Comprehensive analysis of interactions between brain aging and late-onset psychoses using heuristic mapping models

Abstract

Background

Advanced brain aging is closely associated with late-onset psychoses, including bipolar disorder(BD), schizophrenia(SP), and major depressive disorder(MDD). Studies have shown that neuroimmune homeostasis imbalance plays an important role in the progression of these conditions. However, potential pathogenetic features of these different late-onset psychoses in the context of brain aging, as well as the crosstalk between aging and these psychiatric conditions, remain to be fully elucidated.

Methods

To further explore key interrelationships between aging and psychoses systematically, a series of work was designed: First, machine learning was used to identify aging and disease-related features and an aging score was devised. These results were further used to optimize a series of mapping models to explore potential relationships between aging and disease. Subsequently a number of bioinformatics methods, including sensitivity analysis, enrichment analysis, and network analysis were conducted to explore crucial links between aging and these tested diseases.

Results

Shared features and specific characteristics of these psychoses in the context of brain aging are identified. The neuroimmune homeostasis is highlighted, and changes affecting cellular homeostasis, vascular homeostasis and the heart-brain axis are also integrated. Furthermore, various potential pathogenic features characterize the individual conditions. These results support the importance of neuro-immunosenescence and vascular-lymphatic aging theories.

Conclusions

By systematically exploring the aging index in different late-onset psychoses, this work integrates various risk factors and presents an approach to apply to other aging-related diseases. It provides a computational framework for future therapeutic interventions targeting the interplay between aging and psychosis.

Plain language summary

Psychosis is a mental state in which people lose contact with reality, and features symptoms such as hallucinations and delusions. Late-onset of psychoses occur in older people, and the increased occurrence is associated with brain aging. In this work, we designed computational models to study key interactions between aging and development of disease. We found that an imbalance in the immune system in the brain and various risk factors associate with specific causes of psychosis. Our computational framework presents an approach that could be applied to other aging-related diseases to enable further study of crucial interactions between aging and disease.

Ren et al. Identify key interactions between aging and late-onset psychoses by designing a sereis of heuristic mapping models. Both shared features and specific characteristics of these psychoses in the context of brain aging are integrated.

Introduction

Given the rapid aging of the global population, interest in late-onset psychotic conditions is increasing^1–3. In individuals over 45^1–3, the prevalence of bipolar disorder (BD) and schizophrenia (SP) is 0.1–0.5%, while major depressive disorder (MDD) also occurs commonly^4–6. Late-onset mental illness leads to considerable disability, has a poorer outcome, and is associated with a higher risk of suicide and mortality⁷. Additionally, people suffering from mental disorders have an increased risk for developing multiple chronic diseases⁸. However, as the same psychiatric diseases also occur in younger people⁹, potential specific pathogenic features of late-onset psychoses merit additional research.

Studies of late-onset psychoses have shown that they are often accompanied by accelerated brain aging^10–13. BD, SP, and MDD also exhibit aging-related features such as chronic low-level inflammation, oxidative stress, and genomic instability^14–16. For example, BD is associated with elevated cytomegalovirus IgG titers and an expansion of senescent and regulatory T cells¹⁷, a rapid age-related decline of gray and white matter volume was described in SP¹⁸, while a deficiency of brain-derived neurotrophic factor in MDD is thought to affect neural stem cell proliferation and differentiation¹⁵. On the other hand, late-onset mental illness appears to accelerate the process of aging, resulting in a greater risk of patients developing age-related diseases^14,15,19.

Substantial evidence supports the role of altered neuroimmune homeostasis in the pathologic processes seen in late-onset psychoses²⁰. It has been suggested that the development of BD, SP, and MDD is associated with dysregulated immune responses in the Central nervous system(CNS) that affect microglia activation, the production of pro-inflammatory cytokines, the emergence of self-reactive T cells, and the functional state of the blood-brain barrier²⁰. It was also observed that these immunologic alterations can disrupt synaptic transmission, neuroendocrine regulation, and impair neuroplasticity¹³. For example, elevated cerebrospinal fluid cytokine levels were shown to hinder nerve regeneration in BD, SP, and MDD^13,21,22. Furthermore, the cortical neuroinflammation in SP often results in immunologic abnormalities also detectable in systemic immune responses²¹. Brain aging disrupts various aspects of the neuroimmune homeostasis, thus increasing the risk of psychosis, as senescent CNS cells, especially microglia and astrocytes, drive excessive inflammatory responses that result in impaired synaptic plasticity²³. In conclusion, we hypothesize that various forms of imbalance in neuroimmune homeostasis contribute to the development and progression of late-onset psychoses in the aging brain (Fig. 1A). Therefore, we embarked on a systematic exploration of the relationship between the underlying mechanisms involved in aging and late-onset psychoses.

Fig. 1. The workflow.

A Diagram illustrating our hypothesis. B The computational pipeline used in this work. C The data flow of the mapping model. D Process of integration of different generators used to identify key relationships between aging and each disease.

The use of artificial intelligence opens up powerful approaches in the analysis of various ‘omics’ profiles obtained in a range of mental disorders. The use of machine learning (ML) in psychiatric research has seen an exponential increase recently²⁴ and the combination of machine learning and bioinformatics has been used to identify differential genes of BD, SP, and MDD²⁵. In addition, the generative adversarial network (GAN) approach has been employed extensively to establish a diagnostic neuroimaging model in SP²⁶. Functional magnetic resonance imaging (fMRI) data was also analyzed using GAN to predict the clinical trajectory of MDD²⁷. In short, potential pathological characteristics could be further investigated by generating pseudo-samples from random values with the help of relative optimization rules. Moreover, network analysis can be used to identify key molecular features of complex psychiatric diseases with multiple etiologies. As an example, analyzing transcriptome data allowed investigators to establish key pathogenic biomolecular networks in MDD patients²⁸. Nevertheless, the relationships between molecular mechanisms seen in brain aging and late-onset psychoses remain to be fully explored. We hypothesized that developing effective prediction models of late-onset psychoses in the context of brain aging would facilitate the identification of molecular pathways shared between, or specific to, BD, SP, and MDD.

Despite growing evidence linking aging to psychosis, the precise molecular mechanisms driving this association, and whether they are shared or distinct across different disorders, remain poorly understood. A systematic, data-driven approach is urgently needed to dissect these complex interactions. In this work, we explore potential relationships between brain aging and late-onset mental diseases by optimizing a series of heuristic mapping models, where key interactions between aging and disease are identified with the help of sensitive analysis. This work integrates crucial aging-related pathways for each disease, highlighting the importance of neuroimmune homeostasis. It also provides a systematic analysis of risk factors that are shared between or specific to individual late-onset psychiatric conditions.

Methods

A brief description of this work

This work aimed to identify key interactions between aging and each late-onset psychosis, and explore potential interactive mechanisms between them (i.e., the neuroimmune homeostasis, Fig. 1A). As a result, a series of computational methods were integrated as follows (Fig. 1B): (1) Markers of BD, SP, MDD, and aging were identified using ML. (2) An aging score was calculated to verify the correlation between molecular patterns seen in advanced aging and late-onset psychoses. (3) A series of mapping models was designed to study any potential relationship between aging and psychoses (Fig. 1C and Supplementary Fig. 1). (4) Sensitive analyses were performed to identify key relationships between aging and disease development (Supplementary Fig. 2). (5) Enrichment analysis and network analysis were used to systematically identify interactive mechanisms between aging and major psychotic diseases.

Data and preprocessing

The DNA methylation profiles were obtained from the GEO database (Supplementary Data 1 and 2), including GSE38873, GSE41037, GSE41169, GSE611431, GSE89707, GSE112179, GSE152026, GSE157252, GSE129428, GSE198904, and GSE88890. These datasets were obtained from three different platforms: GPL21145, GPL13534, and GPL8490. In this work, the dataset with age-matched (age ≥ 45) normal samples>10 was retained for the following normalization step. As a result, BD, MDD, and SP were analyzed.

The steps of obtaining DNA methylation profiles were as follows:

1. Samples without age or phenotype indices (i.e., late-onset mental illness versus control) were excluded.

2. CpG sites with missing values ≥ 25% were deleted.

3. According to different brain regions, the k-nearest neighbor algorithm (k = 10, with the Euclidean distance) was used to supplement the missing values.

4. Individuals without the age or phenotype (i.e., normal or disease) index were deleted.

5. Patients with early-onset psychoses (age < 45) were removed from the late-onset group, but still used for their own models.

6. The z-score normalization was performed based on healthy individuals.

7. The Singular Value Decomposition (SVD) method was used to eliminate inter-sample variation based on the top three principal components in healthy individuals.

8. The z-score was then utilized to normalize all individuals accepting the mean and the standard deviation of the healthy individuals.

In total of 2986 analyzable samples were collected, including 394 control individuals with age > 50 years, (270 in training datasets + 124 test datasets), 949 younger controls (age ≤ 50, 640 training set + 309 test data), 192 patients suffering from BP(128 early onset + 64 late onset), 1031 individuals with SP (784 early onset + 247 late onset), and 420 MDD patients (240 early onset + 180 late onset). The DNA methylation dataset included 24,313 CpG sites (Supplementary Data 3).

We also obtained the gene expression profiles from the GEO platform (Supplementary Data 4 and 5), including GSE35974, GSE208338, GSE62191, GSE5388, GSE5389, GSE210064, GSE62333, GSE145554, GSE76826, and GSE98793. These datasets were obtained from three different platforms: GPL17077, GPL20844, GPL4133, GPL5188, GPL570, GPL6244, and GPL96. The gene expression profiles were processed as follows:

1. Samples without age or health-related information (i.e., late-onset mental illness versus control) were deleted.

2. The gene expression matrix for each dataset was integrated by aligning the probe number within the corresponding gene symbol.

3. Missing gene expression values were filled with a value of 0.

4. Data processing was performed on the summary matrix to remove genes with ≥25% missing expression data.

5. The gene expression matrix was logarithmically transformed if it contained outliers.

6. Based on the mean and standard deviation values of gene expression level values in control individuals, z-score normalization was performed for the MS and control samples.

7. The SVD method was performed to eliminate the inter-sample variation based on the top three principal components of the control samples.

8. The z-score was then utilized to normalize all samples based on the mean and the standard deviation of the control samples.

The process above yielded 838 analyzable samples, including 314 healthy individuals with an age > 50 years old: 148 samples, and age ≤ 50, 166, and patients suffering from BD (72 early onset + 79 late onset), SP (80 early onset + 101 late onset), and MDD (51 early onset + 134 late onset), respectively. The studied gene expression profiles included 13,796 gene symbols (Supplementary Data 6).

Modeling the aging as well as each disease predictor

The whole DNA methylation profiles were used to build a series of classification models as follows:

1. The normal aging model, discriminating normal young (age ≤ 50) and normal old (age > 50).

2. The late-onset BD model, discriminating late-onset BD (age ≥ 45) individuals and age-matched control individuals (age ≥ 45).

3. The late-onset MDD model, discriminating late-onset MDD (age ≥ 45) individuals and age-matched control individuals (age ≥ 45).

4. The late-onset SP model, discriminating late-onset SP (age ≥ 45) individuals and age-matched control individuals (age ≥ 45).

5. The early-onset BD model, discriminating early-onset BD (age < 45) individuals and age-matched control individuals (age < 45).

6. The early-onset MDD model, discriminating early-onset MDD (age < 45) individuals and age-matched control individuals (age < 45).

7. The early-onset SP model, discriminating early-onset SP (age < 45) individuals and age-matched control individuals (age < 45).

As a result, the whole profiles could be divided into 9 parts as follows:

1. normal young (age < 45, 520 training data + 250 test data);

2. normal young (45 ≤ age ≤ 50, 120 + 59);

3. normal old (age > 50, 270 + 124);

4. early-onset BD (age < 45, 80 + 48);

5. late-onset BD (age ≥ 45, 45 + 19);

6. early-onset MDD (age < 45, 160 + 80);

7. late-onset MDD (age ≥ 45,125 + 55);

8. early-onset SP (age < 45, 525 + 259);

9. late-onset SP(age ≥ 45,165 + 82).

Therefore, each one of these 9 parts was randomly split into a training and a test dataset, where the training dataset was used to model the classification model as well as 10-fold cross-validation, and the test dataset was used to evaluate the classification ability.

In this work, the proportion of each test dataset was set from 0.3 to 1/3, and the ratio of each training set was set from 2/3 to 0.7. The additional reason for the details of the number of each training dataset was divided by ten as far as possible, which was suitable for the 10-fold cross-validation; otherwise, the size was set as divided by five. To warrant that there were no overlaps between training and test datasets, each individual was used as training data or test data, but not both. We also re-checked the training set to ensure that it represented a wide spectrum of brain regions and blood.

