Abstract
White matter of the human brain is influenced by common genetic variations and shaped by neural activity-dependent experiences. Variations in microstructure of cerebral white matter across individuals and even across fiber tracts might underlie differences in cognitive capacity and vulnerabilities to mental disorders. The frontoparietal and cingulo-opercular networks of the brain constitute the central system supporting cognitive functions, and functional connectivity of these networks has been used to distinguish individuals known as “functional fingerprinting.” The frontal aslant tract (FAT) that passes through the two networks has been implicated in executive functions. However, whether FAT can be used as a “structural fingerprint” to distinguish individuals and predict an individual's cognitive function and dysfunction is unknown. Here we investigated the fingerprinting property of FAT microstructural profiles using three independent diffusion MRI datasets with repeated scans on human participants including both females and males. We found that diffusion and geometric profiles of FAT can be used to distinguish individuals with a high accuracy. Next, we demonstrated that fractional anisotropy in different FAT segments predicted distinct cognitive functions, including working memory, inhibitory control, and relational reasoning. Finally, we assessed the contribution of altered FAT microstructural profiles to cognitive dysfunction in unmedicated patients with obsessive-compulsive disorders. We found that the altered microstructure in FAT was associated with the severity of obsessive-compulsive symptoms. Collectively, our findings suggest that the microstructural profiles of FAT can identify individuals with a high accuracy and may serve as an imaging marker for predicting an individual's cognitive capacity and disease severity.
SIGNIFICANCE STATEMENT The frontoparietal network and cingulo-opercular network of the brain constitute a dual-network architecture for human cognitive functions, and functional connectivity of these two networks can be used as a “functional fingerprint” to distinguish individuals. However, the structural underpinnings of these networks subserving individual heterogeneities in their functional connectivity and cognitive ability remain unknown. We show here that the frontal aslant tract (FAT) that passes through the two networks distinguishes individuals with a high accuracy. Further, we demonstrate that the diffusion profiles of FAT predict distinct cognitive functions in healthy subjects and are associated with the clinical symptoms in patients with obsessive-compulsive disorders. Our findings suggest that the FAT may serve as a unique structural fingerprint underlying individual cognitive capability.
Introduction
White matter of the human brain, constituting the structural backbone of functional communications between brain regions (Suárez et al., 2020), is influenced by common genetic variations (Zhao et al., 2021) and is shaped by neural activity-dependent experiences (Sampaio-Baptista and Johansen-Berg, 2017). Growing evidence suggests that the microstructure of cerebral white matter has high heritability and varies across individuals (Croxson et al., 2018). In particular, white matter tracts connecting association cortices show higher individual variability than those connecting primary sensorimotor regions (Sydnor et al., 2021). These intersubject variations in white matter structure might be linked to individual differences in behavior and cognition in a healthy population (Glahn et al., 2013) and are predictive of clinical outcomes in patients with psychiatric or neurologic disorders (Keller et al., 2017; Wang et al., 2021). It is, therefore, important to characterize the uniqueness of white matter tracts, or structural connectome fingerprinting, for individuals (Yeh et al., 2016) and to understand its relationship with cognitive functions and mental disorders (Genon et al., 2022).
The frontal aslant tract (FAT), identified by diffusion MRI (dMRI) tractography, is a white matter tract that connects the pars opercularis and pars triangularis in the inferior frontal gyrus (IFG) and pre-supplementary motor area (pre-SMA) of the superior frontal gyrus (SFG), respectively (Catani et al., 2012). These brain regions are core constitutes of the frontoparietal network (FPN) and cingulo-opercular network (CON). Functional connectivity profiles of FPN and CON have potentials in distinguishing individuals or serving as “functional fingerprints” (Finn et al., 2015; Badhwar et al., 2020; Jin et al., 2020). These two parallel networks also constitute a dual-network architecture underlying top-down cognitive control (Dosenbach et al., 2008; McTeague et al., 2017). The structural features of FAT have been linked to speech and cognitive control functions (Catani et al., 2013; Dick et al., 2019), and impairments of FAT disrupt working memory and inhibitory control abilities (Li et al., 2019; Rutten et al., 2021). However, whether the microstructural properties of FAT can serve as “structural fingerprints” of individuals, including identification of individual subjects and association with individual cognitive ability, has not been systematically investigated.
dMRI is a noninvasive tool to measure microstructure of white matter (Wandell, 2016) using diffusion tensor models with metrics such as fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD). Among these metrics, FA has been used as the primary metric of interest in many studies because of its robustness in measuring the overall integrity and directionality of white matter fibers, and its high sensitivity in detecting changes in structural connectivity (Grieve et al., 2007). Moreover, tractography based on dMRI has been used to capture interindividual variations in white matter tracts and link white matter phenotypes to cognition (Glahn et al., 2013).
In this study, we first investigated the fingerprinting properties of FAT using three independent datasets, including 44 individuals with two separate dMRI scans from the Human Connectome Project (HCP), 192 individuals with two separate dMRI scans from the Southwest University Longitudinal Imaging Multimodal (SLIM), and three individuals from the MultiCenter study, each scanned at 10 sites with 12 repeated sessions. We addressed the question of whether FAT structural profiles, measured by dMRI, can be used to uniquely identify individuals (structural fingerprinting). Next, using a cohort of 466 individuals with a single dMRI and functional MRI scan, as well as cognitive assessments, from the HCP S500 dataset, we investigated whether FAT structural properties can be used to predict individual cognitive control ability. Finally, we assessed the association of FAT structural properties with clinical symptoms related to cognitive control impairments using a dMRI dataset including 68 unmedicated patients with obsessive-compulsive disorders (OCDs) and 80 demographically matched healthy control subjects. Findings from this study would establish white matter microstructural features as potential structural fingerprints that not only distinguish individuals but also predict cognitive capacity and mental disease severity.
