Dissecting causal relationships between cortical morphology and neuropsychiatric disorders: a bidirectional Mendelian randomization study

Abstract

Brain cortical morphology, indexed by its surface area and thickness, is known to be highly heritable. Previous research has suggested a relationship of cortical morphology with several neuropsychiatric phenotypes. However, the multitude of potential confounders makes it difficult to establish causal relationships. Here, we employ Generalized Summary-data-based Mendelian Randomization and a series of sensitivity analyses to investigate causal links between 70 cortical morphology measures and 199 neuropsychiatric, behavioral, and metabolic phenotypes. We show that total brain cortical surface area (TSA) has significant positive causal effects on 18 phenotypes. The strongest effects include TSA positively influencing cognitive performance, while reverse analyses reveal small effects of cognitive performance on TSA. Global mean cortical thickness (MTH) exhibits significant causal effects on five phenotypes, including schizophrenia. MTH reduces schizophrenia risk and bidirectional causality is found between MTH and smoking initiation. Finally, in regional analyses we detect positive influences of the transverse temporal surface area on cognitive performance and negative influences of transverse temporal thickness on schizophrenia risk. Overall, our results highlight bidirectional associations between TSA, MTH, and neuropsychiatric traits. These insights offer potential avenues for intervention studies aimed at improving brain health.

Introduction

Cortical thickness and surface area are highly heritable^1,2 and are strongly associated with neuropsychiatric phenotypes, including substance use disorders.^3–7 To improve our understanding of neurobiological determinants and neuropsychiatric disorders, it is critical to dissect the interplay between cortical morphology and neuropsychiatric phenotypes.

Establishing causality between brain measures and neuropsychiatric phenotypes is challenging due to numerous possible confounding factors, e.g., medication use⁸ and comorbid conditions.⁹ Mendelian Randomization (MR) may overcome several challenges pertinent to other lines of research, such as reverse causation, confounding and other types of bias.¹⁰ Of particular relevance for neuropsychiatric disorders is the possible impact of (psychotropic) medication on brain morphology, ^11,12 which is circumvented in MR.

Recent MR studies have examined causal links between brain measures and psychiatric disorders, including substance use.^13,14 One study identified brain measures with causal influences on the risks of schizophrenia (SCZ), anorexia nervosa, and bipolar disorder in the UK Biobank (UKB).¹¹ Another MR study detected negative associations between global cortical thickness and alcohol drinking behaviour.¹² Considering the importance of cortical morphology for cognitive functioning and the previously detected associations of cortical morphology with neuropsychiatric traits,^15,16 we set out to further elucidate its role and how it is related to mental health.^15–17

Given the co-occurrence of inflammation, metabolite changes, and type 2 diabetes mellitus (T2D) with neuropsychiatric phenotypes,^9,18,19 it is of interest to examine cortical morphology changes in the context of metabolic phenotypes. For example, inflammation may contribute to structural brain changes through the activation of microglia and/or astrocytic dysfunction in various neuropsychiatric disorders.²⁰ Moreover, alterations in cortical thickness have been observed in middle-aged patients with T2D, which may be related to diabetes-related brain damage.²¹ Additionally, cortical thinning was observed in relation to body mass index (BMI) and visceral adipose tissue (VAT), likely involving a mechanism of adipose tissue-related low-grade inflammation.²² Other findings suggest that C-reactive protein (CRP) and vitamin D exert causal effects on region-specific cortical thickness and point to a negative causal relationship between serum CRP levels and the thickness of the lingual region.²³ Therefore, examining causal relationships between a range of metabolic traits and cortical morphology could provide clues about the underlying mechanisms of various (neuropsychiatric) disorders.

Here, we hypothesized that specific cortical morphology measures have causal relationships with neuropsychiatric, behavioral, and metabolic phenotypes. By leveraging the availability of recently published large genome-wide data, we aimed to shed light on causal mechanisms using a thorough and stringent methodology, namely Generalized Summary-data-based Mendelian Randomization (GSMR). By incorporating the heterogeneity in dependent instruments (HEIDI) method,^24,25 GSMR is powerful and conservative in identifying and removing pleiotropic instrumental variables.²⁶ In addition, we used cortical morphology data from Enhancing Neuro-Imaging Genetics through Meta-Analysis (ENIGMA),²⁷ which has several advantages over the use of genome-wide association studies (GWASs) based only on UK Biobank participants. For example, ENIGMA is a brain imaging-focused study that collected data from many sites across the globe, thus increasing the chances of generalizability.²⁸ Finally, by including an unprecedented set of neuropsychiatric, behavioral, and metabolic phenotypes, we aimed to refine the understanding of directions of effect between cortical morphology and such phenotypes. We thus searched for any bidirectional causal relationships between 70 cortical morphology measures (both global and regional measures)²⁸ and 199 phenotypes spanning neuropsychiatric, behavioral and metabolic phenotypes. We highlight several new unidirectional and bidirectional relationships that extend our understanding of the relationships between neuropsychiatric disorders and cortical morphology. The findings provide actionable insights that could inform the development of targeted lifestyle interventions for neuropsychiatric disorders.

Results

Overview: analysis set-up & genetic instruments

We performed forward GSMR analyses of total surface area (TSA) and mean cortical thickness (MTH, N=~0.5 million subjects) on the 199 phenotypes explained in the methods section (N=~13.5 million subjects; Figure 1; also see STable 1–2, Supplementary tables ST1–ST5. Of those phenotypes, 172 fulfilled the criterion of two or more instrument variables (IVs), allowing us to also perform reverse GSMR analyses on cortical morphology measures (STable 2, Supplementary tables ST3, ST5), along with the respective sensitivity analyses for these 172 phenotypes (Supplementary tables ST6, ST7). As the instrument strength was strong for all phenotypes tested (F-statistics in forward and reverse MR analyses ranging from 28.11 to 98.89), we did not find any evidence of weak instrument bias. For all 199 GSMR tests of MTH effects, the number of genetic instruments was ≥10, except for 9 tests for which the numbers of instruments ranged from 7–9 (Supplementary tables ST4). Out of these four tests, two were significant: the protective effects of MTH on SCZ (using 9 instruments) and on number drinks per week (using 8 instruments). Finally, in the regional analyses, we observed that the global-corrected method yields less potential causal relationships than its non-global-corrected counterpart.

Figure 1. Schematic overview of our study design.

