Cdkn2a/p16Ink4a loss impairs Spatial memory independently of Alzheimer’s-associated genetic pathways in young adult mice

Abstract

Background

The cyclin-dependent kinase inhibitor CDKN2A/p16INK4A (p16INK4A) is a central regulator of cellular senescence, widely studied in ageing and cancer. Although its expression increases with age and disease, its role in the adult brain remains poorly defined. Recent evidence suggests a possible contribution to neuroinflammation and cognitive decline in Alzheimer’s disease (AD), but underlying mechanisms are unclear. We investigated the functional impact of p16Ink4a inactivation on spatial cognition and hippocampal gene expression in adult mice under normal and amyloidogenic conditions using the App^NL−G−F AD model.

Methods

Spatial learning and memory were assessed using the Morris Water Maze (MWM). Hippocampal transcriptomes from control, p16INK4A^knockout (p16Ink4a^KO), App^NL−G−F, and double mutant mice were analysed by RNA sequencing. We performed differential gene expression analysis, gene set enrichment analysis (GSEA), and interaction modelling to assess the molecular consequences of p16Ink4a loss alone and in combination with amyloid pathology.

Results

Loss of p16Ink4a resulted in modest and short-lasting alterations in spatial learning and early memory retrieval, without exacerbating early deficits in App^NL−G−F mice. Transcriptomic analysis indicated that p16Ink4a^KO upregulated metabolic, mitochondrial, and translational pathways, while downregulating synaptic and cytoskeletal genes. In contrast, App^NL−G−F mice displayed strong immune activation. The double mutants showed additive transcriptional changes, yet GSEA indicated non-linear interactions in synaptic and immune-related pathways. Unexpectedly, p16Ink4a deletion alone was associated with enrichment of senescence-related signatures.

Conclusion

Constitutive p16Ink4a loss shows modest effects on spatial memory and broad hippocampal transcriptomic changes, independently of amyloid pathology. These findings point to possible non-canonical roles for p16Ink4a in neuronal homeostasis and emphasise the need for temporally controlled models to better dissect the brain-specific functions of senescence regulators in ageing and neurodegeneration.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00018-026-06079-3.

Introduction

The cyclin-dependent kinase inhibitor CDKN2A, which encodes the tumour suppressor p16INK4A, is a central regulator of cell cycle arrest and a well-established marker and effector of cellular senescence (CS). Acting as a potent and conserved enforcer of G1 phase arrest, p16INK4A plays a pivotal role in maintaining the senescent state in response to diverse cellular stressors, including DNA damage, oxidative stress, oncogene activation, and chronic inflammation [1]. CS is a stable and irreversible cell cycle arrest, first described by Hayflick in the 1960s [2]. The cells cycle arrest is accompagned by distinct phenotypic and metabolic features, most notably the senescence-associated secretory phenotype (SASP). While transient senescence supports tissue remodelling, regeneration, and tumour suppression. In contrast, the chronic accumulation of senescent cells, promotes inflammation, disrupts tissue homeostasis, and contributes to ageing-related diseases [3–6]. Growing evidence in human post-mortem brains and mouse models implicates the CS and its factor p16INK4A as a contributor to β-amyloid (Aβ)-induced pathology in Alzheimer’s disease (AD) [7–11].

In APP/PS1 double transgenic mice (Mo/HuAPP695swe and PS1-dE9), Aβ plaques were closely associated with oligodendrocyte progenitor cells (OPC) and disease-associated microglia (DAM) displaying a senescence-like phenotype (e.g., upregulation of p21Cip/waf1 and p16Ink4a). Senolytic therapies in these models reduced Aβ burden, attenuated neuroinflammation, and improved cognitive performance [7, 8]. In a recent study, impaired autophagic flux was identified as a key contributor to intraneuronal Aβ42 accumulation in the senescence-accelerated 5xFAD; Terc⁻/⁻ mouse model. Interestingly, these double mutants displayed significant neuronal loss in brain regions where Aβ aggregates and senescence markers co-localized, thus linking senescence to Aβ toxicity through defective protein clearance mechanisms and CS [9]. However, senolytic treatment, using dasatinib and quercetin, shows context- and sex-specific effects. In APP^NL−F mice, treated females mice benefited from reduced senescence markers, decreased amyloid burden, and improved cognition, while treated males showed exacerbated senescence without cognitive improvement [10]. Conversely, in normally aging C57BL/6 mice, the same treatment caused detrimental effects in females (increased SASP, metabolic dysfunction, cognitive impairment) but minimal response in males [11]. These findings underscore the context- and sex-dependent effects of senolytic interventions in both physiological and pathological ageing. Deciphering these distinctions is especially important considering the first-in-human clinical trial (SToMP-AD, NCT04685590), which evaluated the senolytic combination of dasatinib and quercetin in patients with Alzheimer’s disease [12]. Thus, the understanding of mechanisms driven by senescence mediators is a major challenge.

While p16INK4A has been extensively studied in oncology, its physiological roles in the adult, non-pathological brain remain poorly characterised. Recent evidence reveals that p16INK4A expression is dynamic across development and ageing, playing distinct roles in neuronal differentiation versus age-related neuroinflammation. Indeed, p16INK4A is expressed in the embryonic and early postnatal brain without clear co-localization with CS markers, suggesting a classical role in cell cycle inhibition and neuronal differentiation. With age, however, p16INK4A expression increases and becomes accompanied by SASP activation, contributing to neuroinflammation [13]. Elevated p16Ink4a in the aged subventricular zone (SVZ) correlates with declined neural stem/progenitor cell (NSPC) proliferation. Genetic ablation of p16Ink4a restored neurogenic potential in aged mice, highlighting its role in restricting neurogenesis during ageing. Interestingly, this effect was region-specific, as p16Ink4a^knockout (p16Ink4a^KO) did not enhance baseline neurogenesis in the subgranular zone (SGZ) of the dentate gyrus [14]. However, under neurogenic stimulation such as physical exercise, p16Ink4a^KO mice showed increased NSPC activation in the SGZ, albeit without a corresponding rise in mature neurone production due to heightened apoptosis [15, 16]. Together, these findings underscore the complex role of p16Ink4a in maintaining neural stem cell quiescence and neuronal homeostasis. The present study aimed to investigate the brain function of p16Ink4a. For this an inactivation of p16Ink4a, was explored in both normal and pathological contexts. First, we analysed p16Ink4a^KO mice on a Wt background to uncover the physiological roles of p16 in adult brain homeostasis. In parallel, we generated a double mutant mouse model by crossing p16INK4A^KO mice with the App^NL−G−F AD model, which exhibits progressive β-amyloid deposition and early cognitive impairment [17]. Cognitive performance was assessed using the Morris Water Maze (MWM), and bulk RNA sequencing profiled transcriptomic changes in the hippocampus. Together, this dual approach provides novel insights into the neurobiological functions of p16Ink4a and its potential contributions to cognitive ageing and neurodegeneration.

Materials and methods

Mouse lines

We used two mouse lines for the study. The App^tm3.1Tcs allele [18], also named App^NL−G−F, was kept on the C57BL/6JCrl genetic background in our centre for more than 20 generations. The Cdkn2a^tm2.1Rdp allele [19], was originally maintained on a mixed genetic background composed of 50% C57BL/6J, 25% FVB/N, and 25% 129/SvPas. To increase C57BL/6J contribution, this line was backcrossed twice with C57BL/6J mice, followed by one cross with App^NL−G−F mice prior to experimental cohort generation. This breeding strategy resulted in an estimated genetic background of 93.75% C57BL/6J, 3.125% FVB/N, and 3.125% 129/SvPas. Mice were housed at the Institut Clinique de la Souris (ICS, Illkirch-Graffenstaden, France) in compliance with French Ministry of Agriculture guidelines (law 87–848). Animals were kept under controlled conditions (22 ± 2 °C, 40–60% humidity, 12 h/12 h light/dark cycle) with food and water available ad libitum. Mice were group-housed (up to 5 females or 4 males per cage). All cages were ventilated and equipped with nesting material. Small biopsies were used for genotyping. The four relevant genotypes for the study, wild-type (Wt), p16Ink4a^KO/KO (noted p16Ink4a^KO), App^{NL−G−F/NL−G−F} (noted App^NL−G−F), and App^{NL−G−F/NL−G−F};p16Ink4a^KO/KO (noted App^NL−G−F;p16Ink4a^KO),were produced from four different crosses: [1] App^{NL−G−F/Wt}; p16Ink4a^KO/KO x App^{NL−G−F/Wt} p16Ink4a^KO/KO [2], App^{NL−G−F/Wt};p16INK4^Wt/Wt x App^{NL−G−F/Wt};p16INK4^Wt/Wt, [3] App^{NL−G−F/NL−G−F};p16Ink4a^KO/Wt x App^{NL−G−F/NL−G−F};p16Ink4a^KO/Wt [4], App^Wt/Wt;p16Ink4a^KO/Wt × App^Wt/Wt;p16Ink4a^KO/Wt. Each cross produced at least two genotypes of interest, each representing 25% of the offspring and being littermates. This strategy was chosen instead of a double-heterozygous cross to increase the yield of animals with the desired genotypes, as the latter would have resulted in only 6.25% of offspring carrying each of the four correct mutation combinations. Then individuals were tested at 4 months of age (± 2 weeks) and three experimental batches. To minimize potential batch effects, each batch included at least three animals per genotype. The detailed batch composition is provided in Table 1. Behavioural testing was conducted at the ICS, following approval from the French Ministry of Higher Education and Research (DAP#43632), and following the Animal Research: Reporting of In Vivo Experiments (ARRIVE) guidelines [20].

Table 1.

