Resting‐State and Task Functional Magnetic Resonance Imaging Network Topology Metrics With no Threshold Selection to Predict Cognition

ABSTRACT

Network topology measures characterise brain networks' organisation. Graph theoretical approaches have shown fMRI topology metrics' association with cognitive performance. Because arbitrary connectivity threshold selection biases such metrics, alternatives including the minimum spanning tree (MST) and novel measures following principles of persistent homology were proposed. The present study compared alternative and graph theoretical metrics in association with cognition for resting‐state and task‐fMRI. Functional connectivity matrices were computed from Human Connectome Project (Young Adult) fMRI scans during resting‐state, working memory (WM), gambling, language, motor, relational processing, social cognition, and movie‐watching conditions. Global efficiency, clustering coefficient (at three thresholds), diameter, leaf fraction (LF), backbone strength (BS), and cycle strength were measured. Each was tested in association with cognitive test scores. ResultsBS significantly predicted general cognitive performance, specifically progressive matrices score, composite fluid and crystallised cognition, vocabulary, spatial orientation, and WM. Diameter significantly predicted WM. WM task BS outperformed the predictive performance of graph theory measures, but not at rest, where MST LF outperformed other measures. Stronger associations were observed between cognitive test scores and topology measures derived from task‐based fMRI, especially the N‐Back task, as opposed to resting‐state fMRI. Among task‐based topology measures, BS was the most strongly related to cognition.

Key Points

• Topology metrics obtained from task‐based fMRI scans, especially during a N‐Back working memory task, are more associated with cognition than metrics from resting‐state fMRI scans.

• Backbone strength and minimum spanning tree diameter can predict specific cognitive measures in healthy young adults.

• Backbone Strength especially outperforms other topology metrics in the prediction of cognition in the working memory fMRI task.

Minimum spanning tree and persistent homology derived metrics are alternatives to conventional brain topology measures that avoid threshold selection and were able to predict cognitive functions in healthy young adults. This was especially true for metrics from task‐based fMRI scans, in particular for working memory tasks, compared to resting state.

1. Introduction

Network neuroscience studies in the past decades have investigated brain networks using mathematical principles of graph theory, representing networks (graphs) in terms of points (nodes) with connections between them (edges), from which can be inferred non‐random topological properties such as the formation of clusters, distant and local integration of nodes through mutual connections, the existence of hubs (central, highly connected nodes), etc. (Bullmore and Sporns 2009; Farahani et al. 2019; Fornito et al. 2013; Han et al. 2025). The structure of a graph gives a theoretical estimate of how “efficient” a brain network can be at transmitting information across its regions: minimising paths required to be crossed for the information to flow, while maximising the information exchanged within a “specialised” segregation of clusters of nodes, also known as the “small‐world” property (Bullmore and Sporns 2009; Farahani et al. 2019; Han et al. 2025). Following the rationale that a more efficient and integrated brain network should facilitate mental functions, topological measures have been found to be associated with cognitive functions (Farahani et al. 2019) such as working memory (Stanley et al. 2015), attention (Qian et al. 2013; Sun et al. 2014), and neurological disorders such as Alzheimer's Disease, Multiple Sclerosis or Attention‐Deficit/Hyperactivity Disorder, which are understood to alter the brain network's organisation (Bullmore and Sporns 2009; Farahani et al. 2019; Han et al. 2025).

There exists a variety of ways through which the brain can be represented in a connectome (i.e., a structure with nodes and edges). Functional magnetic resonance imaging (fMRI), which estimates the activation of populations of neurons via the concomitant blood‐oxygen‐level‐dependent (BOLD) signal, has been used to infer “functional” networks (Bullmore and Sporns 2009; Farahani et al. 2019; Fox and Raichle 2007). A conventional model defines nodes as delineated brain regions; and edges as the statistical association between the activations of pairs of regions across time, deduced to reflect “functional connectivity” (FC) between them (Farahani et al. 2019; Fornito et al. 2013; Fox and Raichle 2007). FC reflects the neural communication between regions during different mental states, when the individual is at rest or completing a task (Farahani et al. 2019; Fox and Raichle 2007; Smallwood et al. 2021). Given the relatively high degree of noise and artefacts associated with the fMRI/FC data, it has been a common practice to apply thresholding based on arbitrary FC values: removing edges with low FC weight, either via selecting edges with a weight higher than a fixated value or selecting a proportion of the highest‐weighted edges (Fornito et al. 2013). However, that selection complicates generalizability and reproducibility in the literature due to its relative arbitrariness. Furthermore, thresholding a network has a direct effect on topological measures, and biases group comparisons when FC weights or densities of edges above threshold differ (Garrison et al. 2015; Tewarie et al. 2015; van den Heuvel et al. 2017). Testing multiple thresholds may also suffer from an arbitrary range of thresholds, require selecting the best threshold(s) based on ad hoc criteria (e.g., statistical sensitivity of the resulting topology measure), and/or require additional corrections/summarising measures to account for the inflated multiplication of comparisons across the thresholds (Drakesmith et al. 2015; Fornito et al. 2013; van den Heuvel et al. 2017).

Here, we focus on alternative approaches to conventional graph theoretical measures, which avoid arbitrary proportional or absolute FC thresholding selections (Blomsma et al. 2022; Kuang et al. 2018; Ryu et al. 2023; Tewarie et al. 2015). While not threshold‐free in the sense that edges are still selected, the selection follows standardised algorithms that guarantee common network structures and edge densities, without requiring a subjective tuning of the passing FC weights. One example is the Minimum Spanning Tree (MST), a tree‐like network classically built via an algorithm which connects all nodes with the minimum number of edges (N nodes—1), starting from the strongest edges and explicitly omitting edges that form loops, step‐by‐step, guaranteeing unique networks with equal densities, while prioritising the most reliable edges and minimising the effect of noise (Blomsma et al. 2022; Kruskal 1956; Tewarie et al. 2015). Expanding on that, approaches derived from “persistent homology” (PH) were applied on MST‐like network building (i.e., “graph filtration”) by Ryu and colleagues (Ryu et al. 2023). PH preserves information about the excluded loop‐forming edges (i.e., “cycles”, hence the homology) by tracking changes throughout the building process (hence persistent) (Kuang et al. 2018; Liang et al. 2021; Ryu et al. 2023). Such use of PH can help identify properties both for edges making the fundamental “backbone” (or MST) of a network, and its cycles (loop‐forming edges) (Ryu et al. 2023). Because the loop‐forming edges are also the strongest in the building order, the most reliable edges are also prioritised here, the weakest edges being distinguished as only addable to the backbone after it is already complete. PH complements information from the highly connected/integrated nodes of the backbone with their interconnecting cycles through which information also flows (Kuang et al. 2018; Liang et al. 2021; Ryu et al. 2023). Figure 1 depicts an MST building process and describes MST‐based and PH‐based measures which can be derived.

