Ethics statement

For the Human Connectome Project (HCP) dataset, ethical approval was granted by the Washington University Institutional Review Board (IRB, 201204036). The Brain Genomics Superstruct Project (GSP) was approved by the Partners HealthCare IRB and the Harvard University Committee on the Use of Human Subjects in Research. The UK Biobank (UKB) study was conducted under a protocol approved by the North West Multi-centre Research Ethics Committee (reference: 16/NW/0274). All data included in the ADHD-200 dataset were obtained under IRB approvals at the respective contributing institutions, including Peking University (PKU), Kennedy Krieger Institute (KKI), NeuroIMAGE, New York University Child Study Center (NYU), Oregon Health and Science University (OHSU), University of Pittsburgh, Washington University in St. Louis, and Brown University. According to the International Neuroimaging Data-Sharing Initiative (INDI) protocol, no additional IRB approval was required for secondary analyses of de-identified publicly available data. The original studies included in the Autism Brain Imaging Data Exchange (ABIDE II) dataset were approved by the IRBs at each participating site (full site list available at https://fcon_1000.projects.nitrc.org/indi/abide/abide_II.html). In accordance with INDI data usage policies, no additional IRB approval was required for secondary analyses of de-identified publicly available data. The Bipolar-Schizophrenia Network for Intermediate Phenotypes (BSNIP-1) study protocol was approved by the IRBs of Hartford Hospital, the University of Texas Southwestern Medical School, the University of Maryland, the University of Chicago, Wayne State University, and Harvard University. The MDD dataset was approved by the Ethics Committees of Beijing Anding Hospital, West China Hospital of Sichuan University, the First Affiliated Hospital of Zhejiang University, and Henan Mental Hospital of Xinxiang. The AUD and TUD datasets were approved by the University of New Mexico Human Research Review Committee. The Alzheimer’s Disease Neuroimaging Initiative (ADNI) protocol was approved by the Institutional Review Boards of all participating institutions. Detailed information regarding the participating ethics committees and institutional review boards is available at https://adni.loni.usc.edu/about/governance/#core-details. Written informed consents were obtained after providing a complete description of the research process to each participant. For participants younger than 18 years of age, written informed consent was obtained from their parents or legal guardians, with assent obtained from the participants where applicable.

Participants

Neuroimaging data for the training of brain age prediction were obtained from 3 consortia, including HCP (n = 2,489) [16], GSP (n = 1,034) [17], and UKB (n = 39,679) [18]. ADHD (n = 344) and age-gender-quantity matched HC (n = 344) were obtained from ADHD-200 project [19]; ASD (n = 484) and matched HC (n = 484) from ABIDE II [20]; SZ (n = 152) and matched HC (n = 152), BP (n = 143) and matched HC (n = 143) from BSNIP-1 [21]; MDD (n = 258) and matched HC (n = 258) from Beijing Anding Hospital, West China Hospital of Sichuan, First Affiliated Hospital of Zhejiang and Henan Mental Hospital of Xinxiang [22]; AUD (n = 155) and TUD (n = 144) from Albuquerque [23]; AD (n = 361) and matched HC (n = 361), MCI (n = 657) and matched HC (n = 657) from ADNI [24]. The diagnosis of ADHD, ASD, SZ, BP, and MDD were based on the Structured Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) [25]. Participants with Alcohol Use Disorder Identification Test (AUDIT) > 7/Fagerström Test for Nicotine Dependence (FTND) > 7 were categorized as AUD/TUD [26,27] (A&TUD were from AUD and TUD with AUDIT > 7 and FTND > 7). The diagnosis of MCI and AD was based on the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association (NINCDS/ADRDA) criteria [28]. More detailed information on the diagnostic criteria and medication status for healthy participants and patients can be found in Text A in S1 Appendix. Demographic information of individuals participated in this study can be found in Table 1. The scatterplot of ages used in each control subgroup can be found in Fig A in S3 Appendix. Gray matter volume (GMV) from sMRI generated by Statistical Parametric Mapping version 12 (SPM12) was used for brain age prediction. Details on imaging parameters and preprocessing pipeline can be found in Text B in S1 Appendix and Table A in S2 Appendix. In addition, details on the impact of 1.5T scanner acquisitions on predictive performance and PAD difference can be found in Text C in S1 Appendix and Tables B and C in S2 Appendix. The research program was approved by the institution review board at each local site. Written informed consents were obtained after providing a complete description of the research process to each participant.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Demographic information of individuals participated in this study.