Then, the ReliefF algorithm was used to identify key features. The first 500 models were studied for the training of predictors. The optimal model was selected by 10-fold cross-validation. Finally, to verify the accuracy of a predictor, the selected model was verified using the test dataset.

1. In establishing the aging model, the healthy individuals with an age of over 50 were associated with a label of 1; those healthy individuals with an age of ≤50 had a label of 0. In the three late-onset disease models, the late-onset patients (age ≥ 45) had a label of 1, and the control group (age ≥ 45) received a label of 0. In the three early-onset disease models, the late-onset patients (age < 45) had a label of 1, and the control group (age < 45) received a label of 0.

   The reason for the threshold (the 50-year mark) in the aging model was as follows: First, the maximum age in the public dataset (i.e., GEO) was close to 100 years (in this work was 96 years). As a result, the 50-year-old was used to split normal young and old with a minimum hypothesis(without other additional assumptions). Second, we also try other thresholds (e.g., 20, 30, 40, 60, and 70). The results showed that the thresholds from 40 to 70 were with enough accuracy (Supplementary Data 7). However, the imbalance between young and old will be accelerated if the threshold is too large or too small. Last, there is a series of recent research indicating the 50-year (from 40 to 60) as a significant threshold for aging^29,30. In sum, the age of 50 could be used as the threshold for the normal aging model.

2. The 13,614 CpG sites (ignoring any CpG sites with missing values) were ranked by the ReliefF algorithm.

3. To reduce bias associated with sample imbalance, the large group (healthy younger individuals in the aging model, or the control group in each disease model) was divided into a series of subgroups, although the desire to balance the number of samples in the subgroups necessitated overlaps between the subgroups based on the age of the individuals.

4. In the aging model, the healthy younger individuals were divided into three subgroups: 0–32 years (327 samples), 24–41 years (323 samples), and 32–50 years (335 samples).

For late-onset disease models, such as the BD model, the control group was divided into 9 subgroups: 45–48 years of age (66 samples), 46–49 years (68 samples), 49–52 years (79 samples), 51–55 years (72 samples), 55–58 years (63 samples), 57–62 (65 samples), 60–67 years (67 samples), 64–76 years (69 samples), and 68 years to maximum survival (68 samples). For the late-onset SP model, the age ranges between the 3 subdivided control groups were: 45–55 years (186 samples), 51–63 years (177 samples), and 57 years to survival (180 samples). The three age groups of control individuals in the late-onset MDD model were: 45–53 years (152 samples), 51–62 years (165 samples), and 58 years to survival (162 samples).

For early-onset disease models, such as the BD model, the control group was divided into 15 subgroups: 16–19 years (40 samples), 19–20 years (44 samples), 21 years (32 samples), 22–23 years (48 samples), 24–25 years (39 samples), 26–27 years (37 samples), 28–29 years (36 samples), 30 years (26 samples), 31–32 years (45 samples), 33–34 years (40 samples), 35–36 years (42 samples), 37–39 years (31 samples), 40–41 years (27 samples), 42–43 years (39 samples), 43–44 years (33 samples). For the late-onset MDD model, the age ranges between the 3 subdivided control groups were: 16–25 years (183 samples), 26–33 years (164 samples), and 34–44 years (173 samples).

The ReliefF algorithm was run before and after the subdivision of the major class during the creation of the aging model. The predictor was created using kNN (k = 5, cosine distance). The optimal model was identified by 10-fold cross-validation, selecting the model with the highest accuracy.

(5) The identified features were considered aging or disease markers, respectively (Supplementary Data 8–11).

Calculating the age score

The age score was used to compare the degree of aging between the disease group and the healthy control group. The score was calculated as follows:

1. The Pearson correlation coefficient was calculated between the normal training set and the chronological age based on biomarkers of aging.

   $$\rho =corr\left(age,aging\_mar\ker \right)$$ 1
2. We calculated the product of the methylation values of all samples with ρ, calculated their mean, and obtained the original aging score:

   $$origin\_score=\sum aging\_mar{\ker }_i*{\rho }_i$$ 2
3. To adjust the original score based on age, first we calculated the transformed age:

   $$transformed\_age=\frac{1}{1+\exp \left(-\frac{age-50}{50}\right)}$$ 3
4. The adjusted score was calculated based on the transformed age, where b is the regression coefficient obtained for an aging score:

   $$adjusted\_score=origin\_score-b*transformed\_age$$ 4
5. The Kruskal–Wallis tests(one-way) was performed on the adjust_score and phenotype (healthy (label:0) to disease (label:1)) to verify the aging acceleration in the three diseases.

   $$p=Kruskal-Wallis\_test\left(adjusted\_score,phenotype\right)$$ 5

   where non-parametric H-statistics were used to calculate the rank-sum based values as follows:

   $$H=\frac{12}{N*\left(N+1\right)}{\sum }_{i=1}^k\frac{R_i^2}{n_i}-3*\left(N+1\right)$$ 6

   where N was the number of groups, n_i was the number of individuals within the i-th group, R_i was the rank-sum value within the i-th group.

   Then, the H-statistics were used in the χ² test to calculate the p-value.

   In addition, the F-test (one-sided) was also performed based on the F-statistics:

   $$p=F-test\left(adjusted\_score,phenotype\right)$$ 7

   where the F-statistics were used to evaluate the differential pattern between disease and control groups:

   $$F=\frac{{{\mathrm{s}}_1}^2}{{{\mathrm{s}}_2}^2}$$ 8

   where S₁² and S₂² were the sample standard deviation for disease and control, respectively.

6. The disease score could also be calculated by summarizing the score within each subgroup (see “Data and preprocessing” section: each subgroup in the control group combined with the disease group) based on the disease markers:

$$\mathrm{score}=\mathrm{mean}\left(\sum \mathrm{corr}\left(\mathrm{phenotype},\mathrm{disease}\_{\mathrm{marker}}_{\mathrm{i}}\right)*\mathrm{disease}\_{\mathrm{marker}}_{\mathrm{i}}\right)$$ 9

The reason for the difference in the detail of the correlation for the disease score different from the aging score is that aging is a progressive process where the correlation should be related to the chronological age, but the disease status has a binary outcome (i.e., disease versus control) in this work, which is equivalent to the point-biserial correlation.

The data flow of the mapping model

As shown in Fig. 1C, the mapping models were designed to systematically investigate potential regulatory (mapping) relationships between different genetic features (i.e., DNA methylation CpG sites), concentrating on the relationship between markers of aging and disease (i.e., BD, SP, or MDD).

The overarching principle of the mapping model was to generate a series of “pseudo” samples (containing the combination of aging and disease markers) from random column vectors (10 columns, using uniform distribution where the range was from −1 to 1), containing two steps as follows:

1. Preliminary mapping step, from the random values to pseudo-samples, by using the mapping/weighting network (defined as the weights from the random column vector to the “pseudo” samples/matrix).

   Assuming that there were N fake samples to be generated, including m markers (including the union of aging markers and disease markers), the pseudo matrix was F_N*m (m was the number of CpG sites in this work). Meanwhile, the matrix of the original random value was R_N*10. As a result, the pseudo matrix could be generated/optimized from the original random value multiplied by the relative weighting matrix W_10*m: F_N*m = R_N*10*W_10*m.

2. Extracting principal components step, where the first 10 principal components were extracted from the “pseudo” CpG matrix (samples).

   Obtaining the coefficient matrix, C_m*10, of the first 10 principal components allowed the calculation of the principal components as P_N*10 = F_N*m*C_m*10.

3. Re-mapping step, mapping to the pseudo matrix from the principal components. In this step, the pseudo matrix could be recalculated by these 10 principal components: F_N*m = P_N*10*Q_10*m, where Q is the new weighting matrix related to P (the first 10 principal components).

4. Relationship summary using the interaction matrix I_m*m, based on the coefficient matrix C of the first 10 principal components and its weighting matrix Q: I_m*m = C_m*10*Q_10*m. In addition, potential relationships between aging and disease could be investigated: if a row in I_m*m was related to an aging marker and a column was related to a disease marker, a potential relationship “from aging to disease” was presented; or “from disease to aging” could also be presented vice versa.

In short, the “pseudo” matrix was generated as follows:

1. Preliminary mapping from random columns (i.e., F_N*m = R_N*10*W_10*m);

2. The first 10 principal components were extracted (P_N*10 = F_N*m*C_m*10);

3. The 10 principal components were remapped as (F_N*m = P_N*10*Q_10*m);

4. Multiplying the coefficient matrix of the 10 principal components and the relative weighting matrix (I_m*m = C_m*10*Q_10*m) resulted in the Interaction Matrix I_m*m between different markers.

The weighting matrix W was related to the original random values, and the weighting matrix of Q could be optimized by iteratively adjusting pseudo-samples as follows:

During the “preliminarily generating/mapping” step, W was optimized based on the adjusted pseudo-samples and the original random values. The 5-nearest actual control samples were selected based on the first 10 principal components of the aging markers. The 5-nearest actual samples were also selected based on the first 10 principal components of the disease markers (i.e., BD, SP, or MDD markers). Next, 5 out of the 10 nearest samples were randomly selected using the kNN search method (with the cosine distance). At this point, the mean values of the randomly selected samples were calculated as the object sample to be optimized for each of the “pseudo” samples, respectively. As a result, W was optimized (from random values to the object samples).

During the “re-generating/re-mapping” step, the generated samples were adjusted by comparison with k-nearest actual healthy control samples, actual disease samples, and other generated samples, where the value of k was set as 5. In addition, the 5-nearest samples (from actual healthy control, actual disease patients, or other generated samples) were randomly selected from 10 nearest samples using the kNN search method (with the cosine distance); then, the mean values of the randomly selected samples were calculated as the object sample to be optimized for each of the “pseudo” samples. This process optimized Q derived from the first 10 principal components for the object samples.

Versions of the mapping model

As shown in Supplementary Fig. 1, there were five versions of “pseudo” samples:

1. “aging-disease−”: no accelerated aging in healthy control (m = 2192);

2. “aging+disease−”: accelerated aging in healthy control (m = 2192);

3. “aging+BD+”: accelerated aging in BD patients (m = 1104);

4. “aging+SP+”: accelerated aging in SP patients (m = 895);

5. “aging+MDD+”: accelerated aging in MDD patients (m = 1300).

A series of classification models was used to evaluate these “pseudo” samples:

1. The real disease predictors (i.e., BD, SP, and MDD, explained in “Modeling the aging as well as each disease predictor” section) were used to discriminate the phenotype of the “pseudo” samples,

2. Both aging and disease scores were used to find the nearest real samples to each “pseudo” sample,

3. A simulated predictor (based on the combination of aging and disease markers) was also used to evaluate the generative power of each variety of “pseudo” samples.

In the “aging-disease−” group, the following approaches were used to evaluate the “pseudo” samples:

1. Testing the ability of real disease predictors to classify the sample into the normal group, and

2. Using simulated predictors to classify the pseudo sample into the actual sample group.

During the “re-mapping” step, the “pseudo” samples were adjusted/optimized by a series of methods:

1. proximity to the real “aging-disease-” samples, where 5 out of 10 nearest real samples were randomly selected and used based on the adjusted aging score, which was subtracted from the minimum value in the training individuals of the aging model. As a result, the value of the processed aging score was from 0 to 1 in this work. The processed aging score was then used to randomly select based on a uniform distribution probability as formula (10), and then as the weight to summarize the target patterns as formula (11) as follows:

   $$p_i=\frac{1-aging\_scor{e}_i}{{\sum }_{k=1}^{10}\left(1-aging\_scor{e}_k\right)}$$ 10

   $$weigh{t}_i=\frac{1-aging\_scor{e}_i}{{\sum }_{k=1}^5\left(1-aging\_scor{e}_k\right)}$$ 11

   The reason the aging score was subtracted by 1 was that real samples with a lower aging score were more likely to be selected in the “aging-” model. Both the probability in formula (10) and the weight in formula (11) were normalized by calculating the summarized value.

   In addition, the reason for the linear proportional weight was as follows:

   First, the distribution of the adjusted aging score was concentrated in a special range in this work, from −0.03291 to 0.02865 (shown in Supplementary Fig. 3, also as from 0 to 0.06156 after subtracting the minimum value). Therefore, there was no necessity to transform into other complex weighting strategies (i.e., log(x) or e^x).

   Second, the “aging−” model tends to select a lower aging score, and “aging+” models tend to select a higher aging score. As a result, the simplest linear proportional scheme was used in this work, only considering the aging score without any additional non-linear assumptions.

   Last, these mapping models adjusted pseudo-samples with the help of probability as well as weighting schemes by searching and evaluating various nearest related real samples, but not restricted to limited samples. It might lead to uncertain results because of the random scheme. However, a series of related real samples with matching aging score could be considered, without restricting to limited searching results by additional non-linear transformations(for example, avoiding/alleviating the “Mode collapse” similar to relative generative models).

