BrainEnrich: Revealing Biological Insights for Imaging‐Derived Phenotypes Through Transcriptomic Enrichment

ABSTRACT

While the field of imaging transcriptomics is evolving rapidly, several methodological challenges persist in functional enrichment analysis. Here, we introduce BrainEnrich, an R package that integrates whole‐brain gene expression profiles from the Allen Human Brain Atlas (AHBA) with in vivo imaging‐derived phenotypes (IDPs). By offering a suite of flexible association methods, aggregation strategies, comprehensive lists of predefined gene sets, and both competitive and self‐contained null models, the package enables researchers to examine the spatial coupling between molecular profiles and IDPs at both group and individual levels. A novel feature of BrainEnrich is its individual‐level enrichment analysis, which mapped individual IDPs onto a molecular coordinate framework, capturing the molecular signature of individual IDPs and enabling a deeper exploration of inter‐individual variability. Its statistical power was examined through extensive simulation studies based on linear regression with different combinations of test statistics and null models. The results suggest that with appropriate null models, this approach effectively controlled Type 1 error while retaining sensitivity to detect associations between molecular profiles and phenotypic data. Two case studies were performed to demonstrate the utility of the package. In the group‐level enrichment analysis, the effect size map of case–control comparison in cortical thickness of major depressive disorder patients was associated with molecular pathways such as synaptic signaling, lipid regulation, and steroid hormone balance, providing candidate molecular annotations for group‐level IDPs. A separate case study found that synaptic gene set scores showed nominal associations with multiple cognitive measures, demonstrating its utility for individual‐level molecular annotations to explore associations with phenotypic variables. Collectively, BrainEnrich provides a flexible framework for integrating macro‐level IDPs with micro‐level transcriptomic profiles for molecular contextualization of IDPs.

BrainEnrich provides a flexible framework for integrating imaging‐derived phenotypes (IDPs) with transcriptomic data from the human brain. By supporting both group‐level and individual‐level enrichment analyses, it helps uncover the molecular signatures and biological pathways associated with neuroimaging profiles.

1. Introduction

Enrichment analysis is a widely used method in high‐throughput genomics studies to annotate differentially expressed genes to specific biological pathways and processes (Zhao and Rhee 2023). The whole‐brain gene expression maps from the Allen Human Brain Atlas (AHBA) have enabled the enrichment analysis of spatial associations between ex vivo transcriptional patterns and in vivo imaging‐derived phenotypes (IDPs) (Hawrylycz et al. 2012). This analysis examines whether the spatial patterns of gene expression from predefined sets of functionally related genes are significantly associated with interregional variations in IDPs. Typically, predefined gene sets are sourced from curated databases like Gene Ontology (Gene Ontology Consortium 2004) and Kyoto Encyclopedia of Genes and Genomes (Kanehisa and Goto 2000), which categorize genes based on their functions and roles in specific pathways. Alternatively, gene sets may be derived from empirical studies that identify biological profiles, such as cell type markers or findings from GWAS studies (Lake et al. 2018; Seidlitz et al. 2020). The transcriptional analysis of IDPs has given rise to the emergence of imaging transcriptomics (Arnatkeviciute et al. 2022), an interdisciplinary field that yielded numerous novel insights into the molecular architectures underlying IDPs related to health and disease (Arnatkeviciute et al. 2022; Cao et al. 2025; Norbom et al. 2021; Paus 2023).

While the field of imaging transcriptomics is evolving quickly, several methodological challenges remain in functional enrichment analysis. Our previous study showed that the choice of test statistics (i.e., how to aggregate the spatial correlation within predefined gene sets) and the type of null models (i.e., spinning brain regions versus resampling genes) could influence the enrichment results (Cao et al. 2024). Current tools often lack flexibility to accommodate these critical parameters, and rigid implementations of inappropriate statistical tests may lead to false‐positive findings (Cao et al. 2024). In addition, the predefined gene sets used in the previous studies varied in version and preparation, which hampered the reproducibility of the imaging transcriptomics findings. Although standardized resources like org.Hs.eg.db are available (Carlson 2015), there is a notable lack of a toolbox that effectively integrates these resources to provide consistent predefined gene sets for imaging transcriptomics. Moreover, identifying core genes that contribute to the enriched terms helps with interpretation and visualization of the results (Subramanian et al. 2005; Wu et al. 2021). However, it has been largely overlooked in previous functional enrichment analysis in the field. Finally, most imaging transcriptomics studies have predominantly focused on enrichment analysis of group‐level IDPs such as effect sizes from case–control comparisons on cortical thickness, or standardized loading of principal component analysis of cortical thickness (Cao et al. 2023). However, this approach was not able to capture individual differences, thereby constraining the potential for integrating transcriptomics with neuroimaging data.

To bridge these gaps, we introduced the R package BrainEnrich which offers flexible options for test statistics and null models, includes comprehensive predefined gene sets from curated resources, provides utilities for identifying core genes within enriched terms, and supports the enrichment analysis at the individual level. Although functional enrichment analysis was typically used to contextualize group‐level IDPs in previous studies, we extended its use to individual‐level IDPs and assessed the statistical performance of linear regression analysis of phenotypic differences using individual gene set (GS) scores derived from this novel approach. Additionally, we demonstrated the application of both group‐level and individual‐level enrichment analyses through realistic case studies.

2. Methods

2.1. BrainEnrich Analytical Framework

2.1.1. Group‐Level Enrichment Analysis

As illustrated in Figure 1A, the group‐level enrichment analysis began by correlating a group‐level IDP (e.g., an effect size map from a case–control comparison of cortical thickness) with the transcriptional profiles of each gene in the AHBA. This process yielded a spatial coupling vector, indicating the spatial association of each gene with the IDP. The gene‐specific associations were then summarized within predefined gene sets (Figure 1C) corresponding to specific biological profiles as a GS score. The GS score represented the collective associations of genes within a shared biological context and could be calculated using various test statistics (as described below). To determine the significance of the associations between the predefined gene set and the group‐level IDP, the empirical GS score was compared against that from either the self‐contained or competitive null models (Figure 1D).

FIGURE 1.

Overview of the BrainEnrich workflow for imaging transcriptomics. (A) A group‐level IDP (e.g., effect size map) is correlated with AHBA transcriptional profiles, yielding a spatial coupling vector that is aggregated within the predefined gene sets (C). The resultant GS score vector is tested against the null GS generated by the null models (D). (B) A similar GS scoring approach is applied to individual IDPs, generating a spatial coupling matrix that is aggregated within the predefined gene sets (C). The resultant participant‐specific GS scores are used for downstream analyses, and the results are tested against the null GS scores generated by the null models (D). (C) Predefined gene sets from various sources and user‐defined sets aggregate gene‐specific associations into GS scores for either group‐level or individual‐level analyses. (D) Self‐contained (spinning brain regions) or competitive (resampling genes) approaches are used to generate the null GS scores. (E) Simulation utilities evaluate Type 1 error and power across different test statistics, null models, and sample sizes for the individual‐level enrichment analysis. (F) A leave‐one‐out (LOO) procedure identifies genes that substantially influence GS scores (in group‐level analysis) or the resultant test statistic (in individual‐level analysis), highlighting the primary contributors to enriched terms.