https://doi.org/10.1371/journal.pmed.1004860.t001

Brain age prediction with age correction

Age-range matched training HCs were selected from HCP, GSP, and UKB (641 for ADHD, 1,061 for ASD, 2,164 for SZ, 2,177 for BP, 2006 for MDD, 1912 for AUD, 1912 for TUD,1805 for A&TUD, 34,308 for AD and MCI) generating 10 training sets for each diagnostic group. Patients were not included in the training or cross-validation procedures. Details on samples construction for training and testing can be found in Fig B in S3 Appendix. The sample overlap between the corresponding training sets of different diagnostic groups can be found in Table D in S2 Appendix. Schaefer atlas [29] with 1,000 cortical regions of interest (ROIs) plus 16 subcortical ROIs from Melbourne subcortical brain atlas [30] were combined to generate 1016-ROIs as the “augmented Schaefer (AS)-1016” [31]. Averaged GMV based on augmented Schaefer-1016 brain atlas were used as features in brain age prediction (Fig 1a). The extreme gradient boosting (XGBoost) algorithm based on gradient boosted decision trees was chosen to predict the brain age, with chronological age of individuals in the healthy training set serving as the prediction target. The parameters were optimized using a 10-fold cross-validation conducted within the healthy training set. The optimal number of training iterations was determined by evaluating the prediction error across 2,000 rounds and early stopping was applied if performance did not improve for 20 consecutive rounds [32]. The hyperparameters of XGBoost were set up with a maximum depth of 4 and learning rate of 0.05, with other parameters set to default. The performance of predictive model was estimated by correlating the predicted brain age and chronological age, as well as by computing the mean absolute error (MAE) and coefficient of determination (R²), in the healthy training set. After optimization, the trained models were applied to independent patient samples to generate predicted brain age and calculate the PAD (deviation between predicted brain age and chronological age, Fig 1b and 1c).

Accuracy of brain age prediction suffers from an age-dependent bias, which leads to the uncertainty in clinical interpretation [33]. Thus, a bias-adjustment scheme was conducted during brain age prediction [34]. Specifically, this scheme relies on the slope (α) and intercept (β) of a linear regression model of PAD against chronological age from the training set. For each test sample with chronological age of Ω, the offset is defined in equation (1).

(1)

Then, the offset is subtracted from the individual estimated brain age to achieve a bias-free brain age estimation.

PAD group difference comparison

Group effects of PAD between each patient group and HC were estimated by computing the Cohen’s d effect sizes using a linear model [35] accounting for age, age², sex, and site (Fig 1c), as shown in equation (2).

(2)

where t denotes contrast statistics, n₁ and n₂ are the numbers of sample size in two groups, and df represents the degrees of freedom used for a corresponding t value in a linear model.

Interpretable brain patterns of PAD

Shapley Additive Explanations (SHAP) [36] is a game-theoretic approach for additive feature attribution that was used to determine the marginal contributions from predictions. Specifically, the changes in model outcome relative to the baseline can be apportioned to individual features, as shown in equation (3).

(3)

where f(x_i) represents the predicted brain age for a single testing sample x_i; E_train[f(x)] represents the fixed baseline (average chronological age in training set); M represents the number of input features and is the SHAP value which represents the contribution of feature j to the model prediction of participant i. Moreover, to mitigate potential bias arising from correlations among input features, SHAP values were computed using TreeSHAP with the interventional feature perturbation scheme, which anchors the perturbation process in the empirical data distribution rather than relying on the independence assumption used in conventional perturbation strategies, thereby reducing attribution bias when features are correlated [37]. Then, a SHAP value matrix with sample size × number of input features (n × M) for testing set was generated. We incorporate interaction terms (SHAP × group) in multiple linear regression models to identify the brain features for PAD difference in each diagnostic group, where the age-corrected PAD was the dependent variable, while age, age², sex, site, group, SHAP value, and SHAP × group were the independent variables. Subsequently, PAD difference T-map was obtained according to statistical t-values of the interaction coefficients. The brain regions with significant interaction coefficients (p_FDR < 0.05) were considered as the brain regions associated with the PAD difference (Fig 1d).

