|> 0.1). d Violin plots depicting Runx1
expression levels in microglia under sham, hypoxia-only, and H-I
conditions. e Gene regulatory network (GRN) of Runx1 and its
upregulated downstream targets in microglia. f Heatmaps illustrating
differential intercellular signaling patterns quantified as changes in
information flow between hypoxia-only vs sham (left) and H-I vs sham
(right). Information flow, derived from CellChat analysis, represents
the aggregate strength and probability of ligand-receptor interactions
between cell populations (see methods). The color scale indicates the
increased (red) and decreased (blue) changes in information flow. g
Circular plots depicting signaling probability differences for VEGFA
pathway via Flt1/Kdr receptors (left) and ANGPT2 pathway via integrin
α5β1 receptors (right). Line color indicates the increase (red) or
decrease (blue) in signaling probability
The significantly increased DEGs were also identified in microglia,
fibroblasts, and endothelial cells following H-I, with the majority
being upregulated (Fig. [195]5a, Suppl. File 3). Microglia responded
specifically to H-I injury by upregulating genes related to
phagocytosis (e.g., Dab2, Ctsb, Ctsl), response to external stimuli
(Runx1, Tgfb1, Prkch), and wound healing (Tgfbi, Tgfb1). SCENIC
analysis identified Runx1 as a key transcriptional regulator in the
microglial response following H-I injury (Fig. [196]5c). Notably, Runx1
itself showed significant upregulation in both expression levels and
the proportion of microglia expressing it (Fig. [197]5d). Runx1
exhibited increased regulatory activity on diverse downstream targets
encompassing several crucial aspects of microglial function
(Fig. [198]5e). These targets include genes involved in immune
activation (e.g., Cd86, Ptprc/Cd45, C3ar1, Ccl4, Cd84, Cd83, Il1a),
phagocytosis and debris clearance (Fcrls, Sirpa, Lrp1, C1qb), and
cytoskeletal remodeling and migration (Cyth4, Picalm, Arhgap25,
Arhgap30, Cyfip1). The extensive influence of Runx1 across these
functional categories underscores its potential role as a master
regulator orchestrating multiple aspects of the acute microglial
response to H-I injury in the developing brain. Additionally,
intercellular signaling analysis suggests increased autocrine signaling
of microglia through Csf, especially post-H-I, while Csf1 regulatory
activity was mostly increased by Runx1 (Fig. [199]5f).
DE and SCENIC analyses revealed a complex transcriptional response of
endothelial cells post-ischemia. We observed transcriptional
downregulation and decreased activity of the Tead4 and Lef1 TFs,
involved in developmental gene regulation and cellular differentiation
[[200]14, [201]88] (Fig. [202]5c). Conversely, there was increased
activity in the Otx2, Nr3c1, Myc, Bach2, and Elf1 regulons, and key
angiogenic genes such as Angpt2, Vegfa, Pecam1, Itgav, Edn1, and S1pr1
were transcriptionally upregulated (Fig. [203]5b). These findings
suggest that endothelial cells respond to ischemia by downregulating
certain developmental transcription factors while simultaneously
activating angiogenic pathways to promote vascular repair.
Additionally, DE analysis identified upregulation of pathways related
to wound healing and positive regulation of locomotion, with increased
expression of genes such as Dock5, Nedd9, S1pr1, Itgav, and Itga6,
suggesting that endothelial cells are initiating processes related to
vascular remodeling and maintenance of blood–brain barrier integrity in
response to hypoxic stress (Fig. [204]5b).
Fibroblast-like cells exhibited upregulation of genes associated with
migration and collagen metabolism—processes important for vascular
remodeling and wound healing after ischemia (Fig. [205]5b). At 3 h
post-ischemia, increased expression of genes such as Fgf2, Itga5, Cd44,
and Tns1 suggests initiation of signaling pathways related to
migration, even though significant cell movement is unlikely to occur
at this early stage. Similarly, upregulation of collagen metabolism
genes—including Mmp19, Mmp28, Ctsl, Col5a3, Col6a1, and Hspg2—indicates
early activation of extracellular matrix remodeling and synthesis.
These findings suggest that fibroblast-like cells begin engaging in
pathways associated with tissue repair soon after ischemic injury.