2.1.2. Individual‐Level Enrichment Analysis

Although enrichment analysis was typically used to contextualize group‐level IDPs, we extended its use to individual‐level IDPs. As illustrated in Figure 1B, a similar GS scoring approach was applied to individual IDPs (e.g., cortical thickness of a group of participants). This process resulted in a spatial coupling matrix that represented the spatial correlations between gene expression and each participant's IDPs, with rows corresponding to genes and columns to participants. These gene‐specific associations were then summarized within predefined gene sets (Figure 1C), yielding a GS score matrix that captured the collective associations of genes within a shared biological context for each participant. Then, these GS scores could be used in downstream analyses, such as exploring correlations with other phenotypes, predicting phenotypic outcomes, or identifying participant subtypes. Conceptually, the participant‐specific GS score quantifies the degree to which an individual's IDP spatial pattern aligns with the AHBA‐derived spatial expression pattern of genes within a predefined gene set. Therefore, individual‐level BrainEnrich scores can be interpreted as molecular annotations of individual imaging phenotypes. They allow downstream inferences about whether participants whose IDPs show stronger alignment with a given molecular system also differ in cognition, behavior, clinical symptoms, or other phenotypic outcomes. Importantly, these scores do not represent participant‐specific gene expression or pathway activity; instead, they summarize the spatial correspondence between each participant's imaging phenotype and a reference transcriptional architecture. The resulting metrics (e.g., summary statistics, model performance) from the downstream analysis could be compared against those from the null models (Figure 1D) to assess the significance of the results. Because individual‐level GS scores are subsequently used in participant‐level statistical models, BrainEnrich also includes normality diagnostics for these scores. These diagnostics are intended to flag aggregation methods or gene sets that may yield strongly non‐normal individual GS score distributions before applying parametric downstream models. For the downstream analysis, we focused on linear regression and made it available in the package. This was due to its relative simplicity compared to other downstream analyses that may involve additional parameters and settings, as well as its popularity in literature. The linear regression workflow also reports normality diagnostics for the GS scores of each term, providing users with additional information for assessing whether parametric modeling assumptions are appropriate. When strong non‐normality is detected, users should interpret the linear regression results with caution and consider post hoc sensitivity analyses.

2.1.3. Transcriptional Profiles

The AHBA (http://www.brain‐map.org) provided maps of human postmortem brain gene expression through a whole‐brain microarray survey across six donors (Hawrylycz et al. 2012). Guidelines for preparing AHBA gene expression data for imaging transcriptomics analysis have been established (Arnatkevic̆iūtė et al. 2019), and a standardized procedure is available through the open‐access toolbox “abagen” (Markello et al. 2021). Therefore, BrainEnrich did not include utilities for preparing the transcriptional data. To facilitate analysis with parcellated data with the Desikan‐Killiany (Desikan et al. 2006) and Schaefer atlas (Schaefer et al. 2018), we included prepared transcriptional profiles for these atlases sourced from the ENIGMA‐TOOLBOX (Larivière et al. 2021). These profiles were preprocessed following established guidelines and parcellated with the corresponding atlas using “abagen” (Markello et al. 2021). BrainEnrich provides AHBA expression matrices prepared under multiple gene differential‐stability thresholds (r _stability > 0.2, r _stability > 0.4, and r _stability  > 0.6), which retained 12,669, 6514, and 1646 genes, respectively. Differential stability quantifies the reproducibility of a gene's regional expression pattern across the six AHBA postmortem donors, with higher values indicating more consistent spatial expression profiles across donors (Hawrylycz et al. 2015). Accordingly, stricter thresholds retain fewer genes but prioritize more reproducible spatial expression patterns.

2.1.4. Predefined Gene Sets

To reduce dependencies and avoid complicating the installation process (especially since some of these packages are large and span across species, which is unnecessary for the purpose of our package), we did not include utilities that fetch predefined gene sets directly using existing R packages. Instead, we prepared several predefined gene sets using these packages or from the source:

1. Gene Ontology (GO). GO is a comprehensive knowledge base that describes the functions of genes in terms of biological processes, molecular functions, and cellular components (Gene Ontology Consortium 2004). The GO gene sets were obtained using the “get_GO_data” function from the R package clusterProfiler (version 4.12.0) (Wu et al. 2021), which is based on data from “org.Hs.eg.db” (version 3.19.1).

2. DisGeNet. DisGeNet is a comprehensive platform that integrates information on gene‐disease associations from various sources, providing valuable insights into the genetic basis of human diseases (Yu et al. 2014). The DisGeNet gene sets were obtained using the “get_DGN_data” function from the R package “DOSE” (version 3.30.1).

3. Kyoto Encyclopedia of Genes and Genomes (KEGG). KEGG is a widely used biological database initiated in 1995 to understand biological systems through genome data (Kanehisa and Goto 2000). It features manually curated pathway maps and the KEGG Orthology system, aiding in the analysis of gene functions and molecular networks. KEGG has expanded to include applications in health sciences, integrating drug and disease data, and is now a key resource in biological research (Kanehisa 2019; Kanehisa et al. 2021, 2023, 2019). The KEGG pathway and modular gene sets for Homo sapiens were prepared using the “prepare_KEGG” function from the R package “clusterProfiler” (version 4.12.0), which retrieved data through the REST API (https://rest.kegg.jp).

4. WikiPathways. WikiPathways (https://www.wikipathways.org/) is a community‐driven platform that curates pathways across various biological systems, covering a wide range of pathways related to metabolism, signaling, gene regulation, and disease mechanisms (Agrawal et al. 2024; Martens et al. 2021). The WikiPathways gene sets for Homo sapiens were prepared using the “prepare_WP_data” function from the R package “clusterProfiler” (version 4.12.0).

5. Reactome. Reactome (http://www.reactome.org) is an open‐source, curated bioinformatics database focused on human pathways and reactions, developed through a collaboration among several institutions (Croft et al. 2010). The Reactome gene sets for Homo sapiens were prepared using the “get_Reactome_DATA” function from the R package “ReactomePA” (version 1.48.0).

6. Medical Subject Headings (MeSH). MeSH is the controlled vocabulary developed by the U.S. National Library of Medicine (NLM) for indexing articles in MEDLINE/PubMed (Lipscomb 2000). It comprises over 25,000 clinical and biological terms, with a size nearly twice that of GO. MeSH has been shown to be a valuable resource for interpreting biomedical data (Yu 2018). We prepared the MeSH gene sets for Homo sapiens using R package “AnnotationHub” (version 3.12.0). The correspondence (gene2pubmed) between MeSH IDs and Entrez Gene IDs was established and curated by NCBI's expert teams. Additionally, the relationships between MeSH terms and PubMed IDs were derived from PubMed data and subsequently merged to establish the final correspondence between Gene IDs and MeSH IDs.

7. SynGO. SynGO (https://syngoportal.org) is a public platform for synapse research, with the original release providing detailed annotations for 1112 genes related to synaptic function and location (Koopmans et al. 2019). SynGO offers a valuable resource for interpreting large‐scale biological data, with all information based on expert‐curated, published evidence. The SynGO gene sets were prepared from source data files available at https://www.syngoportal.org/#downloads, with the release version‐2023‐12‐01.