Correlation between PAD and symptoms

The linear models accounting for age, age², sex, and site were used to correlate PAD with symptoms in patient groups. Symptom scores are available for SZ including positive and negative scores of the Positive and Negative Syndrome Scale (PANSS); BP including Montgomery-Asberg Depression Rating Scale (MADRS) and Young Mania Rating Scale (YMRS); MDD including Hamilton Depression Rating Scale (HDRS); AUD including AUDIT; TUD including FTND, and AD and MCI including the Mini-Mental State Examination (MMSE). Then, the t-statistics of the linear models were converted to r, as shown in equation (4).

(4)

where df represents the degrees of freedom used for a corresponding t value in the linear model. Thus, correlation taking covariates into account reflect a partial correlation between PAD and symptoms.

Gene enrichment analyses

Post-mortem gene expression data from Allen Institute for Brain Science (AIBS) and PAD difference T-map were used to identify the biological function of genes relatively over- or underexpressed in association with this brain pattern. Specifically, we processed the regional microarray expression data from the Allen Human Brain Atlas (AHBA) [38] using the abagen toolbox (https://github.com/netneurolab/abagen) to map the transcriptomic data onto 1,016 parcellated brain regions from the augmented Schaefer-1016 (details on transcription data processing can be found in Text D in S1 Appendix) to generate regional gene expression matrix (1,016 ROIs × 15,632 genes) as the predictive variable. PAD difference T-map (1,016 ROIs × 1) of each diagnostic group was used as response variable to identify the gene expression component related to these T-maps by partial least square (PLS) regression. The statistical significance of the variance explained by each component was assessed by permuting the response variables 1,000 times. The significance component with the highest explained variance for PAD difference T-map (PLS_T) was selected for subsequent analysis. The error in estimating PLS_T weight for each gene was evaluated by bootstrapping, and the ratio of each gene’s weight to its bootstrap standard error was used to calculate the Z scores. Genes with Z > 3 and Z <−3 (p_FDR < 0.05) were defined as PLS_T+ and PLS_T− gene sets, respectively. Metascape [39] (https://metascape.org/gp/index.html) was used to identify gene ontology (GO) terms that significantly enriched in these gene sets. Redundancy elimination and visualization of GO term lists (REViGO) [40] was used to summarize and visualize gene ontology (http://revigo.irb.hr) by removing redundant terms (Fig 1e).

Artificial intelligence statement

We confirm that no artificial intelligence tools or technologies were used in the conduct of the study or in the preparation of the article content.

This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 Checklist). This study had no prospective protocol or analysis plan.

The performance of brain age prediction

The sample size and age distributions of training and testing sets for each diagnostic group are shown in Fig 2a, with no group differences (p = 0.79–0.99, two-sample t test) in age distribution between patient and HC. For each diagnostic group, an independent XGBoost model was built to predict brain age from averaged GMV extracted from augmented Schaefer-1016 brain atlas (1,000 cortical and 16 subcortical ROIs) through 10-fold cross-validation (details can be found in “Brain age prediction with age correction” section). Results showed excellent predictive performance of the model (without age correction: r = 0.68–0.88, mean absolute error-MAE = 1.65–4.97, coefficient of determination-R² = 0.45–0.76; after age correction: r = 0.89–0.93, MAE = 1.15–3.75, R² = 0.76–0.84, Fig C in S3 Appendix), with consistently high performance and low coefficient of variation across folds (without age correction:0.9%–12.1%, after age correction:0.5%–10.5%, Tables E and F in S2 Appendix, details can be found in Text E in S1 Appendix), indicating the effectiveness of the brain age prediction model and highlighting the necessity of age-bias correction. Then, the group-specific XGBoost models were applied to the corresponding independent 9 diagnostic groups to generate the predicted brain age, which was then correlated with age. Results showed that the correlation between predicted brain age and chronological age, as well as the MAE and R², were consistently better in HC than in patients across all clinical groups (r = 0.75, MAE = 1.56, R² = 0.57 for ADHD, and r = 0.77 MAE = 1.37, R² = 0.60 for HC; r = 0.82, MAE = 3.26, R² = 0.63 for ASD and r = 0.83, MAE = 3.09, R² = 0.67 for HC; r = 0.85, MAE = 6.10, R² = 0.59 for SZ and r = 0.89, MAE = 4.56, R² = 0.75 for HC; r = 0.88, MAE = 5.21, R² = 0.72 for BP and r = 0.89, MAE = 4.56, R² = 0.75 for HC; r = 0.79, MAE = 5.24, R² = 0.58 for MDD and r = 0.83, MAE = 4.64, R² = 0.66 for HC; r = 0.75, MAE = 4.52, R² = 0.51 for AUD and r = 0.89, MAE = 3.60, R² = 0.75 for HC; r = 0.79, MAE = 4.67, R² = 0.59 for TUD and r = 0.89, MAE = 4.26, R² = 0.69 for HC; r = 0.79, MAE = 4.74, R² = 0.52 for A&TUD and r = 0.84, MAE = 3.22, R² = 0.70 for HC; r = 0.88, MAE = 3.36, R² = 0.71 for AD and r = 0.90, MAE = 2.59, R² = 0.77 for HC; r = 0.89, MAE = 2.81, R² = 0.70 for MCI and r = 0.91, MAE = 2.59, R² = 0.79 for HC, Fig 2b). We also investigated the impact of sample size on the performance of brain age prediction and observed that the larger training sample size the better performance and the smaller variability (Fig D in S3 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Age distributions and the performance of brain age predictions.