Batch composition

Batch	Sex	Wt	p16Ink4a KO	App NL−G−F	AppNL−G−F;p16Ink4aKO
1	F	2	2	2	2
	M	2	1	4	2
2	F	2	4	2	2
	M	3	2	3	3
3	F	4	2	2	2
	M	3	3	1	2
	Total	16	14	14	13

Morris water maze (MWM)

Animals were habituated to the testing room for 30 min prior to each session. To avoid potential olfactory interference between sexes, males were tested in the morning and females in the afternoon. All experiments were carried out between 9 a.m. and 7 p.m. Behaviour was recorded using Noldus EthoVision XT software (v14; NOLDUS, Wageningen, The Netherlands) [21]. The test took place in a 150 cm circular pool, virtually divided into four quadrants (NE, SE, NW, SW). The water was rendered opaque using non-toxic white paint. For the acquisition phase, the hidden platform was submerged 1.5 cm below the surface in the NE quadrant. Distal visual cues were placed on the surrounding walls. Mice were trained over six consecutive days, with four 90-second trials per day. In each trial, mice were released from different starting points (N, S, E, W), with the order randomised daily. After each trial, a rest period of 15 to 30 min was provided. If a mouse succeeded in locating the platform, it was retrieved after remaining there for 15 s, giving it time to encode the platform’s location. If the mouse failed to find the platform within 90 s, it was gently guided to the platform and retrieved again after 15 s. Total path length and latency to reach the platform were recorded and analysed to assess spatial learning performance. Swimming velocity was also measured to control for potential locomotor impairments. In addition, initial trajectory error (ITE) was assessed using the Rtrack package (version 2.0.3) to evaluate the accuracy of the mice’s initial navigation strategy [22]. ITE is defined as the distance between the endpoint of the animal’s initial trajectory and the centre of the target (i.e., the platform). The length of the initial trajectory corresponds to the theoretical optimal and direct path from the starting point to the platform [22]. For the probe test on Days 7 and 14, the platform was removed, and a 60-second probe trial was conducted to assess spatial memory retrieval. Mice were released from the SW quadrant on Day 7 and from the S point on Day 14. Latency to enter the target quadrant, as well as time spent in this quadrant, were measured to assess spatial memory performance. In addition, the number of crossings of the platform area, defined as the frequency with which the animal crosses the exact former platform location, was analysed using Noldus EthoVision XT software (v14; NOLDUS, Wageningen, The Netherlands) [21]. A visual cue test was conducted on Day 15, consisting of four 60-second trials. In this test, the platform was made visible using a flag, and all distal visual cues were removed from the surrounding environment to eliminate spatial references. For each trial, the platform was placed in a different quadrant, and the starting position was selected to maximise the distance to the platform. Escape latency was measured as an indicator of visual and motivational abilities, independently of spatial memory. All statistical analyses were performed using R (v2024.04.2 + 764). To analyze mean swim distance, escape latency, and ITE, a linear mixed-effects model (LMM) was fitted using the lmer() function from the R package lme4 (v1.1.35.5) [23]. LMM were chosen because they can accommodate unbalanced designs and missing data, unlike repeated-measures analysis of variance (RM ANOVA) [24]. This approach allows retaining individual animals in the analysis even if outliers or missing values are present on a given day, thereby improving statistical power and robustness. Before statistical analyses, the normality of the data was assessed separately for each genotype and testing day using the Shapiro–Wilk test. A p-value (p) < 0.05 indicating a significant deviation from normality. For normally distributed groups, outliers were identified using Grubbs’ test (p < 0.05). For non-normally distributed groups, extreme values were detected using the interquartile range (IQR) method, considering as outliers any observations below Q1 − 1.5 × IQR or above Q3 + 1.5 × IQR. The LMM models included genotype, day, sex and their interaction as fixed effects. Mouse ID and batches were included as random effects to account for repeated measures and variability related to batches and sex. The random effect structure was selected based on the lowest restricted maximum likelihood (REML) criterion and the absence of singular fit, ensuring an optimal balance between model fit and complexity [23]. Model assumptions were assessed by inspecting residual plots and quantile-quantile (Q–Q) plots to evaluate normality and homoscedasticity of residuals. To assess the statistical significance of the fixed effects, a Type III ANOVA with Satterthwaite’s approximation for degrees of freedom was conducted using the anova() function from the lmerTest package (v3.1.3). Post hoc comparisons were carried out on estimated marginal means (EMM) using the emmeans() function from the emmeans package (v1.10.5), with Tukey correction applied to adjust for multiple testing [25]. For the probe tests (Days 7 and 14), spatial memory was assessed by analysing latency to the target quadrant, time spent in the target quadrant, and the number of platform zone crossings. Depending on the distribution of each variable, statistical comparisons between genotypes were performed using either two-way ANOVA (TW ANOVA) followed by Tukey post hoc tests or KW tests with Dunn’s post hoc tests corrected using the Holm method. To evaluate whether each genotype displayed a spatial preference, one-sample t-tests were used to compare the time spent in the target quadrant against chance level (15 s, i.e., 25% of the trial). To analyze the number of platform zone crossings during probe trials, a generalized linear model (GLM) with a Poisson distribution was fitted using the glm() function in R. Outliers were identified and removed within each genotype group using the IQR method. The dispersion of the Poisson model was evaluated by dividing the sum of squared Pearson residuals by the residual degrees of freedom to assess goodness-of-fit. A cutoff value of 1.5 was used to determine whether overdispersion was present and if correction was required. Genotype and sex were included as a fixed effect. For the visible platform test (Day 15), latency to reach the flagged platform was compared across genotypes using OW ANOVA test. A summary of the statistical analysis is presented in Table 2 and Detailed statistical outputs are available in Supplementary data 1.

Table 2.

Summary of the statistical analysis of the Morris water maze variables

Variable	Statistical test
Mean Path Lenght	Type III ANOVA results for fixed effects in the LMM	Sum Sq	Mean Sq	NumDF	DenDF	F value	Pr(> F)
	Genotype	788,310.10	262,770.03	3.00	47.42	3.54	0.021
	Day	23,113,197.87	4,622,639.57	5.00	233.67	62.26	0.000
	Sexe	53,526.68	53,526.68	1.00	47.41	0.72	0.400
	GenotypexDay	1,227,813.89	81,854.26	15.00	233.56	1.10	0.355
	GenotypexSexe	112,005.89	37,335.30	3.00	47.42	0.50	0.682
	DayxSexe	1,159,540.11	231,908.02	5.00	233.67	3.12	0.009
	GenotypexDayxSexe	1,717,095.29	114,473.02	15.00	233.56	1.54	0.092
Mean Escape Latency	Type III ANOVA results for fixed effects in the LMM	Sum Sq	Mean Sq	NumDF	DenDF	F value	Pr(> F)
	Genotype	743.65	247.88	3.00	47.67	1.22	0.313
	Day	66,118.53	13,223.71	5.00	232.35	65.00	0.000
	Sexe	257.95	257.95	1.00	47.67	1.27	0.266
	GenotypexDay	3,904.96	260.33	15.00	232.26	1.28	0.216
	GenotypexSexe	104.75	34.92	3.00	47.67	0.17	0.915
	DayxSexe	3,916.08	783.22	5.00	232.35	3.85	0.002
	GenotypexDayxSexe	5,757.61	383.84	15.00	232.26	1.89	0.025
Mean ITE	Type III ANOVA results for fixed effects in the LMM	Sum Sq	Mean Sq	NumDF	DenDF	F value	Pr(> F)
	Genotype	1,552.60	517.53	3.00	46.82	5.68	0.002
	Day	10,700.80	2,140.16	5.00	237.45	23.48	0.000
	Sexe	8.92	8.92	1.00	46.12	0.10	0.756
	GenotypexDay	2,503.99	166.93	15.00	237.43	1.83	0.031
	GenotypexSexe	171.82	57.27	3.00	46.60	0.63	0.600
	DayxSexe	729.71	145.94	5.00	237.45	1.60	0.160
	GenotypexDayxSexe	1,011.89	67.46	15.00	237.43	0.74	0.742
Mean Velocity	Type III ANOVA results for fixed effects in the LMM	Sum Sq	Mean Sq	NumDF	DenDF	F value	Pr(> F)
	Day	153.06	30.61	5.00	237.85	7.42	0.000
	Genotype	11.98	3.99	3.00	47.58	0.97	0.416
	Sexe	37.51	37.51	1.00	47.57	9.09	0.004
	DayxGenotype	29.08	1.94	15.00	237.84	0.47	0.954
	DayxSexe	31.36	6.27	5.00	237.85	1.52	0.184
	GenotypexSexe	25.81	8.60	3.00	47.58	2.08	0.115
	DayxGenotypexSexe	55.93	3.73	15.00	237.84	0.90	0.561
D7_TQLatency	TW ANOVA	Sum Sq	Mean Sq	NumDF		F value	Pr(> F)
	Genotype	209.59	69.86	3.00		3.30	0.029
	Sexe	44.38	44.38	1.00		2.09	0.155
	GenotypexSexe	27.27	9.09	3.00		0.43	0.733
	Residuals	975.22	21.20	46.00
D7_TQDuration	TW ANOVA	Sum Sq	Mean Sq	NumDF		F value	Pr(> F)
	Genotype	458.66	152.89	3.00		2.23	0.098
	Sexe	140.58	140.58	1.00		2.05	0.159
	GenotypexSexe	39.37	13.12	3.00		0.19	0.902
	Residuals	3,157.56	68.64	46.00
D7_PTFFreq	Poisson results	Effect	Estimate	Std, Error	z value		Pr(>|z|)
Genotype	(Intercept)		0.94	0.16	6.03		0.000
Genotype	Genotypep16p16Ink4aKO		- 0.56	0.28	- 2.02		0.043
Genotype	GenotypeAppAppNLGF		- 0.56	0.28	- 2.02		0.043
Genotype	GenotypeAppAppNLGFp16p16Ink4aKO		- 0.33	0.26	- 1.27		0.205
Genotype x Sexe	(Intercept)		0.83	0.25	3.31		0.001
Genotype x Sexe	Genotypep16Ink4aKO		- 0.54	0.43	- 1.25		0.213
Genotype x Sexe	GenotypeAppNLGF		- 0.69	0.43	- 1.60		0.109
Genotype x Sexe	GenotypeAppNLGF; p16Ink4aKO		- 0.21	0.37	- 0.56		0.578
Genotype x Sexe	SexeF		- 0.02	0.34	- 0.05		0.963
Genotype x Sexe	Genotypep16Ink4aKO: SexeF		0.48	0.55	0.88		0.381
Genotype x Sexe	GenotypeAppNLGF: SexeF		0.49	0.58	0.85		0.398
Genotype x Sexe	GenotypeAppNLGF; p16Ink4aKO: SexeF		- 0.02	0.55	- 0.03		0.978
D14_TQLatency	Kruskal Wallis Test	Statistic		df			p_value
	Sexe pooled	0.10		3.00			0.992
	Female	1.82		3.00			0.611
	Male	1.12		3.00			0.772
D14_TQDuration	TW ANOVA	Sum Sq	Mean Sq	Df		F value	Pr(> F)
	Genotype	277.01	92.34	3.00		1.65	0.191
	Sexe	0.24	0.24	1.00		0.00	0.948
	GenotypexSexe	183.81	61.27	3.00		1.10	0.360
	Residuals	2,571.76	55.91	46.00
D14_PTFFreq	Poisson results	Effect	Estimate	Std, Error	z value		Pr(>|z|)
Genotype	(Intercept)		0.55	0.20	2.81		0.005
Genotype	Genotypep16Ink4aKO		0.11	0.27	0.39		0.698
Genotype	GenotypeAppNLGF		- 0.34	0.32	- 1.08		0.281
Genotype	GenotypeAppNLGF; p16Ink4aKO		- 0.24	0.32	- 0.74		0.459
Genotype x Sexe	(Intercept)		1.01	0.21	4.75		0.000
Genotype x Sexe	Genotypep16Ink4aKO		- 0.50	0.38	- 1.31		0.189
Genotype x Sexe	GenotypeAppNLGF		- 0.88	0.41	- 2.13		0.033
Genotype x Sexe	GenotypeAppNLGF; p16Ink4aKO		- 0.76	0.40	- 1.92		0.055
Genotype x Sexe	SexeF		- 0.89	0.40	- 2.26		0.024
Genotype x Sexe	Genotypep16Ink4aKO: SexeF		1.14	0.56	2.02		0.043
Genotype x Sexe	GenotypeAppNLGF: SexeF		0.76	0.69	1.10		0.273
Genotype x Sexe	GenotypeAppNLGF; p16Ink4aKO: SexeF		1.56	0.58	2.70		0.007
D15_VisualCue	TW ANOVA	Sum Sq	Mean Sq	Df		F value	Pr(> F)
	Genotype	28.80	9.61	3.00		0.56	0.643
	Sexe	3.80	3.82	1.00		0.22	0.639
	Genotype: Sexe	33.60	11.20	3.00		0.66	0.584
	Residuals	804.10	17.11	47.00

Transcriptomic data processing

7 days after behavioral testing, mice were sacrificed via cervical dislocation at ~ 5 months of age. For transcriptomic analysis, both hippocampi were collected from a subset of randomly selected animals and stored at − 80 °C. Tissue homogenisation was performed using the Precellys 24 Touch (Bertin Instruments, Montigny-le-Bretonneux, France) at 4500 rpm for 20 s. RNA extraction was carried out using the RNeasy Lipid Tissue Mini Kit (74804, Qiagen), following the manufacturer’s instructions. RNA concentration was measured using the Qubit RNA Assay Kit (Thermo Fisher Scientific), and RNA integrity was assessed using the Agilent Bioanalyzer 2100 (Agilent Technologies, Les Ulis, France).