TSA: total surface area, MTH: mean cortical thickness, GSMR: Generalized Summary-data-based Mendelian Randomization, IVW: Inverse Variance Weighted fixed-effects method, MR-Egger method, WM: Weighted Median MR method, MR-PRESSO: MR-pleiotropy residual sum and outlier method.

Relationships between surface area and other phenotypes

Forward analyses

We identified significant causal effects of TSA on 18 phenotypes (Figure 2A, Supplementary Figure 1A). Further details on beta values, standard errors and p-values can be found in the Supplementary tables ST2, where the significant results are marked in bold. TSA had significant causal effects with a positive effect direction on eight outcomes, namely cognitive performance (CP),²⁹ educational attainment (EA),³⁰ height,³¹ cross disorders group (CDG),³² age of smoking initiation 2022 (AgeSmk#2),³³ high-density lipoprotein cholesterol (HDL),³⁴ bipolar disorder (BIP),³⁵ and age of smoking initiation 2019 (AgeSmk).³⁶ Furthermore, TSA had significant negative causal effects on 10 outcomes, namely smoking initiation 2022 (SmkInit#2),³³ log-transformed triglycerides (logTG),³⁴ smoking initiation 2019 (SmkInit),³⁸ attention-deficit/hyperactivity disorder (ADHD2019),³⁷ smoking cessation 2022 (SmkCes#2),³³ total cholesterol (TC),³⁴ non-high-density lipoprotein cholesterol (nonHDL),³⁴ Alzheimer’s disease (ALZ),³⁸ type II diabetes 2017 (T2D2017), and type II diabetes 2020 (T2D2020).³⁹ The three most significant effects were found for the following phenotypes: CP (β=0.56, se=0.04, p._fdr=8.2 × 10⁻⁴¹⁾, EA (β=0.40, se=0.03, p._fdr=9.1 × 10⁻³²), and height (β=0.35, se=0.04, p._fdr=2.7 × 10⁻¹⁵). A forest plot summarizes the above-mentioned findings using beta values in Figure 2A and odds ratios (ORs) in SFigure 1A.

Figure 2. Forest plots of GSMR models with significant results (p._fdr < 0.05) for ΤSA and ΜTH.

(A) Forward MR results of ΤSA on outcomes. (B) Reverse MR results of outcomes on ΤSA. (C) Forward MR results of ΜTH on outcomes. (D) Reverse MR results of outcomes on ΜTH.

Significant results (p._fdr < 0.05) are colored by category: red= behavior, purple=quantitative phenotype and blue=(neuro)psychiatric phenotype (except for T2D2017, TSD2020). The non-significant results (p._fdr > 0.05) are colored in grey. In Fig.2B, there were not enough significantly associated SNPs for age of smoking initiation 2019 (AgeSmk) to be extracted as instruments and hence the result is unavailable.

(A & B) CP: cognitive performance, EA: educational attainment, Height, CDG: cross disorders group, AgeSmk#2: age of smoking initiation (2022), HDL: high-density lipoprotein cholesterol, BIP#2: bipolar disorder excluding participants from the UK biobank (2021), AgeSmk: age of smoking initiation (2019), T2D2020: type II diabetes (2020), T2D2017: type II diabetes (2017), ALZ: Alzheimer’s disease, nonHDL: non-high-density lipoprotein cholesterol, TC: total cholesterol, SmkCes#2: smoking cessation (2022), ADHD: attention deficit hyperactivity disorder, SmkInit: ever smoked regularly (2019), logTG: log-transformed triglycerides, SmkInit#2: ever smoked regularly (2022)

(C & D) SCZ3: schizophrenia freeze 3, DrnkWk#2: drinks per week (2022), SCZ2: schizophrenia freeze 2, SmkInit#2: ever smoked regularly (2022), BDSCZ: meta-analysis bipolar disorder and schizophrenia, BIP: bipolar disorder (2019), BMI: body mass index, Bis.DB.ratio: ratio of bisLallylic bonds to double bonds in lipids, CH2.in.FA: CH2 groups in fatty acids, Bis.FA.ratio: ratio of bisLallylic bonds to total fatty acids in lipids, otPUFA: other polyunsaturated fatty acids than 18:2, DB.in.FA: double bonds in fatty acids, Crea: creatinine, CH2.DB.ratio: CH2 groups to double bonds ratio.

Of note: The dots (centre for the error bars) represent beta coefficients in MR study. The error bars represent 95% Confidence Intervals of beta coefficients. In multivariable (MVMR) analyses, the result for BIP had the same direction of effect but was no longer significant after FDR-correction for multiple testing (STable 7B). Furthermore, 2 tests had <10 instruments, namely SCZ3 (number of instruments = 9) and DrnkWk#2 (number of instruments = 8). Statistical tests were two-sided, and p-values were FDR-corrected to adjust for multiple testing.

Regarding the sample size of GWASs, ΤSA, MTH= 51,665; CP=257,828; EA= 3 million, Height=253,288, CDG= 438,997; HDL, nonHDL, TC, logTG =1,320,016; BIP#2= 413,466; AgeSmk, SmkInit = 1.2 million; T2D2020=23,326; T2D2017=659,316; ALZ=200,853; SmkCes#2, SmkInit#2, DrnkWk#2= 3.4 million; ADHD= 53,293; SCZ3=161,405; SCZ2=105,308; BDSCZ=41,653; BIP=31,710; BMI=322,154; Bis.DB.ratio, CH2.in.FA, Bis.FA.ratio, otPUFA, DB.in.FA, Crea, CH2.DB.ratio= 24,925.

Reverse analyses

We identified two phenotypes - height and cognitive performance - with significant causal effects on TSA (Figure 2B for beta values and SFigure 1B for ORs). We thus detected positive bidirectional causal relationships between TSA-CP (forward: β=0.56, 95%CI=0.48–0.64, se=0.04, p._fdr=8.2 × 10⁻⁴¹, reverse: β=0.07, 95%CI=0.05–0.09, se=0.01, p._fdr=3.53 × 10⁻¹²) and TSA-height (forward: β=0.35, 95%CI=0.27–0.34, se=0.04, p._fdr=2.7 × 10⁻¹⁵, reverse: β=0.05, 95%CI=0.04–0.06, se=0.004, p._fdr=1.7 × 10⁻²³). Statistical details can be found in Supplementary tables ST3.