2. Deviation from actual “aging+disease−” samples, where 5 out of 10 nearest real samples were randomly selected and used.

3. Deviation from actual “aging+disease+” (including “aging+BD+”, “aging+SP+”, and “aging+MDD+”) samples, where 5 out of 10 nearest real samples were randomly selected.

4. Deviation from other synthetic “aging-disease−” samples, selecting 5 out of 10 nearest generated samples.

To summarize, the synthetic samples were adjusted as follows:

$$\begin{array}{c}\mathrm{fake}\mathrm{sample}\leftarrow \mathrm{fake}\mathrm{sample}\hfill \\ +\mathrm{k}*\left(near{R}_{aging-disease-}-fakesample\right)\hfill \\ +\frac{d_{aging-disease-}}{d_{aging-disease-}+d_{aging+disease-}}*\left(near{R}_{aging-disease-}-near{R}_{aging-disease-}\right)\hfill \\ +\frac{d_{aging-disease-}}{d_{aging-disease-}+d_{aging+\mathrm{disease}+}}*\left(near{R}_{aging-disease-}-near{R}_{aging+disease+}\right)\hfill \\ +\frac{d_{aging-disease-}}{d_{aging-disease-}+d{F}_{aging-disease-}}*\left(near{R}_{aging-disease-}-near{F}_{aging-disease-}\right)\hfill \end{array}$$ 12

where k was the error rate of the synthetic samples based on the simulated predictors, using leave-one-out cross-validation in the training data.

The “aging+disease−” synthetic samples were evaluated as follows:

1. Would actual disease predictors classify them into the healthy control group and

2. Would simulated predictors classify them as actual samples.

During the “re-mapping” step, synthetic samples were adjusted/optimized according to:

1. Their proximity to actual “aging+disease-” samples, selecting 5 out of 10 nearest actual randomly selected samples. The adjusted aging score was used to randomly select based on a uniform distribution probability as formula (13), and then as the weight to summarize the target patterns as formula (14).

   $$p_i=\frac{aging\_scor{e}_i}{{\sum }_{k=1}^{10}aging\_scor{e}_k}$$ 13

   $$weigh{t}_i=\frac{aging\_scor{e}_i}{{\sum }_{k=1}^{5}aging\_scor{e}_k}$$ 14

   In other words, the real samples with greater aging scores were more likely to be selected in “aging+” models. Both the probability in formula (13) and the weight in formula (14) were also normalized by calculating the summarized value.

2. Deviation from actual “aging-disease-” samples, using 5 out of 10 nearest actual samples.

3. Deviation from actual “aging+disease + ” (including “aging+BD+”, “aging+SP+”, and “aging+MDD+”) samples, again, by selecting 5 out of 10 nearest actual samples.

4. Deviation from other synthetic “aging+disease-” samples, using 5 out of 10 randomly selected nearest synthetic samples.

During this step, synthetic samples in this group were adjusted as follows:

$$\begin{array}{c}\mathrm{fake}\mathrm{sample}\leftarrow \mathrm{fake}\mathrm{sample}\hfill \\ +\mathrm{k}*\left(near{R}_{aging+disease-}-fakesample\right)\hfill \\ +\frac{d_{aging+disease-}}{d_{aging+disease-}+d_{aging-disease-}}*\left(near{R}_{aging+disease-}-near{R}_{aging-disease-}\right)\hfill \\ +\frac{d_{aging+disease-}}{d_{aging+disease-}+d_{aging+\mathrm{disease}+}}*\left(near{R}_{aging+disease-}-near{R}_{aging+disease+}\right)\hfill \\ +\frac{d_{aging+disease-}}{d_{aging+disease-}+d{\mathrm{F}}_{aging+disease-}}*\left(near{R}_{aging+disease-}-near{F}_{aging+disease-}\right)\hfill \end{array}$$ 15

where k was the error rate of synthetic samples based on simulated predictors, using leave-one-out cross-validation of the training data.

The “aging+BD + ” synthetic samples were evaluated based on:

1. whether actual BD predictors would classify them into the BD group and

2. if simulated predictors would classify them into the actual sample group.

To achieve this aim, during the “re-mapping” step, synthetic samples were adjusted/optimized according to their:

1. proximity to randomly selected actual “aging+BD+” samples, using 5 out of 10 nearest actual samples. The adjusted aging score was used to randomly select based on a uniform distribution probability as formula (13), and then as the weight to summarize the target patterns as formula (14).

2. deviation from actual “aging-disease-” samples, using 5 out of 10 random nearest actual samples,

3. deviation from actual “aging+disease-” samples, using 5 out of 10 random nearest actual samples,

4. deviation from other synthetic “aging+BD+” samples, using 5 randomly selected synthetic samples out of the nearest 10.

The summary of this adjustment process was as follows:

$$\begin{array}{c}\mathrm{fake}\mathrm{sample}\leftarrow \mathrm{fake}\mathrm{sample}\hfill \\ +k*\left(near{R}_{aging+BD+}-fakesample\right)\hfill \\ +\frac{d_{aging+\mathrm{BD}+}}{d_{aging+BD+}+d_{aging-disease-}}*\left(near{R}_{aging+BD+}-near{R}_{aging-disease-}\right)\hfill \\ +\frac{d_{aging+BD+}}{d_{aging+BD+}+d_{aging+disease-}}*\left(near{R}_{aging+BD+}-near{R}_{aging+disease-}\right)\hfill \\ +\frac{d_{aging+BD+}}{d_{aging+BD+}+d{\mathrm{F}}_{aging+BD+}}*\left(near{R}_{aging+BD+}-near{F}_{aging+BD+}\right)\hfill \end{array}$$ 16

where k was the error rate of synthetic samples based on the simulated predictors, using leave-one-out cross-validation of the training data.

Similarly, the “aging+SP+” synthetic sample set was evaluated based on:

1. whether a real SP predictor would classify them into the SP group and

2. if a simulated predictor classified them as actual samples.

Similar to the process described above, during the “re-mapping” step, these synthetic samples were adjusted/optimized based on:

1. Their proximity to actual “aging+SP + ” samples, where 5 randomly selected actual samples out of 10 were used. The adjusted aging score was used to randomly select based on a uniform distribution probability as formula (13), and then as the weight to summarize the target patterns as formula (14).

2. Deviation from actual “aging-disease-” samples, again using 5 out of 10 nearest actual samples.

3. Deviation from actual “aging+disease−” samples, using 5 out of 10 nearest actual samples.

4. Deviation from other synthetic “aging+SP+” samples, using 5 out of 10 nearest synthetic samples.

These adjustments were completed as follows:

$$\begin{array}{c}\mathrm{fake}\mathrm{sample}\leftarrow \mathrm{fake}\mathrm{sample}\hfill \\ +k*\left(near{R}_{aging+SP+}-fakesample\right)\hfill \\ +\frac{d_{aging+\mathrm{SP}+}}{d_{aging+SP+}+d_{aging-disease-}}*\left(near{R}_{aging+SP+}-near{R}_{aging-disease-}\right)\hfill \\ +\frac{d_{aging+SP+}}{d_{aging+SP+}+d_{aging+disease-}}*\left(near{R}_{aging+SP+}-near{R}_{aging+disease-}\right)\hfill \\ +\frac{d_{aging+SP+}}{d_{aging+SP+}+d{\mathrm{F}}_{aging+SP+}}*\left(near{R}_{aging+SP+}-near{F}_{aging+SP+}\right)\hfill \end{array}$$ 17

where k was the error rate of synthetic samples based on simulated predictors, using leave-one-out cross-validation of the training data.

Finally, the “aging+MDD + ” synthetic samples were evaluated based on

1. whether a real MDD predictor would classify them into the MDD group and

2. if simulated predictors would classify them into the actual sample group.

To this end, during the “re-mapping” step, synthetic samples were adjusted/optimized according to:

1. Their proximity to actual “aging+MDD+” samples, using 5 out of 10 nearest randomly selected actual samples. The adjusted aging score was used to randomly select based on a uniform distribution probability as formula (13), and then as the weight to summarize the target patterns as formula (14).

2. Deviation from actual “aging-disease-” samples, using 5 out of 10 nearest actual samples.

3. Deviation from actual “aging+disease-” samples, using 5 out of 10 nearest actual samples.

4. Deviation from other synthetic “aging+MDD+” samples, compared to 5 out of 10 nearest synthetic, randomly selected samples.

The adjustment of this group of synthetic samples can be summarized as follows:

$$\begin{array}{c}\mathrm{fake}\mathrm{sample}\leftarrow \mathrm{fake}\mathrm{sample}\hfill \\ +k*\left(near{R}_{aging+MDD+}-fakesample\right)\hfill \\ +\frac{d_{aging+\mathrm{MDD}+}}{d_{aging+MDD+}+d_{aging-disease-}}*\left(near{R}_{aging+MDD+}-near{R}_{aging-disease-}\right)\hfill \\ +\frac{d_{aging+MDD+}}{d_{aging+MDD+}+d_{aging+disease-}}*\left(near{R}_{aging+MDD+}-near{R}_{aging+disease-}\right)\hfill \\ +\frac{d_{aging+MDD+}}{d_{aging+MDD+}+d{\mathrm{F}}_{aging+MDD+}}*\left(near{R}_{aging+MDD+}-near{F}_{aging+MDD+}\right)\hfill \end{array}$$ 18

where k was the error rate of synthetic samples based on simulated predictors, using leave-one-out cross-validation of training data.

The above iterative optimization process was repeated until the mapping network models converged, or the number of iterations exceeded 30. This allowed us to establish five interaction matrices. In addition, the mapping models of early-onset were also designed to refer to relatively late-onset models, where the adjusted aging score was used to calculate the weight to summarize the target patterns as formula (11) or (14), respectively.

Sensitivity analysis using the MCMC method (Supplementary Fig. 2)

To explore key relationships between features associated with late-onset psychosis and brain aging(in the context of an advanced aging brain), an overall sensitivity analysis was carried out based on MCMC. This method can be used for sampling from certain posterior distributions following a given probability background in high-dimensional space. The key step in MCMC is to construct a Markov chain whose equilibrium distribution is equal to the target probability distribution. The steps were conducted as follows:

1. First, a transition kernel of an ergodic Markov chain was constructed. In this study, based on the parameters of each mapping model (normal aging+, BD, SP, MDD), the prior distribution of each parameter was a normal distribution. As a result, 1000 groups of random parameters were generated using the normal distribution based on the parameters from the first to the 30th iteration, and then were further evaluated, including 1000 sets of Im*m = (Cm*10*Q10*m, in “5.4”) and Jm*m = (Cm*10*W10*m, in “5.4”).

2. Simulate the chain until it reaches equilibrium. The Metropolis-Hastings sampling method is used to determine whether the new sample (θ*) was acceptable based on the α value:

$$\alpha =\frac{P\left({\theta }^{*}\mid X\right)q\left({\theta }^n\to {\theta }^{*}\right)}{P\left({\theta }^n\mid X\right)q\left({\theta }^{*}\to {\theta }^n\right)}$$ 19

where P(θⁿ | X) and P(θ^*|X) are the posterior probabilities of the nth accepted sample and the new sample, q(θⁿ → θ^*) is the transition probability from the nth accepted sample to the new sample, and q(θ^*→θⁿ) is the transition probability from the new sample to the nth accepted sample.

In this work, the aging score and disease score were used to evaluate the simulated samples, and 1000 groups of random parameters were generated as conditional samples. Then, the random parameters were selected as follows:

1. The first group of parameters was retained.

2. For the next group of parameters, if the score was larger than the previous group of parameters, they were retained;

The “aging+disease+” group (i.e., BD, MDD, or SP) was evaluated as follows:

$$\begin{array}{c}\mathrm{ne}xt\_aging\_score+next\_disease\_score>\hfill \\ previous\_aging\_score+previous\_disease\_score\hfill \end{array}$$ 20

The “aging+disease−” group was evaluated as follows:

$$\begin{array}{c}next\_aging\_score+1-next\_disease\_score\hfill \\ >previous\_aging\_score+1-previous\_disease\_score\hfill \end{array}$$ 21

otherwise, they were retained by a 0−1 uniform distribution probability.

$$\frac{next\_score}{previous\_score}>random\_value$$ 22

where, for the “aging+disease+” group, was calculated as follows:

$$\mathrm{ne}xt\_sccore=next\_aging\_score+next\_disease\_score$$ 23

$$previous\_score=previous\_aging\_score+previous\_disease\_score$$ 24

and, for the “aging+disease−” group, was calculated as follows:

$$\mathrm{ne}xt\_sccore=next\_aging\_score+1-next\_disease\_score$$ 25

$$previous\_score=previous\_aging\_score+1-previous\_disease\_score$$ 26

Both original aging score and disease score were transformed using the sigmoid function:

$$score=1/\left(1+\exp \left(-score\right)\right)$$ 27

(3) The first 20th retained groups of parameters were deleted.