Library preparation was performed using the Illumina Stranded mRNA Prep Ligation kit (Reference Guide – PN 1000000124518). RNA-Seq libraries were generated from 350 ng of total RNA following the manufacturer’s instructions (Illumina Stranded mRNA Prep, Ligation kit and IDT for Illumina RNA UD Indexes, Illumina, San Diego, USA). After polyA selection, mRNA was fragmented at 94 °C for 8 min. DNA libraries were then amplified using 12 PCR cycles, and excess primers were removed by two rounds of purification using SPRIselect beads (Beckman-Coulter, Villepinte, France). Library quality and concentration were assessed using the Bioanalyzer 2100 system (Agilent Technologies, Les Ulis, France). Sequencing was performed on an Illumina NextSeq 2000 platform using single-end 50 base reads. Image analysis and base calling were conducted using RTA (v4.12.2) and BCL Convert (v4.2.7). Raw sequencing reads were preprocessed to remove adapter sequences, polyA tails, and low-quality reads (defined as those with a Phred score < 20). Reads shorter than 40 bases after trimming were excluded from downstream analysis. Preprocessing was performed using Cutadapt (v4.2) [26]. Reads were first aligned to ribosomal RNA (rRNA) reference sequences from RefSeq [27] and GenBank [28] using bowtie (v2.3.5) [29], and rRNA-matching reads were discarded. Remaining reads were then aligned to the Mus musculus GRCm38 genome assembly using STAR (v2.7.10b) [30]. Gene body coverage was assessed across all samples using the geneBodyCoverage function from RSeQC (v5.0.1) [31], and transcript integrity number (TIN) values were calculated for transcripts with at least 10 uniquely aligned reads. Strand specificity was also evaluated using RSeQC (v5.0.1) [31]. Gene expression levels were quantified using uniquely aligned reads with STAR (v2.7.10b) [30], based on Ensembl annotation version 102. Only reads unambiguously assigned to a gene were retained for downstream analysis. To reduce noise from lowly expressed genes, those with fewer than 15 reads in more than 10% of the samples were excluded, yielding a final set of 18,768 genes. Outlier samples were identified using principal component analysis (PCA) on variance-stabilised expression data (VST)-normalised using DESeq2 (v1.45.3) [32]. The number of PCs retained was determined using the Elbow Method, automated with the FindPC package (v1.0) [33], which identifies the point of maximal deviation from a straight line between the first and last PC. Mahalanobis distances were then calculated based on the selected PCs to detect statistical outliers. Squared Mahalanobis distances were compared to the 95th percentile of the Chi-squared distribution; samples exceeding this threshold were classified as outliers.

Transcriptomic data analysis

Differentially expressed genes (DEG) were identified using the DESeq2 (v1.45.3) [32]. The comparisons are the following: p16Ink4a^KO vs. Wt, App^NL−G−F vs. Wt and p16Ink4a^KO; App^NL−G−F vs. Wt. Analyses were performed using the full model, including all genotype groups simultaneously. To stabilize log2 fold change (LFC) estimates and improve interpretability, apeglm shrinkage (Adaptive Shrinkage using Empirical Bayes) was applied [34]. Genes were considered differentially expressed if they met the following criteria: adjusted p-value (p.adj) < 0.05 and |LFC| > 0.1. Genes with LFC > 0.1 were classified as upregulated, and those with LFC < −0.1 as downregulated. p.adj was calculated using the Benjamini-Hochberg method to control for false discovery rate (FDR). We selected this lenient |LFC| threshold in combination with apeglm shrinkage to avoid excluding biologically meaningful but modest transcriptional changes, particularly for genes with low counts. Statistical assessment of DEG overlap between comparisons was performed using Fisher’s exact test on 2 × 2 contingency tables (shared vs. non-shared genes). p-values (p) were reported to evaluate the significance and direction of the associations between differential signatures. Protein–Protein Interaction (PPI) networks were constructed using the STRING database (v12.0) [35]. The analysis included both physical and functional associations, only interactions supported by experimental data and curated databases were retained. A medium confidence threshold (score ≥ 0.4) was applied, and confidence scores scaled edge thickness to reflect interaction strength. First- and second-shell interactors were excluded to focus strictly on direct interactions among DEG. To identify gene modules within the STRING network, the Markov Cluster Algorithm (MCL) was applied using STRING’s integrated clustering tool [36]. The inflation parameter was set to 1.5. Only clusters containing more than two genes were retained for downstream functional annotation. The relative proportions of major hippocampal cell types were inferred from normalised gene expression data using the R package BRETIGEA (Brain Cell Type-Specific Gene Expression Analysis (v1.0.4) [37]. This algorithm is based on a singular value decomposition (SVD) approach applied to sets of cell type–specific marker genes. The analysis was performed using the 50 most robust markers for each cell type, specifying the mouse species and applying the SVD decomposition method. The inferred cell types included astrocytes, endothelial cells, microglia, neurones, mature oligodendrocytes, and OPC. Cell-type proportion distributions were assessed by genotype. Data normality was tested using the Shapiro-Wilk test, and variance homogeneity using Levene’s test. Global differences across conditions were evaluated with the non-parametric KW test, followed, when significant, by post hoc Dunn’s tests with Holm correction for multiple comparisons. To explore interactions between the App^NL−G−F and p16Ink4a^KO genotypes, a likelihood ratio test (LRT) was performed using the DESeq2 (v1.45.3) [32]. This approach compares a full model including the interaction term (~ App^NL−G−F + p16Ink4a^KO + App^NL−G−F: p16Ink4a^KO) with a reduced model containing only the main effects (~ App^NL−G−F + p16Ink4a^KO). Genes were considered significantly associated with an interaction effect if they showed an adjusted p-value (p.adj) < 0.05 and an |LFC| > 0.1. In addition to the interaction term, the main effects of App^NL−G−F and p16Ink4a^KO were also extracted from the full model. Gene set enrichment analysis (GSEA) was conducted using gene sets from MSigDB Mouse Collections (v2024.1.Mm), including gene ontology (GO) biological process (GOBP), GO cellular component (GOCC), GO molecular function (GOMF) [38], Hallmark [39], and Reactome pathways [40]. To explore senescence-related programs, the SenSaul SASP signature [41] and the Reactome CS pathway (R-MMU-2559583), including sub-branches (e.g., p53-, p16INK4A-, oncogene-, and SASP-associated senescence), were analysed. Additionally, to assess microglial state-specific signatures, GSEA was performed using a curated set of 2,054 genes defining homeostatic, activated, and interferon (IFN)-responsive microglial populations [42]. Enrichment was performed with the GSEA() function from clusterProfiler (v4.14.6) using fgsea (v1.32.4) [32, 33], with set.seed(123456789) for reproducibility. Only gene sets containing 10–500 genes were tested. Significance was determined at the FDR < 0.05 (Benjamini-Hochberg correction). An unweighted scoring scheme was applied (exponent = 0, eps = 0). To minimize redundancy and improve interpretability, enriched terms were filtered using the Jaccard similarity index (threshold ≥ 0.5), retaining terms with higher enrichment scores. GO terms were further categorized into higher-level functional classes to facilitate biological interpretation and cross-comparison across conditions. Table 3 summarizing the enrichment analyses performed across different models and data sources. Supplementary Data 5 provides the full results of these enrichment analyses.

Table 3.

Overview of gene set enrichment analyses

Comparison	Data source	Gene Set
p16Ink4aKO vs. Wt	DEG	Hallmark, Reactome, GO, Senescence, Microglia
AppNL−G−F vs. Wt	DEG	Senescence, Microglia
AppNL−G−F;p16Ink4aKO vs. Wt	DEG	Senescence, Microglia
Main effect of p16Ink4aKO	Interaction Model	GO
Main effect of AppNL−G−F	Interaction Model	GO
Interaction	Interaction Model	GO

Results

Cdkn2a/p16Ink4a inactivation induced mild spatial learning impairments without affecting the App^NL−G−F phenotype

To assess the impact of p16Ink4a^KO on spatial learning and memory deficits, double homozygote App^NL−G−F;p16Ink4a^KO mutants, with single mutant and control animals, were tested using the same MWM paradigm. The experimental protocol consisted of three distinct phases: an acquisition phase, a probe test, and a visual control test (Fig. 1a). During the acquisition phase, mice were trained to locate a submerged hidden platform over six consecutive days. Spatial learning was assessed by tracking the progressive reduction in escape latency and path length across days. A LMM analysis followed by Type III ANOVA, used to evaluate the significance of each factor, genotype, day, sexe, revealed significant main effects of genotype and dat for the path length (Table 2; Type III test, F(3, 47.42) = 3.54, p = 0.021) but only day for the escape latency (Table 2; Type III test, F(3, 47.67) = 1.22, p = 0.31). Post hoc analyses showed that p16Ink4a^KO mice exhibited significantly longer travelled distances compared to Wt controls (EMM, Tukey-adjusted, estimate = −161.58, SE = 54.5, p = 0.023). No other genotype comparisons reached statistical significance (Fig. 1b, c; and Supplementary data 1a, b). A strong effect of training day was observed for both measures (path length: Type III test, F(5, 233.67) = 62.26, p = 4.52 × 10^{− 41}; latency: Type III test, F(5, 232.35) = 65, p = 3.09 × 10^{− 42}), reflecting progressive learning over time. No significant genotype × day interaction was found for either variable, suggesting that all groups followed a similar learning trajectory (path length: Type III test, F(15, 233.56) = 1.10, p = 0.35; latency: Type III test, F(15, 232.26) = 1.28, p = 0.22) (Table 2; Supplementary data 1a, b). These results indicate that while all mice learnt the task, the deletion of p16Ink4a alone may slightly impair the efficiency of spatial navigation without any detectable sex-related effect. This impairment was not observed in the double mutant App^NL−G−F;p16Ink4a^KO mice. Analysis of swim velocity showed significant variation across training days (Type III test, F(5, 237.85) = 7.42, p = 1.75 × 10^{− 6}). Still, no genotype effect was detected (Type III test, F(3, 47.58) = 0.97, p = 0.42) or genotype × day interaction (Type III test, F(15, 237.84) = 0.47, p = 0.95), indicating that motor performance was unaffected by genotype (Supplementary Fig. 1a, Table 2 and Supplementary data 1 d). Additionally, performance in the visible platform test on Day 15 did not differ between genotypes, sexes, or their interaction (TW ANOVA: genotype, F [3, 41] = 0.56, p = 0.64; sex, F [1, 41] = 0.22, p = 0.64; genotype × sex, F [3, 41] = 0.66, p = 0.58), confirming that all groups were equally able to locate the flagged platform using proximal visual cues (Supplementary Fig. 1b, Table 2 and Supplementary Data 1 h). These results indicate that neither motor nor visual deficits contributed to the behavioral differences observed in the hidden platform task, supporting the interpretation that the increased path length in p16Ink4a^KO mice reflects subtle cognitive alterations rather than sensorimotor impairments. Interestingly, a significant sex effect was observed for the swim velocity (Table 2; Type III test, F(1, 47.57) = 9.10, p = 0.0041), with males swimming slightly faster than females (EMM, Tukey-adjusted, estimate = −2.05, SE = 0.68, p = 0.004). In line with this, significant day × sex interactions were found for both path length (Type III test, F(5, 233.67) = 3.12, p = 0.0095) and escape latency (Type III test, F(5, 232.35) = 3.85, p = 0.0023), as well as a genotype × day × sex interaction for escape latency (Type III test, F(15, 232.26) = 1.89, p = 0.025). However, these effects were inconsistent across training days and genotypes (Table 2; Supplementary Data 1 h).