FIGURE 1.

Illustration of conventional topological network measures and concepts. More detailed definitions of topological measures and their interpretations are provided in the methods' subsection “Topological measures”.

More importantly, measures derived from PH principles were found to be more sensitive in predicting cognitive performance in a lifespan sample (in functions including fluid reasoning, episodic memory and vocabulary); compared to overall FC (mean weight of the whole network) and conventional graph theory‐based measures (Ryu et al. 2023). Given a future use of topological measures could be to develop biomarkers (Blomsma et al. 2022; Bullmore and Sporns 2009), seeking the most sensitive measure is an important aim for network topology studies. Identifying features that are reproducible and reliably associated with cognition helps toward potentially indicating normal or pathological processes, which can in turn integrate biomarkers that assist clinicians in their guidance, diagnostic processes or treatment decisions (Blomsma et al. 2022). Furthermore, cognitively reliable properties contribute to the development of measures that are meaningful and accurate: because they are commonly interpreted in terms of information flow, efficiency, integration, etc. of brain networks, and thought to support cognitive ability (Bullmore and Sporns 2009; Ryu et al. 2023), metrics that can best predict cognition are more interpretable. Evidence for differences in predictive performance between fMRI network measures is still limited, especially for task‐based fMRI, but task‐derived FC is considered to engage processes that match more with phenotypes that are studied in cognitive assessments (Zhao et al. 2023). Previous literature suggests task‐fMRI functional connectomes predict behavioural outcome measures better than resting‐state fMRI (Greene et al. 2018; Zhao et al. 2023), but this difference remains to be investigated for the brain network topology measures.

In the present study, we tested the effects of MST, PH, and graph theory derived topological measures in resting‐state and task‐based fMRI data on a wide range of cognitive tasks. In light of the above discussion, we hypothesised that topological measures based on standard non‐arbitrary edge selection (PH and MST‐based) perform better in predicting cognitive performance compared to their conventional graph theory counterparts, and that prediction is higher for measures from task‐based fMRI connectomes compared to resting‐state fMRI connectomes.

2. Materials and Methods

2.1. Dataset

This study used the Human Connectome Project Young Adult (HCP‐YA) neuroimaging dataset by (Van Essen et al. 2013) comprising 1200 healthy participants, of which 881 were included because they had valid fMRI data (i.e., without excessive head motion: mean framewise displacement > 0.3). The HCP study was approved by the Institutional Review Board (IRB) at Washington University (#201204036). Data was obtained by the present authors in agreement with the WU‐Minn HCP Consortium Open Access Data Use Terms. We included participants that had completed all fMRI tasks and neuropsychological measures: two subgroups were created based on data acquired in the 3 T scanner (N = 805) or 7 T scanner (N = 169), because they proposed different fMRI tasks. The two subgroups mostly overlapped, with 32 participants being unique to the 7 T scanner subgroup. The distributions of age and sex are displayed in Figure 2.

FIGURE 2.

Age and sex distributions from the dataset.

2.2. Cognitive Measures

The cognitive measures, several of which were from the NIH Toolbox, are briefly described in Table S1.1, and the full set of assessments laid out in the ‘WU‐Minn HCP 1200 Subjects Data Release Reference Manual’, chapter 5. Scores unadjusted by age were selected over adjusted scores when applicable, due to age being controlled in the models subsequently across all measures, including those that did not have an age‐adjusted score available. Measures included:

• Visual Episodic Memory: Picture Sequence Memory

• Cognitive Flexibility: Dimensional Change Card Sorting

• Inhibition: Flankers test

• Fluid Intelligence: Penn Progressive Matrices (Total Correct Responses)

• Oral Reading Recognition

• Vocabulary Comprehension: Picture Vocabulary

• Processing Speed: Pattern Comparison Processing Speed

• Self‐regulation/Impulsivity: Delay Discounting (Area Under the Curve (AUC) for Discounting of $200)

• Spatial Orientation: Variable Short Penn Line Orientation Test (Total Correct Responses)

• Sustained Attention: Short Penn Continuous Performance Test (True Positives)

• Verbal Episodic Memory: Penn Word Memory Test (Total Correct Responses)

• Working Memory: List Sorting Test

• Cognition Fluid Composite

• Cognition Crystallised Composite

• Emotion Recognition: Penn Emotion Recognition Test (Total Correct Responses)

2.3. Neuroimaging Acquisition Protocol

The EPI‐BOLD neuroimaging acquisition protocols are summarised in Table 1. Participants were scanned at Washington University (St. Louis, United States). The 3 T scanner used a 32‐channel Siemens receive head coil and a body transmission coil. The 7 T scanner used a Nova32 32‐channel Siemens receive head coil and a head‐only transmit coil surrounding the receiver head coil.

TABLE 1.

Neuroimaging protocol.

N participants	Scanner	TR (s)	TE (ms)	Flip angle (°)	Echo spacing (ms)	FOV (mm)	N slices	Voxel size (mm)	Multiband factor	BW (Hz/Px)
805	3 T Siemens Connectome Skyra	720	33.1	52	0.58	208 × 180	72	2 × 2 × 2	8	2290
169	7 T Siemens Magnetom	1000	22.2	45	0.64	208 × 208	85	1.6 × 1.6 × 1.6	5	1924

Note: See more details in the “WU‐Minn HCP1200 Subjects Data Release Reference Manual” https://humanconnectome.org/storage/app/media/documentation/s1200/HCP_S1200_Release_Reference_Manual.pdf.

Abbreviations: BW = bandwidth, FOV = field‐of‐view, HCP‐YA = Human Connectome Project Young Adults, TE = Echo time, TR = repetition time.

2.4. fMRI Resting‐State and Task Conditions

Using the 3 T scanner, participants were recorded at rest via passive viewing of a fixation cross during four runs of approximately 15 min.

fMRI tasks included:

• Language: Developed by Binder and colleagues (Binder et al. 2011), the Story‐Math protocol proposed story stimuli combined with mathematical stimuli matching linguistic auditory and phonetic features or the word stimuli to control for non‐semantic processes, both engaging the participant's attention. Two runs interleaving 4 blocks of a story task and 4 blocks of a math task (each block is 0 s on average), and additional math trials at the end to complete each 3.8 min total run. The story blocks presented auditory stories (5–9 sentences) adapted from Aesop's fables, followed by a 2‐choice question about the topic of the story. The math blocks presented auditive addition and subtraction problems, followed by “equals” and then a 2‐choice question. Participants pushed a button to select an answer.

