Tests for the robustness of results

To test the robustness of the major findings, we re-analyzed the data by assessing the functional connectomes using the Schaefer atlas which partitioned the brain into 400 areas (Schaefer et al., 2018). Then we assessed individual identifiability along the language hierarchy based on 400 by 400 FC matrices.

To validate the main results, we adopted a different strategy of condition pair to assess individual identifiability, which allowed us to use more time points to compute FCs. Given the hierarchical nature of language, it is reasonable to assume that, in addition to the state-independent intrinsic activities, those state-specific processes engaged in the low-level task (e.g., listening to the backward-played story) are also engaged in the high-level task (e.g., listening to the intact story) (as illustrated in Figure 1A). Thus, the degree of similarity of subjects’ FCs between the low- and the high-level tasks should mainly depend on the low-level task, which in turn would largely determine whether a subject can be identified between the conditions. Nevertheless, we noted that while the shared components of low- and high-level tasks can improve the accuracy through shared state-specific contribution, differences between state-specific activities could reduce the performance. In this way, changes in identification accuracy between conditions may underestimate the contribution of the targeted processes (the shared state-specific processes).

Following the above logic, we combined one dataset of the intact condition with the dataset of the rest, the backward, the sentence-scrambled, and the second intact conditions separately. Then identification was conducted between the two datasets for each of the four pairs. We predicted that the identifications of the rest-intact, backward-intact, scrambled-intact and intact-intact conditions should follow a similar pattern as the identification of the rest-task, backward-backward, scrambled-scrambled and intact-intact sessions.

For each condition, a total of 64 time points were extracted to compute FCs. For the backward and sentence-scrambled conditions, this was done by concatenating the corresponding time series from two scan sessions. For the intact condition, this was done by concatenating the corresponding time series from two blocks. For the resting state, the first 64 time points were extracted. Time series were normalized within a session before the concatenation.