2.1 Participants

The IBC dataset refers to neuroimaging data collected in 13 individuals³ (11 males, 2 females; age mean/SD = 34 ± 5 years) with no previous history of psychiatric, neurological or any other medical disorders that can change brain function. Handedness was determined with the Edinburgh Handedness Inventory (Oldfield, 1971); results of the survey indicate that members of the cohort are predominantly right handed (range: [0.3, 1]; mean ~0.8).

The experiments were carried out with the understanding and formal consent of the participants, in accordance with the Helsinki declaration and the French public health regulation.

For more information about demographic data of the cohort, consult Table A1 of Appendix A.

2.2 Materials

2.3 Experimental procedure

The task-fMRI data of the IBC-dataset first release were collected per participant in four different MRI sessions. Upon arrival to the research center, participants were instructed about the execution and timing of the tasks referring to the upcoming session. Every session was always composed of several runs dedicated to one or more tasks and each task-related run was always repeated in multiples of two, alternating its phase-encoding direction (see Section 2.5 for details).

Full information on the structure of the MRI sessions, namely which tasks were acquired in every session, can be found in Sections 2.3 and 2.5 and Tables 2 and 3 of Pinho et al. (2018) as well as in Section 2 of the IBC documentation (see Section 2.6 of this article for more details about the IBC documentation).

Additionally, a detailed description about the procedures that were undertaken toward handling and training the participant before each MRI session are also provided in Section 2.3 of Pinho et al. (2018).

2.4 Experimental paradigms

The first release of the IBC dataset comprises 12 tasks. Their majority pertains to already-existing paradigms developed and validated in previous studies. They were thus combined in four different MRI sessions according to this criterion. For further details about the organization of the MRI sessions, consult Table 2 of Pinho et al. (2018). Concretely, these 12 tasks refer to the reproduction of the ARCHI and HCP batteries, with minor adaptations, as well as to a new task addressing language comprehension and based on a RSVP paradigm. A summary of the experimental paradigm of each task is given next. Each design is categorical, that is, it relies on the principle of “cognitive subtraction” as means to isolate effects-of-interest across experimental conditions within every task. All conditions per task are listed and described in Appendix B. More information about adaptation and implementation of the corresponding software protocols can be found in Section 2.4 of Pinho et al. (2018).

2.5 Imaging-data acquisition

2.6 Data storage

The online access of all data is assured by the Human Brain Project (HBP) EBRAINS platform as well as the OpenNeuro public repository (Poldrack et al., 2013) under the accession number ds000244 (dataset DOI: 10.18112/openneuro.ds000244.v1.0.0).

The individual and unthresholded z-maps, obtained from all contrast maps of the aforementioned experimental conditions (see Section 2.4 and Appendix B), can be found in the NeuroVault repository (Gorgolewski et al., 2015), under the collection with the id = 4438: https://identifiers.org/neurovault.collection:4438.

The scripts used for data analysis are publicly available under the Simplified BSD license: https://github.com/hbp-brain-charting/public_analysis_code.

The documentation of the dataset, containing a full description of the experimental designs, acquisition parameters and analysis pipeline, is accessible on the IBC website: https://project.inria.fr/IBC/data/.

For further details, consult the Data and Code Availability Statement provided as Supplementary Material.

2.7 Data analysis

2.7.6 Dictionary learning

To obtain individual topographies that summarize the spatial representation of the effects-of-interest, dictionary learning was applied to statistical maps featuring the main contrasts of each experiment, in each subject. This sparse decomposition method extracts spatially distinct components from the joint distribution of contrast values. It thus provides a low-dimension representation of the overlapping neural support of the contrasts probed in the dataset.

We used a multiple-subject dictionary learning model that captures functional correspondence across datasets, as described in Varoquaux et al. (2013). Such hypothesis allows the factorization of the individual contrast maps into a dictionary of cognitive profiles (common to all subjects) plus subject-specific spatial maps (also known as loadings). The sparsity was enforced with an ℓ₁−norm penalty on the loadings of the components, together with a nonnegative constraint.