• Working memory: Two runs. In each run, participants were presented with 8 blocks of trials (10 trials of 2.5 s each, for 25 s) and 4 fixation blocks (15 s). Each trial presented the stimulus for 2 s, with a 500 ms inter‐task interval (ITI). Each stimulus consisted of pictures of places, tools, faces, or body parts (non‐mutilated parts of bodies with no “nudity”), in separate blocks. Half of the blocks were presented with a 2‐back task and the other half with a 0‐back task. Each block started with a 2.5 s cue indicating the 2‐back task or a target for 0‐back.

• Gambling: Adapted from Delgado et al. (2000). Two runs with 4 blocks of trials (8 trials) and 4 fixation blocks (15 s). Participants were asked to guess the number on a mystery card (represented by a ‘?’) to win or lose money. The ‘?’ was shown for up to 1500 ms (if the participant responded before, a fixation cross was displayed for the remaining time). Participants were informed numbers ranged from 1 to 9 and asked to indicate if they thought the number was more or less than 5 by pressing one of two buttons. Upon pressing, a 1 s feedback was shown with the number on the card and (1) a green up arrow with ‘$1’ for reward trials, (2) a red down arrow with ‘−$0.50’ for loss trials; or (3) the number 5 and a grey double headed arrow for neutral trials. Each block had 8 trials, with a 1000 ms ITI, that were either mostly reward (6 reward trials pseudo randomly interleaved with either 1 neutral and 1 loss, 2 neutral, or 2 loss trials) or mostly loss (6 loss trials interleaved with either 1 neutral and 1 reward, 2 neutral, or 2 reward trials). Each run had 2 mostly reward and 2 mostly loss blocks.

• Motor: Adapted from Buckner et al. (2011). Two runs with 13 blocks of 10 movement trials of 12 s, and 3 fixation blocks (15 s). Participants were presented with a visual cue (3 s) asking to either tap their left or right fingers, squeeze their left or right toes, or move their tongue. Each block consisted of 10 movements for 12 s, preceded by a 3 s cue. Each block consisted of 2 of tongue, 4 hands (2 right and 2 left), and 4‐ft movements (2 right and 2 left).

• Relational Processing: This task, adapted from Smith et al. (2007), consisted of two runs with 3 “relational” blocks, 3 “matching” blocks, all 18 s, and 3 fixation blocks (16 s). Participants were presented with 6 different shapes with 1 of 6 different textures. In the relational processing condition, participants were shown 2 pairs of shapes, with one pair at the top of the screen and one at the bottom. They were tasked to decide what dimension (shape or texture) differed across the top pair of objects and then whether the bottom pair differed along that same dimension. In the control matching condition, participants were shown two objects at the top of the screen and one object at the bottom, and a word in the middle (either ‘shape’ or ‘texture’). They were told to decide whether the bottom object matched either of the top objects on the word's dimension. Answers were given by pressing one of two buttons. For the relational condition, the stimuli were 3500 ms long (500 ms ITI), with 4 trials per block. In the matching condition, stimuli were 2800 ms long (400 ms ITI) with 5 trials per block.

• Social Cognition (Theory of Mind): Two runs of 5 video blocks and 5 fixation blocks (15 s). Participants were presented with video clips (20s each) of objects (squares, circles, triangles) that either interacted in a social way or moved randomly. Videos were developed by Castelli et al. (2000) or Wheatley et al. (2007). After each video, participants judged whether (i.e., yes/no/not sure) the objects had a mental interaction (i.e., the shapes appear to account for each other's feelings and thoughts). The first run showed 2 Mental and 3 Random videos, the second run 3 Mental and 2 Random videos.

Using the 7 T scanner, resting‐state and movie‐watching scans were selected from the same first session: two resting‐state acquisitions of approximately 16 min, and two movie‐watching sessions, which were 15:12 and 15:18 min, respectively.

• Movie: Participants were watching short (1–4.3 min) clips concatenated into videos of 11.9–13.7 min, including a Vimeo repeat validation clip. The first movie watching session contained clips of freely available independent films under Creative Commons licencing and the second Hollywood movies from Cutting et al. (2012). Before each clip and at the end of the whole movie there were 20 s of ‘rest’ (black screen with ‘REST’ in white text). Details of the sources of each clip and timings were made available in ConnectomeDB on the S1200 Project page (‘7T Movie Stimulus Files’, ‘Task Resources’ heading) (https://db.humanconnectome.org/data/projects/HCP_1200).

2.5. fMRI Scans Preprocessing

The fMRI data obtained from the HCP‐YA provided preprocessed derivatives generated as surface meshes in fsLR 32 k space (‘*_space‐fsLR_den‐91k_desc‐denoised_bold.dtseries.nii’), which were post‐processed using the XCP‐D algorithm v0.11.0 (Mehta et al. 2024) with the ‘linc’ mode set of recommended parameters further defined in XCP‐D documentation (https://xcp‐d.readthedocs.io/en/latest/workflows.html#processing‐modes). As part of the post‐processing, timeseries were denoised by regressing out the six motion parameters, mean global signal, mean white matter signal, mean cerebrospinal fluid signal with their temporal derivatives, and quadratic expansion of six motion parameters, tissue signals, and their temporal derivatives. Band‐pass filtering was applied with a second‐order Butterworth filter retaining signals between 0.01 and 0.08 Hz. The post‐processed timeseries were parcellated into 200 cortical regions‐of‐interest (ROIs) from the Schaefer‐200 atlas (Schaefer et al. 2018), plus 19 subcortical ROIs from FreeSurfer's aseg parcellations (Fischl 2012), generating an N × 219 matrix for each run, where N is the number of frames. Similar results (see Supporting Informations S2) were obtained when these analyses were repeated using the Schaefer‐400 atlas for constructing the functional connectome.

2.6. Functional Connectome Building

FC networks were determined as the fisher Z‐transformed Pearson r correlations between the average BOLD signal timeseries of each of the 219 parcellations. A FC matrix was computed from concatenated timeseries of each task (all events combined). Due to the ambiguity of negative edges' interpretation, negative weights were set to zero. If strongly negative weights are interpreted as strongest edges and integrated in the MST/Backbone, then it is difficult to argue that the latter is a representation of the most important connections or represents the main ‘information flow’ in the network.