Fig. 1.

Spatial learning and memory in the Morris Water Maze (MWM). (a) Schematic of the MWM protocol. Mice were trained over 6 days to locate a hidden platform using distal cues. Probe tests were conducted on Days 7 and 14, followed by a visible platform test. (b–d) Spatial learning during training. Data were analysed using linear mixed-effects models (LMM), with day, genotype, sex, and their interactions included as fixed effects. Batch and mouse ID were modelled as random effects. LMM revealed significant effects of day on (b) path length (Type III test, F(5, 233.67) = 62.26, p = 4.52 × 10^{− 41}), (c) escape latency (Type III test, F(5, 232.35) = 65, p = 3.09 × 10^{− 42}), and (d) initial trajectory error (ITE) (Type III test, F(5, 237.45) = 23.48, p = 2.2 × 10^{− 16}). (b) A significant main effect of genotype was also observed for path length (Type III test, F(3, 47.42) = 3.54, p = 0.021), with p16Ink4a^KO mice exhibiting significantly longer path lengths than Wt controls (estimate marginal means (EMM), Tukey-adjusted, estimate = −161.58, SE = 54.5, p = 0.023). (d) For ITE, there was a significant main effect of genotype (Type III test, F(3, 46.82) = 5.67, p = 0.0021), with p16Ink4a^KO (EMM, Tukey-adjusted, estimate = −6.39, SE = 1.77, p = 0.004) and App^NL−G−F;p16Ink4a^KO (EMM Tukey-adjusted, estimate = −5.98, SE = 1.81, p = 0.0095) mice showing significantly higher ITE values compared to Wt indicating impaired accuracy of initial goal-directed navigation. A significant genotype × day interaction was also observed (Type III test, F(15, 237.43) = 1.83, p = 0.031). Wt and App^NL−G−F mice showed significant ITE improvement by Day 3 (Wt: EMM, Tukey-adjusted, estimate = − 14.26, SE = 3.38, p = 0.0005; App^NL−G−F: EMM, Tukey-adjusted, estimate = − 11.42, SE = 3.38, p = 0.031) whereas p16Ink4a^KO and App^NL−G−F;p16Ink4a^KO exhibited delayed improvement (first significant at Day 6 for p16Ink4a^KO: estimate = − 13.87, SE = 3.65, p = 0.0025; and at Day 5 for App^NL−G−F; p16Ink4a^KO: estimate = − 16.26, SE = 3.86, p = 0.0005). No sex effects were detected (sex : Type III test, F(1,46.12) = 0.10, p = 0.76; genotype × sex: Type III test, F(3,46.60) = 0.63, p = 0.60; day × sex: Type III test, F(5,237.45) = 1.60, p = 0.16; genotype × day × sex: Type III test, F(15,237.43) = 0.74, p = 0.74). (Wt, n = 16 (8 M, 8 F); p16Ink4a^KO, n = 14 (6 M, 8 F); App^NL−G−F, n = 13 (7 M, 6 F); App^NL−G−F;p16Ink4a^KO, n = 12 (7 M, 5 F)). (e–g) Day 7: Spatial memory retrieval. (e) Genotype affected latency to reach the target (Two-Way (TW) ANOVA, F [3, 43] = 3.30, p = 0.029) with p16Ink4a^KO mice were significantly slower than Wt controls (Tukey’s tests, p.adj = 0,021). No significant effects of sex (TW ANOVA, F [1, 43] = 2.09, p = 0.155) or genotype × sex interaction (TW ANOVA, F [3, 43] = 0.43, p = 0.733) were found (Wt, n = 15 (7 M, 8 F); p16Ink4a^KO, n = 14 (6 M, 8 F); App^NL−G−F, n = 12 (7 M, 5 F); App^NL−G−F; p16Ink4a^KO, n = 13 (7 M, 6 F)). (f) No significant differences were detected between genotypes, sex, or their interaction in the time spent in the target quadrant (TW ANOVA: genotype, F [3, 43] = 2.23, p = 0.098; sex, F [1, 43] = 2.05, p = 0.159; genotype × sex, F [3, 43] = 0.19, p = 0.902). When data were pooled across sex, Wt (one-sample (OS) t-tests, t [14] = 3.53, p = 0.003) and p16Ink4a^KO mice (OS t-test, t [13] = 2.33, p = 0.037) spent significantly more than 25% of the trial duration in the target quadrant. When analysed separately by sex, both male (OS t-test, t [8] = 2.39, p = 0.048) and female Wt mice (OS t-test, t [7] = 4.41, p = 0.0045) showed a significant preference, whereas in the p16Ink4a^KO group only males reached significance (OS t-test, t [6] = 3.01, p = 0.030). (Wt, n = 15 (8 M, 7 F); p16Ink4a^KO, n = 14 (6 M, 8 F); App^NL−G−F, n = 13 (7 M, 6 F); App^NL−G−F; p16Ink4a^KO, n = 12 (7 M, 5 F)). (g) Additionally, a Poisson model analysis of platform crossings revealed a significant reduction in the number of target zone entries for p16Ink4a^KO (Poisson, β = −0.56, p = 0.043) and App^NL−G−F (Poisson, β = −0.36, p = 0.043) mice compared to Wt (Wt, n = 16; p16Ink4a^KO, n = 13; App^NL−G−F, n = 13; App^NL−G−F; p16Ink4a^KO, n = 12). (h–j) Day 14: Long-term memory recall. (h) No significant genotype effects were observed for latency (KW, χ² [3] = 0.11, p = 1). (Wt, n = 15; p16Ink4a^KO, n = 14; App^NL−G−F, n = 12; App^NL−G−F; p16Ink4a^KO, n = 13). (i) No significant effect of genotype, sex, or their interaction was detected for the time spent in the target quadrant on Day 14 (TW ANOVA: genotype, F [3, 43] = 1.65, p = 0.19; sex, F [1, 43] = 0.004, p = 0.95; genotype × sex, F [3, 43] = 1.10, p = 0.36). Only Wt mice displayed a significant quadrant preference (OS t-test, t [15] = 2.25, p = 0.04) (Wt, n = 15 (8 M, 7 F); p16Ink4a^KO, n = 14 (6 M, 8 F); App^NL−G−F, n = 13 (7 M, 6 F); App^NL−G−F;p16Ink4a^KO, n = 12 (7 M, 5 F)). (j) No differences were detected in platform crossings between genotypes (Poisson, all p > 0.05) (Wt, n = 15; p16Ink4a^KO, n = 14; App^NL−G−F, n = 13; App^NL−G−F;p16Ink4a^KO, n = 11)

Increased initial trajectory errors highlight genotype-dependent deficits in Spatial cue integration

To further assess spatial learning strategies beyond latency and path length, we analysed the ITE, defined as the distance between the end point of the initial swim segment and the centre of the goal platform. This measure captures the precision of the animal’s initial heading and provides insight into early goal-directed navigation. The LMM revealed significant main effects of genotype (Table 2; Type III test, F(3, 46.82) = 5.67, p = 0.0021) and day (Type III test, F(5, 237.45) = 23.48, p = 2.2 × 10^− 16). Post hoc analyses indicated that p16Ink4a^KO and double mutant App^NL−G−F;p16Ink4a^KO mice exhibited significantly higher ITE compared to Wt mice (p16Ink4a^KO: EMM, Tukey-adjusted, estimate = −6.39, SE = 1.77, p = 0.004; App^NL−G−F;p16Ink4a^KO: EMM, Tukey-adjusted, estimate = −5.98, SE = 1.81, p = 0.0095), suggesting less accurate goal-directed navigation (Fig. 1d, Supplementary data 1c). A significant genotype × day interaction was also observed (Table 2; Type III test, F(15, 237.43) = 1.83, p = 0.031). Notably, Wt (EMM, Tukey-adjusted, estimate = 14.26, SE = 3.38, p = 0.0005) and App^NL−G−F (EMM, Tukey-adjusted, estimate = 11.42, SE = 3.04, p = 0.031) mice both showed a significant improvement in ITE by Day 3, indicating early engagement of spatial cues (Fig. 1d, Supplementary data 1c). Wt mice maintained this improvement, as evidenced by a consistent reduction in ITE from Day 1 to later days (Day 1–Day 4: EMM, Tukey-adjusted, estimate = 19.0, SE = 3.38, p < 0.0001; Day 1–Day 5: EMM, Tukey-adjusted, estimate = 21.67, SE = 3.38, p < 0.0001; Day 1–Day 6: EMM, Tukey-adjusted, estimate = 27.34, SE = 3.44, p < 0.0001). However, App^NL−G−F mice did not show significant improvement beyond Day 3, with non-significant changes in ITE between Day 1 and Day 4 (EMM, Tukey-adjusted, estimate = 9.24, SE = 3.76, p = 0.14) and between Day 1 and Day 5 (EMM, Tukey-adjusted, estimate = 8.40, SE = 3.76, p = 0.23) (Supplementary data 1c). This plateau might suggest a limited ability to consolidate or optimise spatial strategies over time. In contrast, p16Ink4a^KO and App^NL−G−F;p16Ink4a^KO mice showed a delayed improvement, with the first significant reduction in ITE observed at Day 6 (EMM, Tukey-adjusted, estimate = 13,87, SE = 3.65, p = 0.0025) for p16Ink4a^KO and at Day 5 (EMM, Tukey-adjusted, estimate = 16.26, SE = 3.86, p = 0.0005) (Fig. 1d, Supplementary data 1c). ITE was not influenced by sex or by any sex-related interaction (sex : Type III test, F(1,46.12) = 0.10, p = 0.76; genotype × sex: Type III test, F(3,46.60) = 0.63, p = 0.60; day × sex: Type III test, F(5,237.45) = 1.60, p = 0.16; genotype × day × sex: Type III test, F(15,237.43) = 0.74, p = 0.74) (Supplementary data 1c). These findings indicated that all mutant genotypes exhibit impairments in early goal-directed navigation. Given that no significant differences were detected between the mutant groups, specifically between App^NL−G−F and App^NL−G−F;p16Ink4a^KO, combining p16Ink4a^KO in the App^NL−G−F did not result in a synergistic effect on the precision of spatial orientation.