Formally, consider the set of brain maps obtained for c = 51 contrasts in a subject s ∈ [n]. By enumerating the values across a mesh of vertices, each is a p−dimensional vector, where p is the number of vertices; X^s is thus a matrix of size p × c. Functional-correspondence dictionary learning solves the following minimization problem for λ > 0:

where U^s ≥ 0,  ∀ s ∈ [n]. Here, denotes the set of matrices with row norm smaller than 1. U_s matrices have shape p × k, whereas the functional-loading matrix V has shape k × c, k being the number of components. Herein, we used k = 20. On the other hand, the V matrix defines the functional characteristics of the components. The estimated subject-specific spatial components (U^s), s ∈ [n] can be interpreted as individual topographies; these components may overlap, although their values are zero in most regions. The λ parameter was calibrated in order to yield a sparsity of around 75%. As the estimation problem is nonconvex, initialization matters; here, we created an initial V matrix by clustering the voxels across subjects into k = 20 clusters and took the normalized average of the cluster signal. The implementation relies on the mini-batch k-means and the dictionary-learning methods of scikit-learn v0.21.3 (Pedregosa et al., 2011), a Python machine-learning library (https://scikit-learn.org/stable/).

For display purposes, a hard assignment of vertices to components was undertaken by ascribing each vertex to the component that gets the largest loading (note that all U^s are nonnegative):

Finally, group maps were obtained by first computing the median of the components, as follows:

Then, the same labeling strategy, as the one described above, was employed at every vertex. Note that by construction, a label was assigned to a given vertex if, at least, half of the subjects obtained a nonzero loading for the corresponding component at that vertex. Because components were labeled to outline their functional definition, we chose the contrast label with the highest value in the dictionary—row of V matrix—for that component.

The stability of this decomposition was thus tested by reestimating the dictionary, using only AP- and PA-based activation maps; spatial correspondence of the components was measured through their intrasubject and intersubject correlations. A priori, it was hypothesized that such correlations would be higher within than between subjects.

2.7.7 Prediction of contrast maps across tasks

Because functional activations are overlapping between the IBC tasks due to common underlying cognitive components, we hypothesized that such similarities allow for predicting these contrast maps from the contrast maps of other tasks.

Keeping the notation of the previous section, let us consider the problem of predicting one contrast of one subject, which refers to the column of X^s, j ∈ [c]. We shall also consider the remaining columns that represent brain activity from the remaining contrasts, as well as all data from the other subjects Xⁱ, i ≠ s. To this end, we used Ridge regression:

in which λ was chosen by generalized cross-validation. The prediction output was defined as follows:

The success of the prediction is quantified by the cross-validated coefficient of determination that is computed at each location i:

In practice, a leave-three-subjects-out cross-validation experiment was employed, that is, training sets were composed of the contrasts from 10 subjects and 3 were left for the test set. Contrasts from the same task were predicted jointly, as they are correlated; hence, predicting one from the other would lead to optimistic bias. Ward's clustering method was used for parceling the brain volume into 100 subregions. The aforementioned analysis was performed in these subregions, similarly to what was done in Tavor et al. (2016). Prediction accuracy at every voxel was measured in terms of R²−scores obtained from each data partition. The maximum of this statistic across tasks was reported afterward. The reason for taking the maximum relates to the fact that, at a given location, it is unlikely to obtain robust responses for more than one task; this means that the baseline of the predictive score is zero,⁴ as confirmed by the analogous computation performed with a dummy classifier. It thus becomes more appropriate to select the maximum rather than the average or the median of the accuracy predictions across tasks, since the maximum outlines the existence of, at least, one latent cognitive dimension explaining activity in a certain brain location.

Finally, the same experiment was repeated but scrambling the subject correspondence between train and test through a random permutation. This aims at quantifying the subject specificity of these functional relationships.

2.7.8 Region-of-interest analysis

The functional-response profiles of a set of six brain regions, previously reported to participate in the language network, were obtained from a subset of contrasts pertaining to language processing. Six regions-of-interest, based on the statistical maps obtained from Pallier, Devauchelle, and Dehaene (2011), were selected a priori.