2.7. Topological Measures

Network topology calculations were achieved using R v.4.3.2 (R Core Team 2023). Graph theoretical measures were calculated using the igraph package (Csardi and Nepusz 2006), repeated for three purposefully arbitrary proportional FC thresholds (10%, 20%, and 30% highest edges):

• Global Efficiency (GE): the average inverse shortest path length, i.e., the number of edges required to connect each pair of nodes. Increased global efficiency reflects more global integration.

• Weighted Clustering Coefficient (CC, also named transitivity): the fraction of triangles formed in the network scaled by the triangles' edges FC weight, following Barrat's calculation (Barrat et al. 2004; Csardi and Nepusz 2006). Increased clustering coefficient reflects more network segregation.

The MST/backbone was computed using single‐linkage hierarchical clustering from R's stats package, which iteratively merges clusters with the smallest distance between nodes. To apply it, FC matrices were inversed (i.e., 1‐edge value) so that strongly correlated edges were prioritised (Ryu et al. 2023; Tewarie et al. 2015), and converted to distance matrices (indicating, for each pair of nodes, the sum of weights of the links/edges separating them; the lower the inverse weight the “shorter” the distance). Hierarchical clustering, starting from the “closest” nodes in the distance matrix, links nodes together until they make one cluster network, adding edges one at time. Because it only adds new nodes to clusters, this skips loop‐forming edges as is explicitly done in Kruskal's MST algorithm (Kruskal 1956). This ends up producing an acyclic tree‐like network with density N nodes −1, that is, the MST.

Two other commonly analysed measures which are strongly associated with graph measures (Tewarie et al. 2015) and commonly used in MST studies (Blomsma et al. 2022) were calculated:

• Diameter: the longest path between a pair of nodes in the MST, divided by the total number of edges. A lower diameter reflects an increased global integration.

• Leaf Fraction (LF): the fraction of leaf nodes (nodes with one edge) in the MST divided by the maximum possible number (i.e., N nodes −1) of edges (leaves). A lower leaf fraction reflects a less centralised topology, and a higher leaf fraction reflects a higher reliance on central “hub” nodes.

In a similar manner to graph theory's small worldness, an efficient brain network can be characterised as a departure from a line‐like topology (low leaf fraction and large diameter) towards a more star‐like topology (high leaf fraction and low diameter) (Blomsma et al. 2022).

The above process mirrors PH‐based ‘filtration’ of a simplicial complex (representation of data as nodes and edges), in which components (isolated nodes or clusters of connected nodes) are gradually merged when edges connect them together, every time the distance threshold for an edge to be added increases. This ‘filtrates’ topological structures with components disappearing (death, by merging into a bigger cluster) and other components appearing (birth, e.g., loops and higher order structures) each step of the way (Kuang et al. 2018; Liang et al. 2021; Ryu et al. 2023). In this study, every time the next “closest” edge in the distance matrix is loop‐forming, a cycle appears, so its weight is stored in a birth weight set (W10), while when the next edge connects components (clusters “dying” to a larger cluster), their weight is stored as a death weight set (W0), until there is only one merged component connecting all nodes. Summarising these weights, PH‐based measures were used following Ryu et al. (2023):

• Backbone strength (BS): mean weight of edges in W0, that is, the edges of the backbone/MST network, reflecting the core brain‐integrated processing.

• Cycle strength (CS): mean weight of edges in W10, that is, the cycles captured from the MST building, reflecting information flow between the integrated hubs during the processing.

• Backbone Dispersion (BD): the range of elements in W0, that is, weights of the backbone/MST (the largest value minus the smallest)

Another control to topological networks, overall FC (or mean connectivity), was defined as the mean of all weights in the matrix. It has been shown to have an impact on topological network measures (van den Heuvel et al. 2017) and will have an effect on selected PH measures as they are also weight‐based. Mean connectivity was also tested separately on its own.

2.8. Statistics

Statistical modelling was also run in R. All numerical predictors were scaled to allow model comparability.

To examine the effect of network topology on cognition globally, each network topology measure was first tested for the prediction of scores merged together into one cognitive score, with scores within each measure transformed into a Z‐score to make scales comparable. For each topology measure and fMRI condition, a separate mixed‐effect linear model (nlme package v3.1.164) was run as the following formula: cognitive performance ~age + sex + cognitive test type + topology measure + overall FC + (1|participant_id). Cognitive test levels were allowed to keep their own residual variance. Overall FC was dropped if it created multicollinearity with the topological measure (r > 0.09, VIF > 10, or tolerance < 0.02, as per criteria given by (Mayers 2013; Menard 2001; Myers 1990)). Independence of errors was verified using the Durbin–Watson test. A semi‐partial R‐squared in all mixed‐effect models was computed for each topological network measure using the Jaeger et al. approach in the r2glmm package (v.0.1.3) (Jaeger et al. 2017).

To investigate the effect of network topology on cognitive measures individually, a set of models was run for each cognitive measure and fMRI condition separately. For each cognitive test score, a linear model was fitted to test the effect of the topological measure on the cognitive score, controlling for age, sex, and overall FC (if no multicollinearity): Cognitive performance~age + sex + topology measure + overall FC. and a partial R squared (pR ²) was extracted with the rsq package v.2.7 (Zhang 2024).

To investigate the relative predictive performance of network measures, the difference in occurrence of significant effects (at p < 0.05) was calculated via a Chi‐square test (chisquare package v.1.1.1 in R (Alberti 2024)), with Monte Carlo Simulation at 2000 replicates, across topology measures. All cells were to have at least 1 count (Campbell 2007), and topology measures without a single significant effect were removed from the analysis. For a post hoc analysis, standardised residuals adjusted across cells with the moment‐correction approach (García‐Pérez and Núñez‐Antón 2003) were extracted. Because of the overlapping samples and varying sample sizes, two separate Chi‐square tests were performed for the 3 and 7 T fMRI task results. p‐values were corrected for false discovery rate (for the positive occurrence of significant tests, one‐tailed) across 3 and 7 T findings.

Finally, two regression models were run for the 3 and 7 T scanners separately, to test whether pR ² values significantly differed between topology measures across all models. The models tested for the effect of topological measure (as a categorical variable), fMRI task (rest and other conditions), and their interaction, on their respective pR ² values, controlling for cognitive measure type (three categorical predictors). The generalised least squares (GLS) method was used with the nlme package to account for heteroscedasticity, log‐transforming the pR ² and allowing the fMRI task categorical levels to keep their own residual variance. The regression model can be summarised in this equation:

$$\log \left({\mathrm{p}R}^2\right)~\text{topology measure}+\text{cognitive measure}+\text{fMRI task}$$

The baseline intercepts for the cognitive measures and topology measures categorical variables were coded via sum contrast (nlme's contr.sum option), where the grand mean was the baseline reference instead of an arbitrary level. Resting‐state was chosen as the baseline intercept for the fMRI task categorical variable.