Spatial memory retrieval reveals genotype-dependent impairments

To further assess memory retrieval and determine whether impairments could be detected at the spatial memory recall level, probe tests were conducted 24 h (Day 7) after the final training session for short-term and 7 days later (Day 14) for long-term memories. On Day 7, TW ANOVA revealed a significant main effect of genotype on latency to reach the former platform location (Table 2; TW ANOVA, F [3, 43] = 3.30, p = 0.029), but no significant effects of sex (TW ANOVA, F [1, 43] = 2.09, p = 0.155) or genotype × sex interaction (TW ANOVA, F [3, 43] = 0.43, p = 0.733). Post hoc Tukey’s tests revealed that p16Ink4a^KO mice exhibited significantly longer latencies compared to Wt controls (Tukey’s tests, p.adj = 0.021), while other pairwise comparisons were not significant. No significant differences were found in the time spent in the target quadrant for genotype (Table 2; TW ANOVA, F [3, 43] = 2.23, p = 0.098), sex (TW ANOVA, F [1, 43] = 2.05, p = 0.159) or the genotype × sex interaction (TW ANOVA, F [3, 43] = 0.19, p = 0.902) (Fig. 1f, Supplementary data 1f). When sex was not considered, OS t-tests showed that Wt (OS t-test, t [14] = 3.53, p = 0.003) and p16Ink4a^KO mice (OS t-test, t [13] = 2.33, p = 0.037) spent significantly more than 25% of the trial duration in the target quadrant. Interestingly, sex-stratified analyses revealed that both male (OS t-test, t [8] = 2.39, p = 0.048) and female (OS t-test, t [7] = 4.41, p = 0.0045) Wt mice displayed a significant quadrant preference, whereas in the p16Ink4a^KO group only males reached significance (OS t-test, t [6] = 3.01, p = 0.030) (Fig. 1f, Supplementary data 1f). Additionally, a Poisson model analysis of platform crossings revealed a significant reduction in the number of target zone entries for p16Ink4a^KO (Table 2; Poisson, β = −0.56, p = 0.043) and App^NL−G−F (Poisson, β = −0.36, p = 0.043) mice compared to Wt, suggesting less focused searching behaviour in these genotypes (Fig. 1g, Supplementary data 1 g). When sex was included in the model, neither genotype, sex, nor the genotype × sex interaction reached significance (Supplementary Data 1 g). On Day 14, no significant difference was observed between genotypes for either the latency to reach the target quadrant (KW : χ² [3] = 0.11, p = 1) or the time spent in the target quadrant (TW) ANOVA: F [3, 43] = 1.65, p = 0.19) (Fig. 1h, i and Supplementary data 1 h, i). When analyzed separately by sex, KW tests revealed no significant genotype effect on latency in either females (KW, χ² [3] = 1.82, p = 0.61) or males (KW, χ² [3] = 1.12, p = 0.77). Similarly, a TW ANOVA on time spent in the target quadrant showed no significant main effect of sex (TW ANOVA, F [1, 43] = 0.004, p = 0.95) or genotype × sex interaction (TW ANOVA, F [3, 43] = 1.10, p = 0.36), indicating that both male and female mice exhibited comparable search performance across genotypes (Supplementary data 1i). Notably, only Wt mice spent significantly more than 15 s in the target quadrant (OS t-test, t [15] = 2.25, p = 0.04), suggesting that long-term spatial memory was maintained in this group (Fig. 1i, Supplementary data 1i). When analyzed separately by sex, no genotype reached significance in females. In contrast, male App^NL−G−F;p16Ink4a^KO mice spent significantly less than 15 s in the target quadrant (OS t-test, t [4] = − 8.54, p = 0.001) (Supplementary data 1i). Poisson regression analysis of platform crossings showed no significant effect of genotype when sex was not considered (all p > 0.0.5) (Fig. 1j, Supplementary Data 1j). When sex was included in the model, a significant main effect of sex was observed (Poisson, β = −0.89, p = 0.024), with females making fewer platform crossings overall. In contrast, significant genotype×sex interactions were found for p16Ink4a^KO (Poisson, β = 1.14, p = 0.043) and App^NL−G−F;p16Ink4a^KO (Poisson, β = 1.47, p = 0.015) mice, indicating that these female groups exhibited relatively more crossings than their male counterparts. Interestingly, a significant main effect was also detected for the App^NL−G−F genotype (β = −0.88, p = 0.033) (Supplementary Data 1j). These results suggest that genotype alone did not affect platform crossing performance, but that sex-related or time-of-day–related factors may have modulated the observed effects in specific genotypes. Overall, the MWM data indicate that p16Ink4a inactivation causes a modest, short-lasting reduction in retrieval precision, while overall spatial preference remains intact. As expected in the App^NL−G−F model, the spatial memory impairment was characterised by increased latency to the target and reduced search focus. However, the double mutant App^NL−G−F;p16Ink4a^KO mice did not show an exacerbated phenotype compared to single mutants. Thus, while p16Ink4a^KO alters certain aspects of spatial search behaviour, it does not aggravate the memory deficits already present in App^NL−G−F mice. Although multiple sex-related differences were noted, they were inconsistent and did not persist across analyses. Because sex and testing time were confounded (males tested in the morning, females in the afternoon), the variability observed cannot be confidently attributed to biological sex effects. Moreover, sex-stratified analyses further reduced statistical power, making it difficult to determine whether the observed differences reflect true sex effects or sampling variability.

p16Ink4a loss drives dominant transcriptional changes, partially attenuated by the App^NL−G−F mutation

To gain further insight into the molecular mechanisms underlying these behavioural alterations and to determine whether interactions may exist at the transcriptional level, we next performed transcriptomic analyses in the hippocampus. This approach aimed to identify gene expression changes associated with p16 deletion, either in isolation or combined with the App^NL-G-F. Volcano plot analysis revealed that p16Ink4a^KO induced the most extensive transcriptomic alterations, with a marked predominance of upregulated genes (388 upregulated, 136 downregulated) (Fig. 2a, Supplementary Data 2). In comparison, the App^NL-G-F model exhibited a milder transcriptional response (187 upregulated, 44 downregulated) (Fig. 2b, Supplementary Data 2), while the App^NL-G-F;p16Ink4a^KO double mutant showed an intermediate profile (198 upregulated, 81 downregulated) (Fig. 2c, Supplementary Data 2). Overlap analysis demonstrated that the double mutant shared significantly more DEG with p16Ink4a^KO than with App^NL-G-F (Fisher’s exact test, p = 0.0005; odds ratio = 0.55), indicating that p16 loss exerts a dominant effect on the transcriptomic landscape (Fig. 2d, Supplementary Data 2). Notably, the reduced number of DEG in the double mutant compared to p16Ink4a^KO suggests that the App^NL-G-F background may partially attenuate the transcriptional consequences of p16 inactivation.

Fig. 2.

Transcriptional and functional rewiring across p16INK4A^KO, App^NL−G−F, and App^NL−G−F; p16INK4A^KO models (a–c) Volcano plots of differentially expressed genes (DEG) in hippocampi of p16INK4A^KO (a), App^NL−G−F (b), and App^NL−G−F; p16INK4A^KO (c) mice versus wild-type (Wt) controls. Red and blue points represent significantly upregulated and downregulated genes, respectively (adjusted p-value (p.adj) < 0.05; absolute log2 fold change (|LFC|) > 0.1). (d) Venn diagram showing the overlap of DEG across genotypes. The App^NL−G−F; p16INK4A^KO double mutant shares more genes with p16INK4A^KO than with App^NL−G−F. (e–g) DEG-derived protein-protein interaction (PPI) networks for each genotype, constructed using STRING (v12.0) and clustered via p16INK4A^KO network the Markov Cluster Algorithm (MCL). Nodes represent DEG and are color-coded by module. Upregulated genes are shown in full circle, while downregulated genes are depicted as circles. Edges represent PIP. (e) The p16INK4A^KO network is enriched for modules involved in translation, oxidative phosphorylation, and RNA processing, synaptic signalling and cytoskeletal organization. (f) The App^NL−G−F network displays prominent immune-related modules (e.g., myeloid activation, chemokine signaling, TLR response), along with metabolic and lysosomal remodeling. (g) The App^NL−G−F; p16INK4A^KO network combines immune and metabolic modules from each parental genotype and includes distinct modules associated with anti-inflammatory signaling, oxidative detoxification and cholesterol biosynthesis

To further resolve genotype-specific transcriptional programs, DEG-based PPI networks were generated using STRING (v12.0) [35] and clustered via MCL [36] to identify functional modules potentially driving phenotypic outcomes. This analysis revealed coordinated pathway-level dysregulation rather than isolated transcript changes, underscoring coherent perturbations across functionally related gene sets. Robust and convergent pathway-enrichment signals were identified across multiple analytical frameworks and biological replicates. In the p16Ink4a^KO model, the network was enriched in modules composed predominantly of upregulated genes associated with ribosomal function and translation (Module 1), RNA processing (Module 3), oxidative phosphorylation (Module 2), and amino acid metabolism (Modules 14–15), indicative of enhanced biosynthetic and bioenergetic activity. Additional modules related to protein degradation (Modules 4, 9), mitochondrial import (Module 16), and cellular stress responses (Module 8) were also enriched in upregulated genes. Conversely, modules linked to synaptic signalling (Modules 7, 10), cytoskeletal organisation (Modules 5, 12, 13), and neurogenesis were largely composed of downregulated genes (Fig. 2e, Supplementary Data 3a). In the App^NL-G-F model, module analysis revealed a strong immune signature, with Module 1 enriched for myeloid regulators (e.g., Itgax, Csf1r, Spi1), and Modules 3, 5, 7, and 10 capturing pathways involved in chemokine signalling, phagocytosis, complement activation, and Toll-like receptor signalling, the last one suggesting concurrent inflammation. Additional modules linked to Rho GTPase signalling (Module 4) and DAP12-associated microglial pathways (Module 5) further supported immune activation. Upregulation of translation (Module 2) and lysosomal/membrane remodelling pathways (Modules 8–9) reflected elevated inflammatory and metabolic demand (Fig. 2f, Supplementary Data 3b).The double mutant App^NL-G-F;p16Ink4a^KO recapitulated features of both models, integrating p16Ink4a^KO-associated metabolic modules (e.g., Module 1, Module 2) and App^NL-G-F-linked immune modules (e.g., Module 3, Module 4, Module 11). Notably, distinct modules emerged, including those associated with anti-inflammatory signalling (Module 7), oxidative detoxification (Module 8), and cholesterol biosynthesis (Module 12), potentially reflecting compensatory mechanisms to the dual genetic perturbation. Altogether, these findings support a primarily additive transcriptional architecture, with a subset of functionally meaningful interactions (Fig. 2g, Supplementary Data 3c).