8. Cell Type. Integrating spatially cell‐specific transcriptomic data with IDPs has provided insights into the cellular correlates underlying brain development and psychiatric disorders (Martins et al. 2021; Seidlitz et al. 2020). We included gene sets for seven cell types: (1) astrocytes, (2) endothelial cells, (3) microglia, (4) oligodendrocyte progenitors, (5) excitatory neurons, (6) inhibitory neurons, and (7) oligodendrocytes, which were curated in a previous study (Seidlitz et al. 2020). Additionally, other gene sets of cell types curated or identified from separate studies were also included (Lake et al. 2018; Martins et al. 2021).

In sum, BrainEnrich provides 26 predefined annotation collections, comprising 76,780 predefined gene sets before analysis‐specific filtering. The number of gene sets varied across collections, ranging from 7 (cell‐type gene sets) to 30,160 (DisGeNet gene sets). A summary of the number and size distribution of predefined gene sets in each collection is provided in Table S7. Because downstream analyses depend on the available background genes and user‐defined gene set size thresholds, the final number of tested gene sets may differ from these raw collection‐level counts. For genes coded using Entrez IDs, conversion to gene symbols was performed using the “mapID” function from the R package AnnotationDbi (version 1.66.0), based on data from “org.Hs.eg.db” (version 3.19.1). These external predefined gene sets are hosted on a GitHub repository and will be automatically downloaded when the user first accesses the relevant data. Our package was designed to be universal and flexible with user‐defined predefined gene sets. A tutorial on how to prepare predefined gene sets from various sources, including CSV, XLSX and GMT files, is available with the package's documentation.

2.1.5. Association Methods

The package provided multiple methods for assessing associations between brain imaging phenotypes and transcriptional profiles, including both univariate and multivariate approaches. Pearson's correlation was used by default to examine linear associations between each gene's expression and the brain map, with an optional Fisher's z‐transformation for standardization. Additionally, Spearman's correlation was available as a rank‐based measure of association that is robust to outliers. To account for multivariate structure within the datasets, the package included partial least squares (PLS) regression, which identified latent components that captured maximum covariance between transcriptional profiles (predictors) and the IDP (response). The first component from PLS, which explained the greatest variance in gene expression data and maximizes correlation with IDPs, was often of particular interest in imaging transcriptomics research (Hansen et al. 2022; Li et al. 2021; Morgan et al. 2019; Romero‐Garcia et al. 2020; Xue et al. 2023). Two PLS1‐based measures were available for the subsequent aggregation: regression coefficients of the PLS1 (PLS1c), which represented the weights or importance for each gene in predicting the imaging phenotype, and loading weights of the PLS1 (PLS1w), which highlighted each gene's contribution to the primary latent factor. Notably, the PLS approach was designed for analyzing multiple genes against a single IDP, rather than multiple genes against multiple IDPs for individual‐level analyses. In BrainEnrich, PLS is implemented as an optional association method for deriving gene‐level statistics, including PLS1 regression coefficients and loading weights. These PLS‐derived gene‐level scores can then be aggregated within predefined gene sets and evaluated using the same competitive or self‐contained null models as other association methods. Thus, the implemented null models support statistical testing of gene set enrichment based on PLS‐derived scores. However, these enrichment‐level null models are distinct from PLS‐specific procedures used to evaluate the significance or stability of latent variables themselves. If users aim to interpret PLS components directly, additional PLS‐specific permutation or bootstrapping procedures may be warranted. To accommodate diverse analytical needs, the package also supported user‐defined association functions including non‐linear or advanced association techniques beyond those provided.

2.1.6. Aggregation Methods

In enrichment analysis, test statistics aggregated all associations within a predefined gene set as gene set (GS) scores to represent the collective association with IDPs. In BrainEnrich, the GS score refers to the selected aggregation statistic of gene‐level spatial coupling values within a predefined gene set. Depending on the aggregation method, it can be calculated as the mean, median, mean absolute value, maxmean, or a KS‐based statistic. This differs from classical GSEA terminology, where the enrichment score (ES) refers to a specific rank‐based statistic and the normalized enrichment score (NES) refers to its null‐normalized form (Subramanian et al. 2005). In BrainEnrich, significance is assessed by comparing the observed GS score, or the downstream statistic derived from individual GS scores, with scores generated from competitive or self‐contained null models. Different aggregation methods (i.e., test statistics) provided various perspectives on these associations in neuroimaging transcriptomics. Our package offered multiple aggregation methods to capture the overall association strength, including the mean (Mean), median (Median), and mean absolute (Meanabs) values. The mean (Mean) statistic summarized the average association within the gene set, while the median (Median) provided a measure of central tendency that was less influenced by extreme values. The mean absolute value (Meanabs) captured the general strength of association by aggregating the absolute correlation values, thus reflecting the overall magnitude of correlations within the gene set. Additionally, the max mean (Maxmean) statistic calculated the average of the positive values and the average of the negative values within the gene set, selecting the one with the larger absolute value to highlight dominant patterns in the data (Efron and Tibshirani 2007). The package also included two Kolmogorov–Smirnov (KS)–based statistics, the traditional KS test and its variant (Subramanian et al. 2005). The traditional weighted KS test, commonly used in gene set analysis, evaluated the enrichment of specific gene sets within ranked lists by calculating a cumulative sum across ranked correlations. The cumulative sum increased for genes within the set and decreased for those outside it, weighted by correlation strength (Subramanian et al. 2005).

When genes within a gene set are represented at both extremes, traditional KS‐based statistics could yield a bimodal distribution, which was not ideal for downstream analysis in individual‐level enrichment. To address this issue, we adopted a modified KS statistic (KS sum) that produced individual GS scores with an approximately normal distribution by calculating the difference between the largest positive and negative deviations from zero. This measure was originally proposed for gene set variation analysis (GSVA) that examined pathway activity across different samples (Hänzelmann et al. 2013). Moreover, Maxmean and Meanabs were explicitly excluded for individual‐level analysis, as these aggregation methods were more likely to yield non‐normally distributed individual‐level GS scores (see Figure S1). It is worth noting that our package allowed user‐defined aggregation methods, providing flexibility in how associations were summarized for both group‐level and individual‐level enrichment analysis.

2.1.7. Null Models

The concepts of “competitive” and “self‐contained” null models were originally described in gene set analysis (Goeman and Bühlmann 2007) and were later introduced to the imaging transcriptomics field in our previous work (Cao et al. 2024). Both types of null models have been used in previous group‐level enrichment analyses; each has its advantages and disadvantages (Cao et al. 2024; Khatri et al. 2012; Maleki et al. 2020). Self‐contained null models evaluated each gene set in isolation, enabling researchers to test whether a specific gene set was significantly associated with IDPs independently from other genes. This model was built by spinning brain regions while preserving their spatial autocorrelation (Alexander‐Bloch et al. 2018; Markello and Misic 2021; Váša et al. 2018) and was particularly useful when the interest lies in understanding the spatial correspondence of a gene set with an IDP, without comparison to other genes. In contrast, competitive null models were built by resampling genes to create null gene sets of the same size and were advantageous when the goal was to reveal the prominent gene sets. Our package provided both competitive and self‐contained null models, allowing users to choose the approach that best aligns with their research goals (see Figure 1D).