To test how predictive performance specifically differed across topological measures and fMRI tasks, post hoc estimates of the predicted marginal means of log(pR ²) across all interaction terms were extracted via the emmeans R package v.1.10.6 (Lenth 2024), comparing measures pair‐wise across topological measures: PH measures against Graph theory measures against MST measures (not comparing measures from the same broader category). Correction for false discovery rate was applied across all contrasts from both 3 and 7 T scanners.

3. Results

3.1. Global Effect of Network Topology on Cognitive Performance

Figure 3 depicts the effect of each model per network topology and fMRI task on all cognitive HCP‐YA measures scores grouped together. All effect sizes are low (pR ² < 0.0196 as per Cohen's criterion (Cohen 1988), see the discussion section for its contextualisation in the brain network literature). Topology measures that significantly predicted overall cognition were BS and Diameter in the working memory task, and BS, GE (10%), and overall FC in the language task.

FIGURE 3.

Effect of topological network measures on cognitive performance across all measures and fMRI conditions. Models testing network measures' effect on the relative score of all cognitive measures, controlling for age, sex, and if applicable, overall FC, for each fMRI condition. Small ✱: significant coefficient at p < 0.05. Large ✱: significant coefficient at p < 0.05 after adjustment for false discovery rate.

3.2. Task‐Specific Effect of Network Topology on Cognitive Performance

Figure 4 depicts the effects of network topology measures on each cognitive measure scores and fMRI task where at least 1 topology measure's effect survived FDR correction. Following FDR correction across all models together, the following effects remained significant a p < 0.05: working memory and language tasks' BS on fluid intelligence (Penn's matrices); working memory task's BS and Diameter on the cognition fluid composite score; working memory task's BS on the cognition crystallised composite score, vocabulary comprehension, spatial orientation, working memory and (visual) episodic memory scores; working memory task's CS on (verbal) episodic memory; motor task's BS on fluid intelligence (Penn's matrices); resting‐state's LF on self‐regulation; language and motor tasks' BS on working memory scores. See Figures S1.1–S1.3 for the full report for all cognitive test measures and fMRI tasks.

FIGURE 4.

Effects of topological network measures on all cognitive scores across fMRI conditions. Models testing network measures' effect on specific cognitive measure scores, controlling for age, sex, and if applicable, overall FC, for each fMRI condition. Small ✱: significant coefficient at p < 0.05. Large ✱: significant coefficient at p < 0.05 after adjustment for false discovery rate. The thick grey bars separate the 3 T scan‐based tasks (tested for 805 participants) from the 7 T scan‐based tasks (169 participants). See Figures S1.1–3 for the full report for all cognitive test measures and fMRI tasks.

3.3. Relative Performance of Topological Networks Across Models

The Chi‐square test returned significant differences before FDR correction in occurrence of significant effects (all depicted in Figures S1.1–3), per network measure across 3 T fMRI tasks (χ ² = 99.937, Cohen's w = 0.282, p = < 0.001) but not for 7 T fMRI tasks (χ ² = 9.677, Cohen's w = 0.180, p = 0.44). Moment‐corrected residuals for each network measure are laid out in Table 2. After FDR correction across all residuals, BS had a significantly higher chance of returning significant effects across all cognitive measures and fMRI tasks in 3 T scans, BD and CC (at thresholds 10%, 30%) were less likely to return significant effects in 3 T scans, and GE (30%) was more likely to return significant effect in the 7 T scanner.

TABLE 2.

Moment‐corrected standardised residuals of the Chi‐square test for distribution of significant main effects of topology measure on cognitive functions.

Topology measure	3 T fMRI tasks		7 T fMRI tasks
	NS effects	Sig. effects	NS effects	Sig. effects
Backbone dispersion	0.979	−2.805*	NA	NA
Backbone strength	−4.365	12.495***	−0.537	1.954
Clustering coefficient (t.10%)	0.979	−2.805*	0.028	−0.103
Clustering coefficient (t.20%)	0.216	−0.619	0.310	−1.131
Clustering coefficient (t.30%)	1.133	−3.242*	0.028	−0.103
Cycle strength	0.521	−1.493	0.310	−1.131
Diameter	0.064	−0.182	0.310	−1.131
Global efficiency (t.10%)	−0.394	1.129	NA	NA
Global efficiency (t.20%)	0.369	−1.056	0.310	−1.131
Global efficiency (t.30%)	0.827	−2.368	−0.818	2.983*
Leaf fraction	−0.394	1.129	−0.253	0.926
Mean connectivity	0.064	−0.182	0.310	−1.131

Note: NS = non‐significant effects. Sig. = significant effects. After correction for false discovery rate across rates of Sig. NA = could not be tested as no significant effect were present at all (empty Sig. cells).

^* p < 0.05 significant.

^*** p < 0.001 significant.

The GLS model testing the effects of topological measure, fMRI task and their interaction on log(pR ²), controlling for the type of cognitive measure, shows multiple main effects (adjusted for the interaction) and interaction effects.

• For main effects of the cognitive test measures for the 3 T fMRI tasks: Fluid intelligence (Penn Progressive Matrices) (β = 0.84, SE = 0.21, t = 3.97, p(FDR) = 0.002), Vocabulary Comprehension (β = 0.70, SE = 0.21, t = 3.28, p(FDR) = 0.017), Self‐regulation (β = 0.62, SE = 0.21, t = 2.92, p(FDR) = 0.042), and the Cognition Crystallised Composite (β = 0.60, SE = 0.22, t = 2.78, p(FDR) = 0.048) scores were significantly associated with higher pR ² (relatively to the grand mean, averaged across fMRI tasks and topological measures), while Cognitive flexibility (β = −0.94, SE = 0.21, t = −4.37, p(FDR) = <0.001) and Sustained Attention (β = −0.84, SE = 0.22, t = −3.95, p(FDR) = 0.002) were significantly associated with lower pR ². For 7 T fMRI models, Cognitive Flexibility was also significantly associated with lower pR ² (β = −1.56, SE = 0.49, t = −3.22, p(FDR) = 0.020).

• For main effects of fMRI task in models from the 3 T scanner, the language task was significantly associated with higher log(pR ²) (β = 0.94, SE = 0.23, t = 4.17, p(FDR) = <0.001) (relative to the resting‐state baseline, averaged across topological measures). No main effect of the Movie fMRI task was found for 7 T fMRI models.

