Individual differences in childhood cognition significantly impact academic performance, quality of life, and long-term health outcomes, including social, physical, and mental health in adulthood. Cognitive deficits in youth are linked to increased risks for psychopathology, risk-taking behaviors, cardiovascular disease, and mortality. Understanding the emergence of these cognitive differences is crucial for promoting healthy neurocognitive development. While neuroimaging studies show that complex cognitive tasks engage large-scale brain networks, the relationship between individual cognitive differences and the spatial layout of functional networks (functional topography) on the cortex remains unclear. Research has faced two key challenges: accounting for person-specific variation in functional topography, especially pronounced in association cortices, and achieving reproducible results which typically requires large samples. This study addresses these challenges using advanced machine learning to identify individual-specific functional brain networks in large discovery and replication samples. The overarching hypothesis is that the functional topography of association networks is associated with individual differences in cognitive function in children. Traditional fMRI studies often employ a "one-size-fits-all" approach with standardized network atlases, assuming a 1:1 correspondence between structural and functional neuroanatomy across individuals. However, substantial inter-individual heterogeneity in functional topography has been demonstrated, particularly in association cortices crucial for higher-order cognition and implicated in cognitive impairments in adult psychiatric illnesses. Precision functional mapping techniques offer an alternative by deriving individually-defined networks that capture each brain's unique functional topography. These personalized functional networks (PFNs) have shown high stability within individuals and predict individual activation patterns on fMRI tasks. Networks supporting higher-order cognition and exhibiting the greatest variability in functional topography are often located along the sensorimotor-association (S-A) axis, a hierarchical cortical organization gradient. Individual variation in association network topography, including the fronto-parietal network, ventral attention network, and default mode network, is believed to affect cognitive differences. Prior research suggested a link between greater total cortical representation of fronto-parietal PFNs and better cognitive performance, with models trained on functional topography predicting cognition in unseen data. However, the replication of these findings remained a challenge due to limitations in sample size and single-site data collection. This study aims to overcome these limitations by employing two large, matched samples from the ABCD Study, totaling n=6972, to rigorously investigate the relationship between functional topography and individual cognitive differences. The study uses spatially-regularized non-negative matrix factorization (NMF) to identify PFNs, replicating and extending prior work by investigating if PFN topography predicts general or specific cognitive abilities and exploring the alignment of PFN-cognition associations with the S-A axis.

The introduction thoroughly reviews the existing literature on the relationship between individual differences in cognition and brain structure and function. It highlights the limitations of previous studies, such as the use of "one-size-fits-all" approaches that ignore inter-individual variability in functional brain network topography. The authors cite numerous studies demonstrating the importance of individual differences in cognition, the role of large-scale brain networks in cognitive processes, and the challenges of achieving reproducibility in brain-behavior studies. The review emphasizes the need for precision functional mapping techniques and the use of large samples to overcome these challenges. The authors refer to previous studies that have established the existence of a sensorimotor-association (S-A) axis of cortical organization and posit that individual variation in the functional topography of networks located at the association pole of this axis might be particularly important for understanding individual differences in cognition. The literature review sets the stage for the current study by clearly outlining the gaps in existing knowledge and the rationale for the chosen methodology.

This study utilized data from the Adolescent Brain Cognitive Development (ABCD) Study baseline sample, consisting of 6972 children aged 9-10 years. To account for inter-individual heterogeneity in functional brain network spatial layout, precision functional mapping using spatially-regularized non-negative matrix factorization (NMF) was employed. A previously defined group atlas served as a prior for generating personalized functional networks (PFNs) for each individual. The NMF process yielded 17 PFNs per individual, represented as matrices of network weights across vertices (soft parcellation for primary analyses) and non-overlapping networks (hard parcellation for visualization and secondary analyses). The average S-A axis rank was computed for each PFN to determine its position along the cortical hierarchical organization. To assess the predictive power of PFN topography on cognitive performance, ridge regression models were trained. General cognition was defined as the first principal component from a Bayesian probabilistic principal components analysis (BPPCA) of nine cognitive tasks. Models were trained using the cortical representation of each PFN (network loadings at each vertex), controlling for age, sex, site, and head motion. Matched discovery (n=3525) and replication (n=3447) samples were used for out-of-sample testing, with nested cross-validation for parameter tuning in the training samples and out-of-sample testing in the held-out sample. Repeated random cross-validation (100 repetitions) was performed to ensure results were not sample-dependent. Prediction accuracy was evaluated across three major cognitive domains: general cognition, executive function, and learning/memory. Linear ridge regression models were trained independently on each PFN to determine individual network contributions to prediction accuracy. Spearman's rank correlations were used to assess the relationship between prediction accuracy and a PFN's position on the S-A axis, and spatial permutation testing (spin tests) verified the significance of this relationship. To investigate associations between PFN topography and cognitive performance, linear mixed-effects models were used. Total cortical representation of each PFN (sum of network loadings across vertices) was used as a measure of spatial extent. Models included fixed effects for age, sex, head motion, and a random intercept for family ID, controlling for site using ComBat harmonization. Sensitivity analyses were conducted to control for psychotropic medication use and socioeconomic status. Statistical significance was assessed after Bonferroni correction for multiple comparisons.