Materials and Methods
Participants and MRI data acquisition
Three separate datasets with repeated dMRI scans were used in the current study for the fingerprinting analysis. The primary dataset was from the HCP (N = 44; Van Essen et al., 2012). The other two datasets [i.e., SLIM (N = 192; Liu et al., 2017) and MultiCenter (N = 3, repeated at 10 sites; Tong et al., 2020) cohorts] were used for validating the reproducibility of fingerprinting properties across different cohorts and across different scanning sites, respectively. For the OCD study, 68 unmedicated OCD patients and 80 healthy control subjects were scanned at the Shanghai Mental Health Center.
HCP data.
The HCP “Test-Retest” cohort dMRI data contain data from 44 healthy adults (20 males; mean ± SD age, 30.2 ± 3.44 years; age range, 22–35 years). Each subject had two scans at an interval of 139.3 ± 68.2 d. In addition, the HCP S500 data (n = 466) were used to explore the relationship between FAT diffusion measures and 11 cognitive control measures (out-scanner assessments including List sorting, Card sorting, Flanker, processing speed task scores; in-scanner assessments including relational task accuracy and reaction time, working memory task accuracy and reaction time, and gambling task accuracy and reaction time for larger choice). The HCP dMRI data were acquired using a 2D spin-echo single-shot multiband echoplanar imaging (EPI) sequence on an HCP-customized Siemens 3.0 T Connectome Skyra scanner at Washington University in St. Louis with the following scanning parameters: repetition time (TR) = 5520 ms; echo time (TE) = 89.5 ms; resolution = 1.25 × 1.25 × 1.25 mm3; field-of-view (FOV) = 220 mm; b-values = 1000/2000/3000 s/mm2; and a total of 270 diffusion directions were distributed equally over these three shells with 18 images with b = 0 s/mm2. T1-weighted (T1w) structural image data were acquired using the magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence with the following parameters: TR = 2400 ms; TE = 2.14 ms; inversion time (TI) = 1000 ms; FOV = 224 mm; resolution = 0.7 × 0.7 × 0.7 mm3; and flip angle = 8°. All HCP fMRI data for three cognitive control tasks (i.e., working memory, relational reasoning, and gambling tasks) were acquired using a multiband gradient echo EPI sequence (TR = 720 ms; TE = 33.1 ms; resolution = 2.0 × 2.0 × 3.0 mm3; flip angle = 52°; multiband factor = 8).
SLIM data.
The SLIM dataset contains 192 healthy young adults (86 males; mean ± SD age = 19.92 ± 5.87 years; age range = 17–25 years) scanned on a Siemens Trio 3 T scanner at Southwest University Center for Brain Imaging. All subjects had their repeated scans after 302.8 ± 92.6 d. The dMRI images were acquired with the following scanning parameters: TR = 11,000 ms; TE = 98 ms; resolution = 2.0 × 2.0 × 2.0 mm3; FOV = 256 mm; 30 directions (b-value = 1000 s/mm2) with one nondiffusion frame scanned three times repeatedly. T1w images were acquired using the MPRAGE sequence with the following parameters: TR = 1900 ms; TE = 2.52 ms; TI = 900 ms; FOV = 220 mm; resolution = 1.0 × 1.0 × 1.0 mm3; and flip angle = 9°.
Multicenter data.
The MultiCenter data include three healthy subjects (one male; age range, 23–26 years) scanned on 10 different 3 T Siemens Prisma scanners, with one scan on nine of the scanners and three scans on one scanner. The dMRI images were acquired using a simultaneous multislice spin-echo EPI sequence with the following parameters: TR = 5400 ms; TE = 71 ms; resolution = 1.5 × 1.5 × 1.5 mm3; FOV = 220 mm; b-values = 1000/2000/3000 s/mm2; and 90 diffusion directions distributed uniformly over the three shells. T1w images were acquired using the magnetization-prepared two rapid acquisition gradient echo sequence with the following parameters: TR = 5000 ms; TE = 2.9 ms; TI = 700, 2500 ms; FOV = 211 × 256 × 256 mm3; resolution = 1.2 × 1.0 × 1.0 mm3.
OCD data.
Sixty-eight righted-handed unmedicated OCD patients (37 males; mean ± SD age = 28.94 ± 5.87 years; age range, 20–42 years) and 80 demographically matched healthy control subjects (37 males; mean ± SD age = 28.08 ± 6.13 years; age range, 19–45) were included in our study. The patients all met DSM-IV (Diagnostic and Statistical Manual of Mental Disorders, 4th edition) criteria for OCD and had not been under any medication for >8 weeks. The clinical symptom severity of the patients was assessed using the Yale–Brown Obsessive Compulsive Scale (Y-BOCS). All the participants were scanned on a 3 T Siemens Verio scanner. The dMRI images were acquired using the spin-echo EPI sequence with the following parameters: TR = 10,100 ms; TE = 110 ms; resolution = 2.0 × 2.0 × 3.0 mm3; FOV = 220 mm; b-values = 1000/2000 s/mm2; 128 directions. The T1w images were acquired using the following MPRAGE sequence: TR = 2300 ms; TE = 2.96 ms; TI = 900 ms; FOV = 256 mm; resolution = 1.0 × 1.0 × 1.0 mm3; flip angle = 9°. Written informed consents were obtained from all participants. The study was approved by the Institutional Review Board of Shanghai Mental Health Center. Differences in demographic and clinical characteristics between healthy control subjects and unmedicated OCD patients were assessed using independent two-sample t tests for continuous variables and χ2 test for categorical variables.
Data preprocessing and FAT structural profile quantification
dMRI data preprocessing.