• For main effects of topological measure from the 3 T fMRI models, MST Diameter was significantly more likely to get higher pR ² (β = 1.58, SE = 0.56, t = 2.82, p(FDR) = 0.048) while CC (10%) was likely to get lower pR ² (β = −0.96, SE = 0.21, t = −4.48, p(FDR) = <0.001) (relatively to the grand mean, at the resting‐state baseline and averaged across cognitive measures). No main effect was found for 7 T fMRI models.

• Interaction effects found for the 3 T scanner‐based models: BS was significantly associated with higher log(pR ²) in the gambling (β = 2.12, SE = 0.76, t = 2.79, p(FDR) = 0.048) and working memory (β = 3.23, SE = 0.76, t = 4.27, p(FDR) = <0.0001) fMRI tasks; while Diameter (β = −3.32, SE = 0.75, t = −4.42, p(FDR) = <0.001) and LF (β = −2.21, SE = 0.75, t = −2.94, p(FDR) = 0.042) in the language task and Diameter in the motor task (β = −3.11, SE = 0.73, t = −4.25, p(FDR) = <0.001) had significantly negative associations with log(pR ²). In the 7 T scanner, CS in the movie‐task was significantly associated with higher log(pR ²) (β = 2.46, SE = 0.87, t = 2.85, p(FDR) = 0.048).

Figure 5 details the significant differences in estimated mean log(pR ²) directly compared across all topological measures and fMRI task combinations, for every pair of contrasts across task‐based and rest‐based conditions, after FDR correction. Figure S1.4 also details all pairs of contrasts with between‐tasks comparisons.

FIGURE 5.

Pairwise contrasts in predictive performance across topological measures in each fMRI task. BD = Backbone Dispersion, BS = Backbone Strength, CLUSTERING = Clustering coefficient, CS = Cycle Strength, DIAM = Diameter, GLOB_EFF = Global efficiency, LEAF = Leaf fraction, LG = Language, MST = Minimum Spanning Tree, PH = Persistent Homology, RAW_FC = Mean functional connectivity, SE = standard error, WM = Working memory. 0.1, 0.2 and 0.3 indicate the proportional threshold (10%, 20%, 30%) of matrices graph measures were computed from. Values are the contrasts in estimated log(partial R‐squared) between conditions. Marginal means of log(pR ²) were estimated from a generalised least squares model testing the effect of topological measure and its interaction with fMRI task condition, controlling for cognitive measures. Each dot is the difference for the condition named first on the x axis relatively to the condition named second. * significant at p < 0.05 **p < 0.01 ***p < 0.001 (FDR corrected). Figure S1.4 also details all pairs of contrasts with between‐tasks comparisons.

4. Discussion

Network topology measures have been used in the brain imaging field to assess the efficiency and integration of brain networks, characterising how information flows across regions, preserving the brain's functionally specialised clusters while also facilitating long‐range connections. The present study focused on topological measures that relied on algorithms which do not involve an arbitrary thresholding selection, that is, MST and PH‐derived measures, assessing their association with cognitive performance across multiple cognitive measures. Further, the study explored differences in associations between topology measures and cognition for networks derived from resting‐state fMRI compared to task‐based fMRI. This study's findings suggest that the PH‐derived BS measure is the most reliable predictor of cognitive performance followed by MST diameter, both for overall performance and specific cognitive tests, especially for networks recorded during the working memory fMRI task. GE or CC's associations were frequently threshold‐specific, and not sensitive across all three selected thresholds, which reflects the instability problem previously flagged in the literature (Garrison et al. 2015; van den Heuvel et al. 2017). The study thus supports the use of BS and MST's Diameter, as they are not subject to arbitrary thresholds and to the instability that thresholding selection creates in more traditional topological measures in graph theory approaches. It also suggests that network measures, both conventional and without arbitrary thresholds, are better associated with cognition in working memory and language tasks fMRI compared to resting‐state fMRI. Such observation could not be verified for Movie‐watching conditions.

This is the first study to our knowledge to compare between task‐fMRI and rest‐fMRI network measures for MST and PH‐derived topology measures. The findings give support to using task‐based fMRI, especially N‐Back and Story‐Math paradigms, over resting‐state scans, to investigate functional networks in relation to cognition. Previous analyses of FC data (including the HCP), highlighted a better prediction of cognition for task‐derived FC over rest‐based FC (Greene et al. 2018; Zhao et al. 2023). Furthermore, working memory‐based functional connectomes, compared to other task‐based FC, best predicted cognition (Greene et al. 2018); this is consistent with the results of our topology network analyses. Previously, FC derived from a semantic N‐Back fMRI task had been shown to elicit GE and less local efficiency or modularity compared to resting‐state fMRI, suggesting the working memory task recruits long‐range connections across the brain; N‐Back fMRI FC also produced more “connector hubs”, that is, nodes connecting subnetworks between each other, and this was associated with accuracy of responses in the task (Cohen and D'Esposito 2016). Because working memory increases integration across brain networks, it is plausible that such networks relate broadly to mental tasks which may depend on them. However, GE in the present picture‐based N‐Back fMRI scan was only significantly associated (at 10%) for spatial orientation, before FDR correction. Given the findings for BS, it is likely long‐range brain integration may be especially relevant for a backbone of highly connected edges, edges that network thresholding is not sensitive enough to discriminate from redundant, cycle‐making edges.

Working memory fMRI tasks may activate functional brain networks whose backbones are especially relevant to cognitive performance: when testing prediction of all cognitive measures together, after FDR correction, working memory task‐derived BS topology predicted global cognition. This finding suggests mental processes and the core backbone of the brain network active during that task (N‐Back) may be serving many cognitive functions. This was further supported by cognitive measure‐specific models, most of which (11 out of 15) were significantly predicted by working memory‐based BS, several before FDR correction (processing speed, oral reading recognition, verbal episodic memory, inhibition) and several after (Penn's progressive matrices, the fluid and crystallised cognition composite scores, visual episodic memory, vocabulary, spatial orientation, working memory). BS may effectively capture relevant connections that other topology networks do not. The working‐memory task‐based MST Diameter, which describes the size of the backbone, also predicted cognitive fluid composite score. The MST's size was therefore more important to predicting cognition than GE, which also captures the span size (inverse of characteristic path length) of the whole network, but without discriminating strongest and loop‐forming edges. Effects of BS remain relatively stable using a larger atlas (Schaefer‐400, Supporting Informations S2).