Relationships between cortical thickness and other phenotypes

Forward analyses

MTH had significant negative causal effects on five outcomes (see Figure 2C for beta values and SFigure 1C for ORs), namely schizophrenia (SCZ freeze 3 (SCZ3)⁴⁰ and SCZ freeze 2 (SCZ2)⁴¹), alcoholic drinks consumed per week 2022 (DrnkWk#2),³³ smoking initiation 2022 (SmkInit#2)³³ and bipolar disorder and SCZ considered together as one phenotype (BDSCZ).⁴² The most significant result was the causal effect of MTH on SCZ3 (β=−0.42, 95%CI=−0.54 – −0.29, se=0.06, p._fdr=1.02 × 10⁻⁸; statistical details in Supplementary tables ST4).

Reverse analyses

We identified 10 significant phenotypes with causal effects on MTH (Figure 2D for beta values, SFigure 1D for ORs). We detected positive causal effects of blood metabolites (Bis.DB.ratio, Bis.FA.ratio, DB.in.FA, otPUFA, Crea)⁴³ and bipolar disorder (BIP)³⁵ on MTH, along with negative causal effects of smoking initiation 2022 (SmkInit#2)³³, CH2.DB.ratio,⁴³ CH2.in.FA,⁴³ and BMI⁴⁴ on TH. The most significant result was found for the ratio of bisLallylic bonds to double bonds in lipids (Bis.DB.ratio, β=0.12, 95%CI=0.07–0.17, se=0.026, p._fdr=0.001). The full names of all abbreviated metabolic traits are in STable 3. Based on the forward and reverse analyses, significant negative bidirectional causal effects between MTH and smoking initiation were established (forward: β=−0.32, 95%CI=−0.44– −0.22, se=0.06, p._fdr= 1.5 × 10⁻⁵, reverse: β=−0.04, 95%CI=−0.06– −0.02, se=0.01, p._fdr=0.01; statistical details in Supplementary tables ST5). Furthermore, to detect whether the causal effects of smoking initiation and bipolar disorder still existed when accounting for BMI and the significantly associated blood metabolites mentioned in this paragraph, two multivariable MR models were conducted. Smoking initiation showed a robust relationship with SmkInit#2 (β=−0.03, 95%CI=−0.08 – −0.004, se=0.01, P= 0.007) in the MVMR-IVW model, whereas bipolar disorder causal effects on MTH disappeared (P=0.084) when jointly analyzed with the BMI and candidate blood metabolites (Supplementary tables ST 7B).

Results from the sensitivity analyses

Most of the sensitivity MR analyses confirmed the directions of effects detected using GSMR (Figure 3). We first discuss the results of the additional MR models that we ran as sensitivity analyses. The statistics pertinent to those results -including ORs, confidence intervals, p-values, F-statistics, I² statistics and residual heterogeneity by Cochran’s Q- test can be found in Supplementary tables ST6 and ST7 (any non-significant results are highlighted in grey). As these were confirmatory analyses, we checked for consistency in the directions of effects rather than significance, so we did not correct for multiple testing. When using Inverse-Variance Weighted (IVW) and Pleiotropy RESidual Sum and Outlier (PRESSO) MR models, we obtained similarly significant causal effects (same direction of effect and nominal p values < 0.05) for all significant results from GSMR. The weighted median method, even though it yielded some non-significant results, was also consistent in all directions of effects. When using the MR Egger model, however, which is the most conservative among the models used here,⁴⁵ we observed non-significant results and several cases where the direction of effect deviated from the consensus (i.e., the direction determined by the rest of the models, see ST6, ST7A, where non-consistent MR Egger results are written in red). All those cases were nominally non-significant (p > 0.05) and therefore did not affect the reliability of our results or our conclusions (STable 4). Additionally, with regards to the I² statistics of MR-Egger, there was high variability, ranging from 0 to 81.6 for TSA- and 0 to 96.7 for MTH-related causal associations. In general, for MTH we observed either very high heterogeneity (I² > 90%) or no heterogeneity at all (I²=0), whereas for TSA the I² statistics were relatively moderate, typically ranging around 65%. High variability in I² statistics suggests that the studies used for those associations have diverse results or effect sizes, which might be attributed to differing study designs or populations. Thus, overall, except for some divergent, non-significant results from the MR-Egger and some convergent, non-significant results from the weighted median, the sensitivity MR methods we employed mainly yielded similar significant results as the GSMR model.

Figure 3. Scatter plots of bidirectional MR analyses using several models to examine causal relationships between total surface area (TSA) and cognitive performance (CP)²⁹ (Fig. 3A & 3B), as well as global mean cortical thickness (MTH) and schizophrenia (freeze3, SCZ3)⁴⁰ (Fig. 3C & 3D).

A. Total SA on cognitive performance.

B. Cognitive performance on total SA.

C. Mean TH on schizophrenia.

D. Schizophrenia on mean TH.

(A) Forward MR results of TSA on CP. (B) Reverse MR results of CP on TSA. Based on Fig.3A and 3B, TSA and CP show a bidirectional causal relationship. (C): Forward MR results of MTH on SCZ3. (D): Reverse MR results of SCZ3 on MTH. Based on Fig. 3C and 3D, there is only a forward causal relationship between MTH and SCZ3. Note: The error bars represent 95% Confidence Intervals of beta coefficients. The five models applied are all denoted: Black: GSMR, red: fixed-effect IVW, green: weighted median, blue: MR Egger, pink: MR PRESSO. The first method served as our main analysis, while the rest of the methods were used for sensitivity analyses. Before using the instruments, we detected and removed outliers with the HEIDI test. SA: total cortical surface area, TH: global mean cortical thickness, CP: cognitive performance,²⁹ SCZ3: schizophrenia PGC GWAS freeze3.⁴⁰

Regarding the sample size of GWASs, ΤSA, MTH= 51,665; CP=257,828; SCZ3=161,405.

Next, when applying the P-value threshold of P<10e-7 for inclusion of genetic instruments to all of the significant findings in GSMR analyses, for the TSA forward analyses, 3 out of 18 causal effects had the same direction of effect but were no longer significant: CDG, nonHDL, and TC (Supplementary tables ST2). All other effects had the same direction of effect and remained significant after correction for multiple testing. In TSA reverse analyses, both positive significant causal effects of height and cognitive performance remained significant (Supplementary tables ST3). Additionally, all 5 causal effects of MTH were confirmed with the new threshold, while 9 out of 10 reverse causal effects of MTH remained similarly significant (Supplementary tables ST3&4), except for the reverse causal effects of MTH on bipolar disorder (p=0.10). As for sub-regional results, the majority of the results were the same when using the new threshold (Supplementary tables ST8–10). However, some results, particularly those related to MTH subregional corrected global thickness (Supplementary tables ST11), were no longer significant due to limited statistical power in the underlying GWASs.