(3) Next, a global sensitivity analysis was performed. The K-S statistic was used to evaluate the differential/sensitivity index of each generated parameter (i.e., “aging+disease−” was compared with each of those in the disease subtypes: “aging+BD+”, “aging+SP+” and “aging+MDD+”, respectively):

$$K-S=\sup \mid F_1-F_2\mid$$ 28

where F₁ was the cumulative distribution of samples in a group (i.e., “aging+disease+”), and F₂ was the cumulative distribution of samples in another group (i.e., “aging+disease-”)

In this work, the interaction matrix I_m*m = (C_m*10*Q_10*m, in “5.4”) is used. Then, the statistic of another matrix “J_m*m” = (C_m*10*W_10*m, in “5.4”) was used as a background relationship. The sensitivity relationship was selected as follows:

① K-S statistic in I_m*m > 0.25 between “aging+disease+” and “aging+disease−”.

② K-S statistic in I_m*m − J_m*m > 0.25 between “aging+disease+” and “aging+disease−”.

③ K-S statistic of the differences between I_m*m and J_m*m (I_m*m − J_m*m) > 0.25 of “aging+disease-”.

Drug-gene interaction analysis

We obtained GSE119291 as the drug distribution profiles, where samples labeled with “schizophrenia” or “control” were used to summarize the difference of the effects between each drug and the DMSO-treated experiment, respectively. In this work, the difference of the similarity (from “schizophrenia” to “control”) between each drug and DMSO was investigated, based on the sensitive “aging-disease” pairs for each disease type:

$$\mathrm{p}=\mathrm{Kruskal}-\mathrm{Wallis}\left(\mathrm{corr}\left(\mathrm{aging},\mathrm{disease}\right){\mathrm{drug}}^{\mathrm{disease}}.*\mathrm{corr}\left(\mathrm{aging},\mathrm{disease}\right){\mathrm{drug}}^{\mathrm{normal}},\mathrm{corr}\left(\mathrm{aging},\mathrm{disease}\right){\mathrm{DMSO}}^{\mathrm{disease}}.*\mathrm{corr}\left(\mathrm{aging},\mathrm{disease}\right){\mathrm{DMSO}}^{\mathrm{normal}}\right)$$ 29

Then the Benjamini–Hochberg FDR < 0.05 was considered as significantly disturbed, coefficient values in a disease had the opposite sign (i.e., + or −) to that in the control were highlighted.

Further, the DGIdb platform (https://www.dgidb.org/) was used to identify potential drugs related to genes in sensitive “aging-disease” pairs for each disease type. Then, the drug-sensitive scores were calculated using the oncoPredict software (https://github.com/maese005/oncoPredict). As a result, the difference of the similarity (from each disease individual to normal) between each drug and control was investigated, based on the sensitive “aging-disease” pairs for each disease type:

$$\mathrm{p}=\mathrm{Kruskal}-\mathrm{Wallis}\left(\mathrm{corr}\left(\mathrm{aging}.*\mathrm{disease},\mathrm{drug}\_\mathrm{sensitive}\right){\mathrm{drug}}^{\mathrm{disease}}.*\mathrm{corr}\left(\mathrm{aging}.*\mathrm{disease},\mathrm{drug}\_\mathrm{sensitive}\right){\mathrm{drug}}^{\mathrm{normal}},\mathrm{corr}\left(\mathrm{aging},\mathrm{disease}\right){\mathrm{null}}^{\mathrm{disease}}.*\mathrm{corr}\left(\mathrm{aging},\mathrm{disease}\right){\mathrm{null}}^{\mathrm{normal}}\right)$$ 30

Then the Benjamini–Hochberg FDR < 0.05 was considered as significantly changed, coefficient values in a disease had the opposite sign (i.e., + or −) to that in the control were highlighted.

Constructing the differential network

To reveal the relationship between features in aging and late-onset psychosis, a differential DNA methylation network was constructed using the following steps:

1. To compare the relationship between each gene pair in the context of aging, the Pearson correlation coefficient and the partial correlation coefficient based on the aging score were calculated, comparing all three diseases and the control groups.

2. The Benjamini–Hochberg FDR method was used to adjust the p-values of the correlation coefficient and the partial correlation coefficient.

3. If the Benjamini–Hochberg FDR was <0.05 within a group, the edges between each gene pair were retained. Then we performed an XOR operation on the edges of the disease group and the control group.

4. The differences between the correlation coefficient and the partial correlation coefficient within the group were calculated. If the signs of the difference between the normal group and the disease group were opposite, and the requirements of (3) were met, the edges between each gene pair were retained ultimately.

5. The scale-free characteristics of the differential networks were verified by the power-law distribution (Supplementary Fig. 4).

6. The shortest path between each pair of aging and disease markers was selected based on each differential network using the Dijkstra algorithm.

7. The established network was used in further analyses (identifying shortest paths, exploring biological functions between aging and disease).

The differential co-expression network was constructed as follows:

1. The Pearson correlation coefficient for each pair of genes was calculated based on each disease type and the control group, respectively.

2. The Benjamini−Hochberg FDR method was used to adjust the p-values of the correlation coefficients.

3. If the Benjamini–Hochberg FDR < 0.05 within the group, the edges between each gene pair were retained. Then we performed an XOR operation on the edges of the disease group and the control group.

4. The relationship between each gene pair was retained if the coefficient value in the disease had the opposite sign (i.e., + or −) to that in the control.

5. Each pair of aging and disease markers identified based on the DNA methylation profiles was also used to calculate the shortest path, if both of them had gene symbols. In such cases, relative functional analyses were also performed.

Enrichment analysis

The biological function of genes was explored by enrichment analysis of the shortest pathway. GO terms and KEGG pathways for the GSEA platform were downloaded from http://software.broadinstitute.org/gsea/downloads.jsp (version 2023.1). The hypergeometric distribution was used to test the degree of enrichment of the GO, BP, and KEGG pathways, using the following formula:

$$P\left(X\ge x\right)=1-{\sum }_{k=0}^{x-1}\frac{C_M^k-C_{N-M}^{n-k}}{C_M^k}$$ 31

where N is the total number of genes in the gene set, M is the number of known genes (corresponding to KEGG pathway or BP terms), which is the number of genes identified in each shortest pathway, and k is the number of common genes between known genes and candidate genes identified in each “aging-disease” shortest pathway. The p-value of each path was controlled using the Benjamini–Hochberg method. Only pathways with the Benjamini–Hochberg FDR < 0.05 were retained in the final analysis. The enrichment score based on FDR was calculated as follows:

$$\mathrm{s}core={\sum }_{FDR<0.1}\left(1-FD{R}_i\right)$$ 32

Identifying network markers

Subnetwork with the shortest pathways among the selected “aging-disease” pairs was constructed, and genes in the subnetwork were sorted according to their betweenness in descending order. To test whether the highest-ranked genes were hubs in the background network, we ran a permutation to count the number of times the top-ranked genes occurred in the shortest paths between randomly selected genes. These tests were run containing the same number of “aging-disease” pairs: 30,036 for BD, 36,825 for SP, and 22,970 for MDD when analyzing DNA methylation profiles, and 11,821 for BD, 22,705 for SP, and 14,051 for MDD when comparing gene expression profiles. We repeated this process 1000 times, and the p-value was calculated as the proportion of times a gene occurred as a top-ranking betweenness gene in 1000 permutations.

For example, the real betweenness of cg21870884 (GPR25) was 21, from aging to BD markers, where there were 16,108 sensitive pairs based on the DNA methylation profiles. Therefore, the permutation test was performed 1000 times, where 16,108 random pairs were selected for each time, and then the simulated betweenness was calculated based on these 16,108 random pairs for each of the 1000 times, respectively. As a result, there are 2 out of 1000 times with betweenness >21, then the p-value of the permutation test was 2/1000 = 0.002.

Estimate the false discovery rate for multiple hypothesis testing

The classical Benjamini–Hochberg False Discovery Rate method was used in this work³¹. Suppose there were a series of p-values (i.e., 24,313). The FDR values were calculated as follows:

(1) The p-values were sorted from lowest to highest.

(2) For each p-value, the FDR could be adjusted as:

$$FD{R}_i=p_i*m/i$$ 33

where i was the rank of the p-value in the total p-values, and m was the number of the p-values (i.e., 24,313).

(3) In order to simplify, the FDR values were used instead of the original p-values, or as adjusted p-values.

Ethics approval and consent to participate

This study analyzes publicly available datasets; therefore, it has been approved by the Ethics Committee of China Medical University. We believe that related human research participants, material, or data in this work are in accordance with the Declaration of Helsinki.

We can confirm that the cited (original) study has IRB approval, participant consent, and permission for secondary use.

We can also confirm that the current study under consideration with us has IRB approval.

Results

Classification models and identifying relevant biomarkers

DNA methylation profiles were obtained from the Gene Expression Omnibus (GEO) database and included 1834 samples and 24,313 CpG sites (Supplementary Data 1–3). These CpG sites were ranked using the ReliefF algorithm, and subsequently, an aging model and three disease-specific models were built using the k-nearest neighbors (kNN; k = 5 with the cosine distance) algorithm, optimized by the 10-fold cross-validation (Supplementary Data 8–11). Summarizing the specificity and the sensitivity of the ReliefF ranking results (i.e., the top-ranked CpG site, the first two CpG sites, or the first three CpG sites… the first 500 CpG sites, shown in Fig. 2A, E, I, M), in the test set, the accuracy of predictors for aging, BD, SP, and MDD were 0.7027, 0.9208, 0.7045, and 0.7525, respectively (Supplementary Data 12 and 13). The corresponding area under the curve (AUCs) values for predictors were 0.63971 (aging), 0.82641 (BD), 0.64829 (SP), and 0.68097 (MDD) (Fig. 2B, F, J, N). The corresponding AUCs of the receiver operating characteristic (ROC) curves were 0.77108, 0.81788, 0.75134, and 0.98058, based on the aging and disease scores (Fig. 2C, G, K, O). The average precision(AP) of the precision and recall (PR) curves was 0.90019, 0.9855, 0.88641, and 0.9941 (Fig. 2D, H, L, P). Overall, these results showed that our predictors worked with sufficient accuracy. In addition, the classical epigenetic aging markers were also investigated based on the profiles used in this work³². However, it could not present enough accuracy in either age group classification or age regression (Supplementary Data 14). It was perhaps that the relative profiles used in this work were specialized in brain aging, where related markers needed to be identified. In brief, both aging and disease (i.e., BD, MDD, and SP) were identified by relative classification models, where the aging markers were further used to summarize into quantified scores³³ (further shown in “Exploring key relationships between aging and psychoses by sensitivity analysis” and “Calculating the age score” sections).

Fig. 2. Machine learning results for the classification model.

A, D, G, J Learning curves for the training dataset. B, E, H, K Sensitivity and specificity (similar to receiver operating characteristic) curves. C, F, I, L The ROC curve for the test dataset. D, I, L, P The PR curve for the test dataset. (A–D) correspond to the aging model, (E–H) to the BD model, (I–L) show the SP model, and (M–P) represent the MDD model.

Furthermore, the identified aging and disease markers had meaningful biological significance. For example, PLK3 (polo-like kinase 3, cg27287808, weight = 0.033823572), the top identified aging marker, is known to contribute to G1/S phase transition, and can induce apoptosis in a p53-dependent manner in response to replicative stress or genotoxic insults³⁴. PLK3 also phosphorylates the spindle-associated protein, SPAR, that affects synaptic plasticity in neuronal dendritic spines³⁴.

CLN8 (Ceroid-Lipofuscinosis, Neuronal 8, cg23833896, weight = 0.0628) was the top BD marker. The deficiency of this protein has been reported to impair lysosome biogenesis, disrupt cellular homeostasis, and cause abnormal dendritic cell development³⁵. Additionally, CLN8 could play a role in cell proliferation during neuronal differentiation and may protect against cell death³⁶.

As for SP, the top identified marker was ADAP1 (Putative MAPK-Activating Protein PM25, cg00969446, weight = 0.0175). This molecule reduces dendritic spine density and spine plasticity, thus affecting memory formation in the hippocampus³⁷. ADAP1 can also regulate neuronal polarity and axon specificity³⁸. In addition, this molecule may also have a role in T-cell signaling, although the details of the involved mechanism(s) remain to be understood³⁹.

