Fig. 5: Experiential features’ independent contributions to predicting mental image fMRI data.

To estimate which experiential features independently contributed to fMRI prediction, and where in default mode network (DMN) was predicted, a variance partitioning method was used. To estimate unique feature contributions, an fMRI encoding model was first fit to the sentence dataset—namely, a set of regression mappings to predict the average activation of each DMN Schaefer ROI, based on the twenty crowd-sourced feature ratings (see Supplementary Fig. 8). The encoding model was transferred to predict the mental image fMRI data with the participant-specific mental image feature ratings. To then isolate the predictive contribution of a feature Y, a second encoding model was fit on the sentence fMRI data to the nineteen other features (excluding Y). The contribution of Y was estimated as the difference in predicted variance (estimated as r²) between the full set of twenty featuresand the subset of nineteen that excluded Y. This value is presented on the cortical maps as a pseudo correlation estimate: sqrt[r²(full) -r²(subset)]. Predictive contributions are displayed only if prediction accuracies derived from all twenty features were significantly greater than the subset of nineteen. Significance was determined using signed-ranks tests on each DMN Schaefer ROI (fifty participants, one-tailed). P-values were corrected according to the false discovery rate (FDR)³⁵. Values are displayed for ROIs where FDR(p) < 0.05. A comparative analysis of all 1000 ROIs is in Supplementary Fig. 10. To assist visual interpretation, the encoding model beta coefficients for the twenty-feature model are in Supplementary Fig. 11. Brain images were generated using BrainSpace v1.10⁵⁷.