All HCP dMRI data underwent minimal preprocessing using FSL (version 5.0.9; https://fsl.fmrib.ox.ac.uk/fsl/fslwiki), including b0 intensity normalization, EPI distortion correction, eddy-current and head-motion correction, and gradient nonlinearity correction (Glasser et al., 2013). For the MultiCenter dataset, the dMRI data were preprocessed using the following pipeline: Gibbs ring removal by MRtrix3 (version 3.0.2; https://github.com/MRtrix3/mrtrix3); EPI distortion correction; and eddy-current and motion correction (Tong et al., 2020). For SLIM dataset and OCD study data, the dMRI images were preprocessed using denoising and Gibbs ring removal, followed by eddy-current and motion correction by MRtrix and FSL. The preprocessed dMRI data were linearly registered to the averaged b0 image by mrDiffusion (http://github.com/vistalab). The voxel-wise fiber orientation distribution (FOD) was calculated using constrained spherical deconvolution with up to eight spherical harmonics. Whole-brain probabilistic fiber tracking was then conducted by MRtrix, using the following parameters (Grotheer et al., 2019): step-size = 0.2 mm; minimum fiber length = 4 mm; maximum fiber length = 200 mm; stopping FOD amplitude = 0.1; stopping maximum angle = 13.5°; maximum number of streamline fibers = 500,000. Seeds for tractography were randomly placed within the gray matter–white matter interface (Smith et al., 2012).
Extraction of FAT.
Bilateral FATs were extracted (Fig. 1a), and the diffusion profiles were measured using Automated Fiber Quantification (https://github.com/yeatmanlab/AFQ). The cortical surface was reconstructed and parcellated using the Desikan–Killiany (DK) atlas in FreeSurfer (version 7.1.1; https://surfer.nmr.mgh.harvard.edu/), based on T1w images. FAT was extracted only if it passed through the two predefined regions of interest (Kronfeld-Duenias et al., 2016). Furthermore, fibers were discarded if their end points were not located within the SFG, parstriangularis (PTRI), or parsopercularis (POPE) regions in the DK atlas. This was done to separate them from the neighboring superior-middle and middle-inferior tracts (Desikan et al., 2006; Catani et al., 2012). After extraction, the fiber tract was refined by removing outlier fibers. Specifically, any fibers with a three-dimensional Gaussian covariance larger than five times the SDs from the center of the tract were discarded. The resulting bilateral FAT tracts were selected for further quantification analysis.
Measurement of FAT diffusion and geometric profiles.
The microstructural properties are not uniformly distributed along the white matter tract (Safadi et al., 2018; Grotheer et al., 2022). To measure the FAT diffusion properties along FAT, each fiber streamline was sampled at 100 equally spaced nodes. The diffusion metrics (i.e., FA, AD, RD, and MD, derived from b = 1000 only) of each node were calculated as the weighted average of the diffusion measurements of each individual fiber. To measure the FAT geometric profiles, we included two along-the-fiber metrics (i.e., span and curvature), and two across-the-fiber metrics (i.e., radius and volume) in our analyses. Following previous work (Glozman et al., 2018; Yeh, 2020), node-wise geometric measures were obtained along the 100-node centerline curve fitted over the FAT. The coordinates of each node along the fiber streamlines were calculated by averaging the locations.
Task fMRI data processing.
The HCP minimal preprocessing pipeline was used to preprocess the task fMRI data. Subsequent preprocessing procedures were conducted using Analysis of Functional Neuro Images (AFNI; https://afni.nimh.nih.gov), which included quadratic detrending, Gaussian spatial smoothing with a full-width at half-maximum (FWHM) kernel size of 6 mm, censoring the time point where head motion exceeded 0.5 mm, and regression of six head motion time courses. The subjects with mean head motion >0.25 mm would be excluded. In total, working memory task fMRI data from 415 subjects and relational reasoning task fMRI data from 409 subjects were used for further analysis. The general linear model was used to estimate the activation β weights of each run (working memory, 2-back–0-back; relational reasoning, relation–match), and the β weights were then averaged for each task. The whole-brain activation map for each task was thresholded at corrected p < 0.05, determined using the AFNI 3dClustSim algorithm. For the working memory task fMRI, the threshold was set at uncorrected p < 0.001 with a cluster size of 528 mm3. For the relational reasoning task fMRI, the threshold was set at uncorrected p < 0.001 with a cluster size of 536 mm3. The significantly activated clusters within the CON network connected by FAT were delineated. The mediation analysis (age, sex, years of education, and head motion controlled) was conducted to determine whether the activation level of the delineated clusters mediated the relationship between FAT FA and cognitive control task performance (Gu et al., 2019). Significance of the model was set at 0.05.
Identification of individuals
Identification analysis.
The analysis of FAT fingerprinting identification was performed following the procedure by Finn et al. (2015) and is shown in Figure 1b. The bilateral FAT diffusion profiles from each subject were transformed into a vector with length of 200 (100 nodes for each FAT tract). The vectors were then concatenated into an N × 200 matrix, where N is the number of subjects for each session. After that, we computed the similarity index between the two sessions using the Pearson correlation coefficients between the two vectors taken from the target matrix and each of the database matrices. For each subject, the individual identification accuracy is 100% if the maximum similarity is from the same subject, otherwise it is 0. Furthermore, permutation tests were conducted to assess the significance of the identification rates. To account for the family structure present in the HCP Test-Retest cohort, a complex block exchangeability restriction (Winkler et al., 2015), based on three family types [singleton, dizygotic (DZ) twin and monozygotic (MZ) twin], was incorporated during the permutation procedure. In the permutation test, we held the database matrix constant and permuted the rows of the target matrix and repeated the above identification procedure 1000 times.
Reproducibility analysis.
To evaluate the reproducibility of FAT diffusion profiles across repeated sessions. Intraclass correlation coefficient (ICC; Shrout and Fleiss, 1979) was calculated for each node i as follows:
Heritability analysis
The HCP S500 dMRI dataset includes 54 pairs of genetically confirmed MZ twins and 26 pairs of same-gender DZ twins. No significant differences in gender and age were found between DZ and MZ groups. Structural equation modeling was performed using OpenMx software (Boker et al., 2011; Neale et al., 2016) to decompose the additive genetic (A), common environmental (C), and unique environmental (E) influences, which is based on the assumption that MZ twins are genetically identical and DZ twins share half of their DNA. For each FAT segment, heritability (
Extraction of FAT segments
The segments of FAT were identified using a clustering method based on the target regions reached by the streamlines across the nodes. For each subject, we extracted the streamlines passing through each node along the FAT, as shown in Extended Data Figure 1-2. A node-specific structural connectivity matrix, A, was constructed based on the number of streamlines connecting each pair of brain regions defined on the Desikan–Killiany atlas. Next, the node-by-streamline matrix for each individual was generated by concatenating 100 node-specific structural connectivity matrices (Extended Data Fig. 1-2b). The node-by-streamline matrix for each individual was then normalized by the total number of streamlines (Extended Data Fig. 1-2c). A group-averaged node-by-streamline matrix was generated by the average of the normalized matrices obtained from the 44 subjects in the HCP test cohort (Extended Data Fig. 1-2c). A clustering on the group-averaged node-by-streamline matrix was performed based on the Euclidean distance between the streamline connectivity using a K-means algorithm. The optimal number of clusters that maximizing the Silhoutte and point biserial metrics from k ∈ [2, 9] was k = 4. Therefore, the FAT in each hemisphere was divided into four segments, as shown in Extended Figure 1-2d.