The package also provided co‐expression‐matched competitive null models, generated by resampling genes that exhibited similar co‐expression patterns to those in the empirical model. This approach was proposed by Wei et al. (2022) to address the co‐expression bias in the competitive null models and was initially used in aggregation‐independent analysis, where gene expression was averaged within a gene set before correlating with IDPs (Wei et al. 2022). In our previous work, we showed that this approach effectively controlled the impact of co‐expression bias in aggregation‐dependent scenarios (Cao et al. 2024). Therefore, it was included as an option in our package; however, its increased computational demand should be taken into consideration.

2.1.8. Identification of Core Genes

Identification of core genes has proven to be informative for the interpretation and subsequent analysis and visualization (Subramanian et al. 2005; Wu et al. 2021). However, previous imaging transcriptomic studies using functional enrichment analysis overlooked core genes that contribute to the enrichment. Here, we defined core genes as genes within an enriched term that have a relatively large influence on the target statistic of interest. In group‐level enrichment analysis, this target statistic is the resulting GS score. A leave‐one‐out (LOO) procedure was introduced to quantify the influence of each gene on the final GS score and identify core genes based on a specific threshold (see Figure 1F). Specifically, for each iteration of LOO, one gene was excluded, and the GS scoring was performed without it, then the absolute change in the GS score compared to the original GS score was calculated. This process was repeated for all genes within the gene set, resulting in an absolute change value for each gene. Genes with large absolute changes were considered to have a significant impact on the final GS score. Those genes with absolute changes above the threshold were defined as core genes. Quantiles or standard deviations threshold could be used to define core genes. By default, the threshold was set at one standard deviation (i.e., mean + 1 SD). For individual‐level enrichment analysis, the same LOO logic was applied to the downstream statistic derived from individual GS scores. In the linear regression workflow, standardized beta coefficients were used as the target statistic because they quantify the estimated phenotype–GS score association. Thus, for each gene, the procedure recalculated the individual GS scores after removing that gene, reran the downstream linear model, and quantified the absolute change in the standardized beta coefficient. In contrast, t‐statistics were used for null‐model‐based significance testing because they account for both effect magnitude and estimation uncertainty.

2.2. Simulation Studies

The performance of statistical tests using various test statistics and null models for group‐level enrichment analysis has been extensively examined in our previous work (Cao et al. 2024). Here, we evaluated the statistical power of the individual‐level enrichment analysis using linear regression. We focused on linear regression as the downstream analysis due to its relative simplicity as well as its popularity in the literature. The simulation framework (see Figure 1E) described below was available in our package, enabling users to easily replicate or extend these simulation analyses.

2.2.1. Simulation Data and Setup

We selected a subset of unrelated healthy adults (n = 420, 226 females, mean age ± SD = 28.60 ± 3.73 years, range = 22 to 36 years) from the Human Connectome Project (HCP; https://www.humanconnectome.org/study/hcp‐young‐adult/data‐releases) dataset (Van Essen et al. 2012). To ensure that no siblings or twins were included in the subset, participants were chosen by randomly selecting one individual from each family ID. High‐resolution T1‐weighted images were acquired on a Siemens Skyra 3T scanner using a 3D magnetization‐prepared rapid gradient echo (MPRAGE) sequence with a repetition time (TR) of 2400 ms, echo time (TE) of 2.14 ms, field of view (FOV) of 224 × 224 mm², voxel size of 0.7 mm³, and 256 slices. The cortical thickness (CT) of 34 Desikan‐Killiany cortical regions from both hemispheres was extracted from structural T1‐weighted MRI brain scans using FreeSurfer (version 5.3) (Fischl 2012). CBRAIN, a web‐based distributed computing platform, was utilized for preprocessing (Sherif et al. 2014). Following processing, all images were visually inspected according to ENIGMA imaging quality control protocols (http://enigma.ini.usc.edu/protocols/imaging‐protocols/). Participants with any region that failed visual quality control were excluded.

Body mass index (BMI), defined as weight in kilograms divided by the square of height in meters, was used as the phenotypic variable due to its known associations with various health outcomes and its potential effects on brain morphology (Laurent et al. 2020; Medic et al. 2016; Steegers et al. 2021). Additionally, Flanker task performance was chosen as a cognitive phenotype for power analysis due to the hypothetical relevance between inhibition and synaptic processes and its potential for associations with SynGO terms (Bari and Robbins 2013). The Flanker task measured an individual's executive function, specifically focusing on inhibitory control and attention. Participants were required to focus on a target stimulus while ignoring distractor stimuli (flankers). In some trials, the flankers were congruent with the target (pointing in the same direction), while in others, they were incongruent (pointing in the opposite direction). Performance was scored based on a combination of accuracy and reaction time, with higher scores indicating better inhibitory control and attentional focus. Associations of BMI and Flanker task performance with molecular profiles were explored across the three null settings using a linear model that included age, sex, and estimated intracranial volume (ICV) as covariates to control for potential confounding effects. The same two phenotypic variables were used in both the Type 1 error and power simulations (as described below) to keep the empirical imaging and phenotypic context comparable across simulation settings. In the Type 1 error simulations, null scenarios were constructed by disrupting the relevant correspondence among phenotypic variables, individual IDPs, and gene sets, whereas in the power simulations the original empirical correspondence was retained to estimate detection probability under realistic data conditions.

For the gene expression data, only genes consistently expressed across donors (r _stability > 0.4) were included, resulting in a total of 6514 genes across the 66 cortical regions. This relatively stringent differential‐stability threshold was selected to prioritize genes with reproducible spatial expression patterns across AHBA donors and to reduce the computational burden of the large‐scale simulation analyses. Due to limited samples available in the right hemisphere, there were no tissue samples matched to the right frontal pole and right temporal pole in the Desikan‐Killiany atlas; therefore, only the left hemisphere's transcriptional profiles were analyzed. Gene sets from the SynGO database, ranging in size from 20 to 200 genes (55 terms), were used as prior gene sets. For the association methods, we used Pearson's correlation, and the Fisher‐transformed correlations were aggregated using one of the following test statistics: Mean, Median, or KS sum. Their performance was systematically evaluated using both self‐contained and competitive null models. Simulations were performed, in part, on the Vermont Advanced Computing Core.

2.2.2. Type 1 Error Simulations

We first evaluated the Type 1 error rates by simulating null scenarios and testing various null models and test statistics. As shown in Figure 1E, we analyzed false discovery rates (i.e., Type 1 errors) under three distinct null settings: null phenotypes, null IDPs and null gene sets, each with no expected correlation between the individual GS scores and phenotypic variables. These null scenarios were constructed using the empirical HCP data by selectively disrupting one component of the tested association while retaining the remaining data structure. For the null phenotypes setting, phenotypic values were randomized across participants, while individual IDPs and GS scores were kept unchanged, resulting in no expected correlation with individual GS scores. For the null IDPs (i.e., spun brain region) setting, individual regional CT values were randomized while preserving spatial autocorrelation. Specifically, regional CT values were spatially permuted using spin‐based rotations of cortical parcels, thereby disrupting the spatial correspondence between each individual's IDP map and AHBA gene expression while preserving the spatial structure of the cortical map. This procedure eliminated the spatial correspondence between individual IDPs and gene expression, yielding null GS scores with no expected correlation with phenotypic variables. For the null gene sets (i.e., resampled gene) setting, genes within predefined sets were randomized, with each empirical SynGO term replaced by randomly sampled gene sets of the same size from the background gene list. This procedure removed the biological relevance or specificity of the original gene sets. This also resulted in null GS scores with no expected correlation with phenotypic variables.