Furthermore, the Cochran’s Q test in the fixed-effect IVW model and the MR-Egger model suggested that in some cases there was significant (p < 0.05) heterogeneity in the instrumental variables, which may be due to violation of instrumental variable assumptions. However, such findings also depend on the number of the extracted genetic instruments, so significant results in the Cochran’s Q test may also be attributed to high statistical power.

Finally, leave-one-out analyses (SFigure 4 and SFigure 5) showed that no SNPs altered the pooled IVW beta coefficient, confirming the stability of our results.

Brain regions and neuropsychiatric phenotypes

Forward analyses for SA regions

For non-global-corrected SA (Figure 4, Supplementary tables ST8), 27 regional SA phenotypes had positive causal influences on CP,²⁹ with insula as the most significant result (β=0.48, 95%CI= 0.41–0.55, se=0.04, p._fdr= 2.11 × 10⁻³⁴). After global correction (Figure 3, Supplementary tables ST9), only two regions were causally associated with a higher CP, namely insula (β=0.28, 95%CI= 0.19–0.38, se=0.05, p._fdr=5.3 × 10⁻⁵) and transverse temporal surface area (β=0.24, 95%CI= 0.14–0.34, se=0.05, p._fdr=0.002).

Figure 4. Regional plots for SA with cognitive performance (CP).

A. Beta coefficients of MR results of regional SA on cognitive performance (CP).

B. −log10 (p._fdr values) of MR results of regional SA on cognitive performance (CP).

C. Beta coefficients of MR results of cognitive performance (CP) on regional SA.

D. −log10 (p._fdr values) of MR results of cognitive performance (CP) on regional SA.

Visualization in a plot of the Desikan-Killiany atlas, right hemisphere (upper) and left hemisphere (lower). Global-corrected (left) and non-global-corrected (right) MR results between SA regional measures and cognitive performance (CP). (A) Beta coefficients of MR results of regional SA on cognitive performance (CP). (B) −log10 (p._fdr values) of MR results of regional SA on cognitive performance (CP). (C) Beta coefficients of MR results of cognitive performance (CP) on regional SA. (D) −log10 (p._fdr values) of MR results of cognitive performance (CP) on regional SA. (A & C): the color intensity represents the strength of the causal association via beta coefficients (red: strong positive, blue: strong negative). (B & D): −log10 p-value after FDR-correction: the lighter the color blue, the more statistically significant the result. Statistical tests were two-sided, and p-values were FDR-corrected to adjust for multiple testing. Regarding the sample size of GWASs, SA= 51,665; CP=257,828.

Out of the 18 previously identified outcomes that are causally influenced by total SA (Figure 1A, SFigure 1A, Supplementary tables ST2), nine did not show any significant results in the respective regional analyses after global correction (Supplementary tables ST9), namely height,³¹ ADHD,³⁷ Alzheimer’s disease (ALZ),³⁸ age of smoking initiation 2019³⁶ and 2022³³ (SmkInit and SmkInit#2), non-high-density-lipoprotein (nonHDL),³⁴ total cholesterol (TC),³⁴ type II diabetes 2017³⁸ and 2020³⁹ (T2D2017, T2D2020). For the remaining nine outcomes, we detected some region-specific causal influences, as demonstrated in Table 1. The full results can be found in Supplementary tables ST10.

Table 1.

Top ten significant results of the global-corrected regional MR analyses between SA regional measures and the selected phenotypes.

exposure	outcome	nsnps	β	95% CI	se	p.fdr
insula	CP	14	0.28	0.19 0.38	0.05	5.30 × 10−5
precuneus	HDL	22	−0.21	−0.28 −0.13	0.04	5.41 × 10−5
parahippocampal	SmkInit	8	0.37	0.24 0.50	0.07	5.41 × 10−5
lingual	logTG	17	−0.22	−0.30 −0.14	0.04	10.02 × 10−5
precuneus	EA	21	−0.18	−0.26 −0.11	0.07	1.83 × 10−3
transverse temporal	CP	14	0.24	0.14 0.34	0.05	1.90 × 10−3
lateralorbitofrontal	SmkInit	19	−0.18	−0.26 −0.10	0.04	4.48 × 10−3
parahippocampal	logTG	6	0.34	0.19 0.50	0.08	6.01 × 10−3
rostralmiddlefrontal	SmkInit#2	16	−0.19	0.01 0.03	0.05	0.03
temporal pole	CDG	1	−0.90	−0.28 −0.10	0.23	0.03

The results are ranked by significance, i.e., the FDR-corrected p-value (p._fdr). nsnps: number of the extracted genetic instruments, i.e., the independent (r² < 0.01) SNPs, β: beta value, se: standard error. Outcomes: CP: cognitive performance, HDL: high-density lipoprotein, SmkInit: smoking initiation 2019 (ever smoking regularly 2019), logTG: log-transformed triglycerides, EA: educational attainment, SmkInit#2: smoking initiation 2022 (ever smoking regularly 2022), CDG: cross disorders group. Statistical tests were two-sided, and p-values were FDR-corrected to adjust for multiple testing.

Reverse analyses for SA regions

In the reverse, non-global-corrected analyses, we detected CP causal effects on 33 SA regions (Supplementary tables ST8). However, after global surface area correction (Supplementary tables ST9), all these effects were absent and the only significant reverse causal associations were height³¹ on supramarginal SA (β=0.02, 95%CI=0.01–0.03, se=0.005, p._fdr=0.022) and educatioxnal attainment EA³⁰ on inferior parietal SA (β=0.06, 95%CI=0.04–0.09, se=0.015, p._fdr=0.037), as marked in bold in ST9. Neither of these associations are particularly strong (very small beta values). The full results can be found in Supplementary tables ST10.