App^NL−G−F –Driven IFN Microglial Response Is Moderated by p16Ink4a Deletion

Functional network analysis revealed marked activation of innate immune pathways in both App^NL−G−F and App^NL−G−F;p16Ink4a^KO models. This immune-related transcriptional profile pointed to a strong involvement of brain-resident immune cells, particularly microglia. To further explore the contribution of microglia to the observed transcriptomic alterations, we applied BRETIGEA (v1.0.4) to deconvolute bulk hippocampal transcriptomes and estimate the relative abundance of major brain cell populations, including astrocytes, endothelial cells, microglia, neurones, oligodendrocytes, and OPC [37]. The resulting cell-type abundance scores were subjected to statistical analysis to assess whether immune activation was associated with changes in cellular composition. KW testing revealed a significant effect of genotype solely on microglial scores (χ² = 17.4, df = 3, p = 5.78 × 10^{− 4}). Post hoc comparisons (Dunn’s test with Holm correction) showed a significant increase in microglial signatures in both App^NL−G−F (p.adj = 7.9 × 10^{− 4}) and App^NL−G−F;p16Ink4a^KO (p.adj = 0.035) compared to Wt, as well as a significant difference between App^NL−G−F and p16Ink4a^KO alone (p.adj = 0.029), while no difference was observed between p16Ink4a^KO and App^NL−G−F;p16Ink4a^KO (Fig. 3a). These findings confirm a pronounced microglial signature in App^NL−G−F and App^NL−G−F;p16Ink4a^KO models, consistent with the immune activation detected in functional network analyses. Given the well-established heterogeneity of microglial populations, we next investigated their functional diversity by performing GSEA using a curated set of 2,054 genes representing three major microglial states: homeostatic, activated, and IFN-responsive [42]. This analysis revealed significant enrichment of the IFN-responsive microglial signature in both App^NL−G−F vs. Wt and App^NL−G−F;p16Ink4a^KO vs. Wt comparisons, indicating activation of this subpopulation in response to amyloid pathology (Fig. 3b, c). In contrast, no such enrichment was observed in p16Ink4a^KO vs. Wt mice, suggesting that this IFN response is specifically driven by App^NL−G−F mutation. Notably, the NES associated with this signature was attenuated in the double mutant App^NL−G−F;p16Ink4a^KO (NES = 3.89) compared to App^NL−G−F alone (NES = 6.9), suggesting that p16Ink4a loss moderates the intensity of the IFN-driven microglial response without fully suppressing it (Fig. 3b, c). This partial attenuation may reflect an attempt to regulate immune activation modulation associated with p16Ink4a deficiency.

Fig. 3.

p16INK4A ^KO attenuates the interferon (IFN)-responsive microglial phenotype induced by App ^NL−G−Fmutation (a) Relative abundance scores of six major brain cell types (ast: astrocytes, end: endothelial cells, mic: microglia, neu: neurons, oli: oligodendrocytes, and opc: oligodendrocyte precursor cells) inferred using BRETIGEA deconvolution of hippocampal transcriptomes [37]. A significant genotype effect was detected only for microglia (mic) (Kruskal-Wallis (KW), χ² = 17.4, p = 5.78 × 10^{− 4}), with significantly increased microglial signatures in App^NL−G−F and App^NL−G−F;p16INK4A^KO mice compared to Wt (Dunn’s test, Holm correction, p.adj = 7.9 × 10^{− 4} and 0.035, respectively), and a significant difference between App^NL−G−F and p16INK4A^KO (p.adj = 0.029). (b–c) Gene Set Enrichment Analysis (GSEA) plots showing enrichment of INF-responsive microglial gene signatures. Significant upregulation of the INF-related program was observed in App^NL−G−F vs. Wt (NES = 6.90, p.adj = 7.53 × 10^{− 31}) (b) and App^NL−G−F;p16INK4A vs. Wt (NES = 3.89, p.adj = 3.02 × 10^{− 10}) (c)

Interactions between p16Ink4a^KO and App^NL−G−F revealed modulation of synaptic and immune pathways

While the overall transcriptomic profile of the App^NL−G−F;p16Ink4a^KO double mutant appeared largely additive, certain aspects, particularly immune modulation, suggested potential interaction effects. We formally tested this hypothesis by performing a LRT of the goodness-of-fit comparing an additive model (~ App^NL−G−F + p16Ink4a^KO) to the full linear model (~ App^NL−G−F + p16Ink4a^KO + App^NL−G−F :p16Ink4a^KO) [32]. This analysis did not reveal widespread interaction effects, reinforcing the conclusion that most transcriptional changes follow an additive pattern (Supplementary Data 4). We further investigated whether non-additive effects might emerge at the pathway level, even if individual gene interactions were absent. To this end, we extracted the main effects of App^NL−G−F and p16Ink4a^KO, as well as the interaction term (App^NL−G−F :p16Ink4a^KO) from the full linear model and performed GSEA on each coefficient set using GO categories (GOBP, GOCC, GOMF) (Supplementary Data 5a-d). Both p16Ink4a^KO and App^NL−G−F mutations independently altered numerous biological pathways, with p16INK4A^K having a stronger effect (614 vs. 433 pathways). Only 41 pathways were linked to the interaction term, indicating limited non-additive effects. The distribution of functional categories was assessed for each condition. In p16Ink4a^KO, metabolism (17.6%, mostly upregulated) and neuronal pathways (16.8%, mostly downregulated) were most represented. Similarly, App^NL−G−F showed downregulation of neuronal pathways (11.8%) and upregulation of immune and inflammatory pathways (21.7%). For the interaction term, neuronal and immune pathways each accounted for 19.5% of enriched terms, with neuronal pathways upregulated and immune pathways downregulated, suggesting divergent modulation in the double mutant compared to single mutations (Fig. 4a, Supplementary Data 5). The top 10 positively and negatively enriched terms for each condition were selected based on their NES values. GSEA revealed that p16Ink4a^KO strongly upregulated pathways related to mitochondrial function, protein synthesis, and energy metabolism, including “GOCC mitochondrial protein-containing complex”, “GOMF structural constituent of ribosome” and “GOCC aerobic respiration” (Fig. 4b, Supplementary Data 5a). In contrast, negatively enriched terms included “GOMF actin binding”, “GOBP dendrite development”, and “GOCC postsynaptic membrane”, suggesting that p16 loss may compromise synaptic structure and plasticity despite enhanced bioenergetic programs. App^NL−G−F primarily upregulated immune-related pathways, such as “GOBP immune effector process”, “GOBP response to bacterium”, and “GOBP positive regulation of defense response” (Fig. 4b, Supplementary Data 5c), consistent with the immune dysregulation previously described in this model [18, 43]. Conversely, negatively enriched terms indicated RNA dysregulation and synaptic impairment, such as “GOBP mRNA processing”, “GOBP ribonucleoprotein complex biogenesis”, and “GOBP dendrite development” (Fig. 4b, Supplementary Data 5c). Although the transcriptomic profile of App^NL−G−F;p16Ink4a^KO was mostly additive, GSEA on the interaction term uncovered specific non-additive regulation of synaptic and immune pathways. Positively enriched terms included “GOCC presynaptic membrane”, “GOCC synaptic vesicle membrane”, and “GOBP neurotransmitter secretion”, suggesting a relative preservation of neuronal function compared to what would be expected from an additive model. In contrast, negatively enriched interaction terms included translation and immune pathways, such as “GOCC ribosome”, “GOCC structural constituent of ribosome”, and “GOBP cytokine-mediated signalling pathway”, indicating attenuated activation of these processes in the double mutant (Fig. 4b, Supplementary Data 5c).

Fig. 4.

Pathway-level interactions between p16INK4A deletion and App^NL−G−F mutation reveal non-additive modulation of neuronal and immune processes (a) Distribution of functional categories enriched in gene set enrichment analysis (GSEA) based on coefficients derived from the full linear model: the main effect of p16INK4A^KO (left), the main effect of App^NL−G−F (centre), and the App^NL−G−F:p16INK4A^KO interaction (right). Enriched pathways were grouped into high-level biological categories. Pathways related to metabolism and neuronal function were most affected by the main effect of p16INK4A^KO, with metabolic terms predominantly overrepresented (red) and neuronal terms underrepresented (blue). Immune and inflammatory pathways were the most represented in the App^NL−G−F main effect. In contrast, interaction-specific terms revealed opposing regulation: neuronal pathways were enriched among overrepresented terms, while immune pathways were enriched among underrepresented ones. (b) Dot plot illustrating the top 10 positively and negatively enriched GO terms (GOBP: GO Biological Process, GOCC: GO Cellular component, GOMF: GO Molecular function) identified by gene set enrichment analysis (GSEA) on coefficients derived from the full linear model: the main effect of p16INK4A^KO (left), the main effect of App^NL−G−F (centre), and the App^NL−G−F:p16INK4A^KO interaction (right). Dot size reflects the number of contributing genes, and colour indicates the normalised enrichment score (NES), from negative (blue) to positive (red)

p16Ink4a loss in the adult brain promotes metabolic activation and impairs synaptic and structural pathways

To better define p16’s brain-specific functions, we next focused on its effects in an App^Wt background, an approach that also addresses the current lack of knowledge on p16’s roles in the adult brain. Thus, we performed GSEA using GO annotations. Consistent with previous STRING and GSEA analyses on p16Ink4a^KO main effect, the top 10 positively enriched pathways (NES > 0) revealed robust activation of mitochondrial function and components (e.g., GOBP aerobic respiration, GOCC mitochondrial protein containing complex, GOCC oxidoreductase complex), protein synthesis machinery (e.g., GOCC ribosomal subunit, GOMF structural constituent of ribosome), and metabolic processes (e.g., GOBP generation of precursor metabolites and energy) (Fig. 5a). These terms formed a highly interconnected network, with mitochondrial and energy-related pathways closely linked to ribosomal functions (Fig. 5b). These hub terms reflect the coordinated regulation of mitochondrial protein synthesis, a critical step in supporting oxidative phosphorylation and energy metabolism. Altogether, these results indicate that p16 loss induces a coordinated transcriptional program where enhanced mitochondrial biogenesis is matched by increased translational capacity. Conversely, negatively enriched pathways (NES < 0) were predominantly associated with neuronal architecture, synaptic signalling, and both inhibitory and excitatory neurotransmission (e.g., GOBP regulation of synapse structure or activity, GOBP postsynapse organisation, GOCC presynaptic membrane). Additionally, pathways involved in cytoskeletal organisation and intracellular transport (e.g., GOMF actin binding, GOMF cytoskeletal motor activity) were also significantly downregulated and formed a tightly interconnected network (Fig. 5a, c). This suggests shared molecular mechanisms underpinning both structural and synaptic impairments. Of particular interest, even though not among the top 10 ranked pathways, “GOBP learning” and “GOBP cognition” were also integrated within this network (Fig. 5c). Their inclusion directly links the observed transcriptional alterations to potential cognitive dysfunctions. Together, these results position p16 as a critical regulator of neuronal structure, synaptic function, and cognitive flexibility, which remain largely underexplored in the adult brain outside pathological or oncogenic contexts.

Fig. 5.