Association between FAT diffusion profiles and cognitive functions
Sparse canonical correlation analysis.
In our work, we used sparse canonical correlation analysis (sCCA) for the multivariate analysis of the diffusion measures and cognitive control measures. This method aims to identify maximal correlations between linear combinations of the variables in two sets, with sparsity regularization (Witten et al., 2009). In our study, the datasets are the 11 principal components (PCs) of FAT FA profile features, denoted as X, and the 11 PCs of the cognitive control features, denoted as Y. The aim is to find out the optimal loading vectors (i.e., u and v, to maximize the covariance between
Permutation and resampling tests.
Block-restricted permutation tests were performed to assess the significance of the obtained sCCA modes. Fourteen family types [i.e., singleton, DZ twin, MZ twin, two non-twin full sibling (FS), two non-twin half sibling (HS), three FS, three HS, two DZ + one FS, two DZ + one HS, two MZ + one FS, two MZ + one HS, two FS + one HS, four FS, two DZ + two FS] were included in the complex block exchangeability restriction design. During each iteration, the rows of the behavioral or clinical measurements matrix were randomly shuffled based on the sibling status. Subsequently, sCCA was performed on the shuffled data using the same set of regularization parameters. This process was repeated 1000 times to generate canonical correlation coefficients representing the null distribution of correlations. The statistical significance level is determined by the fraction of the permuted correlation coefficients equal to or larger than the original correlation coefficient. False discovery rate (FDR) correction was performed across the selected canonical modes. To identify the features consistently contributing to each sCCA mode, resampling tests were performed 1000 times. During each resampling, two-thirds of the subjects were randomly selected from the original datasets and the remaining were randomly replaced from the ones selected. The features were considered significant if the 95% confidence interval is away from zero.
Prediction analysis.
To assess the robustness of the resulting sCCA modes, we performed prediction analysis using 10-fold cross-validation combined with sCCA and repeated it 1000 times. During each iteration of the cross-validation loop, the dataset was first randomly divided into 10 folds. Ninety percent of the subjects were selected to train in the sCCA model, and the remaining 10% of the subjects were used to test the model. We multiplied the derived sCCA coefficient matrices from the training set with the normalized behavioral and diffusion measures in the testing set to get the predicted cognitive control and FAT FA profile measures. The prediction performance was assessed by Pearson's correlation between the predicted cognitive control and FAT FA profile measures. Significance level was determined by block-restricted permutation tests (1000 times) using the same 10-fold cross-validation procedure and FDR correction.
Statistical analysis for OCD data
Segment-wise comparisons for FA values along bilateral FATs were conducted between unmedicated OCD patients and healthy control subjects using analysis of covariance (ANCOVA) tests, adjusted for age and sex effects. The results were corrected for multiple comparisons using the FDR procedure. To investigate the relationship between the FA value of each FAT segment and patients' Y-BOCS compulsive and obsessive scores, a stepwise linear regression was conducted including age, sex, and symptom duration as additional predictors. Post hoc analysis was conducted to explore the specificity of the association between the FAT and OCD patients. We compared the FA profiles between OCD and healthy subjects along the 16 association or projection tracts as described above.
Results
Fingerprinting property of FAT diffusion profiles
We assessed the stability and uniqueness of structural profiles of FAT, which are critical properties for fingerprinting. Structural features of FAT, including geometric (i.e., span, volume, curvature, and radius) and diffusion (i.e., FA, MD, AD, and RD) measures, were extracted from three independent dMRI datasets (i.e., HCP, SLIM, and MultiCenter datasets) to evaluate their fingerprinting properties. The extracted bilateral FATs are illustrated in Figure 1a, and the procedure of identifying individual participants (Shen et al., 2017) is shown in Figure 1b.
FAT diffusion profiles show high identification rates in HCP, SLIM, and MultiCenter Test-Retest cohorts. a, Bilateral FAT extracted in one representative subject, which connects the SFG and parsopercularis/parsopercularis in IFG. The procedure of extraction of FAT segments was shown in Extended Data Figure 1-2. b, A schematic illustration of the identification procedure, including resampling of the diffusion measures along the FAT for each subject (i) and construction of the identification matrix using the intersession similarity measured by Pearson's correlation coefficients (ii). c, ICCs and identification rates of eight FAT structural measures. For both cohorts, the FA, AD, and RD showed ICCs >0.8 and identification rates >90% across test and retest sessions. d, ROC curves evaluating the capability of separating intrasubject and intersubject similarities using different structural measures of FAT. For HCP, SLIM, and MultiCenter (Extended Data Fig. 1-1) Test-Retest cohorts, the FA, AD, and RD showed high separation performance with AUC values all >0.99. e, Heritability of FAT structural measures from structural equation modeling.
Figure 1-1
Intrasubject versus intersubject similarities in the FAT structural profiles for the MultiCenter Test-Retest cohorts. a, The similarity matrices between any two FAT structural profiles. The FA, AD, and RD profiles could separate well the within-subject similarities from the between-subject similarities. b, ROC curves for the separation of intersubject and intrasubject similarities using the FAT structural profiles. The FA, AD, and RD profiles could distinguish the intrasubject similarities from the intersubject similarities. Download Figure 1-1, TIF file.