Forward analyses for TH regions

Before global correction (Supplementary tables ST10), three regional thickness phenotypes (namely caudal middle frontal, middle temporal and inferior parietal) increased the risk of schizophrenia freeze3 (SCZ3),⁴⁰ whereas three regional thickness phenotypes (namely caudal anterior cingulate, inferior temporal and fusiform) decreased such risk (Figure 5). As for schizophrenia freeze2 (SCZ2),⁴¹ it was causally influenced by only one regional measure, namely caudal middle frontal (SFigure 6). Comparing the results of those two data freezes demonstrates that the more data are available, the more causal associations can be identified, as expected (Supplementary tables ST10). However, after global thickness correction (Supplementary tables ST11), only transverse temporal thickness was associated with a lower risk of schizophrenia (SCZ3) (β=−1.96, 95%CI=−2.30– −0.92, se=0.44, p._fdr= 0.006). Furthermore, we detected negative causal effects of superior frontal TH and inferior temporal TH on alcoholic drinks per week 2022 (DrnkWk#2),³³ along with negative causal effects of cuneus TH on BDSCZ.⁴² We did not detect any specific regional effects on smoking initiation 2022 (SmkInit#2)³³ (SFigure 7). The full results can be found in Supplementary tables ST11.

Figure 5. Regional plots for TH with schizophrenia freeze 3 (SCZ3).

A. Beta coefficients of MR results of regional TH on schizophrenia (SCZ3).

B. log10 (p._fdr values) of MR results of regional TH on schizophrenia (SCZ3).

C. Beta coefficients of MR results of schizophrenia (SCZ3) on regional TH.

D. −log10 (p._fdr values) of MR results of schizophrenia (SCZ3) on regional TH.

Visualization in a plot of the Desikan-Killiany atlas, right hemisphere (upper) and left hemisphere (lower). Global-corrected (left) and non-global-corrected (right) MR results between TH regional measures and SCZ3. (A) Beta coefficients of MR results of regional TH on schizophrenia freeze 3 (SCZ3). (B) −log10 (p._fdr values) of MR results of regional TH on schizophrenia freeze 3 (SCZ3). (C) Beta coefficients of MR results of schizophrenia freeze 3 (SCZ3) on regional TH. (D) −log10 (p._fdr values) of MR results of schizophrenia freeze 3 (SCZ3) on regional TH. (A & C): the color intensity represents the strength of the causal association via beta coefficients (red: strong positive, blue: strong negative). (B & D): −log10 p-value after FDR-correction: the lighter the color blue, the more statistically significant the result. Statistical tests were two-sided, and p-values were FDR-corrected to adjust for multiple testing. Regarding the sample size of GWASs, TH= 51,665; SCZ3=161,405.

Reverse analyses for TH regions

We did not detect any phenotypic causal effects on any specific regional TH. The full results can be found in Supplementary tables ST11.

Discussion

Here, by using data from over 13.5 million participants, we provide evidence that cortical morphology has causal relationships with a range of neuropsychiatric phenotypes and related biological measures. Our findings hint that variation in cortical morphology is largely causal of and not consequential to neuropsychiatric phenotypes. In addition, TSA exhibited a more widespread impact on phenotypes (N=18, of which 5 neuropsychiatric) than MTH (N=5, of which 3 neuropsychiatric). Finally, in the reverse analyses we found that fewer phenotypes (N=2, of which 1 neuropsychiatric) influenced TSA than MTH (N=10, of which 1 neuropsychiatric). Of all the detected causal relationships, several noteworthy findings include bidirectional associations between TSA, MTH and neuropsychiatric traits, as well as protective effects of TSA on cognitive performance and of MTH on schizophrenia.

The novelty of our approach and findings may be summarized as follows. First, we leveraged an unprecedented set of 199 phenotypes spanning neuropsychiatric, behavioral, and metabolic domains. This comprehensive approach allowed us to explore bidirectional causal relationships between 70 cortical morphology measures and a wide range of phenotypes, extending our understanding beyond the scope of previous MR studies. Second, we applied several statistical models to examine whether the findings align. Our use of GSMR, incorporating the heterogeneity in dependent instruments (HEIDI) method, in addition to the four sensitivity analyses, provided a stringent and powerful approach. Third, the use of cortical morphology data from the ENIGMA consortium provided significant advantages over the use of UKB data. Fourth and finally, we highlight several novel findings. We uncovered several new unidirectional and bidirectional relationships between cortical morphology and various neuropsychiatric, behavioral, and metabolic phenotypes. These findings provide new insights into the potential causal mechanisms underlying these relationships and highlight novel areas for further research.

Besides the influence of bipolar disorder (BIP) on MTH, no other psychiatric disorders seemed to affect cortical morphology. Conversely, several psychiatric traits were affected by cortical morphology, thus confirming some observational associations⁵, to which we add evidence about causality and directions of effect. Notably, we detected no causal associations for major depression disorder (MDD) and anxiety-related disorders. As those disorders show the lowest heritability among the most common psychiatric problems,⁴⁷ possibly, insufficient GWAS hits have hitherto been detected to be extracted as instruments and help identify causal relationships.

The more widespread impact of TSA than of MTH on cortical morphology may be explained by the greater number of associated loci that have been identified for TSA than for MTH.^27,48 These differences are likely to have been driven by the divergent evolutionary paths of the two metrics, as also captured by the radial unit hypothesis:^15,49 TSA has expanded several orders of magnitude in humans compared to other mammals, while MTH has remained roughly the same across these species. The scale of the measures may further contribute to the difference in genetic discoverability: MTH is ~3–4 mm, while TSA is dozens or hundreds of cubic mm per region.^50–52 TSA is therefore more likely to have a higher signal-to-noise ratio, thus increasing chances of capturing biological processes. Notably, however, the reverse analyses showed that MTH was causally influenced by more phenotypes than TSA, suggesting that the discovered causal associations reflect true underlying mechanisms and do not result from measurement bias.