Transcriptomic effects of p16NK4A loss in the adult heatlhy brain reveal enhanced metabolic activity and structural/synaptic impairments (a) Dot plot showing the top 10 positively and negatively enriched gene sets across Gene Ontology (GO) terms (GOBP: GO Biological Process, GOCC: GO Cellular Component, GOMF: GO Molecular Function) and Hallmark pathways following gene set enrichment analyses (GSEA) in p16INK4A^KO vs. Wt mice. Dot color indicates normalised enrichment score (NES), and size reflects the –log10 adjusted p-value (p.adj). (b) Network representation of positively enriched GO terms (NES > 0) showing highly interconnected mitochondrial and ribosomal components. Central hub nodes such as GOBP mitochondrial translation and GOCC organellar ribosome highlight coordinated upregulation of mitochondrial protein synthesis and energy metabolism. (c) Network of negatively enriched GO terms (NES < 0) centred around synaptic organisation and neuronal architecture. Terms related to excitatory/inhibitory neurotransmission (e.g., GOBP postsynapse organisation, GOCC presynaptic membrane), cytoskeletal dynamics (GOMF actin binding, GOMF cytoskeletal motor activity), and learning-related processes (GOBP learning, GOBP cognition) are tightly interconnected

Pathway analysis reveals metabolic activation, stress responses, and senescence-associated signatures in p16Ink4a^KO

To further elucidate the physiological consequences of p16Ink4a inactivation in the young adult brain and gain a broader systems-level perspective, we complemented our GO-based GSEA with Hallmark pathway analysis. These curated gene sets, derived from the integration of multiple expression datasets, provide non-redundant representations of core biological processes and disease-relevant transcriptional programs. Consistent with our previous findings, this analysis revealed strong positive enrichment for pathways involved in cellular metabolism and biosynthesis, including “Hallmark Oxidative Phosphorylation”, “Hallmark Glycolysis”, and “Hallmark Fatty Acid Metabolisme”, confirming extensive metabolic reprogramming following p16Ink4a loss (Fig. 5a, Supplementary Data 5 d). Activation of anabolic signalling was further supported by enrichment in “Hallmark MTORC1 Signalling” and “Hallmark MYC Targets”, pointing to enhanced biosynthetic activity. Simultaneously, stress-related pathways such as the “Hallmark Reactive Oxygen Species Pathway”, “Hallmark Unfolded Protein Response”, and “Hallmark p53 Pathway” were significantly enriched, reflecting compensatory responses to increased oxidative and metabolic stress (Fig. 5a, Supplementary Data 5 d). These transcriptomic signatures also suggested a potential shift toward a senescent-like state, as evidenced by the positive enrichment of key senescence-associated pathways (“Hallmark TNF Signalling via NF-κB”, “Hallmark P53 Pathway”, and “Hallmark mtor Signalling”) (Fig. 5a, Supplementary Data 5 d). To investigate the activation of CS in the p16^KO model, we performed a GSEA using two complementary signatures: the “SenSaul” gene set, enriched in SASP-associated inflammatory mediators [41], and the Reactome CS pathway (R-MMU-2559583), including the sub-branches covering canonical aspects of senescence. Analyses were also extended to Ap^NL−G−F and p16^KO; p16Ink4a^KO models. In the p16Ink4a^KO hippocampus, we observed significant enrichment for “Reactome Oncogene-Induced Senescence” and the “Reactome Senescence-Associated Secretory Phenotype”, supporting the induction of a complex senescence program that includes inflammatory, epigenetic, and cell cycle regulators such as Hmga1b, H2az2, Anapc11, Anapc16, and Mapk3 (Fig. 6a and b, Supplementary Data 2, 5 d). In App^NL−G−F mice, GSEA revealed a significant enrichment of the “SenSaul” signature, with overexpression of Ccl3 and Ccl4, which are key SASP components involved in immune cell recruitment [44]. Other upregulated genes, such as Tnfrsf1a [45], a receptor promoting NF-κB signalling and SASP induction, further support the presence of an inflammatory senescence-like phenotype in App^NL−G−F (Fig. 6c, Supplementary Data 2, Supplementary Data 5e). In the App^NL−G−F;p16Ink4a^KO double mutant, enrichment was observed for the “Reactome Senescence-Associated Secretory Phenotype” pathway (Fig. 6d, Supplementary Data 5f). Notably, Cdkn2b, encoding the protein p15, a G1 cell cycle inhibitor and paralog of p16Ink4a located at the INK4/ARF locus, was the only significantly upregulated gene in the core enrichment, suggesting a compensatory mechanism aimed at restoring cell cycle control in the absence of p16Ink4a^KO (Supplementary Data 2) [46].

Fig. 6.

Enrichment of senescence-associated pathways in p16INK4A^KO, App^NL−G−F, and double mutant hippocampi Gene set enrichment analysis (GSEA) plots show significant enrichment of senescence-related gene sets. (a–b) In p16INK4A^KO mice, the “Reactome Oncogene-Induced Senescence” (Normalized Enrichment Score (NES) = 1.99, adjusted p-value (p.adj) = 0.014) (a) and “Reactome Senescence-Associated Secretory Phenotype” (NES = 2.60, p.adj = 6.7 × 10^− 4) (b) pathways were enriched, indicating activation of canonical and inflammatory senescence programs. (c) In App^NL−G−F mice, the “SenSaul” gene set was enriched (NES = 2.33, p.adj = 9.5 × 10^− 3), reflecting an inflammatory senescence-like profile. (d) In double mutants, the “Reactome Senescence-Associated Secretory Phenotype” remained enriched (NES = 2.50, p.adj = 6.93 × 10^− 4)

Discussion

This study aimed to investigate the role of the cyclin-dependent kinase inhibitor p16Ink4a in the adult mouse brain, both under physiological conditions and in the context of Alzheimer’s disease. By combining behavioural phenotyping with hippocampal transcriptomic profiling, we explored how p16Ink4a deletion affects cognitive performance and molecular homeostasis in Wt and App^NL−G−F mouse models. Based on prior studies suggesting a potentially protective role of senescent cell clearance or senescence gene inhibitionin neurodegenerative conditions, we hypothesised that the double mutant App^NL−G−F;p16Ink4a^KO mice might exhibit a partial rescue or attenuation of memory deficits [7–10]. However, our results did not support this assumption. Instead, in the MWM paradigm, p16Ink4a deletion alone resulted in subtle impairments in spatial learning and cue integration, along with a transient reduction in retrieval precision, without altering the behavioural phenotype of App^NL−G−F mice. While p16Ink4a’s role in memory processes may be restricted in amyloidogenic environments or counterbalanced by compensatory mechanisms in AD models, we propose that the cognitive phenotype in the p16Ink4a^KO mice may have plateaued, hindering detection of age-dependent additive effects. Here we described for the first time behavioural alterations induced by the constitutive inactivation of Cdkn2a/p16Ink4a gene, contributing to the understanding of p16Ink4a’s role in the brain. Additionally, we observed sex-dependent differences in some behavioural measures; however, these effects were not consistent. They may reflect limited statistical power and potential time-of-day influences rather than robust biological differences, although true sex effects cannot be fully excluded. Because males were tested in the morning and females in the afternoon, these factors cannot be conclusively disentangled. At the molecular level, p16Ink4a^KO and App^NL−G−F mice exhibit strikingly distinct transcriptomic profiles. p16Ink4a inactivation triggered upregulation of metabolic and translational pathways while suppressing genes involved in synaptic organisation and neurotransmission, suggesting a critical role for p16Ink4ain maintaining neuronal homeostasis. Conversely, the App^NL−G−F model displayed pronounced upregulation of innate immunity-related pathways, consistent with the neuroinflammatory hallmarks of AD [47]. The double mutant exhibited a transcriptomic landscape that was primarily additive and largely driven by the p16Ink4a^KO signature. The immune-related transcriptional profile characteristic of the App^NL−G−F model, particularly the enrichment of the IFN-responsive microglial signature, was also detected in the double mutant, although being attenuated. Despite the largely additive nature of the transcriptomic landscape in the double mutant, App^NL−G−F and p16Ink4a^KO exhibit non-additive interactions at the pathway level. Specifically, we identified interaction effects modulating key neuronal pathways, suggesting that the two mutations do not act entirely independently in shaping the neuronal transcriptome. Additionally, interaction analysis revealed a significant reduction in inflammatory pathway activation in the double mutant compared to App^NL−G−F alone. This partial suppression of immune-related gene expression supports the idea that loss of p16Ink4aplays a compensatory role, mitigating the pro-inflammatory effects driven by amyloid pathology. This interpretation is consistent with emerging evidence that p16Ink4a modulates immune responses. In a recent study, Zhang et al. (2022) demonstrated that targeted elimination of p16Ink4a-expressing cells in aged Wt mice, using the inducible p16^INK/ATTAC system, restored immune homeostasis in the brain and improved cognitive performance. In this model, p16Ink4a-positive cells were primarily resident microglia and infiltrating monocytes exhibiting high expression of SASP-related and major histocompatibility complex (MHC) genes, reflecting a state of chronic inflammation. Strikingly, selective depletion of these cells significantly reduced neuroinflammation and preserved cognition, particularly in female mice [48]. However, in our experiment, the constitutive p16Ink4a^KO model did not lead to behavioural improvement in App^NL−G−F;p16Ink4a^KO. This discrepancy likely stems from fundamental methodological differences. While the p16^INK/ATTAC model enables temporally controlled elimination of senescent cells in adulthood, thus avoiding developmental effects, our model lacks p16Ink4a from early embryogenesis, potentially altering neuronal development and obscuring the benefits of removing this CS mediator. Moreover, Zhang et al.’s study (2022) was conducted in aged Wt animals, where the burden of senescence is naturally high [48]. In contrast, our analysis was performed in young adult mice carrying a pathological mutation, where early responses to cellular stress may prevail over ageing-related mechanisms. In contrast, our analysis was performed in young adult mice carrying a pathological mutation, where early responses to cellular stress may dominate over ageing-related processes. The lack of behavioural exacerbation observed in the double mutant may also reflect a saturation effect, in which key cognitive functions have already reached a threshold of impairment beyond which additional deficits are difficult to detect. Critically, our analyses were conducted at a pre-pathological stage, preceding the full development of chronic disease. The absence of robust APP and p16Ink4a interplay may reflect inadequate temporal exposure to cumulative cellular stressors. These results highlight the necessity for longitudinal investigations to elucidate the time-dependent compensatory mechanisms modulating these genetic interactions and to define the long-term consequences of p16Ink4a^KO on neurodegenerative progression, particularly in contexts of sustained amyloidogenic stress. Despite these differences, both studies converge on a central observation: p16Ink4a is a key regulator of inflammation in the brain, particularly within microglia. This finding is especially relevant in the context of AD, where chronic neuroinflammation driven by dysregulated microglial activity is increasingly recognized as a major contributor to disease progression [49]. Interestingly, several recent studies have linked type I interferon (IFN-I) signalling in microglia to AD pathology progression in mouse models [50–52]. In this light, the attenuation of IFN-responsive microglial signatures observed in our App^NL−G−F;p16Ink4a^KO mice suggests that p16Ink4a loss may temper exaggerated immune responses during the early stages of amyloid pathology. Further elucidating how p16Ink4a^KO shapes microglial identity and function, and whether this occurs through its canonical role as a cell cycle inhibitor or via non-canonical mechanisms such as transcriptional regulation, metabolic control, immune modulation, or senescence-associated pathways, could provide valuable insights into the molecular processes that may limit the neurotoxic effects of sustained inflammation in AD. Our biological pathway analysis in the adult brain of p16Ink4a^KO revealed a profound metabolic reprogramming. This includes the coordinated activation of protein biosynthesis pathways, ribosomal translation, mitochondrial function, and key energy metabolism pathways, notably glycolysis and lipid metabolism. Concurrently, a marked cytoskeletal instability affecting actin filaments, microtubules, and intracellular motor complexes was observed. Network analyses revealed a strong interconnection between these pathways, synaptic transmission deficits and behaviour-related pathways. As such, functional changes are probably the primary driver of the cognitive impairments observed in p16Ink4a^KO mice. Our results highlight the importance of p16Ink4a beyond its well-documented role as a tumour suppressor. p16Ink4a has been extensively studied in oncology, as the INK4/ARF locus is frequently mutated in various cancers [53]. While tumours typically appear in the model only after one year of age in our p16Ink4a^KO [19], our results reveal early molecular alterations, well before the onset of tumourigenesis, suggesting that p16Ink4a is fundamentally involved in maintaining cellular homeostasis in the brain. Moreover, the transcriptomic enrichment patterns observed at 4 months in our p16Ink4a^KO mice closely align with those reported by Micheli et al. (2019) in 12-month-old animals. In that study, physical exercise increased the number of neural progenitor cells in the DG of p16Ink4a^KO mice, yet the number of mature neurones remained unchanged due to elevated neuronal apoptosis [16]. These findings underscore a dual role for p16Ink4a in restraining excessive neurogenesis while preserving neural cell survival. Consistent with our results, transcriptomic analyses from Micheli et al. (2023) revealed enrichment in pathways related to cytoskeletal organisation, neuronal function, and protein synthesis [17]. Here, we propose a mechanistic hypothesis whereby constitutive activation of the mTORC1 pathway acts as a central regulatory node, driving an exacerbated anabolic response. This is evidenced by increased gene expression in protein synthesis, ribosomal biogenesis, and nucleotide metabolism. The resulting biosynthetic overload imposes a substantial energetic burden that may be incompatible with neuronal demands. Neurones, which rely heavily on ATP to maintain ionic gradients, support synaptic transmission, enable axonal transport, and preserve cytoskeletal dynamics, are particularly vulnerable to such energetic imbalances [54, 55]. This chronic stress likely manifests through a cascade of cellular disturbances, including cytoskeletal instability and neuronal dysfunction, which converge toward activating apoptotic, autophagic, and senescence pathways, potentially mediated by p53 signalling. This metabolic hypothesis is supported by Buj et al. (2019), who showed that p16INK4A deletion leads to constitutive activation of mTORC1 and enhanced nucleotide biosynthesis through increased translation of ribose-5-phosphate isomerase A, a key enzyme in the pentose phosphate pathway, in cancer cells [56]. Similarly, Tyagi et al. (2017) reported that p16INK4A increases mitochondrial mass and the expression of mitochondrial respiratory complex proteins. However, this adaptation is accompanied by abnormal mitochondrial morphology, impaired respiratory capacity, and reduced ATP production [57]. Notably, these alterations are independent of p16INK4A’s canonical function in cell cycle regulation, further reinforcing that its loss disrupts critical non-canonical metabolic processes [56, 57]. While our findings support a mechanistic link between p16INK4A loss, metabolic dysregulation, and neuronal vulnerability, our interpretations remain hypothetical and are based primarily on transcriptomic signatures. Direct biochemical and quantitative validation of mTORC1 activation, mitochondrial dysfunction and cytoskeletal alterations will be required to fully evaluate this model. Strikingly, transcriptomic profiling revealed a paradoxical enrichment of senescence-associated pathways in the p16Ink4a^KO model, suggesting that p16Ink4a loss may inadvertently drive stress-induced premature senescence (SIPS) through compensatory mechanisms. Notably, a similar senescence-like transcriptional signature was identified in the App^NL−G−F model. However, whether this signature represents bona fide cellular senescence or an exacerbated inflammatory response remains unresolved, given its overlap with SASP markers. Intriguingly, the App^NL−G−F;p16Ink4a^KO double mutant exhibited elevated expression of Cdkn2b, a paralog within the INK4/ARF locus. This suggests compensatory upregulation of Cdkn2b to counterbalance p16INK4A loss, aligning with prior evidence of functional redundancy between INK4/ARF paralogs in restoring cell cycle arrest under stress. Such compensatory mechanisms may partially offset the functional consequences of p16INK4A deficiency in neurodegenerative contexts [46]. However, to to determine whether the observed transcriptional signatures reflect real senescence upregulation, quantitative assessment of canonical senescence markers (e.g., p21, γH2AX, and key SASP cytokines) will be required.