Altered intercellular communication in the neonatal hippocampus following
hypoxia–ischemia
The cell–cell communication analysis revealed substantial changes in
the signaling information flow among neuronal and non-neuronal cells
following hypoxia-only and hypoxia–ischemia (Fig. [206]5f). Endothelial
cells exhibit the most pronounced alterations in their signaling
pathways following H-I, particularly with an upregulation in
endothelial autocrine signaling involving VE-cadherin (Cdh5) and PECAM1
(Cd31), and downregulation of the Apelin signaling pathway
(Fig. [207]5f). The upregulation of VE-cadherin and PECAM1 signals is
essential for endothelial cell junctions, controlling vascular
stability and permeability. In contrast, the downregulation of Apelin
signaling, which regulates angiogenesis and vascular tone, may reflect
early endothelial adaptations to modulate angiogenesis and stabilize
the vasculature in the injured brain. Additionally, endothelial cells
displayed altered communication with astrocytes, microglia, and other
non-neuronal cells through increased Angpt2 signaling via the integrin
α5β1 receptor (Fig. [208]5g, right panel). Angiopoietin-2 (Angpt2) is
involved in vascular remodeling and inflammation [[209]69], suggesting
that endothelial cells may engage in crosstalk with surrounding cells
to coordinate vascular repair processes and modulate immune responses.
These signaling shifts in endothelial cells and among intercellular
communications highlight the pivotal role of endothelial cells in
modulating the vascular response to hypoxia–ischemia, impacting
neuroprotection, neuroinflammation, vascular remodeling, and recovery
processes.
Strikingly, vulnerable neurons (CA1, CA3, mGCs) exhibited increased
communication with endothelial cells via the vascular endothelial
growth factor (VEGF) signaling pathway. Upregulation of Vegfa signaling
and interactions with its receptors Flt1 (VEGFR-1) and Kdr (VEGFR-2) on
endothelial cells may serve as early signals initiating angiogenesis
(Fig. [210]5g, left panel). This finding complements the observed
upregulation of angiogenesis-related genes in endothelial cells,
suggesting a coordinated response between vulnerable neurons and
endothelial cells to restore oxygen supply following ischemic insult.
Conversely, resilient neurons such as imGC and GABAergic neurons did
not show significant changes in cell communication pathways. This
suggests that their resilience is likely maintained through the
enhancement of existing signaling mechanisms and autonomous protective
strategies, rather than through the activation of new intercellular
communication routes.
Discussion
Neonatal hypoxic-ischemic brain injury is a leading cause of
neurodevelopmental disabilities and is responsible for 23% of infant
mortality [[211]19, [212]26, [213]74]. The hippocampus is particularly
vulnerable to H-I due to its high metabolic demands and critical role
in cognitive functions [[214]15, [215]19, [216]47, [217]90]. In this
study, we provide novel insights into the immediate cellular and
transcriptional responses of the neonatal hippocampus following H-I
injury, using single-nucleus RNA sequencing (snRNA-seq) combined with a
machine learning-based cell annotation framework. By examining the
acute phase, just three hours post-ischemic injury, we uncovered
cell-type-specific vulnerability patterns, adaptive stress responses,
and key transcriptional regulators that may influence long-term
neurological outcomes. By constructing a comprehensive reference atlas
and training our model on it, we achieved high specificity in cell type
identification, enabling the detection of subtle changes in cell states
and regulatory networks that might be overlooked using conventional
methods. This machine learning-based framework enhances the resolution
and reliability of our findings and sets a new standard for
interrogating cellular heterogeneity and dynamics in complex tissues
under various physiological and pathological conditions. Our
hippocampal reference atlas was constructed using datasets from mice
aged 1.5 months and older due to the limited availability of
high-quality annotated data from the neonatal hippocampus. Future work
could incorporate neonatal transcriptomic data as it becomes available
to further refine the atlas and enhance classification accuracy for
developmental studies.