We highlight some divergent effects of TSA and MTH on a variety of neuropsychiatric disorders. For example, TSA had a positive causal effect on bipolar disorder (BIP#2) and a negative causal effect on attention-deficit/hyperactivity disorder (ADHD), and Alzheimer’s disease (ALZ). On the other hand, MTH had negative causal effects on schizophrenia (SCZ2 and SCZ3), as well as on the joint schizophrenia and bipolar disorder phenotype (BDSCZ), which likely reflects a SCZ-driven association given the high proportion of SCZ cases. With regards to BIP, our results build on largely conflicting literature. For instance, one brain imaging study ⁵ found BIP to be associated with less MTH, but not with TSA, while another study reported decreased thickness in the left superior temporal gyrus.^53,54 Furthermore, even though the studies of BIP and BIP#2 are non-independent (BIP#2 includes all BIP samples), the former only causally influences MTH, while the latter is only causally influenced by TSA, curtailing the robustness of the BIP findings. One reason for the heterogeneity in cortical morphology findings for BIP in MR studies may relate to the two clinically and genetically different subtypes of bipolar disorder (types 1 and 2), proportions of which may vary across GWASs,⁵⁵ as well as varying durations of illness.^56,57 Importantly, the studies discussed here for BIP were not MR studies and therefore the associations reported may be subject to disease state influences. As for SCZ, previous research has found genetic associations of SCZ with both SA and TH, with the former association pertaining to the global level and the latter mostly to the regional level.⁵⁸ In our study, both MTH and regional TH measures were causally associated with SCZ. In fact, all three SCZ-related phenotypes involved in this study (SCZ2, SCZ3, BDSCZ) were causally influenced by MTH in the same direction. This observation further adds to the already established associations between MTH and SCZ,^59–61 and we now contribute evidence for negative causation. After global thickness correction, only the transverse temporal thickness showed negative causal effects on schizophrenia (SCZ3), while cuneus thickness (TH) showed negative causal effects on BDSCZ. The transverse temporal gyrus (Heschl’s gyrus) is a cortical structure that is implicated in processing incoming auditory information, with TH decreases reported to be associated with the occurrence of auditory verbal hallucinations,^62–64 a common clinical feature of schizophrenia. Another function of the cuneus is facial emotion recognition and visual memory,⁶⁵ both of which are often impaired in schizophrenia. For the ADHD findings that we report, most of the non-global corrected 17 regions we found to influence risk were consistent with a previous study that found smaller surface area in the frontal, temporal, and cingulate regions.⁶⁶ However, after applying global correction, no regions remained significant. Lastly, our study revealed a causal effect of TSA on Alzheimer’s disease (ALZ). However, it is worth noting that a previous MR study did not find a causal relationship between cortical structure and ALZ.⁶⁷ Discrepancies between the results may have originated from differences in statistical methodology and power. In general, excessive cortical thinning is a very prominent characteristic of ALZ, but it mostly impacts specific brain regions.⁶⁸ Therefore, when conducting the broad MR analyses using the global mean of cortical thickness, it is possible that those regional changes in TH faded out in the broader measurement. However, when performing the regional analyses, ALZ was still not causally influenced by any of the regional TH measures. Conversely, several SA regions were found to have causal effects on ALZ, but only when using the non-global-corrected method. Speculatively, these findings could indicate that changes in TSA are what actually triggers the onset of ALZ, even though cortical thinning eventually takes over and shows very high associations in later stages of the disease. This does not mean, however, that cortical thinning is causally implicated in the pathogenesis of ALZ. Interestingly, a previous MR study by Wu et al.⁶⁷ also highlights the causality of TSA on ALZ but identified different regional TSA measures as causal (precentral and isthmus cingulate). This may be because of the different ALZ GWAS summary statistics that were used, with us leveraging the results of a more recent and powerful GWAS.³⁸

Concerning behavioral traits, TSA had positive causal effects on the age of smoking initiation and negative causal effects on ever smoking regularly. Furthermore, genetically predicted TSA increased the chances of quitting smoking. With regards to MTH, we observed a negative causal effect on the number of alcoholic drinks per week and a negative bidirectional causal effect on ever smoking regularly. Based on these findings and considering that cortical volume is mainly influenced by TSA and MTH,⁶⁹ one may surmise that the larger the cortical volume, the healthier lifestyle choices are made. For regional MR analyses, after global correction for SA, three regions (parahippocampal, lateral orbitofrontal, and rostral middle frontal) showed significant effects on ever smoking regularly. The positive association between the surface area of the parahippocampal gyrus and regular smoking may reflect this region’s role in contextual memory and emotional processing, which may influence smoking habits. Additionally, after global correction, we observed negative causal effects of superior frontal thickness and inferior temporal thickness on weekly alcohol consumption. This aligns with previous research linking frontal lobe dysfunctions, such as impaired working memory and pattern recognition, to alcohol use. Binge drinkers often show deficits in frontal inhibitory control and executive functions.^70,71

Through reverse MR analyses, we identified several phenotypes with causal effects on TSA and MTH, thus providing insights into modifiable factors of relevance for brain health and for clinical practice, where physicians over recent years have developed a growing awareness about lifestyle factors impacting the course of illness of many conditions.⁷² For MTH, reverse analyses yielded several actionable targets, i.e., smoking, blood lipid levels, and BMI. In the adult population, cortical thinning is a normal process of ageing, which has also been associated with numerous neuropsychiatric disorders.^73,74 Even if cortical thinning is unavoidable, some associations indicated that the extent to and speed at which it happens can potentially be modified by external factors, e.g., physical activity, education, or diet.^75–77 We found that regular smoking has negative bidirectional relationships with MTH, a novel finding that adds to the existing literature, which so far to our knowledge has identified only associations and no causal relationships.^78,79 Moreover, our findings indicate a negative causal effect of BMI on MTH. Indeed, overweight individuals have reduced MTH, due to obesity-triggered neuroinflammation, which may lead to neuronal loss.⁸⁰ A study involving children aged 4 to 18 years did not detect any significant associations between TH and BMI,⁸¹ suggesting that the link between BMI and cortical thinning may develop after adolescence. Additionally, we observed that several key fatty acid metabolic traits have causal effects on MTH. Polyunsaturated fatty acids (PUFAs), for instance, were found to positively affect MTH. This could potentially be because of PUFAs’ crucial role in maintaining the fluidity and permeability of neuronal membranes.⁸² In sum, our findings hint that avoiding smoking, reducing BMI, and intake of fatty acids may yield benefits for brain health.

