Neural, genetic, and cognitive signatures of creativity

Creativity supports problem-solving and broader societal advancement. Divergent thinking (DT)—the ability to generate multiple ideas for open-ended problems—is a core operationalization of creative ability and relates to perspective taking, mood, and potential. Prior task-based fMRI studies report both consistent and variable activations during DT, commonly implicating bilateral DLPFC, VLPFC, and right ACC, with additional findings in AG, FG, and MTG. Variability likely reflects DT’s complexity and coordination of multiple component processes. Network-level accounts grounded in dual-process theory posit interactions between spontaneous idea generation (DMN) and controlled evaluation/selection (FPCN). Beyond discrete parcellations, macroscale functional gradients suggest a continuous organizational axis from sensory processing to abstract cognition; DT’s distributed nature may align with such gradients. The present study aimed to construct and validate a whole-brain neural representation of DT using MVPA in two task-fMRI samples, decode its cognitive relevance, test its alignment with cortical gradients, and link it to gene expression and neurotransmitter maps. Finally, we assessed its predictive value for individual DT during task performance and generalized to resting-state functional connectivity across large samples.

Task-based neuroimaging of DT frequently reports engagement of executive and associative regions (bilateral DLPFC/VLPFC, ACC), alongside temporal and parietal areas (AG, MTG, FG), reflecting the integration of memory, semantic association, and control. The dual-process theory frames creative cognition as interplay between spontaneous generation (DMN; mind-wandering, free association) and executive control (FPCN; evaluation, selection). Contemporary accounts emphasize that cortical organization is well-captured by continuous gradients (principal gradient spanning unimodal sensory/motor to transmodal DMN), suggesting DT’s neural architecture may be distributed along this axis rather than confined to discrete nodes. Prior work in imaging genetics and neurotransmission indicates cognition is modulated by gene expression patterns and neuromodulatory systems, motivating integrated analyses linking brain activity to transcriptomic and PET-based receptor maps.

Design and samples: Two independent task-based fMRI samples were collected during an Alternate Uses Task (AUT) variant with Novel Use (NU; DT) and General Use (GU; control) conditions. Sample 1: n=55 (after exclusions; 43 females; mean age 19.23). Sample 2: n=30 (after exclusions; 20 females; mean age 19.30). Post-scan, participants recalled/refined responses and rated originality (1–5). Task paradigm: Sample 1 included 4 runs with 20 trials/run (12 NU, 8 GU). Each trial: fixation (2 s), cue (4 s), object presentation (NU: 12 s; GU: 4 s) for idea generation; keypress and brief written response, then rest (8 s). Sample 2 included 6 sessions (40 NU, 20 GU); post-scan originality ratings identical to Sample 1. MRI acquisition: Siemens 3T Trio; task fMRI GRE-EPI (TR/TE=2000/30 ms, FA=90°, 64×64, FOV 220×220 mm², 3 mm slices, gap 1 mm); T1 MPRAGE (1 mm isotropic). Preprocessing via fMRIPrep: motion correction, slice timing correction, coregistration to T1, normalization to MNI152NLin2009cAsym, computation of FD/DVARS, CompCor regressors; nuisance regression and high-pass filtering; motion outlier annotation. First-level GLM: Subject-level GLMs modeled NU and GU conditions to derive beta maps for MVPA. Single-trial GLMs estimated beta images for each NU trial for RVR prediction. MVPA classification: Whole-brain gray-matter masked linear SVM (C=1) classified NU vs GU separately in each sample using 10×10-fold cross-validation, repeated 10 times with different splits. Performance indices: accuracy, sensitivity, specificity, ROC/AUC. Permutation tests (5000) assessed chance-level performance. Cross-sample validation: each model applied to the other sample without refitting. Bootstrap (10,000 iterations) provided voxelwise p-values; FDR q<0.05 yielded thresholded maps. The final DT brain pattern was the average of thresholded weight maps from both samples after confirming high weight-map correlation. Cognitive decoding: Spatial correlations between the unthresholded DT pattern and meta-analytic activation maps (Neurosynth) to identify associated cognitive terms (FDR-corrected). Network overlap and gradients: Overlap of positive/negative predictive weights with 7 canonical resting-state networks (Yeo et al.). Spatial correlations with the first 10 functional gradients (neuromaps); significance via spin tests with FDR correction. Gene expression: Allen Human Brain Atlas microarray data (6 donors) preprocessed with abagen; expression profiles for 15,633 genes mapped to 574 DT-pattern regions. Pearson correlations between regional DT weights and each gene’s expression profile; significance via spatially autocorrelation-preserving permutations (spin/surrogate). Gene Ontology enrichment (GOrilla) on ranked gene list with FDR control. Neurotransmitter maps: PET-derived receptor/transporter density maps for 19 targets across 9 systems (neuromaps). Spatial correlations with DT pattern; significance via spin tests. Individual-level prediction (task): Relevance Vector Regression (RVR) using single-trial NU beta maps within positive DT-pattern ROIs to predict originality ratings, with 10×10-fold CV (repeated), controlling for gender, age, and motion. Performance: Pearson r and MAE; permutation tests for significance. Resting-state generalization: Three datasets—SLIM (n=410), GBB (n=304), BBP (n=600)—acquired on the same scanner; preprocessing via fMRIPrep (SLIM/GBB) or SPM8/DPABI (BBP) with motion control and band-pass filtering; no GSR. DT pattern divided into 6 clusters (left AG; bilateral OFC; left MTG/ITG; left Precuneus; bilateral PFC/ACC/Thalamus; right cerebellum). ROI-mean time series extracted; 6×6 FC matrices (15 unique edges). On SLIM, RVR with 10×10-fold CV predicted DT (AUT) scores from 15 edges; bootstrap (10,000) identified significant edges (FDR). Validation on GBB and BBP via Pearson correlations between significant edges and DT scores.