SLC38A (Solute Carrier Family 38 Member 4, cg15584813, weight = 0.0734) was the top identified MDD marker. This molecule is an identified ATPase and a trans-regulator of mitochondrial ABC transporters. As such, it plays a crucial role in the maintenance of mitochondrial homeostasis and cell survival⁴⁰. Additionally, SLC38A4 also acts as an N-terminal acetyltransferase that is essential for autophagy, thus affecting cellular homeostasis⁴¹.

Comorbidity mechanisms were also revealed based on the classification models, including 173, 84, and 26 markers shared between each early- and late-onset psychosis, respectively. For example, SHOX2 (Short Stature Homeobox 2, cg06156376) was identified as a top shared marker in BD. Previous evidence suggests that thalamic abnormalities represent a shared neurodevelopmental vulnerability in psychotic disorders⁴². It could be indicated that epigenetic alterations of SHOX2 in this developmental regulator may link early neurodevelopmental deficits to the pathophysiology of late-onset BD. LCN1 (Lipocalin 1, cg23815000) was identified as a top shared marker in SP. As a lipocalin family member, LCN1 functions as a scavenger protein involved in oxidative stress regulation and is essential for olfactory signal processing⁴³. Given that olfactory dysfunction is a well-documented prodromal feature of schizophrenia, the identification of LCN1 methylation highlights a potential molecular link between early sensory deficits and the chronic neuroinflammation driving accelerated brain aging⁴⁴. TOP1MT (DNA Topoisomerase I Mitochondrial, cg12188860) was identified as a top shared marker in MDD. As the sole topoisomerase dedicated to the mitochondria, TOP1MT is essential for resolving DNA topological stress during mtDNA replication and transcription, thereby maintaining mitochondrial genome stability and cellular energy supply⁴⁵. Given that mitochondrial dysfunction and compromised bioenergetics are central to the pathophysiology of depression⁴⁶, the identification of TOP1MT methylation suggests a conserved mechanism, where epigenetic regulation of mitochondrial homeostasis underpins vulnerability to MDD across the lifespan.

In short, the top ReliefF algorithm identified potential biomarkers that have biologically relevant roles in cellular homeostasis, neural plasticity, affecting synaptic and spinal plasticity, and influencing immune cell function. Therefore, finding these molecules, which are closely related to the mechanisms involved in late-onset psychosis in the context of brain aging, boosted our confidence in the results.

In order to further investigate the advanced aging pattern in late-onset psychoses, we calculated an aging score by integrating 498 known aging markers (see “Calculating the age score” section). Since the healthy control group also contained younger individuals, inevitably, the median and mean values were higher in the samples from individuals suffering from the tested diseases (Supplementary Data 15). Therefore, both the Kruskal–Wallis test and the F-test were used to verify whether the detected difference in aging scores reflected a considerable difference between diseased individuals and controls, where relative H- and F-statistics were also included. In addition, the aging score was adjusted for the chronological age (see “Calculating the age score” section for details). The results from this analysis indicated the advanced aging pattern in late-onset psychoses compared to controls (Supplementary Fig. 5). Reassuringly, a similar observation was reported previously¹⁰. The calculated disease scores are shown in Supplementary Figs. 6–9 and Supplementary Data 16. These results further emphasized the accelerated aging pattern in late-onset psychosis patients, where key interactive relationships between aging and disease could be explored based on relative individuals with a greater adjusted aging score.

Exploring key relationships between aging and psychoses by sensitivity analysis

A series of mapping models was built to explore the potential relationship between aging and the investigated psychoses (see “The data flow of the mapping model” and “Versions of the mapping model” sections). In brief, different pseudo-samples were generated from random values, with the help of relative mapping models by using a series of heuristic rules (formulas). In addition, potential relationship between aging and disease markers could be summarized based on the interaction matrix I_m*m (see “The data flow of the mapping model” and “Versions of the mapping model” sections, where m was the number of union of aging and disease markers, as 1104, 895 and 1300 for BD, SP, and MDD, respectively): if the row was related to aging marker and the column was related to disease marker, then the pair is identified as “aging to disease” pair; and the “disease-aging” pair is also identified vice versa. As a result, the generated “synthetic” samples demonstrated sufficient accuracy (Supplementary Data 17). The relative variation patterns of the pseudo-samples were also shown in Supplementary Fig. 10, and the accuracy of the 30-round iterative optimization process for the pseudo-samples was shown in Supplementary Data 18. These results indicated that the mapping model could present pseudo-samples with enough accuracy, where most of the mapping models could present convergence. For example, the accuracies for “aging-disease−”, “aging+disease−”, and “aging+SP+” could present convergence. However, the results of “aging+BD+” and “aging+MDD+” could not present convergence, but still had enough accuracy after 30 rounds of optimization. It is perhaps the optimization process randomly selects the nearest real samples with relative probabilities (see “Versions of the mapping model” section), which might present random results if the number of real samples was not large enough (i.e., BD or MDD). As a result, the follow-up sensitive analysis was critical to these models.

Next, these mapping models were integrated (see “Sensitivity analysis using the MCMC method” section), and the Markov chain Monte Carlo (MCMC) method was used to identify key relationships between aging and specific diseases. In other words, key pairs in the interaction matrix (I_1104*1104, I_895*895, and I_1300*1300 for BD, SP, and MDD, respectively) were further identified by MCMC, including retained parameter groups between each “aging+disease+” and “aging+disease−” were compared using the K-S statistic (>0.25, in “Sensitivity analysis using the MCMC method” section). As a result, 30,036 “aging-BD” (16,108 aging to BD and 13,928 BD to aging), 36,825 “aging-SP” (17,270 aging to SP and 19,555 SP to aging), and 22,970 “aging-MDD” (11,111 aging to MDD and 11,859 MDD to aging) pairs were identified, where the sensitive pair was classified as aging to disease if row was related to aging and column was related disease in I_m*m, and classified as disease to aging vice versa. Additionally, the convergence properties were shown in Fig. 3, indicating the convergence by sensitive analysis.

Fig. 3. The convergence properties in the sensitivity analysis.

A–I Based on accuracy. J–R Based on absolute variation of accuracy. A, J In “aging+BD+” model; D, M In “aging+SP+” model; G, P In “aging+MDD+” model; B, E, H, K, N, Q In “aging+disease−” model; C, F, I, L, O, R In “aging-disease−” model; A–C, J–L Based on the BD predictor; D–F, M–O Based on the SP predictor; G–I, P–R Based on the MDD predictor.

The top aging-related sensitive markers are shown in Supplementary Data 19–21. These tables highlight markers with direct relevance to each disease’s specific or shared pathophysiology, along with their directional relationships (“aging to disease” and “disease to aging”). For example, in “aging to disease” relationships in BD, the top aging marker was CAMK1D (Calcium/Calmodulin Dependent Protein Kinase ID, cg24859228). This kinase is a key component of calcium signaling pathways, which are crucial for neuronal activity, synaptic plasticity, and immune cell activation^47–49. Notably, CAMK1D was highlighted as a shared aging marker in the ‘aging to disease’ direction across all three psychoses, indicating vulnerability that links brain aging to these disorders. In the same direction, a top disease marker was INSL6 (Insulin Like 6, cg11830061), a peptide with neuroprotective potential that can inhibit T-cell proliferation and reduce inflammatory cytokines^50,51. In “disease to aging”, SFRS7 (Serine and Arginine Rich Splicing Factor 7, cg18008766) was a top aging marker, known to directly regulate the alternative splicing of tau (MAPT), a key process implicated in neurodegeneration⁵². A key disease marker in this context was PCDH12 (Protocadherin 12, cg12845808), a cell adhesion molecule whose genetic variants are associated with intermediate phenotypes for psychotic disorders and which plays a pivotal role in neuronal migration and circuit formation^53–55.

In “aging to SP”, a top shared aging marker was LRRC40 (Leucine Rich Repeat Containing 40, cg16666456), a candidate gene for Autism Spectrum Disorder implicated in neural development and synaptic function⁵⁶. Its identification as a top “aging to disease” marker, alongside CAMK1D, in all three conditions reinforces the concept of a shared neurodevelopmental vulnerability. The top disease marker was EDARADD (Ectodysplasin A Receptor Associated Death Domain, cg09809672), an essential adaptor protein for activating the NF-κB pathway, a master regulator of inflammation and immune responses^57,58. In “SP to aging”, the top aging marker was NOL7 (Nucleolar Protein 7, cg16253168), which is critical for maintaining nuclear stability and ribosome biogenesis, and its function is linked to cellular senescence^59–61. A key disease marker was CXCL11 (C-X-C Motif Chemokine Ligand 11, cg08046471), a chemokine central to neuroinflammation that mediates T-cell recruitment into the CNS and is a component of the senescence-associated secretory phenotype (SASP)^62–64.

In “aging to MDD”, the top aging marker was also LRRC40. The top disease marker was GPR25 (G Protein-Coupled Receptor 25, cg04983977), a chemokine receptor that may mediate a chemoaffinity axis linking mucosal immunity in CNS⁶⁵. In “MDD to aging”, a key aging marker was CIDEC (Cell Death Inducing DFFA Like Effector C, cg05684195), which is involved in inflammation and apoptosis by regulating the NF-κB pathway⁶⁶. Key disease markers in this direction included PEO1 (Progressive External Ophthalmoplegia 1, cg19578398), as the mitochondrial helicase essential for mtDNA replication, whose mutations can lead to severe encephalopathy and are associated with depressive symptoms⁶⁷.

Furthermore, the classical 353 CpG sites of the Horvath pan-tissue epigenetic clock were also investigated in the sensitive “aging-disease” pairs³². Notably, we found a significant convergence between our de novo identified markers and this foundational aging signature (Supplementary Data 22–24), underscoring that our sensitivity analysis effectively captures core epigenetic aging processes. Several key markers were shared across diseases, often acting in specific directional pathways. For instance, NTSR2 (cg25657834) emerged as a top aging marker in the “Aging to BD” and “Aging to MDD” pathways, consistent with its specific role in the nervous system and links to mental disorders⁶⁸. Similarly, the m1A demethylase ALKBH3 (cg22637507), another top aging marker for both BD and MDD, was also highlighted in “Aging to Disease”, indicating its role as a potential upstream regulator of RNA metabolism in the onset of these conditions⁶⁹. Conversely, EDARADD (cg09809672), a disease marker found across all three psychoses, was most prominent in the “Aging to SP” and “Aging to MDD” directions. As a crucial activator of the NF-κB pathway, its role underscores the importance of inflammation as a consequence of aging that drives disease pathology⁷⁰.

For disease-specific markers, CAPZB (cg13319175) was the leading disease marker in the “Aging to BD” pathway, suggesting that alterations in cytoskeleton dynamics are an early pathological event driven by aging⁷¹. In the reverse “BD to Aging” direction, CHI3L2 (cg10045881) was the top aging marker, while C10orf99 (cg04126866) was a leading disease marker, collectively implicating the dysregulation of inflammatory signaling as a consequence of the disease^72,73. In SP, beyond the shared markers, FLJ33790 (cg16547529) was a prominent disease marker in the “Aging to SP” pathway. For MDD, other top markers included DOLPP1 (cg01027739), an aging marker in the “MDD to Aging” direction involved in protein N-glycosylation⁷⁴, and SELP (cg01459453), a key disease marker in the same pathway, highlighting that established depression pathology may drive changes in cell adhesion and inflammatory processes⁷⁵.

In addition, these sensitive “aging-disease” pairs for each disease were further investigated to identify relative potential drugs by using gene-drug related datasets. In other words, key drugs could be explored in the aspect of the relationship between aging and disease (the sensitive “aging-disease” pairs in this work). Both drug-induced gene expression profiles (GSE119291) and drug sensitivity results(based on gene expression profiles of late-onset psychoses used in this work, and further analyzed by the “oncoPredict” platform) were used to identify aging-related drugs. For the drug-induced gene expression profiles, the correlation of each sensitive pair of aging and disease markers was compared between “schizophrenia” and “control”, and then these differences were further compared by each drug and DMSO (formula (29)). Moreover, the correlation between the drug-sensitive result and each sensitive pair of aging and disease markers was compared between “disease” and “control”, and then was further compared by each drug and the placebo (formula (30)). As a result, key “aging-disease” pairs were identified, including 12426, 9059, and 14283 pairs sensitive to 128, 132, and 135 drugs in BD, MDD, and SP, respectively. Further, the drug sensitivity results also highlighted that the sensitive pairs (18241, 14051, and 22750) were significantly disturbed by a series of (56, 52, and 66) drugs. The top sensitive “aging-disease” pairs and related drugs were shown in Supplementary Fig. 11 and Supplementary Data 25 (sorting by the summarized K-S statistics for each sensitive “aging-disease” pair, and the Benjamini–Hochberg FDR based on the Kruskal–Wallis test), where coefficient values in a disease had the opposite sign (i.e., + or −) to that in the control were highlighted. For example, the sensitive pair of “ADAM15 ← LAG3” was highlighted in SP induced by bumetanide (Supplementary Fig. 11b and Table 20), indicating the key relationship between the inflammatory response (LAG3) in SP and cellular homeostasis (ADAM15) in brain aging. It is perhaps that the abnormal immune response in SP dysregulated the cellular adhesion, and even promoted the brain aging process, where the cellular homeostasis was highlighted as a potential target for SP in the context of aging.