A principal finding of our study is the acute susceptibility of mature
hippocampal neurons to H-I injury. We observed a significant reduction
in mature CA1 and CA3 pyramidal neurons and mature dentate gyrus
granule cells (mGCs) corroborated by immunohistochemical analyses using
neuronal marker NeuN and mGC-specific marker Calbindin. Consistently,
increased apoptosis was detected in these populations. The rapid loss
of these neurons underscores a critical therapeutic window in which
interventions might prevent irreversible damage and preserve
neurological function. Notably, upregulated ribosome biogenesis was
detected in vulnerable neurons within CA1-3 and GC regions, possibly as
an attempt to restore damaged cellular components. However, this
energy-intensive process [[218]31] may exacerbate neuronal stress under
H-I conditions due to limited energy resources during insult [[219]27,
[220]84]. Dysregulated ribosome biogenesis was shown to contribute to
disease progression and worsen ischemic stroke outcomes by driving
inflammation and neuronal stress [[221]31, [222]76], making it a
promising therapeutic target to mitigate cellular damage and improve
recovery.
In contrast, certain neuronal subtypes exhibited remarkable resilience
to H-I. Immature granule cells (imGCs), GABAergic interneurons, and
Cajal-Retzius (CR) cells maintained their proportions and displayed
distinct transcriptional profiles suggestive of adaptive stress
responses. imGCs upregulated autophagy and proteolysis pathways while
downregulating genes associated with neurogenesis and axon extension.
This suggests a strategic shift towards energy conservation and
survival under stress conditions, potentially attributed to their
developmental stage, as immature neurons often possess greater
plasticity and adaptability to stress [[223]54]. However, while
proteolysis supports cellular maintenance in imGCs, modulating
proteolysis—to balance protein cleanup without promoting excessive
breakdown—is essential, as studies show that inhibition of proteolysis
can reduce neuronal death after H-I [[224]40]. These findings highlight
the potential of targeting imGCs, with carefully modulated proteolytic
pathways, to promote endogenous repair and neurogenesis following H-I
injury. CR cells upregulated genes associated with the Hedgehog
signaling pathway, including components of Smoothened signaling, known
for their roles in neuroprotection, cell survival, and differentiation
[[225]42]. Activation of this pathway may enable CR cells to withstand
hypoxic-ischemic stress, contributing to their resilience.
Through differential expression and gene regulatory network analyses,
we identified key transcription factors (TFs) associated with neuronal
vulnerability and resilience. In vulnerable neurons, we observed
downregulation of TFs such as Hivep2, Bcl11a, Bcl6, and Thra, which are
crucial for neuronal development, synaptic function, and calcium
homeostasis [[226]5, [227]13, [228]66, [229]79]. Altered regulation of
calcium-related genes by these TFs may contribute to excitotoxic
neuronal death, a well-established mechanism in ischemic injury
[[230]51, [231]62]. For instance, dysregulation of Hivep2 and Bcl11a
has been implicated in neurodevelopmental disorders and synaptic
dysfunction [[232]13, [233]66], and our findings suggest their roles
extend to mediating neuronal vulnerability in H-I injury. Notably,
Hivep2 and Bcl6 are critical transcription factors in the development
of adult-born neurons [[234]52], suggesting their downregulation may
impact both initial injury responses and subsequent regenerative
processes. These findings underscore the potential of targeting these
TFs to stabilize calcium homeostasis and mitigate excitotoxicity,
providing a potential therapeutic approach to ischemic injury.
We observed a rapid activation of astrocytes and microglia in response
to H-I, evidenced by increased expression of Gfap and Cd86,
respectively. Astrocytes upregulated cholesterol biosynthesis pathways,
essential for membrane repair and synaptic stability [[235]50], and
genes involved in wound healing and stress responses. The upregulation
of genes regulating apoptosis and the MAPK pathway indicates that
astrocytes are actively modulating cell death and inflammatory
signaling following injury. However, prolonged astrocyte activation
shows both beneficial and detrimental effects on neural plasticity and
functional recovery [[236]49], suggesting that therapeutic strategies
should modulate astrocyte function rather than broadly suppress or
enhance their responses.
Microglia exhibited upregulation of genes associated with phagocytosis,
immune activation, and wound healing, with Runx1 (AML1) emerging as a
key TF orchestrating these responses. Known for regulating microglial
proliferation and differentiation during development, Runx1 exhibited
increased activity on downstream targets involved in immune activation,
debris clearance, cytoskeletal remodeling, and migration [[237]33,
[238]44]. This underscores its role as a master regulator of the acute
microglial response to hypoxia–ischemia injury in the developing brain.
Enhanced autocrine and paracrine signaling of microglia via CSF1 and
VEGFA pathways underscores the dynamic interplay between microglia,
neurons, and endothelial cells in coordinating early immune responses
and initiating repair mechanisms. Modulating Runx1 activity could help
balance beneficial and detrimental microglial functions in neonatal
brain injury.