Figure 1-2
Identification of FAT segments. a, Extraction of streamlines passing through each node of the unilateral FAT. Representative streamlines through different nodes are shown. The streamlines shared in common with other nodes are shown in gray, and the ones unique to the respective node are shown in blue. b, Construction of the connectivity matrices defined by the streamlines in connection with the brain regions. The edges that have the highest number of streamlines are demonstrated for each node. The brain regions are defined by the Desikan–Killiany atlas. c, Clustering of the FAT nodes using the group-averaged node-by-streamline matrix. The node-by-streamline matrix for each individual was generated by concatenating 100 node-specific structural connectivity matrices. The node-by-streamline matrix for each individual was normalized by the total number of streamlines. The group-averaged node-by-streamline matrix was generated by the average of the normalized matrices obtained from the 44 subjects in the HCP test cohort. The clustering was performed based on the Euclidean distance between the streamline connectivity using K-means algorithm. d, Visualization of the brain regions reached by the streamlines through different FAT segments. The thickness and the color of the edge correspond to the percentage of the streamlines passing through each FAT segment. The brain regions reached by the highest number of streamlines are shown as nodes. The color of each node represents the percentage of the streamlines ending in this region. The pie chart demonstrates the streamline constitution of the top 5 brain regions with the highest streamline numbers. l, Left; r, right; SF, superior frontal; PREC, precentral; CMF, caudal middle frontal; INS, insula; IP, inferior parietal; SMAR, supramarginal; MT, middle temporal; PORB, parsorbitalis; CAU, caudate; PUT, putamen; ST, superior temporal; BSTS, bankssts; IT, inferior temporal; SP, superior parietal; THAL, thalamus; CAC, caudal anterior cingulate; PCUN, precuneus. Download Figure 1-2, TIF file.
We first tested the stability of these structural measures within individuals across sessions. For the HCP Test-Retest cohort with 44 participants (the interval between two dMRI scans, 139.3 ± 68.2 d), the FA, AD, RD, span, and volume were robust across the two sessions, with high ICC values (FA, 0.93; AD, 0.83; RD, 0.85; span, 0.97; volume, 0.91), while the MD, curvature, and radius showed relatively lower ICC values (MD, 0.62; curvature, 0.70; radius, 0.64; Fig. 1c). For the SLIM cohort with 192 participants (the interval between two dMRI scans: 302.8 ± 92.6 d), the FA, AD, RD, span, and volume were also stable across the two sessions (ICCs all >0.84; Fig. 1c). Furthermore, we tested the stability of the structural measures across different scanning sites using the MultiCenter dataset, in which three participants underwent 12 repeated scans at 10 sites. The similarity matrices are illustrated in Extended Data Figure 1-1a, showing a clear separation between intrasubject and intersubject similarities. All ICC measures from the three cohorts are summarized in Figure 1c. These results demonstrate that the structural features of FAT, such as FA, RD, AD, span, and volume, possess good (0.75 < ICC < 0.9) to excellent (ICC > 0.9) stability within individuals across different scanning sessions and scanning sites.
Next, we tested the identifiability of individual subjects using the structural features of FAT. For the purpose of evaluation, identification was performed across pairs of FAT structural profiles obtained from two scanning sessions, with one as a “target” set and the other as a “database” set as described in previous functional fingerprinting study (Finn et al., 2015). One individual's FAT profile from the target set was selected and compared against each of the profiles in the database set to find the one with the highest similarity, defined as the maximum of the absolute value of Pearson's correlation coefficient. The identification accuracy was computed as the ratio of the number of subjects that have been successfully identified to the total number of subjects. For the HCP Test-Retest cohort, the FA, AD, RD, curvature, and span showed high identification accuracy (Test-Retest accuracy/Retest-Test accuracy: FA, 0.98/0.98; AD, 0.98/0.98; RD, 0.93/0.91; curvature, 0.89/0.91; span, 0.91/0.96; Fig. 1c), while MD, volume, and radius showed lower identification accuracy (MD, 0.41/0.34; volume, 0.57/0.52; radius, 0.27/0.22). For the SLIM cohort, the identification accuracies of FA, AD, and RD (FA, 0.96/0.97; AD, 0.95/0.96; RD, 0.91/0.93) were much higher than the other metrics (Fig. 1c). Block-restricted permutation testing or nonparametric permutation testing was conducted to compute the statistical significance of the identification rates. The permutation p values for the above identification tests were all <0.001. Collectively, the FA, AD, and RD showed consistently high identifiability of individual subjects across different cohorts.
In addition, the separation between intrasubject and intersubject similarities of FAT structural profiles was assessed using a receiver operating characteristic (ROC) analysis in the three cohorts (Fig. 1d). For the HCP cohort, the area under the curve (AUC) values of separation using FA, MD, AD, and RD of FAT were 0.9957, 0.7800, 0.9986, and 0.9995, respectively (Fig. 1d). The geometric features of curvature, span, radius, and volume had separation AUCs of 0.9869, 0.9561, 0.7692, and 0.9284, respectively. The ROC results of SLIM and MultiCenter cohorts were similar to those of HCP data, as shown in Figure 1d and Extended Data Figure 1-1. Thus, all the imaging metrics except for MD and radius showed good performance in the separation of intrasubject and intersubject similarities. To further explore the contribution of genetic factors to FAT structural properties, we performed the heritability analysis using twins data in HCP. The additive genetic factor was significant across all the eight structural measures (Fig. 1e).
Together, our results revealed that the FA, RD, and AD measures of FAT had excellent reproducibility and high identifiability of individuals, indicating their potentials as structural fingerprints. Our results also showed the contribution of genetic factors to the uniqueness of FAT diffusion profiles for individual subjects.