(a) Age distributions of each training (up) and testing (down) sets. The gray shading under each diagnostic group represents its age-, sex-, and number-matched control group. P value represents the group difference of chronological age between the patient and HC calculated by two-sample t test. (b) Pearson correlations between the predicted brain age and chronological age, along with MAE and R², in independent testing sets. The solid line indicates the linear regression fit. ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; SZ, schizophrenia; BP, bipolar disorder; MDD, major depressive disorder; AUD, alcohol use disorder; TUD, tobacco use disorder; A&TUD, AUD and TUD; AD, Alzheimer’s disease; MCI, mild cognitive impairment; HC, healthy control; MAE, mean absolute error; R², coefficient of determination.

https://doi.org/10.1371/journal.pmed.1004860.g002

PAD difference for each diagnostic group

The PAD difference between patient and HC was calculated using linear regression accounting for age, age², sex, and site as covariates. The PAD was higher in all the diagnostic groups than HC (Fig 3a), in which strong difference effects was observed in AD (d = 0.97, 95% confidence interval (CI) [0.82,1.13]; p < 0.001*) and A&TUD (d = 0.84, 95% CI [0.44,1.23]; p < 0.001*), moderate effects in TUD (d = 0.72, 95% CI [0.49,0.96]; p < 0.001*), AUD (d = 0.62, 95% CI [0.39,0.84]; p < 0.001*), SZ (d = 0.53, 95% CI [0.30,0.76]; p < 0.001*), BP (d = 0.46, 95% CI [0.22,0.69]; p < 0.001*) and MCI (d = 0.45, 95% CI [0.34,0.56]; p < 0.001*), small effects in MDD (d = 0.28, 95% CI [0.11,0.46]; p < 0.001*), and negligible effects in ASD (d = 0.06, 95% CI [−0.04,0.16]; p = 0.36) and ADHD (d = 0.01, 95% CI [−0.14,0.15]; p = 0.98). In addition, fraction amplitude of low-frequency fluctuation (fALFF) from functional MRI (fMRI) was also used to predict brain age under the same prediction framework (Details on fMRI preprocessing pipeline can be found in Text B in S1 Appendix). Results showed that no PAD difference was observed in almost all the diagnostic groups (Fig E in S3 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. PAD difference by diagnosis, age, and sex.

(a) PAD difference between patient and HC. Cohen’s d accounting for age, age², sex and site, 95% confidence interval (CI), and two-sided p values are provided. (b) PAD difference between patient and HC in different age subgroups. (c) PAD difference between male and female. All p values are calculated using linear model-based t-tests. (*) is used to indicate results which pass false discovery rate (FDR) correction for multiple comparisons. ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; SZ, schizophrenia; BP, bipolar disorder; MDD, major depressive disorder; AUD, alcohol use disorder; TUD, tobacco use disorder; A&TUD, AUD and TUD; AD, Alzheimer’s disease; MCI, mild cognitive impairment; HC, healthy control; CI, confidence interval.