To sum up, these results indicated that the sensitive “aging-disease” interaction could present a key index between aging and psychoses at the system level. As a result, crucial interactive mechanisms could be identified by exploring sensitive aging-associated interrelationships. For example, robust links between neuroimmune crosstalk, synaptic integrity, and cellular homeostasis were identified as key themes in the pathogenesis of late-onset psychoses (Supplementary Data 19–21).

Enrichment analysis exploring critical interactions between aging and psychoses

To explore potential interactions linking aging and late-onset psychoses, the shortest path between each aging and disease pair was identified based on differential network analysis. This network analysis compared either DNA methylation or gene expression profiles, applying the Dijkstra algorithm. Data from these networks were subjected to a subsequent enrichment analysis using the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway and the Biological Process (BP) terms derived from Gene Ontology (GO). The results indicated that a series of crucial aging-related pathways were integrated for these late-onset psychoses, where relative previous validation details were highlighted in Supplementary Data 26–37.

During the first stage of this work, we set out to identify common features shared by all three late-onset psychiatric disorders (Fig. 4A). This analysis identified that the “CYTOKINE–CYTOKINE RECEPTOR INTERACTION” was the most important KEGG pathway shared between BD, SP, and MDD samples, indicating that the dysregulation of the immune system and chronic inflammation are common to all three conditions⁷⁶. It is worth noting that there is some data suggesting that cytokines can affect neurotransmitter metabolism, neurogenesis, and the neuroendocrine system⁷⁷. Crucially, our analysis also highlighted the “CALCIUM SIGNALING PATHWAY” as a top shared mechanism (Fig. 4A). Calcium dysregulation is a hallmark of brain aging that impairs synaptic plasticity⁷⁸, and transcriptomic studies have recently confirmed that calcium ion transport pathways are differentially altered in the amygdala across SP, BD, and MDD, linking metabolic deficits to neuronal excitability⁷⁹.

Fig. 4. The common enriched pathways seen in the tested psychoses.

A summarizes KEGG pathways; (B) BP terms.

Most strikingly, the enrichment analysis using BP terms revealed another dimension of neuroimmune crosstalk: “LYMPH VESSEL MORPHOGENESIS” and “LYMPHANGIOGENESIS” emerged as the top enriched terms, particularly in SP, BD and MDD (Fig. 4B). While the CNS was traditionally considered devoid of lymphatic vessels, recent discoveries have characterized the meningeal lymphatic system as a critical route for clearing interstitial metabolic waste and immune cells from the brain⁸⁰. It has been reported that the drainage function of meningeal lymphatics declines significantly with aging, leading to the accumulation of neurotoxic proteins and inflammatory cytokines⁸¹. These results indicated that the accumulation of pathological factors in late-onset psychoses may stem from a shared failure in the brain’s clearance system-the glymphatic and meningeal lymphatic vessels, exacerbated by the aging process.

In addition, features associated with the development of each late-onset disease were also explored by identifying the top KEGG pathways and BP terms (with the minimum FDR) in both gene expression and DNA methylation profiles (Fig. 5 and Supplementary Data 26–37), where enriched functions overlapped with early-onset disease were removed. In the analysis of DNA methylation profiles in BD (Supplementary Data 26 and 29), the top enriched KEGG pathway (FDR = 0.00076597) was “GLYCINE SERINE AND THREONINE METABOLISM.” This finding aligns with metabolomics studies that identified altered plasma levels of glycine and serine across different clinical phases of BD^82,83. Furthermore, recent multi-omics and Mendelian randomization analyses have established a causal link between this metabolic axis and psychiatric susceptibility, likely by modulating glutamatergic neurotransmission via the NMDA receptor co-agonist D-serine^84,85. The top enriched BP term (FDR = 0.0029482) in the same analysis was “MUSCLE CELL DIFFERENTIATION.” This likely reflects the shared genetic etiology between cardiac and neural tissues (the heart-brain axis), consistent with observations that mood-stabilizing drugs, such as lithium and olanzapine, modulate myogenic differentiation pathways⁸⁶.

Fig. 5. Enrichment analysis within each disease.

A–C contain KEGG pathways, while (B, D, F) represent BP terms. A, B show changes in BD, (C, D) in SP, and (E, F) illustrate changes seen in MDD.

For gene expression profiles in BD (Supplementary Data 32 and 35), the top enriched KEGG pathway in “aging to BD” was “LYSINE DEGRADATION” (FDR = 0.0041152), highlighted as a unique metabolic signature distinguishing affective disorders from schizophrenia⁸⁷. Notably, the analysis also highlighted the “HEDGEHOG SIGNALING PATHWAY” in “BD to aging” (Supplementary Data 32). Beyond its developmental role, aberrant Hedgehog signaling in the adult brain directly regulates mitochondrial dynamics (e.g., fission) and neuroinflammation, processes potentially exacerbated by the age-related loss of SIRT1-mediated inhibition^88–91. Meanwhile, the top enriched BP term (FDR = 0.0014464) in “aging to BD” was “NEGATIVE REGULATION OF G0 TO G1 TRANSITION”. This process is critical for maintaining neural stem cell quiescence, and its dysregulation can lead to stem cell exhaustion and accelerated senescence⁹². Additionally, the enrichment of “RESPONSE TO L ASCORBIC ACID” (FDR = 0.0018078) emphasizes the role of oxidative stress defense and neurovascular coupling in the pathophysiology of BD⁹³.

For DNA methylation profiles in SP (Supplementary Data 27 and 30), the top enriched KEGG pathway was “CYTOSOLIC DNA SENSING PATHWAY” (FDR = 0.0010635). According to recent studies, this pathway (specifically cGAS-STING) is activated by misplaced self-DNA originating from damaged mitochondria or genomic instability, which subsequently triggers type I interferon responses and the senescence-associated secretory phenotype (SASP), driving sterile inflammation in the aging brain^94–96. The top enriched BP term was “CARTILAGE DEVELOPMENT” (FDR = 1.25E-05). This association is likely driven by pleiotropic risk genes such as DISC1, which regulates the migration and differentiation of cranial neural crest cells, a common progenitor population for both craniofacial cartilage and forebrain structures^97,98.

For gene expression profiles (Supplementary Data 33 and 36), the top enriched BP term in the “aging to SP” direction was “WOUND HEALING” (FDR = 0.003687). Biological aging has been mechanistically compared to a state of “unresolved wound healing”, where the accumulation of senescent cells impairs tissue repair; in the CNS, this process involves pericyte-derived scarring and dysregulated synaptic remodeling mediated by microglia^99–101. In “SP to aging”, the top enriched KEGG pathway was “BETA-ALANINE METABOLISM” (FDR = 0.0011132). Literature indicates that beta-alanine functions as a neuromodulator that competitively inhibits GABA uptake, and its metabolic dysregulation also compromises the synthesis of carnosine, a critical antioxidant, thereby linking inhibitory neurotransmission deficits with oxidative stress in SP^102–104. Additionally, the enrichment of “NEGATIVE REGULATION OF PROTEIN SERINE THREONINE KINASE ACTIVITY” (FDR = 0.00011678, Table 18) reflects the disrupted homeostasis of key signaling kinases, such as GSK-3 and Akt, which are central to the pathophysiology and pharmacological treatment of SP^105,106.

In MDD (Supplementary Data 28 and 31), the top enriched KEGG pathway in the “MDD to aging” comparison was “STEROID HORMONE BIOSYNTHESIS” (FDR = 0.0013976). Dysregulation of cortisol biosynthesis and the hypothalamic-pituitary-adrenal (HPA) axis is one of the most consistent biological findings in MDD, linking stress sensitivity to systemic metabolic features^87,107. In the analysis of DNA methylation BP terms (Supplementary Data 31), the top term was “ENDOCARDIUM MORPHOGENESIS” (FDR = 0.0015522). This enrichment reinforces the evidence for the “heart-brain axis”, suggesting that shared developmental signaling pathways (e.g., Notch and Wnt) may underlie the established comorbidity between congenital heart defects and psychiatric susceptibility^108–110.

For expression profiles (Supplementary Data 37), the results highlighted critical disruptions in cellular homeostasis and survival. In the “aging to MDD” direction, the top enriched BP term was “CYTOPLASMIC TRANSLATION” (FDR = 0.0033717). Recent studies indicate that brain aging is characterized by aberrant ribosome pausing (“stalling”) and the accumulation of tRNA fragments, processes that disrupt proteostasis and specifically deplete proteins required for synaptic maintenance¹¹¹. In the “MDD to aging” direction, the “NECROPTOTIC SIGNALING PATHWAY” (FDR = 0.0050615) was significantly enriched. Unlike apoptosis, necroptosis is a highly immunogenic form of regulated cell death driven by the RIPK1-RIPK3-MLKL axis. Its activation is suppressed by SIRT1 (via deacetylation of RIPK1) and AMPK, suggesting that the metabolic collapse and loss of SIRT1 observed in MDD may directly unleash neuronal necrosis and inflammation¹¹². Furthermore, the enrichment of “GLUCOCORTICOID SECRETION” (FDR = 0.010845) confirms that flattened diurnal glucocorticoid rhythms drive hippocampal atrophy via alterations in structural gene expression, creating a vicious cycle of stress and degeneration^113,114.

The analysis of network markers revealed additional underlying mechanisms between aging and disease

The top network markers(by summarizing the degree or betweenness based on DNA methylation or gene expression profiles) were shown in Supplementary Data 38–40, where specific markers in each disease or relevant previous validation methods were highlighted. For example, the top 10 key markers showing the highest degree of DNA methylation differences are listed in Supplementary Data 38. The most important aging marker with the highest relevance in the BD network was SPI1 (Spi-1 Proto-Oncogene, cg14088811, degree = 1683). This proto-oncogene affects glial cell proliferation and gliosis, which in turn regulate the immune response^115,116. The top network marker in SP was EDIL3 (EGF Like Repeats And Discoidin Domains 3, cg16773899, degree = 1737), which was suggested as a potential target for the immune checkpoint blockade therapy¹¹⁷. BAZ1B (Bromodomain Adjacent To Zinc Finger Domain 1B, cg14093368, degree = 1498) was top-ranked in MDD. Apart from sensing L-arginine levels and promoting T-cell survival¹¹⁸, this molecule also plays an important role in chromatin remodeling and DNA repair damage¹¹⁹.

Additionally, a betweenness value was also calculated for each “aging-disease” pair, based on the shortest path between them, and relevant permutation tests were also performed to compare the same number of random pairs (running 1000 times), then the p-values were derived as the proportion of times a gene occurred as a top-ranking betweenness gene in 1000 permutations (Supplementary Data 39–40, and Figs. 6, 7, relative details were also discribed in the “Identifying network markers” section). When analyzing DNA methylation profiles, the top network marker in “aging to BD” was CAMK1D (Calcium/Calmodulin Dependent Protein Kinase ID, cg24859228). CAMK1D is a key component of the calcium signaling pathway, regulating CREB-dependent gene transcription essential for synaptic plasticity⁷⁸. RUSC1 (RUN And SH3 Domain Containing 1, cg11503744), the top network marker in “BD to aging”, regulates NGF-dependent neurite outgrowth and vesicular trafficking; its dysfunction may underlie the deficits in axonal connectivity and neuronal differentiation observed in the accelerated aging of the BD brain^120,121.

Fig. 6. The top Network markers with the maximum betweenness based on DNA methylation profiles.

A–C ”aging-disease”, (B, D, F) ”disease-aging.” A, B show networks in BD, (C, D) are SP networks, and (E, F) illustrate those in MDD.

Fig. 7. The top Network markers with the maximum betweenness based on gene expression profiles.

Similarly to Fig. 6, (A–C) show ”aging-disease”, (B, D, F) ”disease-aging” links. A, B show networks in BD, (C, D) are SP networks, and (E, F) illustrate those in MDD.

When the network analysis was conducted on gene expression profiles, PROX1 (Prospero Homeobox 1) was identified as a key marker in “aging to BD”. PROX1 is a master regulator of adult hippocampal neurogenesis in the dentate gyrus, a process critical for mood regulation that is impaired in BD and restored by mood stabilizers via the Wnt signaling pathway^122,123. When the comparison was done in “BD to aging”, PHYHIP (Phytanoyl-CoA 2-Hydroxylase Interacting Protein) emerged as the top network component. PHYHIP is a synaptic protein whose epigenetic alteration has been linked to Lewy body pathology, suggesting that synaptic proteostasis collapse may drive the transition from psychiatric symptoms to neurodegenerative features in older BD patients^124,125.