Our study has several strengths, including a large number of phenotypes and a wide range of powerful statistical methods. These include our main analysis method (GSMR), a stringent approach,²⁶ along with various extra steps to ensure reliability, e.g., a range of sensitivity analyses, the use of several GWASs for single phenotypes, and the HEIDI test. Considering that quality assessment tools have not been adequately validated and established yet,⁸³ we leveraged the aforementioned strategies along with the estimation of several parameters like the F-statistics. We did, however, fill out the Strobe-MR checklist (Supplement), to show how we complied with standards in MR research. Nevertheless, the following limitations should be borne in mind when interpreting our findings. First, a general limitation of MR is that MR assumptions cannot be explicitly proven. We increased chances that such assumptions are met by using powerful and stringent statistical methods, as well as sensitivity analyses. In addition, limited GWAS sample sizes may reduce statistical power in MR research; this may have applied to some of the phenotypes under study. Thus, although by applying multiple testing correction we diminished the type I error likelihood, it is possible that for some phenotypes relevant associations were missed. Furthermore, underrepresentation of various ethnic groups in GWASs hampers the global generalizability of our results. Finally, the limited representation of diverse ethnic groups in GWAS studies, which is a general challenge in psychiatric genetics, affects the broader applicability of our findings. Genetic variants may not have the same impact across populations. Two-sample Mendelian randomization (MR) requires population matching primarily to ensure the validity and accuracy of causal inferences [for more details please see our reply to reviewer’s 2 point 1]. To minimize the risk of bias arising from population stratification, we focused on European populations. Future studies should incorporate more diverse populations to enhance the generalizability of genetic research, while novel methods may address population stratification to better understand the interplay between genetics, environment, and various phenotypes across populations. In addition, longitudinal studies are needed to confirm temporal relationships and causality between cortical morphology and related phenotypes. Moreover, investigating interactions between genetic predispositions and environmental factors such as diet, lifestyle, and socioeconomic status may provide a more comprehensive understanding of these relationships. Finally, intervention studies may examine a range of brain morphology and behavioral outcome measures of lifestyle sessions given to targeted populations, e.g., those who smoke or those with a specific blood lipids profile.

In conclusion, we show that genetically predicted variation in cortical morphology plays causal roles in neuropsychiatric disorders, particularly schizophrenia. We additionally provide new insights into effects of genetically predicted phenotypes on cortical morphology, many of which are actionable.

Materials and methods

Study design

We first conducted a two-sample MR study using GWAS summary statistics of two cortical measures, namely global total surface area (TSA) and mean cortical thickness (MTH), as well as 199 phenotypes of interest (STable 2). The data analyzed in this study are exclusively quantitative, consisting of summary-level statistics from GWAS. Our main method consisted of Generalized Summary-data-based MR (GSMR),²⁶ with extra sensitivity analyses performed thereafter. Every MR analysis was bidirectional, with each cortical morphology trait tested both as a potential causal factor (forward analysis) and a potential outcome (reverse analysis). In the sensitivity analyses, consistent directions of effect across all methods validated our results and further supported the newly discovered significant causal associations.⁸⁴ To further explore those associations and potentially identify any regions that drove these results, we also performed GSMR on a regional level to check for any causal relationships between the significantly associated phenotypes and the cortical measures (SA and TH) of 34 brain regions. Finally, to assess whether any of the cortical regions is relevant on a global level with regards to the selected phenotypes, we used global correction for TSA and MTH. Analyses were performed from January 1 to June 1, 2023 using the latest GWASs.

Data sources

Data sources comprised 70 cortical morphology measures – i.e., TSA and MTH, along with 68 regional surface area (SA) and thickness (TH) measures (derived from the same study population), as defined by the Desikan–Killiany cortical atlas (STable 1) - and 199 phenotypes (STable 2), encompassing neuropsychiatric (N=36), diabetes-related (N=2), behavioral (N=14) and quantitative phenotypes (N=147, primarily metabolic, see STable 3 for full biochemical names). All those phenotypes were derived from a total of 40 GWASs. More details about the exact phenotypes under study along with the sample sizes and respective references for all GWASs are provided in STable 2. Wherever possible, for a single phenotype, we also retrieved summary statistics from two studies. Even though those studies were non-independent, with the latest one including all data from the previous one, we reasoned that comparing their results and confirming that a larger sample size leads to more significant hits would corroborate the robustness of our findings. This was implemented for schizophrenia (SCZ), where we used both Psychiatric Genomics Consortium (PGC) freezes 2 and 3,^40,41 for bipolar disorder,^35,85 for smoking-related phenotypes and for alcohol consumption.^33,36 Although inclusion of multiple or different ancestries would have been preferred, given the dearth of available GWAS studies in such populations, we restricted the inclusion of GWAS summary statistics to European ancestry only (more details about these GWAS studies may be found in Supplementary Table 2). To that end, we manually verified in the respective publications that all summary statistics had been derived from populations of European ancestry. In the event a GWAS matching our phenotypes of interest had been done on diverse or specific non-European ancestries, we obtained the European-only GWAS summary statistics. These details may also be found in Supplementary Table 2.

As for cortical morphology measures, we used summary statistics from the latest GWAS by the ENIGMA consortium.²⁷ The other GWAS summary statistics were selected upon extensive review of the methods sections of the respective studies to avoid sample overlap with ENIGMA.²⁷ For instance, considering ENIGMA includes many participants of the UKB, we also ensured that there were no UKB participants in any of the other 39 GWASs we used. To further prevent sample overlap, we specifically requested summary statistics excluding data from the UKB and, when not possible, we replaced the UKB-inclusive summary statistics with those of previous, unrelated studies. Finally, as an additional step to ensure lack of sample overlap between exposure and outcome data sources, we checked the intercepts from cross-trait linkage disequilibrium score regression (LDSC)⁸⁶ for all significant pairwise MR relations (see Results section). We found that, except for educational attainment (EA) and cognitive performance (CP), for every intercept estimate the confidence interval spanned zero, indicating that substantial sample overlap was unlikely. For EA and CP, the intercepts were 0.015 and 0.03, with confidence intervals being 0.002–0.027 and 0.015–0.044, respectively. This suggested that there may be a small amount of shared ancestry or overlap between the EA, CP, and ENIGMA studies, but this was unlikely to substantially impact the MR results (more details can be found in the Supplementary Methods, with LDSC results presented in Supplementary tables ST1). To ensure the MR results were directly comparable across dichotomous and continuous phenotypes, we converted any GWAS Odds Ratios (ORs) into betas: $beta=ln\left(OR\right)$ (1). Then, the effect sizes (betas) of all GWAS summary statistics were transformed into Z scores before the MR analyses: $Zscore=effectsize\left[beta\right]/standarderror\left[SE\right]$ (2).

Data harmonization, instrumental variable selection, and MR assumptions

To ensure reliability and validity of our analyses, a comprehensive data harmonization and preprocessing strategy was implemented. First, we verified the consistency of the single-nucleotide polymorphism (SNP) identifiers, as well as their respective alleles. To select independent SNPs, we applied LD pruning with an LD threshold of r² < 0.01. For each measure of interest, we selected SNPs associated at p < 10⁻⁶ as instrumental variables (IVs) for both exposures and outcomes, which used the same settings for previous studies.^87,88 A trait was tested as a potential exposure on condition that at least two significant loci had been identified. This is the minimum IV requirement for the selected MR analysis methods.