In this study, we investigated the role of p16Ink4ain the adult mice brain under both physiological and Alzheimer’s disease conditions with some limitations. We found an unexpected function of p16Ink4a with non-canonical effects on brain function independent of classical senescence pathways. We also observed a modulation of immune-related pathways in the App^NL−G−F;p16Ink4a^KO double mutant, yet these molecular changes did not translate to detectable behavioural differences. The pervasive transcriptomic rewiring caused by p16Ink4a^KO in non-pathological contexts complicates the isolation of AD-specific interactions. In addition, the use of a constitutive p16Ink4a inactivation introduces an intrinsic limitation. Because p16Ink4a is absent from early embryogenesis onward, the model does not recapitulate physiological conditions and may induce compensatory or otherwise unexpected mechanisms that differ markedly from the consequences of p16Ink4a loss in adulthood. Such lifelong absence of p16Ink4a complicates the attribution of adult phenotypes solely to its physiological function in the mature brain. These results underscore the necessity for temporally controlled KO models to disentangle p16Ink4a’s developmental versus adult functions in neurodegenerative processes. A second limitation of our study is that sex and time-of-day were confounded, as males were tested in the morning and females in the afternoon. This separation was implemented to minimize stress, odor, and hormonal cross-contamination between sexes, ensuring animal welfare and stable behavioral baselines. However, this design prevents full disentanglement of circadian and sex effects. Future studies should consider randomizing testing times or including time-of-day as an explicit factor to better control for potential circadian influences. The limitations of bulk transcriptomics preclude unequivocal discrimination between bona fide senescence and inflammation-mimicking transcriptional responses. To resolve this ambiguity, future studies should integrate spatially and temporally resolved approaches (e.g., single-cell RNA-seq, spatial transcriptomics). Such granular characterisation is essential to rigorously define senescence before evaluating senolytic or senomorphic therapies in this model. Despite these technical constraints, our work challenges the conventional view of p16Ink4a by revealing its non-canonical roles in neuroimmune modulation and transcriptional regulation, extending beyond its established cell cycle arrest and senescence functions. We demonstrate that p16Ink4a critically regulates metabolic pathways (e.g., glycolysis, mitochondrial respiration) and cytoskeletal organisation in the brain, establishing it as a multifunctional guardian of neuronal homeostasis. Building on these findings, we propose that the observed metabolic reprogramming is driven by mTORC1 hyperactivation and posit that rapamycin-mediated mTORC1 inhibition offers a tractable strategy to test this hypothesis mechanistically. Notably, our work uncovers an unexplored dimension of p16Ink4a biology: its potential role in neurodevelopment. While p16Ink4a has not been directly linked to neurodevelopmental disorders, emerging studies highlight striking overlaps between oncogenic signalling networks (e.g., RAS/MAPK, PI3K/AKT) and pathways implicated in conditions like autism spectrum disorder (ASD). This parallelism raises intriguing questions about p16Ink4a’s involvement in early brain wiring and synaptic refinement [58–60]. Notably, signal strength, rather than mutation status alone, may determine divergent clinical outcomes, promoting proliferation in cancer and impairing differentiation in neurodevelopmental disorders [58]. Within this framework, the role of p16Ink4a in balancing cell survival, differentiation, and plasticity may extend beyond tumour suppression, warranting further investigation into its epigenetic and developmental functions in the brain.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AD

Alzheimer’s Disease

ARRIVE

Animal Research: Reporting of in Vivo Experiments

ASD

Autism Spectrum Disorder

Aβ

β-Amyloid

CS

Cellular Senescence

DAM

Disease-Associated Microglia

DEG

Differentially Expressed Genes

EMM

Estimated Marginal Means

FDR

False Discovery Rate

GLM

Generalized Linear Model

GO

Gene Ontology

GOBP

Gene Ontology Biological Process

GOCC

Gene Ontology Cellular Component

GOMF

Gene Ontology Molecular Function

IFN

Interferon

IFN-I

Type I Interferon

IQR

Interquartile Range

ITE

Initial Trajectory Error

KO

Knock-Out

KW

Kruskal-Wallis

LFC

Log2 Fold Change

LMM

Linear Mixed-Effects Model

LRT

Likelihood Ratio Test

MCL

Markov Cluster Algorithm

MWM

Morris Water Maze

NSPC

Neural Stem/Progenitor Cell

OPC

Oligodendrocyte Progenitor Cells

OS t-test

one-sample t-test

OW ANOVA

One-Way ANOVA

p

p-value

p.adj

Adjusted p-value

PCA

Principal Component Analysis

PPI

Protein–Protein Interaction

Q–Q

Quantile-Quantile

REML

Restricted Maximum Likelihood

rRNA

Ribosomal RNA

SASP

Senescence-Associated Secretory Phenotype

SGZ

The Subgranular Zone

SIPS

Stress-Induced Premature Senescence

SVD

Singular Value Decomposition

TIN

Transcript Integrity Number

TW ANOVA

Two-Way ANOVA

VST

Variance-Stabilized Expression Data

Wt

Wild-Type

Author contributions

Keyes and Herault contributed to the conceptualization of the Study with the funding acquisition. The formal analyses were performed by Stephan and Plassard under the supervision of Herault. The first draft writing was done by Stephan and Herault and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the “Fondation pour la Recherche Médicale”, “Centre National de la Recherche Scientifique (CNRS)”, “Institut National de la Santé et de la Recherche Médicale” (INSERM), the “Université de Strasbourg (Unistra)”, and support from the program Investissements d’Avenir labeled IdEx Unistra (ANR-10-IDEX-0002), a SFRI-STRAT’US project (ANR 20-SFRI-0012), EUR IMCBio (ANR-17-EURE-0023) to YH, INBS PHENOMIN (ANR-10-INBS-07 PHENOMIN) and FRANCE2030 to YH. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

The transcriptomic datasets generated during and analysed during the current study are available in the GEO DATABASE repository under the reference GSE300348.

Declarations

Ethics approval and informed consent

The study was conducted following the Directive of the European Parliament: 2010/63/EU, revising/replacing Directive 86/609/EEC and with French Law (Decret n° 2013 − 118 01 and its supporting annexes entered into legislation 01 February 2013) relative to the protection of animals used in research. YH was the principal investigator of this study (accreditation 67–369) in the PHENOMIN-ICS animal facility (Agreement C67-218-40). The study was approved from the French Ministry of Higher Education and Research (APAFIS #43632-2023042814259627 v4), and following the Animal Research: Reporting of In Vivo Experiments (ARRIVE) guidelines [20]. There was no need for an informed consent.

Competing interests

The authors declare no competing interests.