When analyzing methylation data in SP, the top network marker in “aging to SP” was MOCS2 (Molybdenum Cofactor Synthesis 2, cg21540749). MOCS2 encodes a rate-limiting enzyme for molybdenum cofactor biosynthesis; its deficiency leads to severe structural brain abnormalities (e.g., loss of gray/white matter differentiation) and intractable seizures, highlighting its critical role in maintaining neuronal structural integrity and metabolic homeostasis against aging-related stress^126,127. In “SP to aging”, GPR160 (G Protein-Coupled Receptor 160, cg18741908) was identified as the top network marker. GPR160 has recently been de-orphanized as a receptor for CARTp (Cocaine- and Amphetamine-Regulated Transcript peptide), a neurotransmitter involved in reward processing, energy metabolism, and anxiety; its dysregulation may underlie the complex interplay between psychosis, metabolic syndrome, and neuroinflammation observed in SP^128–130.

When the network analysis was conducted on gene expression data, LRRC40 (Leucine Rich Repeat Containing 40) was the top network marker in “aging to SP”. LRRC40 interacts with Plexin A2 to regulate neural development and has been identified as a candidate gene for Autism Spectrum Disorder (ASD), reinforcing the concept of shared neurodevelopmental vulnerability between SP and ASD^56,131. In “SP to aging”, NOL7 (Nucleolar Protein 7) emerged as the central network marker. NOL7 is essential for maintaining nucleolar stability and ribosome biogenesis; its depletion triggers nucleolar stress and accelerates cellular senescence by promoting the degradation of nuclear lamina proteins (e.g., Lamin B1), providing a molecular basis for the “accelerated aging” phenotype in SP¹³².

For DNA methylation, the top network feature in “aging to MDD” was PROX1 (Prospero Homeobox 1, cg14019317). As a master regulator of neurogenesis in the dentate gyrus, PROX1 is essential for the maturation of granule cells, a process that is suppressed in MDD but restored by antidepressant treatments, thereby linking hippocampal plasticity to mood recovery^122,123. In “MDD to aging”, SHOX2 (Short Stature Homeobox 2, cg06156376) was the top network marker. SHOX2 regulates the firing properties of thalamocortical neurons; its dysregulation may disrupt sleep-wake rhythms and thalamic gating, contributing to the circadian disturbances and cognitive decline frequently observed in older MDD patients¹³³.

When differential gene expression data were analyzed, ALKBH3 (AlkB Homolog 3) was identified as a key marker in “aging to MDD” comparisons (ranking second in betweenness, Table 28). ALKBH3 is a dioxygenase that mediates the demethylation of N1-methyladenosine (m1A) in DNA and RNA to repair oxidative lesions; its dysfunction compromises genomic integrity under stress, potentially accelerating brain aging in MDD^134,135. Meanwhile, in “MDD to aging”, CARD6 (Caspase Recruitment Domain Family Member 6) emerged as the most critical network component. CARD6 is a microtubule-associated protein that interacts with RIPK kinases to modulate NF-κB signaling and necroptosis; its identification as a top hub provides direct molecular evidence that the necroptotic cell death pathway drives neuroinflammation and neurodegeneration in the progression of MDD^136,137.

In addition, classical epigenetic aging markers (353 CpG sites) were also investigated³². In the DNA methylation differential network, a compelling number of high-betweenness markers were identified (Supplementary Fig. 12 and Supplementary Data 41), indicating a remarkable convergence between our network-derived hubs and the established epigenetic aging markers. Several key markers, which were also identified in our sensitivity analysis, emerged as top-ranking bridges in multiple contexts. Notably, NTSR2 (cg25657834) was the top marker for both “Aging to BD” and “Aging to MDD” pathways, underscoring a shared mechanism where this neurotensin receptor, linked to mental disorders, acts as a critical bridge when aging contributes to disease onset⁶⁸. In the reverse direction, CHI3L2 (cg10045881) was the top marker for both “BD to Aging” and “MDD to Aging”, suggesting that the pathological state of these mood disorders may converge on triggering inflammatory responses via this chitinase-like protein⁷². For schizophrenia, the top marker in the “Aging to SP” pathway was TBC1D23 (cg16984944), a protein critical for endosome-to-Golgi trafficking and implicated in neurodevelopmental disorders¹³⁸. In the “SP to Aging” direction, DOLPP1 (cg01027739) was the top marker, another gene also identified as a top sensitive marker and a top bridge marker in the “BD to Aging” pathway, pointing to its broad role in pathways affected by psychosis⁷⁴.

Similarly, the differential co-expression network revealed that the top betweenness genes also overlapped with the Horvath aging clock (Supplementary Fig. 13 and Supplementary Data 42), many of which were also prominent in the sensitivity analysis. ALKBH3 stood out as the top marker for both “Aging to BD” and “Aging to MDD”, reinforcing its role as a key upstream aging-related regulator for mood disorders⁶⁹. The “Aging to SP” was topped by NTSR2, reconfirming its importance as a shared vulnerability across both network types and analyses. In the reverse “Disease to Aging”, CHI3L2 was the top marker for BD, while RHBDD1 was the top marker for SP; the former again highlights a key inflammatory responder, while the latter is a rhomboid family protease regulating cell growth and apoptosis¹³⁹. For MDD, besides the shared ALKBH3, FOXE3 was another top marker in the “Aging to MDD”, a transcription factor whose dysregulation in developmental pathways may be an early event in aging-related psychosis¹⁴⁰. The top marker in the “MDD to Aging” was DOLPP1, underscoring its cross-disease and cross-network relevance. The consistent overlap across different data types and diseases demonstrates a striking functional convergence, not only validates our network-based approach but also pinpoints a core set of genes at the nexus of aging and psychosis.

In summary, potential network makers were identified, including both verified specific markers in each disease and other aging-related risk factors.

Discussion

This work utilized a series of computational approaches to explore molecular mechanisms linking brain aging to the development of three forms of late-onset psychosis. Then, both the common and specific characteristics across different psychoses were compared based on advanced brain aging. As a result, a series of risk factors were integrated at the system level, including both specific and other aging-related pathways in each psychosis. In addition, our work could also study relevant aging-related diseases by exploring the crosstalk between aging and disease based on the aging index.

A series of common features were shared between these diseases based on aging, suggesting the strong involvement of immune dysregulation in the development and progression of late-onset psychiatric disorders. For example, “CYTOKINE–CYTOKINE RECEPTOR INTERACTION” was the most important KEGG pathway common to all three forms of psychoses (Fig. 4A). This finding aligns with extensive meta-analyses demonstrating that elevated peripheral levels of pro-inflammatory cytokines (e.g., IL-1β, IL-6, TNF-α) are a transdiagnostic hallmark of acute and chronic phases in SP, BD, and MDD^141,142. Chronic low-grade inflammation (“inflammaging”) disrupts synaptic plasticity and memory function via cytokine-mediated modulation of neurotransmitter receptors (e.g., NMDA, AMPA) and neurotrophic factors, thereby increasing vulnerability to psychiatric conditions in the aging brain¹⁴³. Crucially, the identification of “HEMATOPOIETIC CELL LINEAGE” and “POSITIVE REGULATION OF LEUKOCYTE PROLIFERATION” (Figs. 4A, B) points to a systemic origin of neuroinflammation. We interpret this enrichment as a molecular signature of the age-related“myeloid bias” where the aging bone marrow shifts toward producing pro-inflammatory cells. Given the direct vascular channels connecting the skull bone marrow to the meninges^144,145, this hematopoietic system may function as a proximate reservoir, potentially driving immune cell influx into the CNS. Consequently, this structural route provides a mechanistic link explaining how this systemic skew in immune cell production may drive the local microglial reactivity and synaptic disruption reported in late-onset psychoses¹⁴⁶.

Changes affecting neural plasticity are critical factors in the development of late-onset psychiatric disorders. Our enrichment analysis identified the “CALCIUM SIGNALING PATHWAY” as a prominent shared mechanism across BD, SP, and MDD (Fig. 4A). This finding aligns with the well-established role of intracellular calcium dysregulation in brain aging, a process known to impair synaptic plasticity and neuronal excitability^78,147. Crucially, large-scale genomic studies have confirmed that voltage-gated calcium channel genes, particularly CACNA1C (encoding the CaV1.2 subunit), constitute a shared genetic basis for these three major psychiatric disorders¹⁴⁸. Consistent with this, our sensitivity analysis identified CAMK1D (Calcium/Calmodulin Dependent Protein Kinase ID) as a key aging-associated marker linked to disease susceptibility. Furthermore, the enrichment of the “MAPK SIGNALING PATHWAY” and “NEUROACTIVE LIGAND–RECEPTOR INTERACTION” (Fig. 4A) points to a broader disturbance in synaptic signal transduction. The MAPK/ERK cascade is essential for coupling synaptic activity to the translational control of synaptic proteins, a process required for the consolidation of long-term plasticity and memory¹⁴⁹. Disruption of this pathway has been shown to hinder the transport of AMPA receptors to synapses, contributing to cognitive decline¹⁵⁰. Additionally, the identification of neuroactive ligand–receptor interactions highlights potential deficits in upstream neurotransmitter systems, such as dopamine and glutamate receptors, which initiate these intracellular cascades^151,152. Collectively, these results suggest that aging-associated calcium instability may compromise the intracellular signaling machinery required for maintaining synaptic integrity and function in late-onset psychoses.

Cellular homeostasis in the CNS fundamentally depends on the maintenance of structural integrity, a process requiring precise coordination between the cytoskeleton and cell adhesion systems. Our enrichment analysis revealed that “REGULATION OF ACTIN CYTOSKELETON” and “CELL ADHESION MOLECULES (CAMs)” (Fig. 4A) were simultaneously enriched across BD, SP, and MDD, implying a systemic collapse of cellular microstructures in the aging psychiatric brain. At the synaptic level, the reduction in dendritic spine density is a transdiagnostic feature observed in SP, BD, and MDD, directly linked to the instability of the actin cytoskeleton^153,154. Mechanistically, cell adhesion molecules function as upstream regulators of these cytoskeletal dynamics. Furthermore, this structural fragility extends beyond the synapse to the systemic barrier of the brain, as the “CELL ADHESION MOLECULES” pathway encompasses both synaptic organizers (e.g., Neurexins and Neuroligins)¹⁵⁵ and components of endothelial tight junctions. Genetic and experimental evidence links Claudin-5, a vital tight junction protein, to the etiology of schizophrenia and depression, where its downregulation leads to increased blood-brain barrier (BBB) permeability and behavioral abnormalities^156,157. We propose that in the context of brain aging, the erosion of these adhesive barriers may facilitate the infiltration of peripheral inflammatory factors, thereby establishing a vicious cycle where structural compromise exacerbates neuroinflammation.

Disruptions in vascular homeostasis and the heart-brain axis emerged as foundational vulnerabilities linking aging to late-onset psychoses. Our analysis revealed a dual enrichment pattern: fluid trafficking pathways, specifically “LYMPH VESSEL MORPHOGENESIS”, “LYMPHANGIOGENESIS”, and “LYMPH VESSEL DEVELOPMENT” (Fig. 4B), alongside striated muscle biology terms, including “CARDIOCYTE DIFFERENTIATION”, “CARDIAC MUSCLE CELL DIFFERENTIATION”, and “POSITIVE REGULATION OF STRIATED MUSCLE CELL DIFFERENTIATION” (Fig. 4B). It could be speculated that these signatures reflect deficits in CNS clearance and neurovascular maintenance. The lymphatic enrichment aligns with the “glymphatic dysfunction” hypothesis, where meningeal lymphatics facilitate both waste drainage and immune surveillance¹⁵⁸. Neuroimaging studies confirm reduced glymphatic diffusivity (DTI-ALPS index) in schizophrenia and bipolar disorder, correlating with cognitive deficits^159,160, while stress-induced lymphatic impairment exacerbates anxiety-like behaviors in depression¹⁶¹. Concurrently, the cardiac/muscle signature likely signifies vascular smooth muscle defects or shared genetic etiologies. Large-scale genomic analyses confirm extensive overlap between schizophrenia and cardiovascular risk factors¹⁶². At the molecular level, we identified GATA4 (a cardiac transcription factor) as a key aging marker; its accumulation in the aging brain triggers the senescence-associated secretory phenotype (SASP), bridging cardiovascular signaling and neuroinflammation¹⁶³. Additionally, the identification of the vascular remodeling promoter OVGP1 in MDD further supports compromised cerebral vascular integrity¹⁶⁴. Collectively, late-onset psychoses appear characterized by a microenvironment vulnerable to both lymphatic insufficiency and vascular aging.