Endothelial cells displayed a complex transcriptional response
post-H-I, characterized by downregulation of developmental TFs (Tead4
and Lef1), increased activity of angiogenic TFs (Myc and Elf1), and
upregulation of pro-angiogenic TFs (Vegfa, Angpt2, and Pecam1). This
suggests an immediate attempt to initiate vascular repair and restore
oxygen supply, shifting from the developmental program. The increased
interaction between vulnerable neurons and endothelial cells via VEGFA
signaling highlights the significance of neurovascular communication in
responding to ischemic injury. Altered signaling involving Angpt2 and
integrins indicates endothelial involvement in modulating vascular
permeability and inflammation [[239]69]. These adaptations may
facilitate revascularization and support the restoration of cerebral
blood flow.
Our findings underscore the critical importance of the early
post-injury period in shaping outcomes following neonatal H-I. The
rapid loss of mature neurons suggests that timely interventions aimed
at preventing excitotoxicity and stabilizing calcium homeostasis could
be particularly effective. Targeting the identified TFs associated with
neuronal survival and death pathways offers potential strategies for
therapeutic modulation. Enhancing the activity of protective TFs or
signaling pathways in vulnerable neurons, or leveraging resilience
mechanisms in GABAergic neurons and CR cells, could mitigate neuronal
loss. Moreover, the activation of astrocytes and microglia presents
both challenges and opportunities. While glial responses are essential
for debris clearance and initiating repair, excessive activation can
lead to inflammation and secondary injury [[240]61, [241]83].
Therapeutic approaches that modulate glial activation to enhance
protective functions while minimizing detrimental effects could improve
recovery outcomes. Supporting endothelial cell function and promoting
angiogenesis might also facilitate the restoration of cerebral blood
flow, reduce tissue damage, and accelerate recovery processes.
This study provides a comprehensive assessment of immediate cellular
and molecular responses within the therapeutic window
post-hypoxia–ischemia, revealing cell-type-specific vulnerability and
neuroprotective mechanisms. It should be noted that our analysis
utilized two mice per condition (balanced for sex), which represents a
limitation in fully capturing biological variability. While this design
allowed identification of robust transcriptional patterns across cell
types, the small sample size may have limited our ability to detect
more subtle cell-type-specific changes. Future studies with larger
cohorts, including both sexes in different conditions, would
substantially enhance our understanding of the heterogeneity in
cellular responses to hypoxic-ischemic injury. Future work will also
incorporate multiple time points post-H-I to capture the temporal
progression of cellular and transcriptional changes. Longitudinal
studies are necessary to understand the dynamics of these responses and
identify critical windows for intervention. An important clinical
extension would be investigating how therapeutic hypothermia—the
current standard of care for neonatal H-I [92]—modulates these early
cellular responses, as this would inform the development of adjuvant
therapies. Additionally, examining whether similar mechanisms exist in
milder H-I injuries could address an important clinical gap.
Integrating multi-omics approaches, such as metabolomics, proteomics,
post-translational modifications (PTMs), epitranscriptomics, and
epigenetics, would offer a more comprehensive understanding of the
underlying regulatory mechanisms.
Conclusion
Our comprehensive single-cell analysis elucidates the immediate
cellular and transcriptional adaptations in the neonatal hippocampus
following hypoxia–ischemia. The selective vulnerability of mature
neurons, contrasted with the resilience of immature granule cells,
GABAergic neurons, and Cajal-Retzius cells, underscores the complexity
of neuronal responses to injury. The rapid activation of glial and
endothelial cells further highlights the multifaceted nature of the
early post-injury environment. By identifying key transcriptional
regulators and signaling pathways, our study provides a foundation for
developing targeted interventions aimed at mitigating damage and
promoting repair in neonatal hypoxic-ischemic encephalopathy. The
developed publicly accessible data and analytical tools will serve as
valuable resources for the neuroscience community to explore
hippocampal cell heterogeneity and pathophysiology.
Supplementary Information
[242]Additional file 1.^ (4.5MB, pdf)
[243]Additional file 2.^ (404.8KB, xlsx)
[244]Additional file 3.^ (10.8KB, xlsx)
[245]Additional file 4.^ (385.6KB, xlsx)
Acknowledgements