Other significant predictions after FDR correction include language task‐based BS, predicting global cognition, language task and motor‐task based BS predicting the Penn Progressive Matrices and working memory scores; language task‐based CS, predicting Penn Progressive Matrices scores; and rest‐based MST Leaf fraction, predicting self‐regulation. ‘Loop’ edges isolated by the CS measure from the MST significantly predicted verbal episodic memory in the language task after FDR correction, involving edges that contribute to the integration of brain regions (Ryu et al. 2023). The language Story‐Math fMRI paradigm, which involved diverse stimuli processing, may activate cycle edges that support distinction of new words from truly memorised words in the Penn Word Memory Test, but particularly verbal ones (as the visual episodic memory test did not get a similar association). This is consistent with other significant language‐CS associations that did not survive FDR correction for crystallised cognition and oral reading recognition scores. MST LF at rest reflects the presence of segregated modules, and this may support self‐regulation in anticipation of delayed reward outcomes characteristic to the Delay Discounting test; other consistent associations of the same metric at rest were with cognitive flexibility and inhibition scores.

The language fMRI task GE (10%), BS, and overall FC are also among significant effects for the models testing topology measures' effects on scores from all cognitive measures together (Figure 3), which suggests the functional brain networks active during the former may relate to performing cognitive tasks in general. The fMRI language task had its own independent positive main effect in the GLS analysis on predictive performance across models. This task's overall FC outperformed a few graph theory measures at rest (CE and CS, Figure 5) and in task‐based fMRI (except BS, Figure S1.4) suggests the heuristically simple mean FC is on par with topological approaches in that task (which combined verbal and mathematical elements), itself likely activating relevant brain connections for cognition. While BS and Overall FC did not significantly differ within the language task (Figure S1.4), only BS and CS achieved significant cognitive measure‐specific predictions in the language fMRI among the topological measures surviving FDR correction (Figure 4). It may be that, by nature, BS and CS narrow down the network to specifically high edges, which are more diluted by the averaging of the Overall FC. Graph measures also appear among the significantly increased contrasts for the language task (Figure S1.4), compared to LF and Diameter in the same task, which suggests graph theoretical measures are, at least in the language‐fMRI task, better compared to MST measures.

Additional task‐based measures predicted global cognition, but only before FDR correction (Figure 3). That included movie‐watching BS, which also predicted fluid intelligence (matrices), vocabulary and spatial orientation (Figures S1.1–3). While the movie task improved the performance of CS (interaction effect), evidence that movie‐task topology outperformed rest was limited and unable to confirm previous findings for FC edges in the same dataset (Finn and Bandettini 2021). Gambling‐task BS also had an effect on global cognition (Figure 3), fluid intelligence (matrices), oral reading recognition, vocabulary, verbal episodic memory, working memory and cognition crystallised composite score (Figures S1.1–3); with a reliably improved effect over rest‐BS as confirmed by the significant GLS interaction, which is consistent with previously found improved effects in FC and the HCP (Greene et al. 2018). Motor‐task BD and BS also predicted global cognition before FDR correction (Figure 3), both also predicting inhibition, fluid intelligence (matrices), and vocabulary. BS effects survived FDR correction for fluid intelligence (matrices) and working memory; which is also consistent with motor networks being omnipresent in FC‐cognition association (Greene et al. 2018) and motor task FC being related to within‐network communication (Cohen and D'Esposito 2016), which the BS metric may be prone to capture, although the MST's Diameter mattered less (negative interaction in the GLS model). Finally, social cognition effects were also mostly tied to BS (Figure 3), with individual effects on visual episodic memory, fluid intelligence (matrices), processing speed, working memory, and the cognition fluid composite score (there was also an association between CS and spatial orientation; Figures S1.1–2). Besides the two motor‐task effects, given none of these associations survived FDR correction, this remains inconclusive.

Within the resting‐state fMRI condition, only MST Diameter was associated with higher predictive performance in the GLS model (and CC (10%) lower), and only for the subsample in 3 T scan. After FDR correction, only resting‐state MST LF significantly predicted a score, self‐regulation (Figure 3). MST Diameter also outperformed rest‐based CC and task‐based CC, GE, CS and BD in the 3 T scanner (Figure 5). For resting‐state conditions, the MST's structure may thus be the most informative about cognitive functions. Given LF and Diameter negatively interacted with the language task, and Diameter in the motor task (both relatively to the resting‐state baseline), MST measures may even be more reflective of cognition‐supporting networks in resting‐state scans compared to these tasks. That may not be true to the working memory task, for which Diameter's only FDR‐corrected significant effect was found (Figure 4), and Diameter and LF outperforming rest‐based CC (Figure 5). Diameter refers to the MST's longest path and LF its fraction of leaf nodes, purely in terms of structure rather than weight (which is what PH focused on). At rest, MST's cognition‐specific significant associations (before FDR, Figure S1.1–3) show that LF positively predicted performance while Diameter negatively, which reflect a more centralised ‘star‐like’ MST (Blomsma et al. 2022). Centralised resting‐state MSTs are related to better predictions in cognitive test scores, but the opposite seems to be true for the working memory task MST. Resting‐state fMRI network topology and its relation to cognition is therefore specific and well distinct from task‐based fMRI. We note that the effects of MST Diameter may be affected by the granularity of the functional atlas. Apart from interaction effects, the MST Diameter‐related predictions and comparisons were no longer significant when running the analyses with the Schaefer‐400 atlas (see Supporting Informations S2).

In Ryu and colleagues' study, resting‐state CS significantly negatively predicted Vocabulary (WAIS‐III subtest (Wechsler 1997)) and Fluid reasoning (Z‐score based on WAIS‐III's Block design task, Letter–Number Sequencing test, and Matrix Reasoning subtests) (Ryu et al. 2023). This was not verified for the present's closest constructs (Penn Progressive Matrices and Picture Vocabulary), which may be explained by the difference in specific cognitive measures and the fact their models had also controlled for mean cortical thickness and education. Instead of resting‐state CS, the present findings report associations for CS in the language, movie, and social cognition fMRI tasks (Figures S1–3, movie CS predicting the fluid intelligence score), and an effect surviving FDR correction for language‐CS's association with verbal episodic memory (Figure 3). Movie‐CS' association with cognition was negative, consistent with associations found in Ryu and colleagues' article. Other effects must however be taken with caution, as most did not survive FDR correction, and all had very low effect sizes (pR ² < 0.0196, discussed further below). Such associations with CS suggest the networks active during movie watching or Story‐math tasks use additional cycle‐forming connections, in addition to the backbone, to perform their reasonings, connections which might be tied to fluid mental functions, which may not be true to resting‐state networks.