• MVPA classification of NU vs GU:
  • Sample 1 (model1): Accuracy 80% ± 3.8% (p < 0.00001); AUC 0.90; sensitivity 80.0% (95% CI: 70–90%), specificity 80.0% (95% CI: 69–89%).
  • Sample 2 (model2): Accuracy 85% ± 4.6% (p < 0.00001); AUC 0.94; sensitivity 96% (95% CI: 89–100%), specificity 73% (95% CI: 56–89%).
  • Cross-sample validation: model1 tested on sample2 accuracy 70% ± 5.9% (p=0.0027); model2 tested on sample1 accuracy 78% ± 3.9% (p < 0.00001).
  • Thresholded SVM weight maps highly correlated across samples (r=0.838, p<0.001), enabling an averaged DT brain pattern.
• DT brain pattern regions:
  • Positive weights: bilateral DLPFC, bilateral DMPFC, left VLPFC, bilateral ACC, bilateral OFC, left AG, left MTG, bilateral thalamus, right cerebellum.
  • Negative weights: right superior parietal lobule, right precuneus, right inferior lateral occipital cortex.
• Cognitive decoding (Neurosynth): Positive weights associated with Emotion, Cognitive, Memories, Judgments, Retrieval, Semantic/Reasoning; negative weights with Visual, Selective, Location, Perception, Sensory (FDR-corrected p<0.05).
• Network overlap: Positive weights overlap with Default Mode, Frontoparietal Control, and Limbic networks; negative weights overlap with dorsal Attention, Visual, and Somatomotor networks.
• Functional gradients: Significant correlations with principal gradient (r=0.66, pFDR=0.0005), 9th gradient (r=0.47, pFDR=0.0005), and negative correlation with 8th gradient (r=−0.40, pFDR=0.0007).
• Gene expression and enrichment:
  • Significant positive spatial correlations for specific genes (e.g., TUNAR r=0.23, p=0.0045; TGFBI r=0.22; PRRX1 r=0.21; TCEA3 r=0.20).
  • GO enrichment (FDR q<0.05): negative regulation of multicellular organismal process; detoxification of copper ion; detoxification of inorganic compound.
• Neurotransmitter receptors/transporters (PET maps): Strongest positive spatial correlations with MOR (µ-opioid receptor; r=0.67, pspin<0.001), CB1 (OMAR; r=0.45, pspin<0.001), H3 (r=0.41, pspin<0.001), mGluR5 (r=0.39, pspin<0.001).
• Individual-level prediction (task-based RVR):
  • Sample 1: average r=0.35 (range 0.21–0.39), ppt<0.001; MAE=0.82.
  • Sample 2: average r=0.14 (range 0.09–0.17), ppt<0.001; MAE=1.82.
• Resting-state generalization:
  • SLIM (n=410): RVR prediction average r=0.15, Ppt=0.002; MAE=0.76. Significant FC edges: left AG–left MTG; left AG–left Precuneus; left MTG–right cerebellum.
  • Validation: left AG–left MTG FC correlated with DT in GBB (r=0.14, p=0.018) and BBP (r=0.10, p=0.017); other edges not significant in validation samples.

The study defines a robust, distributed neural signature of divergent thinking that generalizes across independent samples and analytical approaches. The DT pattern emphasizes DMN and FPCN regions, aligning with dual-process accounts wherein spontaneous associative generation (DMN) couples with executive evaluation/selection (FPCN). Negative weights within primary sensory/attention networks suggest reallocation of resources away from externally oriented processing toward internally guided cognition during DT. Mapping onto the principal functional gradient indicates that DT spans a hierarchical axis from unimodal sensory regions to transmodal associative hubs, supporting the view that creativity entails dynamic integration from concrete sensory representations to abstract, high-level cognition. This gradient-based organization underscores that DT is not merely a set of discrete activations but follows a continuous organizational principle that distinguishes it from other higher-order functions. Biogenetic analyses link the DT pattern to genes involved in neurotransmitter synthesis/release and to PET-derived neurotransmitter systems—particularly MOR, CB1, H3, and mGluR5—implicating neuromodulatory pathways (including dopamine-related signaling) tied to reward, mood, motivation, and learning/memory in creative cognition. These associations converge with literature highlighting the roles of reward and affect in facilitating creative thought. At the individual level, the DT pattern predicts originality ratings during task performance and, when translated to resting-state connectivity among DT-pattern ROIs, generalizes to predict DT across large independent cohorts. Together, findings suggest that the identified pattern captures both a common neural basis of DT and meaningful inter-individual variability in creative ability.

This work establishes a comprehensive, multiscale neural representation of divergent thinking that is validated across independent task-fMRI samples, cognitively decoded via meta-analytic maps, aligned with cortical functional gradients, and biologically grounded through transcriptomic and neurotransmitter associations. The DT brain pattern predicts individual creative performance during task engagement and generalizes to resting-state connectivity features predictive of DT across large samples. Future research should extend these integrative approaches to other DT modalities (e.g., visual), convergent thinking, and creativity across domains (music, art, science), and test more demographically balanced samples to enhance generalizability.

• Task specificity: The study relies on an AUT-based verbal DT task, which, despite adaptations, has known reliability limitations and may not capture all facets of creativity.
• Stimulus and scoring: Although expanded and curated stimulus sets and self-ratings were used, subjective ratings may introduce bias; task design mitigations were implemented.
• Sample composition: Predominantly female samples may limit generalizability; future work should use more balanced cohorts.
• Scope: Focused on verbal DT; results may not fully generalize to other DT modalities or convergent thinking without further testing.