For each null setting, we employed a subsampling procedure to conduct simulation studies using real data. We considered scenarios involving subsamples of different sizes (n = 100, 200, or 300) to quantify the likelihood (psig) that the methods would identify significant correlations in these null settings across 1000 iterations. For instance, if a particular pathway was found to be significantly associated with the phenotypic variable in 300 out of 1000 iterations, the psig (probability of significance) would be calculated as 300/1000 = 0.3. This calculation was performed separately for each subsample size and test statistic. Both parametric and non‐parametric approaches were employed for significance testing. In the parametric approach, p values corresponding to the phenotypic data in the linear model were used. In the non‐parametric approach, the corresponding t‐statistic was compared against null models specific to the null setting (i.e., the self‐contained null models were used for the null IDPs setting; the competitive null models were used for the null gene sets setting). These simulation analyses were designed to determine when parametric tests showed inflated false‐positive rates and whether corresponding non‐parametric null models could mitigate this inflation. Inflation under the null gene sets setting indicated the need for competitive null models to control gene‐set composition effects, whereas inflation under the null IDPs setting indicated the need for spin‐based self‐contained null models to control spatial autocorrelation in IDPs.

2.2.3. Power Simulations

Due to the lack of definitive ground truth for power simulations, we employed a subsampling approach with realistic data to estimate the statistical power. Specifically, we repeated the linear regression analysis with subsample sizes of 100, 200, and 300 participants for 1000 times respectively (see Figure 1E). We calculated the likelihood of identifying significant correlations between a given gene set and phenotypic variables at an alpha level of 0.05. For simplicity, we refer to this as “power” to reflect the method's sensitivity in detecting true associations across different subsample sizes, accounting for the realistic variability of the data. However, in scenarios where the null hypothesis is true and no real association exists, the observed rate of significance should approximate the false positive rate (Type 1 error). For realistic data, some cases would inevitably involve scenarios where the null hypothesis holds. The results of our Type 1 error simulations (see Results section) showed that the individual‐level enrichment analysis with non‐parametric test using null models effectively controlled the false positive rate. This provides confidence that our power estimates were reliable, even when some data would contain null associations. In the realistic data setting, a psig greater than 0.05 meant that the test statistic identified significant associations for a given gene set in more than 5% of the repetitions, suggesting that the method has some sensitivity to detect signals and pick up on patterns that may reflect genuine biological relationships rather than random noise. These simulations were intended to characterize empirical detection probability under the specific molecular‐phenotypic settings examined, rather than to establish a universal ranking of aggregation statistics across all possible phenotypes or gene sets.

2.3. Case Studies

2.3.1. Group‐Level Enrichment Analysis

For the group‐level enrichment analysis, effect size maps were obtained through a meta‐analysis of a case–control comparison of Desikan‐Killiany cortical thickness (CT) from 20 international groups, consisting of a total of 1902 MDD patients and 7658 healthy controls (Schmaal et al. 2017). Briefly, the CT of 34 Desikan‐Killiany cortical regions from both hemispheres was extracted from structural T1‐weighted MRI brain scans using FreeSurfer (version 5.3). Group differences in CT were examined using linear regression models, adjusting for age, sex, and scanner differences. Cohen's d effect sizes were calculated for each cortical region, and final estimates were obtained using an inverse variance‐weighted random‐effect meta‐analysis model.

Genes associated with GO categories of molecular functions, with gene set sizes ranging from 20 to 200, were used as the prior gene sets for the enrichment analysis. We used a self‐contained null model and mean as the test statistic. The false discovery rate correction (Benjamini and Hochberg 1995) was applied across gene sets to account for multiple comparisons. The LOO procedure (as described above) was used to identify core genes that contribute to the enrichment terms, with the default threshold set at the mean absolute change plus one standard deviation (mean + 1 SD).

2.3.2. Individual‐Level Enrichment Analysis

For the individual‐level enrichment analysis, we used CT data from the HCP dataset, as described in the simulation setup. GS scores of SynGO terms with sizes ranging from 20 to 200 were derived for each participant. Then, we conducted linear regression analyses to examine associations between GS scores and a range of cognitive measures from the HCP, while adjusting for participants' age, sex, and intracranial volume (ICV). Individual cognitive measures were derived from the NIH Toolbox, which provided standardized assessments across a variety of cognitive measures, including working memory, processing speed, vocabulary comprehension, reading decoding, and composite scores. For this study, unadjusted scale scores were used, where participant performance was normed to the entire NIH Toolbox Normative Sample (ages 18 and older), irrespective of age or other variables. To determine the significance of the association between SynGO GS scores and these cognition measures, we used the self‐contained null model with 5000 permutations and the KS Sum as the test statistic. For reported terms flagged as non‐normal, we performed post hoc spin‐null sensitivity analyses using heteroskedasticity‐consistent HC3 robust standard errors. Specifically, we refitted the model with the same covariates for both the empirical and spin‐null GS scores, calculated HC3 robust‐SE t‐statistics, and compared the empirical robust‐SE t‐statistic against the corresponding spin‐null distribution. This analysis was used to assess whether the observed associations were primarily driven by ordinary least‐squares standard‐error assumptions.

2.4. Data and Code Availability

The realistic data used for the individual‐level enrichment analysis was obtained from the HCP dataset, available at https://www.humanconnectome.org/. The MDD cortical thickness effect size map used in the group‐level case study was derived from the published ENIGMA‐MDD summary statistics reported previously (Schmaal et al. 2017). The imaging transcriptomics analyses were performed using the R package BrainEnrich 1.0 (https://github.com/zh1peng/BrainEnrich). The enriched GO terms and identified core genes were visualized using the enrichplot R package (https://github.com/YuLab‐SMU/enrichplot). The code used to implement the analyses presented in this study is available at https://github.com/zh1peng/paper_code.