With regards to effect sizes, all of the coefficients representing the association between network measures and cognitive measures collapsed across tests were very small. This is expected because this approach of collapsing cognitive scores across different tests compromises the specificity of the brain‐behaviour associations. The partial contribution to the proportion of variance explained by the significant effects ranged from 0.002 to 0.018, which is very small following Cohen's definition (pR ² = 0.0196 to 0.13 as ‘small effect’ (Cohen 1988)). As recommended by the latter, this classification should be considered within its specific context, and it seems common for fMRI FC‐based cognitive associations to return ‘low’ pR ² values when associated with cognitive measures (< 0.005–0.03 (Ryu et al. 2023), 0.01–0.08 (Hausman et al. 2020), 0.002–0.13 with one medium effect at 0.177 (Iliopoulos et al. 2026), 0.01–0.06 (Choi et al. 2025)). Cognitive measure‐specific findings (Figure 4) are more in line with other reports, with pR ² values up to 0.05, but the majority of effects were below 0.01. It should also be noted that models derived from the 7 T fMRI tasks likely have overestimated pR ² due to their smaller sample sizes compared to those of 3 T, as smaller samples bias R² values positively (Akossou and Palm 2013). Global network measures may thus mostly contribute in a minor way to cognitive performance. It is possible that generally, such indicators of how information flow throughout the brain are too broad or approximate to specifically help identify underpinning processes of specific mental tasks, except, as suggested by the present findings, for the FDR‐corrected significant findings derived from working memory fMRI task.

4.1. Limitations

Limitations in the study include the lack of comparable 7 T fMRI data in relation to the 3 T fMRI tasks. It remains to be verified whether scanner resolution impacts the results more than the difference in sample size and group overlap (given most of the participants for the 7 T scans were a selected portion of the same participants that also had a 3 T scan), which meant that the analyses had to be run separately per scanner, thus making it difficult for one to compare between the 3 and 7 T results. The 7 T's movie and resting‐state conditions' lack of significant associations after FDR correction for cognitive test‐specific regressions may simply stem from the lower sample size and statistical power.

Beyond scanner differences, the present findings may be dependent on the specific dataset and the unique tasks it has gone through in the fMRI scanner, which is a limitation to its replicability: findings that apply with the present picture N‐Back may not apply to other fMRI conditions, as is already the case in the present study for example, the social cognition, relational processing and move watching conditions, which do not demonstrate the differences found for the working memory condition. As such, the results must be interpreted with caution and cannot be presumed to be generalised to fMRI studies and task‐versus‐rest in general.

Regarding the topological measures assessed here, they were also limited in scope, as they reflected broad indicators of network efficiency, but were not informative regarding specific nodes, edges, or sub‐clusters. Conceptually, this lack of specificity is likely linked to the relatively weak brain‐behaviour associations observed. Furthermore, the MST/Backbone's strict selection of edges (N(nodes)‐1) makes it a higher (although less arbitrary) edge threshold than common graph theory approaches. Because it relies on the assumption that these edges form the core, backbone connections of the network, it omits other potentially true or relevant edges and is subject to a similar presupposition to the arbitrary thresholds. Because it follows the number of nodes, the MST building can also be affected by varying network sizes (Blomsma et al. 2022), as we also observe in Supporting Informations S2.

A limitation to the interpretability of our findings also comes with using scan‐wide task FC (Cohen and D'Esposito 2016; Greene et al. 2018; Liu and Yang 2023) which does not distinguish activations related to task‐stimuli, from “background” or attention‐related activations (Cohen and D'Esposito 2016). Task FC can be residualised and task‐specific components isolated (Zhao et al. 2023), and it remains unknown how topology metrics differ across subcomponents. The chi‐square analysis was also limited by the arbitrariness of its cell classification (i.e., the conventional significance at p < 0.05). More objective markers, such as Cohen's small effect size threshold for pR ², did not work as an alternative threshold for classification as many cells were empty (network categories not ever reaching this effect size), making them unsuitable for a chi‐squared analysis.

5. Conclusion

Topology network measures that look at the information flow of brain networks can characterise the integration and efficiency of brain activity during fMRI scans and are related to cognitive performance. While conventional graph theoretical measures are affected by threshold selections of the connectomes, which lack standardisation and are relatively arbitrary, equivalent approaches such as the MST and recently tied to PH help avoid this issue by using systematic algorithms which strictly keep the strongest spanning edges of every network. This study shows that in addition to this, MST/Backbone‐associated metrics may be more reliably associated with cognitive performance than their graph theoretical counterparts. However, evidence for this is stronger in connectomes from task‐based fMRI, especially the working memory N‐Back task, than from resting‐state fMRI.

Author Contributions

Charly Hugo Alexandre Billaud and Junhong Yu: data curation, methodology, software, validation, writing – review and editing. Junhong Yu: funding acquisition, project administration, supervision, resources. Charly Hugo Alexandre Billaud: conceptualisation, formal analysis, investigation, writing – original draft, visualisation.

Funding

The preprocessing of the data used for this work was supported by the Nanyang Assistant Professorship (Award no. 021080‐00001) grant.

The HCP data was made available by the Human Connectome Project, WU‐Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research and by the McDonnell Center for Systems Neuroscience at Washington University.

Supporting information

Table S1:1. Selection of neuropsychological assessments completed by participants of the Human Connectome Project Young Adults.

Figure S1:1. Effects of topological network measures on all cognitive scores across fMRI conditions (1/3).

Figure S1:2. Effects of topological network measures on all cognitive scores across fMRI conditions (2/3).

Figure S1:3. Effects of topological network measures on all cognitive scores across fMRI conditions (3/3).

Figure S1:4. Pairwise contrasts in predictive performance across topological measures in each fMRI task.

Figure S2:1. Effect of topological network measures on cognitive performance across all measures and fMRI conditions with the Schaefer‐400 atlas.

Figure S2:2. Effects of topological network measures on all cognitive scores across fMRI conditions with the Schaefer‐400 atlas.

Table S2:1. Moment‐Corrected Standardised Residuals of the Chi‐square test for distribution of significant main effects of topology measure on cognitive functions with the Schaefer‐400 atlas.

Figure S2:3. Pairwise contrasts in predictive performance across topological measures in each fMRI task with the Schaefer‐400 atlas.

Data Availability Statement

The preprocessed derivatives of the HCP‐YA data are accessible in the BALSA (Brain Analysis Library of Spatial maps and Atlases) library (balsa.wustl.edu) under the WU‐Minn HCP Consortium Open Access Data Use Terms. Topology measures, statistics data, and analyses scripts were made accessible on the following GitHub repository: https://github.com/chabld/Topology_study_2025.