https://doi.org/10.1371/journal.pmed.1004860.g003

The PAD difference between patient and HC in different age range was also investigated by dividing into “young,” “adult,” and “old” subgroups (Fig F in S3 Appendix and Tables G–I in S2 Appendix). Results (Fig 3b) showed that increased PAD relative to HC was most prominently observed in “young” dementia (including AD: d = 1.45, 95% CI [1.16,1.74]; p < 0.001* and MCI: d = 0.59, 95% CI [0.40,0.79]; p < 0.001*) and AUD (d = 0.82, 95% CI [0.41,1.24]; p < 0.001*), and “adult” psychiatric disorders (including SZ: d = 0.72, 95% CI [0.32,1.12]; p < 0.001*, BP: d = 0.48, 95% CI [0.07,0.88]; p = 0.03 and MDD: d = 0.49, 95% CI [0.18,0.80]; p = 2.31e − 03*) and addiction (including TUD: d = 0.82, 95% CI [0.40,1.23]; p < 0.001* and A&TUD: d = 0.99, 95% CI [0.29,1.69]; p = 0.01), with no PAD difference in “young” BP (d = 0.19, 95% CI [−0.22,0.61]; p = 0.39), “old” MDD (d = 0.10, 95% CI [−0.21,0.41]; p = 0.55) and developmental disorders (including ADHD and ASD). We further examined age × group interaction effects on PAD (details can be found in Text F in S1 Appendix) and identified significant interactions for AD and MCI (Table J in S2 Appendix), with estimated divergence ages of approximately 57 years for AD and 68 and 80 years for MCI (Fig G in S3 Appendix). Moreover, PAD difference by sex was also calculated. Results (Fig 3c) showed that female PAD was significantly higher than male in SZ (d = 0.64, 95% CI [0.30,0.98]; p < 0.001*) and AUD (d = 0.62, 95% CI [0.26,0.97]; p < 0.001*), while lower than male in AD (d=−0.43, 95% CI [−0.64, −0.22]; p < 0.001*), with no sex difference in other brain disorders.

PAD difference sequence pattern validations across datasets, atlas resolutions, and prediction models

We validated the robustness of PAD difference sequence pattern among diagnostic categories (dementia > addiction > psychiatric disorders > developmental disorders) when using different training sets (full age range HCP), scales of brain atlas (Schaefer-216/516/816) and prediction models (support vector regression: SVR; back propagation neural network: BPNN; and random forest: RF). Details on the implementation of all these datasets, atlases, and predictive models can be found in Text G in S1 Appendix. Ten-fold cross-validation confirmed the predictive performance of all models (Fig H in S3 Appendix), which were used to predict the brain age in the diagnostic groups, followed by calculating the PAD difference between patient and HC. The Spearman rank correlation coefficient β was used to evaluate the consistency of the PAD difference sequence pattern among diagnostic categories across datasets, brain atlas resolutions, and prediction models (details can be found in Text H in S1 Appendix). Results showed that the identified PAD difference sequence pattern among diagnostic group were highly consistent across training sets (β = 0.90, p = 8.80e − 04, Fig 4a and Table K in S2 Appendix), brain atlas resolutions (β = 1.00/0.98/0.99 and all p < 1.00e − 16 for AS-216/516/816, Fig 4b and Table K in S2 Appendix) and prediction models (β = 0.83/0.91/0.88 and p = 3.18e − 03/3.07e − 04/1.98e − 03 for SVR/BPNN/RF, Fig 4c and Table K in S2 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. PAD difference sequence pattern validation across datasets, atlas resolutions, and different prediction models.

(a) Full age range HCP was used as training sets to test PAD difference. (b) PAD difference under different scales of brain atlas, where Schaefer-216/516/816 were used to extract GMV. (c) PAD difference across different prediction models, where SVR, BPNN, and RF were used to predict brain age. Color bar indicates Cohen’s d accounting for age, age², sex, and site. All p values are calculated using linear model-based t-tests. Significant differences (p < 0.05) are marked with a black box. HCP, Human Connectome Project; AS, augmented Schaefer; SVR, support vector regression; BPNN, back propagation neural network; RF, random forest; ADHD, attention-deficit/hyperactivity disorder; ASD, autism spectrum disorder; SZ, schizophrenia; BP, bipolar disorder; MDD, major depressive disorder; AUD, alcohol use disorder; TUD, tobacco use disorder; A&TUD, AUD and TUD; AD, Alzheimer’s disease; MCI, mild cognitive impairment.