Associations of FAT diffusion profiles with cognitive control functions
To demonstrate the anatomic locations of FAT in relation to FPN and CON, we extracted the cortical areas that FAT connects with and overlaid them onto the cortical surface in FreeSurfer. As shown in Figure 2a, the cortical areas that FAT links to are substantially overlapped with FPN and CON, suggesting that FAT might serve as a structural backbone connecting the functional areas of these two networks. We further investigated the relationship between FAT diffusion profiles and cognitive control functions in 466 healthy adults from the HCP S500 dataset using their diffusion measures and 11 cognitive control assessments (Menon et al., 2020). Since FA is the diffusion measure that showed the highest Test-Retest reliability and the best identifiability of individuals, we focused on investigating FA profiles of FAT in relation to the cognitive control functions. We first examined the FAT diffusion measures in relation to the functional activations of CON and FPN in healthy subjects when performing cognitive control tasks (i.e., working memory and relational reasoning tasks). Significant positive correlations were found between the averaged FA of FAT and functional activation levels of these two networks during working memory task (Fig. 2b). A mediation analysis indicated that the activation levels of IFG and SFG clusters that FAT connected with mediated the relationship between the averaged FA of FAT and the working memory accuracy (bootstrapped p = 0.012), as shown in Figure 2c.
Relation among FAT diffusion measure, functional activation of cognitive control network (CCN), and cognitive control task performance. a, The cognitive control networks and the distribution of the ending cortical points of FAT. i, The frontoparietal network (blue) and cingulo-opercular network (purple) are defined by Power et al. (2011). ii, The probabilistic maps of the ending cortical points of FAT derived from HCP S500. The color bar denotes the percentage of the subjects showing the end points at each cortical voxel. b, Significant correlations between averaged FA of FATs and working memory activation levels were mainly identified in the CCN. c, The relationship between FA of bilateral FATs and working memory (WM) accuracy was mediated by working memory activation levels of IFG and SFG clusters within CCN, controlled for age, gender, years of education, and head motion. Mediation results are shown as the standardized path coefficients (SEs) for each path.
Furthermore, we found that white matter fibers passing through different segments of FAT tend to connect different cortical regions of the brain. For example, the streamlines through the inferior “nodes” of the FAT preferably reach the IFG, while the streamlines through the superior nodes preferably end in the SFG (Extended Data Fig. 1-2b). Therefore, we extracted the FAT segments based on the clustering features of the brain regions reached by the streamlines through different nodes (Extended Fig. 1-2b,c). From a K-means clustering analysis, four segments of the unilateral FAT were extracted (Extended Fig. 1-2d). The inferior-frontal segment of FAT showed maximum passing streamlines from the subregions of IFG, including POPE and PTRI, while the inferior-middle segment showed maximum passing streamlines from the subregions of dorsolateral prefrontal cortex (dlPFC) and IFG, including rostral middle frontal (RMF) and POPE gyrus.
Next, we explored the relationship between FAT subsegment structural features and cognitive control functions using multivariate analysis. sCCA was applied to delineate the sparse and interpretable multivariate patterns of FAT profiles in relation to cognitive control functions while minimizing multicollinearity of the model. Age, sex, and years of education were regressed out to eliminate potential confounding effects from these factors. Based on the covariance explained, the first four canonical modes were selected for further analysis (Extended Data Fig. 3-2). The significance of each mode was determined by the block-restricted permutation test and corrected by FDR for multiple comparisons. The first two modes were significant in our results (mode 1: r = 0.36, pFDR = 0.005; mode 2: r = 0.32, pFDR < 0.001; Fig. 3a). Each canonical mode represents a distinct pattern that relates a weighted set of FAT segments to a weighted set of cognitive control function scores (Fig. 3b). A resampling procedure was used to validate the significance for the two modes. The first mode related the relational task accuracy and processing speed to FAs in the inferior-middle and superior-middle segments of bilateral FATs. The second mode linked higher Flanker task scores and better working memory task accuracy to higher FA in the inferior-frontal segment of right FAT. We measured the cosine similarity index between the loadings and number of crossing streamlines of each segment for the brain regions in Figure 3d. As shown in Figure 3c, the highest cosine similarity index was in right RMF and right POPE for the two modes, respectively. Our post hoc sCCA showed the statistical significance of covariance for the first two modes in the ILF and arcuate fasciculus (Extended Data Fig. 3-3, block-restricted permutation p-value). When further examining the significance of the relationship to specific cognitive control functions, we found that ILF and arcuate fasciculus both showed significant associations with relational task accuracy, and arcuate fasciculus is also related to working memory task accuracy (Extended Data Fig. 3-4). For inhibition function, FAT is the only tract that showed significant relevance. To further test the robustness of our findings, a 10-fold cross-validation was performed to evaluate the generalizability of the two identified brain–behavior associations on unseen data. After 1000 iterations of block-restricted permutations, the averaged canonical correlation coefficients for the two modes were significant (mode 1: out-of-sample r = 0.13, pFDR = 0.021; mode 2: r = 0.23, pFDR < 0.001; Extended Data Fig. 3-5), which consolidated our findings.
Association results between the FAT subsegment diffusion profiles and cognitive control functions evaluated using sCCA. a, Scatter plots of the correlations between the FA profile of FAT and cognitive control scores in the first two canonical modes selected based on the covariance explained (Extended Data Fig. 3-2). Both of the two modes were significant in the block-restricted permutation tests (1000 times) after FDR corrections The schematic illustration of the sCCA pipeline is shown in Extended Data Figure 3-1. The 10-fold cross-validation was performed to evaluate the generalizability of these two sCCA modes on unseen data (Extended Data Fig. 3-5). Results of post hoc sCCA for the other 16 tracts are presented in Extended Data Figures 3-3 and 3-4. b, Loadings of the FA profile of FAT (middle panel) and cognitive control ability (left panel), and the average loadings of FAT segments (right panel) in the two sCCA modes. The features whose 95% confidence intervals did not cross zero in resampling procedure (1000 times) are claimed significant. *Significant features in resampling tests. c, Cosine similarity between the sCCA loadings and the number of crossing streamlines of the FAT segments for the cortical regions. The color bar indicates the cosine similarity index for each brain region.
Figure 3-1
A schematic illustration of the sCCA pipeline linking the FAT diffusion profiles to cognitive control abilities in our study. Download Figure 3-1, TIF file.