The study replicated and extended previous findings demonstrating a robust association between individual differences in functional brain network organization and cognitive abilities in children. Key findings include: 1. **Prediction of General Cognition:** Individualized functional topography accurately predicted out-of-sample cognitive performance in both discovery (r = 0.41, p < 0.001) and replication (r = 0.45, p < 0.001) samples. Repeated random cross-validation yielded consistent results (mean r = 0.44, p < 0.001). These models outperformed models trained solely on socioeconomic status. 2. **Prediction of Other Cognitive Domains:** Individualized functional topography also predicted executive function (discovery: r = 0.17, p < 0.001; replication: r = 0.16, p < 0.001) and learning/memory (discovery: r = 0.27, p < 0.001; replication: r = 0.27, p < 0.001) in held-out data, although with lower accuracy than for general cognition. 3. **Network-Specific Contributions:** Fronto-parietal and ventral attention networks showed the highest prediction accuracies, while somatomotor and visual networks showed the lowest, aligning with prior findings. 4. **Association with Cortical Hierarchy:** The strength of associations between functional topography and cognition aligned with the S-A axis, with association networks contributing most to the prediction of cognitive performance across domains and sensorimotor networks yielding the weakest. Spearman's correlations showed significant relationships between prediction accuracy and S-A axis rank across all three cognitive domains (general cognition: r = 0.601, p = 0.012; executive function: r = 0.547, p = 0.025; learning/memory: r = 0.537, p = 0.028). Spin tests validated this association. 5. **Total Cortical Representation of Fronto-Parietal Networks:** All three fronto-parietal PFNs showed significant positive associations with general cognition in both discovery and replication samples, replicating prior findings. One somatomotor network showed an inverse association with cognition.

This large-scale study successfully replicated and extended previous findings, demonstrating reproducible associations between individual differences in functional brain network topography and cognitive abilities in youth. The use of precision functional mapping, large sample sizes, and rigorous cross-validation methods contributed to the reproducibility and robustness of the findings. The consistent spatial pattern observed, aligning with the S-A axis, underscores the importance of association networks in supporting domain-general cognitive abilities. The stronger association of general cognition with PFN topography compared to other cognitive domains suggests that the spatial representation of these networks may underpin broader cognitive abilities. This observation is consistent with the known dominance of general cognitive abilities in accounting for shared variance across various cognitive tasks and their potential role in mediating the relationship between genetic risk and psychopathology. The alignment of the findings with the S-A axis supports the existence of a perception-cognition processing hierarchy in the brain and highlights the potential role of individual variability in association network topography in predicting cognitive differences. The high degree of inter-individual heterogeneity in association cortices, along with their lower structural and functional heritability and extended developmental plasticity, may explain their strong association with cognitive differences. This extended plasticity offers opportunities for interventions to support healthy cognitive development.

This study provides robust evidence for a reproducible link between personalized functional brain network topography and individual differences in cognition during childhood. The findings highlight the importance of association networks in supporting domain-general cognitive abilities and their alignment with the S-A axis of cortical organization. The results pave the way for future longitudinal studies to investigate the dynamic interplay between functional brain organization and cognitive development, potentially informing interventions aimed at promoting healthy neurocognitive development. Future research should explore the role of subcortical regions, other cognitive domains, and environmental factors in shaping PFN topography.

Several limitations warrant consideration. The cross-sectional design prevents the prediction of future cognitive changes within individuals. Head motion, an ongoing challenge in neuroimaging studies of children, was addressed through preprocessing and covariate inclusion but may still have influenced results. Data were combined across multiple fMRI runs to maximize data quality, which may not perfectly represent purely intrinsic network activity. Differences between the ABCD dataset and the dataset used in the original study might have influenced comparisons. The use of a common reference space for registration might introduce biases due to varying degrees of spatial warping across individuals. Lastly, the analyses focused only on cortical regions, excluding subcortical structures.