https://doi.org/10.1371/journal.pmed.1004860.g004

Brain patterns associated with the PAD difference

To identify brain features contributed to PAD difference in each diagnostic group (ADHD and ASD were excluded due to no PAD difference, A&TUD were excluded to prevent redundancy), we first calculated the SHAP value for each of the 1,016 features. Then, we estimated a multiple linear regression model with age-corrected PAD as the dependent variable, and age, age², sex, site, group, SHAP value, and the interaction term SHAP × group as independent variables. The SHAP × group interaction captures the influence of each brain feature on PAD difference between patient and HC, and PAD difference T-maps were derived from the t-statistics of the corresponding interaction coefficients. PAD difference T-maps for each diagnostic group were displayed in Fig 5a with false discovery rate (FDR) corrected in Fig 5b, whose reproducibility across different scales of brain atlas was validated (Fig I in S3 Appendix). Prefrontal cortex (PFC), middle temporal cortex (MTC), and inferior temporal cortex (ITC) were the shared brain regions with significant SHAP × group interaction within psychiatric group. PFC, superior/middle/inferior temporal cortex (STC/MTC/ITC), superior/middle occipital cortex (SOC/MOC), thalamus, and insula were identified in SZ; PFC, MTC, ITC, and lingual gyrus (LG) in BP; and PFC, MTC, ITC, and MOC in MDD. Within addiction group, PFC, precentral cortex (PC), cingulate cortex (CC), insula, putamen, thalamus, and temporal cortex (TC) were the identified shared brain regions. PFC, PC, CC, insula, putamen, thalamus, STC, MTC, ITC, MOC, inferior occipital cortex (IOC), LG, superior parietal cortex (SPC), and fusiform gyrus (FG) were identified in AUD; PFC, PC, CC, insula, putamen, thalamus, STC, MTC, and ITC in TUD. Within dementia group, PFC, MOC, and FG were the shared brain regions. PFC, SOC, MOC, FG, hippocampus, para-hippocampus, MTC, ITC, PC, anterior cingulate cortex (ACC), amygdala, caudate, insula, and thalamus were identified in AD; PFC, MOC, and FG for MCI. The identified brain regions are summarized in Tables L–N in S2 Appendix. Note that PFC was consistently identified for PAD difference across psychiatric disorders, addiction, and dementia. Moreover, the bootstrapping analysis showed that the identified brain features consistently exhibited high selection frequency across 1,000 resampling conditions, indicating their robustness and stability (details on bootstrapping analysis can be found in Text I in S1 Appendix and Fig J in S3 Appendix).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Brain patterns associated with the PAD difference.

(a) PAD difference T-maps for each diagnostic group. Color bar indicates statistical t-values of the interaction coefficients, and red/blue regions represent stronger positive/negative associations between the contribution of the identified brain features to predicting brain age and PAD in patient groups than in the HC. (b) The FDR corrected PAD difference T-maps. The deep red, deep purple, deep green, and black box represent the shared brain regions within psychiatric disorders (including SZ, BP, and MDD), addiction (including AUD and TUD), dementia (including AD and MCI), and among all patient groups. SZ, schizophrenia; BP, bipolar disorder; MDD, major depressive disorder; AUD, alcohol use disorder; TUD, tobacco use disorder; AD, Alzheimer’s disease; MCI, mild cognitive impairment; PAD, predicted age difference; FDR, false discovery rate; MTC, middle temporal cortex; ITC, inferior temporal cortex; PFC, prefrontal cortex; PC, precentral cortex; CC, cingulate cortex; PUT, putamen; INS, insula; THA, thalamus; TC, temporal cortex; MOC, middle occipital cortex; FG, fusiform gyrus.

https://doi.org/10.1371/journal.pmed.1004860.g005

Relationships between PAD and symptoms

Partial correlation between PAD and symptoms in each diagnostic group were calculated by linear models accounting for age, age², sex, and site. Results (Fig 6) showed that PAD was negatively correlated with MMSE in AD (r=−0.17, p = 0.02) and MCI (r=−0.24, p = 8.12e − 06), and with FTND in TUD (r=−0.21, p = 0.01). However, no association (p > 0.05) was observed between PAD and symptom in other brain disorders (including SZ, BP, MDD, and AUD). The high degree of correlations between the whole brain PAD (predicted by the whole brain 1,016 ROIs) and the identified brain PAD (predicted by the identified brain ROIs) in each diagnostic group (r = 0.73 for SZ, r = 0.49 for BP, r = 0.70 for MDD, r = 0.86 for AUD, r = 0.72 for TUD, r = 0.89 for AD, r = 0.42 for MCI) confirmed the effectiveness of the identified regions. The associations between the whole brain PAD and symptoms were largely replicated in the identified brain PAD, with negative association with MMSE in AD (r=−0.19, p = 5.52e − 03) and MCI (r=−0.27, p = 9.13e − 07) and no association in other brain disorders.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Correlation between PAD and symptoms.