Additionally, multiple cancer-related pathways were also identified during the enrichment analysis (Fig. 4A). Detecting these genes in the course of analyzing psychiatric conditions raises the possibility that the development of cancer and mental illness may involve shared BP during aging. For instance, the PI3K-Akt signaling cascade, a central driver of cell proliferation in prostate cancer¹⁶⁵, is equally indispensable for neuronal signaling and synaptic plasticity, where its integrity is crucial for cell survival. Wnt signaling governs tumorigenesis and cellular homeostasis¹⁶⁶. Furthermore, the enrichment of angiogenic factors points to the dual nature of vascular homeostasis: mechanisms like VEGF signaling, which drive tumor lymphangiogenesis¹⁶⁷, are also critical for neurovascular coupling and brain clearance¹⁶⁸. Notably, specific genetic overlaps, such as the association of the JAZF1 gene with both prostate cancer and schizophrenia¹⁶⁹, further validate the shared genetic architecture underlying these complex disorders.

By investigating the interaction between aging and the tested psychoses, a number of potential pathogenic processes were identified that appeared specific to individual psychiatric conditions. For example, the “RESPONSE TO L ASCORBIC ACID” was significantly enriched in “aging to BD” comparisons (Supplementary Data 29). It modulates neurovascular coupling and cerebral blood flow; acute infusion of ascorbic acid improves endothelium-dependent vasodilation and cardiovagal baroreflex sensitivity, counteracting age-related vascular dysfunction^170,171. Detecting RXRA as a high-degree hub (Supplementary Data 36) seems to support this assumption. Postmortem studies have observed deficits in retinoid signaling components in the prefrontal cortex of elderly patients with mood disorders¹⁷², suggesting that the dysregulation of this nuclear receptor axis may contribute to the accelerated aging phenotype seen in BD.

Distinct metabolic signatures characterizing SP were identified in the context of brain aging. We detected a significant enrichment of “PHENYLALANINE METABOLISM” in the “SP to aging” comparison (Supplementary Data 33), which aligns with clinical findings of impaired phenylalanine kinetics that are specific to SP and distinct from BD or MDD¹⁷³. Mechanistically, elevated brain phenylalanine concentrations competitively inhibit the transport of other large neutral amino acids (e.g., tyrosine, tryptophan) across the blood-brain barrier¹⁷⁴. This metabolic blockade may limit the precursor availability for monoamine synthesis, potentially exacerbating the deficits in neuronal signaling observed in the aging brain. Additionally, the enrichment of “BETA-ALANINE METABOLISM” (Supplementary Data 33) in the “SP to aging” direction is particularly relevant given that β-alanine functions as an endogenous antagonist at GABA receptors¹⁰². This interaction is critical, as cortical GABAergic dysfunction is a well-established pathophysiological feature of SP; moreover, the chronic oxidative stress generated by β-alanine accumulation¹⁰⁴ likely acts as a specific accelerator for the loss of cellular homeostasis in the SP cortex.

In MDD, the analysis highlighted specific alterations in hormonal regulation systems linked to brain aging. We detected a significant enrichment of “STEROID HORMONE BIOSYNTHESIS” in the “MDD to aging” comparison (Supplementary Data 28), which may reflect a deficit in the production of neuroactive steroids (e.g., allopregnanolone) essential for synaptic plasticity^175,176. Concurrently, the enrichment of “GLUCOCORTICOID SECRETION” in the interactions direction points to the potential involvement of HPA axis dysregulation, a well-known driver of hippocampal atrophy in late-life depression^177,178. Notably, our analysis identified NR1D2 (REV-ERBβ), a core circadian clock gene, as a top aging-associated marker (Supplementary Data 37). Given that circadian clocks regulate the diurnal rhythm of cortisol, the identification of NR1D2 suggests that circadian disruption might act as an upstream factor contributing to the sustained HPA axis dysfunction observed in this condition¹⁷⁹.

The work presented here explored systematically potential key biologically relevant links between aging and the development of late-onset psychoses. Our analysis suggests that features associated with vascular stability and the heart-brain axis may represent a foundational vulnerability linking aging and disease. Specifically, lymphatic vessel morphogenesis and development exhibited as “aging to disease” in all three conditions, suggesting that age-related decline in meningeal lymphatic drainage might potentially serve as a predisposing factor for these psychiatric conditions. By contrast, alterations in neuroplasticity appear to be a pathological consequence that may accelerate the aging phenotype. Both the “CALCIUM SIGNALING PATHWAY” and “NEUROACTIVE LIGAND–RECEPTOR INTERACTION” were characterized as “disease to aging” factors across all three psychoses, suggesting that established psychiatric pathology might contribute to accelerated brain aging through synaptic dysregulation. Immune dysregulation displayed a more complex pattern: the source of inflammation, indicated by “HEMATOPOIETIC CELL LINEAGE”, acted as an “aging to disease” factor in BD and SP, while the resulting “CYTOKINE–CYTOKINE RECEPTOR INTERACTION” exhibited bidirectional relationships in BD and SP but was predominantly “disease to aging” in MDD. Finally, cellular homeostasis largely followed the “disease to aging” pattern in MDD, whereas BD and SP showed mixed directional effects. Collectively, our analysis reveals that MDD may predominantly function as a stressor that accelerates aging processes, whereas BD and SP appear to arise from more complex bidirectional interactions between aging and disease. These findings suggest shared and distinct pathways involving disruptions in neuroimmune homeostasis, neural plasticity, cellular integrity, and vascular stability across the spectrum of late-onset psychiatric disorders, as summarized in Fig. 8. In addition, these directed relationships between aging and disease could further present potential medication (Supplementary Fig. 11 and Data 25). For example, ADAM15 was identified as a key sensitive aging marker, indicating a crucial role of the cellular homeostasis during the aging brain. Further, HOXC5 was highlighted as a critical sensitive disease marker, highlighting the key role of the inflammatory response. It is perhaps that the neuroimmune homeostasis was vital to be modulated with proper drugs, along with regulating the cellular homeostasis, where sensitive “aging-disease” pairs were serving as their key role.

Fig. 8. Summary of mechanisms shared between aging and late-onset psychoses.

Rectangles contain genes identified in enrichment analysis, ellipses represent markers detected in sensitivity analysis, and rhomboids represent network markers. The shapes are color-coded according to the biological pathway the genes are involved in: Ribosomal genes: pink, neuroplasticity: green, nuclear stability: blue, immune response: orange, energy metabolism: purple. Light blue indicates functions within SP, light gray indicates functions within MDD, and pale yellow shows functions within BD. If the geometrical shape is drawn with a solid line, the gene was identified by comparing DNA methylation profiles; dashed lines indicate genes showing differences when analyzing gene expression profiles.

Our results could further support the neuro-immunosenescence theory, which posits that aged immune systems drive systemic aging and psychiatric vulnerability through declining protective function and increased pro-inflammatory output^180,181. Crucially, our enrichment analysis highlights the hematopoietic cell lineage (Fig. 4B) and cytokine–cytokine receptor interaction (Fig. 4A) as top shared pathways. This aligns with the concept of age-related ‘myeloid bias’ in the bone marrow, a key driver of systemic inflammaging that disrupts CNS homeostasis¹⁸². Mechanistically, our identification of the calcium signaling pathway and MAPK signaling pathway (Fig. 4A) as common pathogenic factors resonates with established research: dysregulated intracellular calcium (e.g., in macrophages) and overactivated p38-MAPK signaling are known to potentiate NF-κB activation and pro-inflammatory phenotypes^183,184. These molecular disruptions likely fuel the elevated cytokine levels observed in memory-associated brain regions like the hippocampus¹⁸³. In sum, our study suggests that abnormal aging induces late-onset psychoses by disrupting neuroimmune homeostasis via a series of signaling and metabolic axes.

The vascular theory of aging provides a compelling framework for interpreting our findings, positing that age-related vascular deterioration is a primary determinant of systemic aging¹⁸⁵. Our analysis lends support to this perspective by revealing a potential structural link: the enrichment of differentiation pathways associated with cardiac and striated muscle tissues (Fig. 4B) may reflect the molecular signature of vascular aging, particularly the remodeling of vascular smooth muscle cells that leads to arterial stiffness and endothelial dysfunction^186,187. These findings align with clinical observations of vascular biomarker abnormalities in schizophrenia¹⁸⁸ and the prognostic significance of vascular pathology in late-life depression¹⁸⁹. Furthermore, these results indicated an extension to this framework by highlighting the potential involvement of lymph vessel morphogenesis (Fig. 4B) as an ‘aging-to-disease’ driver. This points toward a compromised glymphatic and meningeal lymphatic system, which may fail to effectively clear metabolic waste and inflammatory factors from the aging brain¹⁹⁰. These vascular-lymphatic signatures appear to interact with calcium signaling alterations (Fig. 4A), which are recognized drivers of arterial stiffening¹⁹¹, thereby potentially establishing a microenvironment that exacerbates neuroimmune dysregulation¹⁹².

In summary, our work explored interactive mechanisms between aging and late-onset psychoses at a system level. The results not only confirmed various specific markers in each disease (Supplementary Data 19–40) but also suggested that a series of aging-related pathways (i.e., immune dysregulation, deficits in neuroplasticity, loss of cellular homeostasis, and disruptions in vascular stability) in the context of neuroimmune crosstalk are also key bridges linking brain aging and late-onset mental illness. These risk factors interacted with each other and disrupted neuroimmune homeostasis in the context of aging, which was vital to late-onset psychiatric disorders, where each disease might have its distinct role within the crosstalk with aging (e.g., dysregulation of neurovascular coupling and retinoid signaling in BD; distinct metabolic disturbances involving phenylalanine and beta-alanine in SP; and alterations in steroid hormone biosynthesis and HPA axis regulation in MDD). In brief, our results provide a molecular basis for two key aging theories, neuro-immunosenescence and vascular-lymphatic aging, and can be used to explore the landscape of potential molecular mechanisms in late-onset psychiatric disorders based on the aging brain (Fig. 8).

While we aimed for a comprehensive analysis, this study has limitations: (1) The gender of the patients was ignored during the building of prediction models, although it was taken into account during analysis. (2) The number of patients in the BD group was limited, potentially reducing the validity and generalizability of the reported findings in this regard. (3) Our results, derived from in-silico analysis, would require experimental confirmation. (4) Analyzing single-cell sequence datasets would be crucial to studying the effects of cellular heterogeneity. Given the complexity of the development of psychiatric conditions, disease-related molecular patterns, and the relationships between aging and disease will be needed in the future.

Conclusion

In the presented study, ML methods were used to identify markers associated with aging and various late-onset psychiatric disorders, and aging scores were calculated accordingly. The results showed that certain features of late-onset psychiatric disorders showed notable links to advanced aging. A series of mapping models was constructed to optimize the weighting of features connecting aging and disease development. Further, sensitivity, enrichment, and network analyses were performed to explore key interactions between aging and disease. As a result, both common and specific markers were identified across different psychoses, highlighting disturbances of the neuroimmune homeostasis. Collectively, a series of interactive mechanisms were integrated (including both specific markers in each disease and other aging-associated risk factors) by studying the aging index of each psychosis. We hope that the presented comprehensive analysis of links between aging and late-onset psychiatric conditions will pave the way for future investigations in this field.

Author contributions

Performed the algorithm and analyzed the data: S.R., M.X., and Y.W.; Wrote the manuscript: S.R., W.Y., B.W., and Y.W.; Designed and sponsored the study: Y.L. and Y.W. All authors read and approved the manuscript.

Peer review

Peer review information

Communications Medicine thanks Chang-hyun Park and Xin Niu for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the National Natural Science Foundation of China (32000478 to Y.W.) and the Natural Science Foundation of Liaoning Province of China (2023JH2/20200080 to Y.W.), the CAS Research Fund (Grant No. XDB38050200 to Y.L.), and the Self-supporting Program of Guangzhou Laboratory (Grant No. SRPG22-001 and SRPG22-007 to Y.L.). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

All the DNA methylation and gene expression profiles were downloaded from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) database together with chronological age information. The following profiles were used: (1) DNA methylation: GSE38873, GSE41037, GSE41169, GSE611431, GSE89707, GSE112179, GSE152026, GSE157252, GSE129428, GSE198904, and GSE88890; (2) Gene expression: GSE35974, GSE208338, GSE62191, GSE5388, GSE5389, GSE210064, GSE62333, GSE145554, GSE76826, and GSE98793. Relevant details are shown in Supplementary Data 1–6.

Code availability

The source codes are available at Figshare¹⁹³.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01689-1.