Figure 3-2
Selection of canonical modes based on covariance explained. The first four canonical modes that showed >10% explained covariance were selected. Dashed line denotes the average covariance. Download Figure 3-2, TIF file.
Figure 3-3
The covariance and the significance of the first two modes by post hoc sCCA for the other 16 tracts. Download Figure 3-3, DOCX file.
Figure 3-4
a, b, Loadings of cognitive control ability in the first sCCA modes by ILF (a) and anterior fundus (AF; b) separately. The features whose 95% confidence intervals did not cross zero in resampling procedure (1000 times) are claimed as significant. *Significant features in resampling tests. Download Figure 3-4, TIF file.
Figure 3-5
Distributions of R values between predicted FAT FA diffusion profiles and cognitive control measures using sCCA across 10-fold cross-validation (deep gray) and the null-distribution generated by the block-restricted permutation procedure (light gray). Download Figure 3-5, TIF file.
Associations of FAT diffusion profiles with obsessive-compulsive symptoms of unmedicated OCD patients
Since FAT is closely related to cognitive control function, it is expected that patients with cognitive control deficits would show disrupted microstructure in FAT. To test this hypothesis, we analyzed the dMRI data acquired from a cohort of unmedicated OCD patients (N = 68) and matched healthy control subjects (N = 80). Demographic and clinical characteristics of the included participants are shown in Table 1. There was no significant difference in age, sex, or years of education between the healthy control and patient groups. The diffusion measures in the eight segments of bilateral FAT were compared between two groups. After controlling for age and sex, the patients showed significantly increased FA in the inferior-frontal segment of right FAT (F(1,144) = 9.134, pFDR = 0.05; Fig. 4a) using ANCOVA. We further conducted a stepwise regression analysis to identify the significant predictors of OCD symptom severity. The analysis showed that the FA of the inferior-frontal segment of right FAT was associated with the Y-BOCS obsessive score (R = 0.267, pFDR = 0.028; Fig. 4b), while the FA of the inferior-middle segment of left FAT was associated with the Y-BOCS compulsive score (R = 0.328, pFDR = 0.012; Fig. 4b). Post hoc analysis manifested that the significant differences in FA between the two groups were observed exclusively in the inferior segment of the right FAT (Extended Data Fig. 4-1).
Demographic and clinical characteristics of medication-free OCD patients and healthy controls.
Disrupted FA values in bilateral FAT segments of OCD patients. a, Comparisons of FA in FAT segments between HC and OCD groups. *Statistically significant after FDR correction (pFDR < 0.05). Significantly increased FA was shown in the inferior-frontal segment of right FAT. Post hoc node-wise comparisons for all 18 association or projection tracts were shown in Extended Data Figure 4-1. b, FA values in the inferior-frontal segment of right FAT and in the inferior-middle segment of left FAT were associated with Y-BOCS obsessive (left panel) and compulsive (right panel) scores of OCD patients, respectively.
Figure 4-1
Node-wise comparisons of FA along 18 tracts between HC and OCD groups. The solid line in each plot represents the mean FA measure for each node. The light color band shows the boundaries of the SEM. The light red shaded area represents the section with significant FA differences (FDR corrected). Download Figure 4-1, TIF file.
Discussion
In this study, we first demonstrated the fingerprinting properties of FAT diffusion profiles using three independent datasets with repeated dMRI scans. We next showed that the individual's FAT microstructural property was associated with cognitive control ability in a larger cohort (N = 466) of healthy participants. Moreover, we found that the disrupted FAT microstructure was associated with obsessive-compulsive symptom severity in a separate OCD cohort. These findings suggest that FAT might serve as the structural backbone for FPN and CON underlying the individual differences in cognitive control capability.
FAT diffusion profiles show fingerprinting properties
Uniqueness and stability are important properties of fingerprinting. In our work, stability of the diffusion and geometric profiles of FAT was assessed using ICCs across repeated scans. For the geometric metrics, volume and span demonstrated high ICCs, while radius and curvature showed moderate ICCs, in line with previous findings (Yeh, 2020). For the diffusion measures, we found that ICCs of FA, AD, and RD were consistently high, while that of MD was moderate. This finding aligns with previous studies showing the lowest ICC of MD in whole-brain white matter compared with other diffusion metrics (Duan et al., 2015). A possible reason is that MD is the average of the diffusion tensor eigenvalues, making it more vulnerable to changes in both magnitude of diffusivity and anisotropy of diffusion (Duan et al., 2015).
We further assessed the uniqueness of FAT diffusion metrics and found high identification accuracies (>90%) for FA, AD, and RD of FAT. The identification accuracy using a single FAT in our study is surprisingly comparable to that obtained using whole-brain structural or functional connectome features (Finn et al., 2015; Yeh et al., 2016; Badhwar et al., 2020; Jin et al., 2020). Of note, the cortical regions connected by FAT are largely overlapped with FPN and CON, both showing the highest individual identification rates by their functional connectivity measures (Badhwar et al., 2020; Jin et al., 2020). The FA, AD, and RD are closely related to microstructural properties of axon and myelin sheath (Wozniak and Lim, 2006). During brain development, enhanced individualization of brain connectivity is contributed by both genetic and environmental factors (Yeatman et al., 2012; Kaufmann et al., 2017). Consistent with this, our heritability analysis on diffusion measures in twins confirmed the significant genetic contribution to the fingerprinting property of FAT microstructure.
Different FAT segments contribute to distinct cognitive control functions
Our results showed an association between cognitive control ability and FAT microstructural profiles. More interestingly, functional activation of CON during the working memory task mediated the effect of FAT on the working memory accuracy. Both IFG and pre-SMA regions activated during cognitive control tasks (Cocchi et al., 2013). As the tract connecting these two regions, FAT demonstrated involvement in inhibition, working memory, and visuomotor integration in healthy adults (Aron et al., 2007; Budisavljevic et al., 2017; Varriano et al., 2018). Moreover, microstructural abnormalities of FAT were found in individuals with cognitive control deficits (Li et al., 2019; Rutten et al., 2021).