3. Results

3.1. Type 1 Error Simulations

In the Type 1 error simulations, we evaluated the false discovery rates across three null settings, with each setting representing a scenario with no expected correlation between the GS scores and the phenotypic variables. Tables S1 and S2 present the results of Type 1 error simulations for BMI and Flanker performance respectively. For the randomized phenotypic data (Figure 2A), only the parametric approach was applied, as it effectively controlled the Type 1 error rates (psig) across all test statistics and sample sizes, with psig values remaining close to the nominal false positive rate (0.05). For example, median psig values ranged from 0.047 to 0.057 for BMI and from 0.044 to 0.054 for Flanker performance across test statistics and sample sizes. However, for the resampled genes (Figure 2B) and spun brain regions (Figure 2C) settings, the parametric approach alone generally did not adequately control the Type 1 error rates, showing inflated psig values across most test statistics and sample sizes. This inflation was especially evident for BMI at n = 300, with median psig values of 0.421–0.450 under the resampled‐gene setting and 0.379–0.449 under the spun brain regions setting across test statistics. By comparison, Flanker performance showed lower median psig values under the same settings, ranging from 0.043 to 0.061 for the resampled‐gene setting and from 0.064 to 0.068 for the spun brain regions setting. When we applied the corresponding non‐parametric tests (i.e., the competitive null model for the resampled genes setting and the self‐contained null model for the spun brain regions setting), the psig values remained close to the nominal false positive rate of 0.05, suggesting the effectiveness of these models in mitigating false positives under these scenarios. At n = 300, median non‐parametric psig values ranged from 0.047 to 0.057 across phenotypes, test statistics, and null settings. Together, these results indicate that Type 1 error inflation under parametric testing was more pronounced for BMI than for Flanker performance, whereas the corresponding non‐parametric null models mitigated false positives across both phenotypes.

FIGURE 2.

The probability of significance (psig) in Type 1 error simulations for individual‐level enrichment analysis for BMI (left panel) and Flanker performance (right panel). psig values are shown under three null settings: (A) randomized phenotypic variables, (B) resampled genes, and (C) spun brain regions, with each setting representing a scenario with no expected correlation between the gene set scores and the phenotypic variable. For randomized phenotypic variables (A), only the parametric approach was used, with psig values shown across all test statistics. For resampled genes (B) and spun brain regions (C), results are shown for both parametric and non‐parametric tests, where the competitive and self‐contained null models were applied, respectively. The X‐axis represents psig values, and the dashed line indicates the nominal significance threshold (psig = 0.05). Results are color‐coded by sample size (n = 100, 200, or 300). Subplots in each column denote the results for different test statistics.

3.2. Power Simulations

In the power simulations, we evaluated the sensitivity of different test statistics and null models to detect associations between predefined gene sets and phenotypic variables. In the realistic data setting, a psig greater than 0.05 means that the test statistic identified significant associations for a given gene set in more than 5% of the repetitions, suggesting that the method had sensitivity to detect the associations. Tables S3 and S4 present the results of power simulations for BMI and Flanker performance, respectively. For the parametric approach (Figure 3A) without null models, performance varied across test statistics and sample sizes. Notably, there were SynGO terms with high psig values, indicating they were identified as significant in a large proportion of repetitions, with some terms reaching a psig of 1 for certain test statistics. While the parametric approach showed high psig values for some terms, consistent with the Type 1 error simulations, these results could be false positives.

FIGURE 3.

The probability of significance (psig) in power simulations for individual‐level enrichment analysis for BMI (left panel) and Flanker performance (right panel). psig values are shown under three configurations: (A) parametric test without null models, (B) competitive null models generated by resampling genes, and (C) self‐contained null models generated by spinning brain regions. Subplots in each column denote the results for different test statistics (mean, median, KS sum). Each line on the Y‐axis represents a SynGO term, and different dot colors correspond to analyses with different subsample sizes (n = 100, 200, or 300). The X‐axis represents psig values, and the dashed line indicates the psig value of 0.05.

As shown in Figure 3B,C, both competitive and self‐contained null models demonstrated sensitivity in capturing potential associations between SynGO terms and the two phenotypic variables examined, with psig values above 0.05 observed for specific terms across subsample sizes. However, there was a phenotype‐specific pattern when using BMI versus Flanker performance as the phenotypic variable, with Flanker performance generally showing higher empirical detection probability for SynGO associations and BMI yielding lower psig values. For example, at n = 300, Flanker performance showed a higher number of SynGO terms with psig > 0.05 than BMI under both competitive and self‐contained null models. Under the competitive null model, Flanker performance showed 42/55, 22/55, and 19/55 terms with psig > 0.05 for the Mean, KS Sum, and Median statistics, respectively, whereas BMI showed 0/55, 2/55, and 2/55 terms. Under the self‐contained null model, Flanker performance showed 38/55, 25/55, and 15/55 terms with psig > 0.05 for the Mean, KS Sum, and Median statistics, respectively, whereas BMI showed 0/55, 4/55, and 8/55 terms.

3.3. Group‐Level Enrichment Analysis

Figure 4A shows the spatial pattern of changes in CT of participants with MDD. Using mean as the test statistic with the self‐contained null models, the group‐level enrichment analysis of the MDD effect size map identified molecular functions significantly linked to morphological changes in MDD (see Table S5). Figure 4B presents the top 10 enriched terms, with their respective core genes identified using the LOO approach. Genes with a large impact (greater than 1 standard deviation) on the resultant GS score were defined as core genes. Across the 98 FDR‐significant GO molecular function terms, the LOO procedure identified 836 term‐level core‐gene occurrences, corresponding to 414 unique core genes. Among the top 10 enriched terms shown in Figure 4B, 33 unique core genes were identified. Notable core genes, including GABRB2, GLRA3, and SLC17A6, were identified across multiple terms related to ion channel activities. Figure 4C visualizes the relationships between the enriched GO terms and core genes using a gene‐concept network. Figure S2 shows a tree plot of the top 30 enriched terms, clustered based on semantic similarities. These results demonstrated the use of the BrainEnrich package in identifying and visualizing molecular profiles related to group‐level IDPs.

FIGURE 4.

Group‐level enrichment analysis of the major depressive disorder (MDD) effect size map. (A) The effect size map of group comparison in cortical thickness (CT) between participants with MDD (n = 1902) and healthy controls (7658), shown from lateral and medial views of left‐hemisphere. The color bar denotes the magnitude of the effect sizes. (B) Results of the group‐level enrichment analysis of the MDD effect size map using the mean test statistic and self‐contained null models. The dot plot shows the top 10 enriched GO molecular function terms, with dots sized according to the number of core genes in each pathway and colored by the adjusted p values. The heat map shows core genes associated with each term, colored by their spatial correlation with the effect size map. The leave‐one‐out (LOO) procedure involved sequentially excluding one gene at a time, recalculating the gene set (GS) score for each iteration, and measuring the change in the GS score. Genes that contributed a change in the GS score above the threshold (greater than one standard deviation) were identified as core genes. (C) Gene‐concept network of the top 10 enriched GO terms, showing the relationship between the enriched terms and core genes, with edges indicating gene‐term associations. The gene nodes are colored by their spatial correlation with the effect size map, and term nodes are sized according to the number of core genes.