IVs were extracted using the Heterogeneity in Dependent Instruments (HEIDI)^24,25 test implemented in GSMR,²⁶ which is robust against instrumental outliers. We used the default options in GSMR, including HEIDI testing with a significance threshold of p < 0.01 to detect and exclude instrumental outliers.

The results of every MR analysis rely on three fundamental assumptions with regards to IVs, namely instrumental variable relevance, independence, and exclusion restriction.¹⁰ Generally, while neither of these assumptions can be explicitly proven, we sought to increase the likelihood of them being true in our study. This was achieved by employing powerful and stringent statistical methods along with conducting additional sensitivity analyses, as described in the following sections.

Main analyses: Generalized Summary-data-based MR (GSMR)

The main MR analyses (GSMR) were conducted using the software package “GCTA”^26,89 on the high-performance computing system (HPC) of the University Medical Center Utrecht (UMCU). Two-sample MR analyses were performed using GSMR between total surface area (TSA) or global mean thickness (MTH) on the one hand and neuropsychiatric, behavioral, and quantitative phenotypes on the other. To ensure the MR results were directly comparable across dichotomous and continuous phenotypes, we employed standardized beta values (β). Therefore, ORs were converted into betas: $beta=ln\left(OR\right)$ (2). and all betas were transformed into Z scores: $Zscore=effectsize\left[\beta \right]/standarderror\left[SE\right]$ (2). Finally, we used the false discovery rate (FDR) method to correct the significance threshold for multiple testing to 0.05. In this case, the correction concerns testing multiple exposures against multiple outcomes (e.g., forward analyses: 70 exposures against 199 outcomes and reverse analyses: 199 exposures against 70 outcomes).

Sensitivity analyses

As previously described,⁹⁰ additional MR methods were applied to confirm any significant (p._fdr < 0.05) GSMR findings. First, we verified causality by employing four other MR methods (all bidirectional) as sensitivity analyses with the use of the “TwoSampleMR” package⁹¹ in R (version 4.0.4)⁹². Those methods were: the Inverse Variance Weighted (IVW) fixed-effects method (the default, basic approach in MR),⁹³ MR-Egger (robust to violation of uncorrelated horizontal pleiotropy assumption by all instrumental variables),^94,95 weighted median (WM, robust to violations of assumptions by up to half of the instrumental variables),⁹⁶ and MR pleiotropy residual sum and outlier (PRESSO, robust when the majority of genetic variants are valid instruments and uncorrelated horizontal pleiotropy is limited to a small number of outliers, which are then excluded by the method).⁹⁷ To minimize the risk of violations of MR assumptions impacting our results, such as directional pleiotropy or weak instruments, we compared the results across all the above-mentioned models. As each approach has different assumptions, consistent results across methods strengthen the confidence in MR findings.⁸⁴ In the results section, we therefore report only findings with consistent directions of effect across all sensitivity analyses. Second, we repeated all the FDR-significant GSMR tests using a more stringent threshold for inclusion of genetic instruments of p < 10⁻⁷. Third, two multivariable MR (MVMR-IVW) analyses were conducted to examine whether the effects of genetic liabilities to bipolar and smoking initiation effects on MTH remained significant after adjusting for BMI and other 7 blood metabolic phenotypes that carried significant causal effects on MTH in univariable MR tests. To that end, we conducted multivariable MR (MVMR) analyses to examine which phenotypes remained significant when taking into account pleiotropic effects among other metabolic exposures (Supplemental Methods). MVRM estimates the effects of each exposure on an outcome adjusting for genetic associations between multiple phenotypes and the exposure. MVMR is an extension of MR that may be useful when three or more exposures are involved. Here, MVMR IVW analyses were conducted for the situation where phenotypes that may be biologically linked showed causal effects in the same forward or reverse MTH or SA analysis. This turned out to be the case for smoking initiation (SmokInit#2) and bipolar disorder (BIP) in reverse MTH analyses. To conduct these MVMR analyses, we selected instruments using SNPs from univariable MR tests of BMI, of the 7 significant blood metabolic phenotypes (Figure 2D), of SmokInit#2, and of BIP (n=296 for the SmokInit#2 model and n=132 for the BIP model). Possible residual heterogeneity was examined using the Cochran’s Q test (and the p-value of the Q statistics). Results are reported in Supplementary Table 7B. Furthermore, multiple parameters were estimated using the sensitivity analyses: we calculated instrument strengths using F-statistics, I² statistics (validating the suitability of the instruments in MR-Egger), and residual heterogeneity by Cochran’s Q test (and the p-value of the Q statistics). Finally, we used funnel plots to visualize the distribution of effect estimates and we also applied leave-one-out analyses to evaluate whether any single instrument was driving the results.

Regional MR analyses

First, all phenotypes with significant (p._fdr < 0.05) causal effects of any direction with TSA or MTH were selected. We then applied GSMR to bidirectionally explore associations between those phenotypes and 68 regional cortical measures, i.e., TSA-associated and MTH-associated phenotypes were tested against the SA and TH measures of the 34 regions, respectively (Fig.1). The FDR method was applied over the whole set of analyzed traits to correct the significance threshold for multiple testing to 0.05. Initially, by using non-global-corrected GWAS summary statistics for the regional cortical measures,²⁷ we aimed to identify any regions that could potentially drive the global results. Furthermore, seeking to examine regions with possible significant effects on the phenotypes of interest on a global level, regardless of the previous global results, we repeated the same steps but with global-corrected GWAS data for the regional measures.²⁷ The findings of the regional analyses were visualized with the “ggseg”⁹⁸ package in R.

Supplementary Material

Supplementary methods, tables (STable 1–3) and figures (SFigures 1–7) can be found in SupplementaryMaterial.docx. Supplementary tables (ST1–ST11) can be found in SupplementaryData.xlsx. The STROBE-MR checklist we filled out can be found as a separate supplement.

Data availability

All results data generated and analyzed during this study are included in the supplementary materials accompanying this manuscript. These supplementary materials provide the complete dataset necessary to interpret, verify, and extend the research presented in the article.

For any additional information or access to specific datasets beyond what is provided in the supplementary materials, reasonable requests can be made to the corresponding author.

Code availability

The code used for data analysis is available on GitHub (https://github.com/Bochao1/Brian_NMH).