The diffusion metrics vary along the tract trajectory because of the presence of crossing fibers (Wedeen et al., 2012). Our results showed distinct associations between different FAT segments and cognitive control functions. For example, FAs in the bilateral inferior-middle segments and left superior-middle segment significantly contributed to relational task accuracy. By examining the streamlines passing through these FAT segments, we found that the maximal streamlines reached RMF gyrus, overlapped with rostral lateral prefrontal cortex and dlPFC, which are found to be functionally related to the relational reasoning process in previous literature (Knowlton et al., 2012; Vendetti and Bunge, 2014).
The inferior-frontal segment of right FAT had the largest loading score in association with working memory and inhibition function. The maximum number of streamlines passing through this segment reach the IFG region. Functional IFG activation has been shown in working memory and inhibition control tasks (Nee et al., 2013). Previous work also found that higher FA in right IFG was associated with stronger inhibitory control ability in healthy subjects, in line with our findings that the FAT segment adjacent to the right IFG was associated with inhibition function (Forstmann et al., 2008). Together, our findings suggest that FA in the inferior-middle segment of FAT is related to relational reasoning, while FA in the inferior-frontal segment is associated with working memory and inhibition function.
Disrupted FAT diffusion property is associated with symptom severity of OCD patients
Cognitive control function is commonly disrupted across various mental disorders (McTeague et al., 2017). For example, OCD patients experienced impaired inhibition and working memory function (Norman et al., 2019; Robbins et al., 2019). Disrupted FAT microstructure has been shown in OCD patients (Radua et al., 2014). In addition, the two functional networks (CON and FPN) taht FAT connects have been closely associated with OCD pathology (Stein et al., 2019). The impaired FAT might induce inapposite recruitment of CON and FPN (Hannah and Aron, 2021). In our study, the inferior-frontal segment of right FAT, which is closely linked to IFG and related to working memory and inhibition function, showed increased FA in OCD patients. In individuals with OCD, meta-analyses have revealed longer inhibitory control time and increased inhibitory errors. Moreover, functional studies have shown hyperactivation in brain regions such as the SMA and right IFG during error processing, but hypoactivation in brain regions including IFG and SMA regions during inhibitory control (Norman et al., 2019). In addition, increased functional connectivity between right IFG and SMA was found associated with impaired inhibitory control ability in unmedicated OCD patients (Tomiyama et al., 2022). And our findings provide evidence of disrupted white matter microstructure between right IFG and SMA underpinning the abnormal functional connectivity and brain activation. The abnormality of the structural connection would contribute to the impaired error processing and inhibitory control abilities in OCD patients.
More interestingly, the FA of the inferior-frontal segment of right FAT was associated with the Y-BOCS obsessive score, while the FA of the inferior-middle segment of left FAT was associated with the Y-BOCS compulsive score. The dlPFC and IFG regions, where these segments terminate, are the vital components of goal-directed and habitual inhibitory control networks (Jahanshahi et al., 2015). Previous studies showed that failure of these two networks led to a poor control over obsessive and compulsive symptoms (Banca et al., 2015). Obsession is associated with working memory capacity (Brewin and Smart, 2005; Hofmann et al., 2012), which is often disrupted in OCD patients (Malekshahi Biranvand et al., 2013). In this study, the inferior-frontal segment of right FAT was related to working memory and inhibitory functions and was impaired in OCD patients, consistent with the worsened obsessive symptoms in these patients. Our results showed that FAT was related to cognitive-control functions and specifically involved in the inhibition function. In addition, only right FAT showed altered FA values in OCD patients, which might be related to the inhibitory control function of FAT in OCD psychopathology. Together, we found that the inferior-frontal segment of right FAT, functionally linked to working memory and inhibition capability, was associated with the severity of obsessions in OCD patients.
Limitations and further directions
There are several limitations in our study. First, although our results showed the fingerprinting potentials of the diffusion measures of FAT, the underlying cellular and molecular mechanisms remained to be elucidated. We might need to incorporate other neuroimaging modalities for future studies, such as myelin water imaging or magnetic resonance spectroscopic imaging. Second, our analyses were conducted on young adults' data, but it is unclear whether the fingerprinting properties of FAT profiles remain stable throughout the life span. Further longitudinal studies are needed to characterize the individual trajectory of FAT profiles over time. Third, the lack of cognitive control assessments in OCD patients limited our ability to directly evaluate the relationship between the microstructural alterations in FAT and cognitive control impairments. Fourth, multiple factors like shell number and b-values in dMRI data acquisition might affect tractography accuracy (Jeurissen et al., 2014; Nath et al., 2020). Further studies are warranted to validate the fingerprinting property of FAT tract geometric profiles. Nevertheless, our results showed that the FAT diffusion measures had interindividual variability and different FAT segments were intimately related to distinct cognitive control functions. This might lay a foundation for exploring the therapeutic potential of FAT as an individualized target in mental disorders.
Conclusion
The current study revealed fingerprinting property of diffusion measures along FAT, and individual variations in the diffusion measures of FAT were found to be significantly influenced by genetic factors. Our results also linked diffusion profiles of FAT to cognitive functions in healthy adults, and cognitive impairments in OCD patients. Collectively, these findings shed light on how white matter tracts serve as the structural backbone underlying individual cognitive control capability.
Footnotes
This work was supported by the Shanghai Pilot Program for Basic Research-Shanghai Jiao Tong University (Grant 21TQ1400203); the Program for Professor of Special Appointment (Eastern Scholar) at Shanghai Institutions of Higher Learning; Academic Leader of Health Discipline of Shanghai Municipal Health Commission (Grant 2022XD025); and General Project of Shanghai Municipal Health Commission (Grant 202140054). X.X. and Y.Y. were supported by the Intramural Research Program of the National Institute on Drug Abuse of the National Institutes of Health.
The authors declare no competing financial interests.
- Correspondence should be addressed to Yihong Yang at yihongyang{at}intra.nida.nih.gov or Yao Li at yaoli{at}sjtu.edu.cn