3.4. Individual‐Level Enrichment Analysis

As shown in Figure 5A and Table S6, the individual‐level enrichment analysis identified nominal associations between SynGO terms and cognitive measures, particularly in the domains such as inhibition and processing speed. As shown in Figure 5C, GS scores of SynGO terms such as presynaptic active zone membrane, integral component of presynaptic active zone membrane, presynaptic cytosol, and trans‐synaptic signaling showed nominal associations with individual's Flanker performance, providing candidate molecular annotations for individual differences in inhibitory control. The gene‐concept network shows the relationships between nominally associated SynGO terms and their associated core genes. Figure 5D shows the relationship between individual GS scores for postsynaptic cytoskeleton organization and vocabulary comprehension, presynaptic active zone membrane and Flanker performance, as well as integral components of the presynaptic active zone membrane and reading decoding. For GO:0099523, which was flagged as non‐normal, the post hoc HC3 robust‐SE spin‐null sensitivity analysis yielded a comparable result to the original linear‐model statistic (original spin test: t = 2.313, p _spin = 0.0256; HC3 robust‐SE spin test: t = 2.500, p _spin = 0.0176). Although these results failed to survive the FDR correction across SynGO terms, presynaptic active zone membrane, integral component of the presynaptic active zone membrane, and presynaptic active zone showed a marginally significant correlation with individual Flanker performance (FDR‐adjusted p = 0.08). These results demonstrated the use of the BrainEnrich package in examining and visualizing molecular functions related to individual differences while accounting for covariates like sex, age, and ICV.

FIGURE 5.

Results of individual‐level enrichment analysis. (A) Correlations between individual GS scores for SynGO terms and cognitive measures. Stars indicate nominal associations, and the squares are colored based on the t‐values from the linear model. (B) A schematic illustration of synaptic architecture. (C) The gene‐concept network shows the relationships between nominally associated SynGO terms and their core genes, with term nodes colored by t‐values and sized by the number of core genes. Core genes were identified using the leave‐one‐out (LOO) procedure, with genes that caused substantial changes in resultant standardized beta values defined as core genes. (D) The scatter plots show the distribution of individual GS scores for SynGO terms in relation to specific cognitive measures, with fitted regression lines illustrating the strength and direction of these associations.

4. Discussion

In this study, we introduced the BrainEnrich package that integrates the macro‐level IDPs with micro‐level transcriptional profiles. For group‐level enrichment analysis, BrainEnrich addressed key gaps in functional enrichment analysis in imaging transcriptomics. As highlighted in our previous work, the choice of test statistics and null models can significantly influence enrichment results (Cao et al. 2024). Specifically, competitive null models based on gene resampling may be biased by co‐expression within gene sets, whereas self‐contained null models based on spatial permutation may be sensitive to the distributional structure of gene–phenotype correlations, such as bimodality. In addition, different test statistics interact differently with these null models, leading to variable false‐positive or false‐negative risks. While many existing tools lack flexibility in these aspects, BrainEnrich supported various association and aggregation methods and null models. In addition to widely used pathway databases in imaging transcriptomics such as GO, KEGG, and DisGeNet, BrainEnrich provided other valuable resources like SynGO, Cell Type, WikiPathways, Reactome, and MeSH, allowing for a more comprehensive biological interpretation of IDPs‐gene associations. Moreover, BrainEnrich adopted output format from the widely used package clusterProfiler (Wu et al. 2021), which facilitates visualization with existing tools (e.g., enrichplot). The core gene detection feature was based on a LOO procedure, which enhanced interpretability and visualization of the enrichment results.

In the first case study, the group‐level enrichment analysis of MDD effect size map identified several GO terms whose AHBA‐derived expression profiles spatially covaried with the MDD cortical‐thickness effect size map, showing how BrainEnrich can generate candidate molecular annotations for group‐level IDPs. Sterol transporter activity, steroid binding, and endonuclease activity may at first seem like disparate functions, but they may be linked to overlapping biological processes relevant to neuronal maintenance (Allen et al. 2007; Joëls and Baram 2009). These results suggest that regions showing atrophy in depression are characterized by gene expression profiles that may relate to lipid dysregulation (Parekh et al. 2017), steroid hormone imbalance (McEwen 2005), and propensity for inflammation (Shelton et al. 2011). Core genes such as GABRB2, GLRA3, and SLC17A6 emerged across multiple pathways related to synaptic signaling. Prior studies have suggested that both GABAergic and glutamatergic signaling play key roles in depression (Duman and Aghajanian 2012; Duman et al. 2016, 2019), providing broader biological context for MDD‐related cortical changes. This case study demonstrated the application of BrainEnrich for integrating microscale transcriptional profiles to contextualize imaging findings and provide biologically informed interpretations. However, the analysis was based on spatial correspondence between IDPs and AHBA transcriptional profiles from independent donors. The enriched terms should be interpreted as candidate molecular annotations of IDPs (i.e., MDD‐related cortical thinning), rather than evidence of disease‐related molecular dysregulation or causal mechanisms.

A novel feature of BrainEnrich is its individual‐level enrichment analysis that calculated GS scores on a per‐individual basis and related the molecular profiles with individual differences. This was inspired by gene set scoring methods (e.g., GSVA, ssGSEA) developed for single‐cell RNA‐seq data, which enabled pathway‐level inference at the individual sample (Hänzelmann et al. 2013; Rahmatallah et al. 2016). Notably, ssGSEA and GSVA are applied directly to molecular expression data, such as bulk or single‐cell RNA‐seq profiles, but individual‐level enrichment analysis is designed for imaging transcriptomics, where gene expression is not measured in the participants. Our method maps each participant's IDP spatial pattern onto an external AHBA‐derived molecular coordinate framework. The resulting individual GS scores should therefore be interpreted as participant‐specific molecular annotations of imaging phenotypes, not as direct estimates of pathway activity. To evaluate its performance, we conducted extensive simulations based on the linear regression analysis of the individual GS scores for SynGO terms and two phenotypic variables. For the Type 1 error simulation, we demonstrated that false positives could arise when using parametric tests for null scenarios where genes or brain regions were randomly shuffled. The false positive could be attributed to co‐expression biases or spatial autocorrelation within the data. This aligns with previous findings on group‐level enrichment analysis and highlights the need for adopting proper null models to guard against false positives when using individual GS scores (Cao et al. 2024). The power simulation showed that individual‐level enrichment analysis had statistical power to detect associations between SynGO terms and phenotypic variables, with individual Flanker performance showing stronger and more consistent associations than BMI. The different patterns observed in BMI and Flanker in both Type 1 error and power simulations suggest that the Flanker performance may have a more direct relationship with the SynGO terms than that of BMI. It is worth noting that the simulation framework has been integrated into BrainEnrich, allowing users to perform simulations for individual‐level enrichment analysis with customized test statistics.