Upper: correlation between the whole brain PAD (predicted by the whole brain 1,016 ROIs) and symptoms. Bottom: correlation between the identified brain PAD (predicted by the identified brain ROIs) and symptoms. All p values are calculated using linear model-based t-tests. Significant associations (p < 0.05) are marked with a black box. SZ, schizophrenia; BP, bipolar disorder; MDD, major depressive disorder; AUD, alcohol use disorder; TUD, tobacco use disorder; AD, Alzheimer’s disease; MCI, mild cognitive impairment; PAD, predicted age difference; ROIs, regions of interest; PANSS_pos/neg, positive and negative scores of the Positive and Negative Syndrome Scale; MADRS, Montgomery-Asberg Depression Rating Scale; YMRS, Young Mania Rating Scale; HDRS, Hamilton Depression Rating Scale; AUDIT, Alcohol Use Disorder Identification Test; FTND, Fagerström Test for Nicotine Dependence; MMSE, Mini‐Mental State Examination.

https://doi.org/10.1371/journal.pmed.1004860.g006

Enrichment analyses of genes related to PAD difference T-map

The enrichment analyses (details can be found in the “Gene enrichment analyses” section) were conducted to identify the biological function of genes relatively over- or underexpressed in association with PAD difference T-map. The first (the third for MDD) partial least squares (PLS) component explained 8.1%, 5.4%, 5.1%, 7.2%, 7.0%, 6.3%, and 5.3% variance in SZ, BP, MDD, AUD, TUD, AD, and MCI, which were not observed by chance (permutation test, p = 0.006 for SZ, p = 0.014 for BP, p = 0.048 for MDD, p = 0.019 for AUD, p = 0.046 for TUD, p = 0.047 for AD, and p = 0.048 for MCI, Fig K in S3 Appendix). PLS1 (PLS3 for MDD) scores were positively correlated with the PAD difference T-map in SZ (spatial permutation tests, r = 0.24, p_spin = 0.002), BP (r = 0.21, p_spin < 0.001), MDD (r = 0.22, p_spin < 0.001), AUD (r = 0.21, p_spin = 0.002), TUD (r = 0.22, p_spin = 0.001), AD (r = 0.23, p_spin < 0.001), and MCI (r = 0.22, p_spin < 0.001), respectively (Fig L in S3 Appendix). After correcting for enrichment terms (p_FDR < 0.05) and removing redundant terms, the significant Gene Ontology (GO) biological processes for PLS_T+ and PLS_T− gene sets enriched were shown in Fig 7a and 7b, in which no significant GO terms enriched in PLS_T+ gene set for BP. The top 8 significant GO terms (Tables O and P in S2 Appendix) were labeled for each diagnostic group (only 3 significant GO terms were identified in PLS_T− gene set for MCI). In PLS_T + , the cytoplasmic translation, ribonucleoprotein complex biogenesis, and translation were the shared significant GO terms in psychiatric disorders (SZ and MDD); carboxylic acid catabolic process, small molecule biosynthetic process, and small molecule catabolic process were the shared in addiction (AUD and TUD); protein localization to membrane was the shared in dementia (AD and MCI). For PLS_T − , the endocytosis, import into cell and membrane organization were the shared in psychiatric disorders (SZ, BP, and MDD); inorganic ion transmembrane transport and synaptic signaling were the shared in addiction (AUD and TUD); deoxyribonucleic acid (DNA) damage response and DNA metabolic process were the shared in dementia (AD and MCI).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Significant GO terms for biological processes related to PAD difference T-maps.

Significant GO terms for biological processes that the PLS_T+ (a) and PLS_T− (b) gene sets in each diagnostic group enriched were plotted in semantic space so that similar terms were represented close to each other. The top 8 significant GO terms were labeled for each analysis with redundant GO terms excluded by REViGO. Markers were scaled based on the log10 q-value (q represents the p-value after FDR correction) for the significance of each GO term. Large blue circles were highly significant, while small red circles were less so. SZ, schizophrenia; BP, bipolar disorder; MDD, major depressive disorder; AUD, alcohol use disorder; TUD, tobacco use disorder; AD, Alzheimer’s disease; MCI, mild cognitive impairment; GO, gene ontology.

https://doi.org/10.1371/journal.pmed.1004860.g007