The individual‐level enrichment analysis aggregated spatial coupling within the predefined gene sets to capture how genetic pathways were spatially coupled with individual brain structures or functions. Since AHBA transcriptional profiles were derived from an external source rather than collected from the participants, this approach can be conceptualized as mapping individual IDPs onto a molecular coordinate framework, thereby capturing the inherent molecular signature of each brain phenotype. In the second case study, we showed that SynGO terms involved in the synaptic structure (e.g., membrane, integral components, and cytosol) and trans‐synaptic signaling showed nominal associations with cognitive measures such as processing speed, vocabulary comprehension, and Flanker task performance, suggesting that the regulation of synaptic signaling and structure may be relevant to the variability in cognitive abilities among individuals. Although these associations did not survive FDR correction across SynGO terms, several effects approached corrected significance. This may reflect modest effect sizes as well as the multiple‐comparison burden introduced by testing a relatively broad set of related SynGO terms. These findings should therefore be interpreted with caution, but they also highlight the value of using preselected, hypothesis‐driven gene sets in future applications of individual‐level enrichment analysis when prior biological hypotheses are available. This pattern is broadly consistent with recent large‐scale genetic analyses showing that gene sets involved in synaptic structure and signaling are enriched among the variants influencing general cognitive performance (Huguet et al. 2024; Lam et al. 2017). For instance, a genome‐wide study of > 100,000 individuals found that synaptic cellular components and processes (e.g., regulation of synapse organization and synaptic transmission) were significantly associated with cognitive ability (Lam et al. 2017). Similarly, a recent analysis of copy‐number variants in about 258,000 participants reported that disrupting synaptic genes has major cognitive effects (Huguet et al. 2024). GO analyses identified regulation of the postsynaptic actin cytoskeleton as a top pathway associated with cognitive ability, highlighting that variants affecting spine morphology can modulate cognition (Huguet et al. 2024). In support of these associations, convergent evidence from human postmortem brain studies showed both pre‐ and post‐synaptic protein levels tend to be maintained in cognitively resilient older adults (Bennett et al. 2018; Boros et al. 2017; Carlyle et al. 2021), whereas individuals with cognitive decline or Alzheimer's disease (AD) exhibit pronounced losses of synaptic proteins (Carlyle et al. 2021; Johnson et al. 2020; Ping et al. 2018; Wingo et al. 2019). This case study demonstrated the utility of the BrainEnrich package in exploring molecular‐phenotypic associations through linear regression. Beyond this, individual GS scores derived from the package could facilitate diverse downstream analyses, such as advanced statistical modeling and machine learning.

Several practical considerations may guide the use and interpretation of BrainEnrich. For group‐level IDP maps, mean aggregation with spin‐based self‐contained null models can be used as a practical starting point. For individual‐level enrichment analysis, KS Sum with spin‐based self‐contained null models can be used as a practical starting point. More generally, spin‐based null models are particularly relevant in neuroimaging applications where spatial autocorrelation is a central source of bias, whereas competitive gene‐resampling null models are useful when the research question concerns whether a given gene set shows stronger association than random gene sets of comparable size (Cao et al. 2024). While BrainEnrich offered a comprehensive list of predefined gene sets for enrichment analysis, it is important to note that some of these pathways were not originally designed for imaging transcriptomics. For example, certain GO terms may not be directly relevant to brain structure and function (e.g., pathways related to muscle system GO:0003012), highlighting the need for further refinement of brain‐specific pathway databases. Researchers should assess pathway relevance carefully and interpret the results within the appropriate biological context. For individual‐level analyses, participant‐level GS score distributions should also be inspected using the normality diagnostics provided by BrainEnrich, and terms flagged as strongly non‐normal should be interpreted cautiously or evaluated with post hoc sensitivity analyses. Moreover, the AHBA remains a valuable resource for imaging transcriptomics, offering unique spatial and transcriptional data that are difficult to obtain from other sources. However, its limitations, including a small donor sample size, age range, limited spatial coverage, and sensitivity to preprocessing choices, should be considered when interpreting results. Because BrainEnrich uses AHBA reference expression profiles rather than gene expression measured in the imaging participants, enriched terms should be interpreted as spatial molecular annotations of IDPs rather than direct evidence of participant‐specific gene expression, disease‐related molecular dysregulation, or causal mechanisms. Similarly, core genes are operationally defined as genes with large influence on the GS score or downstream regression coefficient; they should therefore be interpreted as influential contributors to the enrichment statistic rather than validated causal genes. Integrating multi‐omics data, such as single‐cell RNA sequencing, epigenomics, or proteomics, or conducting in vitro experiments (e.g., gene knockout models), can further enhance the biological interpretation of these findings by providing complementary molecular insights.

5. Conclusion

In conclusion, BrainEnrich provides a flexible toolbox for integrating IDPs with AHBA‐derived transcriptomic profiles. By offering multiple association methods, aggregation strategies, predefined gene‐set resources, and competitive or self‐contained null models, it supports enrichment analysis for imaging transcriptomics at both group and individual levels. Simulation studies showed that appropriate null models controlled Type 1 error while retaining sensitivity to detect associations in individual‐level analyses. The case studies demonstrated its utility for generating candidate molecular annotations of group‐level IDPs and for using individual‐level molecular annotations to explore associations with phenotypic variables. Collectively, BrainEnrich provides a flexible framework for integrating macro‐level IDPs with micro‐level transcriptomic profiles for molecular contextualization of IDPs.

Author Contributions

Zhipeng Cao: conceptualization, methodology, formal analysis, writing – original draft, visualization, funding acquisition. Dekang Yuan: methodology, writing. Guilai Zhan: writing, funding acquisition. The remaining authors contributed to manuscript editing and provided feedback throughout the process.

Funding

This work was supported by the National Natural Science Foundation of China (32400925), the Key Medical Discipline Program of Shanghai Xuhui District (SHXHZDXK202308), and the Medical Research Project of Shanghai Xuhui District (SHXH202305, SHXH202504, SHXH202401).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Normality assessment of individual gene set (GS) scores of 55 SynGO terms from 420 HCP participants using the Kolmogorov–Smirnov (KS) Test. For each of six aggregation methods (mean, median, maxmean, meanabs, ks weighted, and ks sum), the distributions of individual‐level GS scores of SynGO terms were tested for normality using the KS test. Each point represents the p value for a single SynGO term, with the red dashed line indicating the 0.05 significance threshold. Lower p values (below 0.05) suggest stronger evidence of deviation from normality. Based on these results, the maxmean, meanabs, and ks weighted methods, whose GS scores exhibited a higher proportion of non‐normal distributions, were excluded from the subsequent linear regression analyses.

Figure S2: Top 30 enriched categories of molecular functions. The tree plot shows hierarchical clustering of the enriched categories based on their pairwise similarities (Wang et al. 2007) using Ward's clustering method (Murtagh and Legendre 2014). The size and color of the circles scale with the magnitude of the gene set score. Three keywords for each cluster are extracted and shown on the right panel. The plot is generated using the R packages GOSemSim (Yu et al. 2010) and enrichplot.

Table S1: The probability of significance (psig) in Type 1 error simulations for individual‐level enrichment analysis for BMI.

Table S2: The probability of significance (psig) in Type 1 error simulations for individual‐level enrichment analysis for Flanker performance.

Table S3: The probability of significance (psig) in power simulations for individual‐level enrichment analysis for BMI.

Table S4: The probability of significance (psig) in power simulations for individual‐level enrichment analysis for Flanker performance.

Table S5: Results of group‐level enrichment analysis of MDD effect sizes map.

Table S6: Results of individual‐level enrichment analysis of cortical thickness.

Table S7: Summary of predefined gene sets.

Data Availability Statement

The data that support the findings of this study are available in GitHub at https://github.com/zh1peng/BrainEnrich. These data were derived from the following resources available in the public domain: Human Connectome Project (HCP) Young Adult Open Access Data, https://db.humanconnectome.org/; Allen Human Brain Atlas (AHBA), https://human.brain‐map.org/; ENIGMA‐TOOLBOX, https://github.com/MICA‐MNI/ENIGMA.