To better account for the differences inherently associated with the precise location and shape of these regions in every participant, subject-specific counterparts of these regions-of-interest were computed by applying a dual-regression approach adapted from Nickerson, Smith, Öngür, and Beckmann (2017). Formally, consider the set of six regions-of-interest's represented as binary vectors (masks) with p voxels aggregated in a matrix R of size 6 × p. For a subject s ∈ [n], its individual projection R^(s) is a matrix with the same size as R, computed as follows:

in which matrix X^(s), with size j × p, represents the set of j individual contrast maps, where j ∈ [c] obtained for c = 43; pinv(X^(s)) denotes the pseudoinverse matrix of X^(s). One should note that pinv(X^(s)) X^(s) is a linear projection operator that transforms R into R^(s), according to the subject-specific functional information embedded in X^(s). This approach is an adaptation of the two stages featuring the dual-regression approach (consult Figure 1 and equations (1) and (3) of Nickerson et al. (2017)) and its notation is adapted to the type of data used in the present context. The number of contrasts maps c amounts to 43 because these are the 51 main IBC contrast maps from the first release minus those 8 selected to draw their functional profiles (see Section 3.5). In order to generate the corresponding subject-specific region masks, the maps obtained in R^(s) were then transformed into binary maps according to a prespecified threshold; this threshold was set in order to yield the same number of voxels in individual regions as in the group region.

Afterward, for each region-of-interest, the mean of the z-scores at the voxels of the selected contrasts, belonging to the corresponding subject-specific region mask, was computed in every subject. Then, for each contrast, the average across subjects and its 95% CI was estimated in order to create the functional fingerprint specific of that region.

Moreover, a leave-one-group-out cross-validation experiment was set in order to verify whether voxels can be correctly assigned to one of these six regions-of-interest, according to the profile of their functional activations. This experiment was run on data from all participants; voxels belonging to pairs of regions-of-interest were classified to their respective region—based on the functional contrasts—using a linear support vector classifier. Training sets were composed of data from 12 participants and prediction was performed in the data from the remaining participant; each group was thus referring to the data from one participant. Predictions were made between all possible combinations of two regions from the original collection of six regions-of-interest. Accuracy prediction at every voxel was measured in terms of classification accuracy per region pair. This analysis was performed using scikit-learn. To compute the chance level, the same procedure was employed using the scikit-learn's dummy classifier.

3.1 Reproduction of ARCHI and HCP tasks

Figure 1 displays the correlation of group-average z-maps between corresponding contrasts obtained from: (a) IBC versus ARCHI datasets and (b) IBC versus HCP datasets. The diagonal of the matrix refers to homologous contrasts and, therefore, it shows how successfully the maps of ARCHI and HCP were reproduced in IBC. The high correlations present on the diagonal confirm that the functional signatures of the same contrast maps were preserved across different datasets, providing clear evidence that the original results obtained in these large datasets are reproduced using the IBC dataset. One exception stands for the punishment-reward contrast map of the HCP Gambling task; nonetheless, results for this contrast map were also very variable within and between subjects for the original dataset (see Figure A1 in Appendix C). Additionally, correlations between different contrast maps provide a quantitative measure of the similarity of their neural correlates. For instance, left hand versus right hand contrast map from the ARCHI Standard task and left hand versus any motion contrast map from the HCP Motor task are highly correlated because both functional signatures mostly refer to movements of the left hand. As another example, social-interaction motion versus random motion contrast maps from both ARCHI social and HCP social tasks are also highly correlated since their functional signatures both pertain to the social judgment of motion-specific properties of a triangle-shape clip art.

3.2 Variability of functional signatures within and between participants

Figure 2 shows the individual functional signatures obtained for the Working Memory task adapted from the HCP protocol. It displays a qualitative visual description of the variability of such functional signatures across participants. Although the main features of functional responses are preserved at a coarse scale across individuals, there are clear differences in the precise locations of the active regions. For instance, this is the case for the place image versus any images contrast that displays an active focus in the vicinity of the transverse occipital sulcus for all participants; yet, the size of the cluster and its MNI-space position vary conspicuously among them. Similarly, the body image versus any image and face image versus any image contrasts display activity along the fusiform gyrus, as well as face image versus any image and tool image versus any image in the lateral occipital complex. However, their expressions differ across subjects, resulting into different functional preferences at the corresponding locations in MNI space. It remains to be checked whether such variations represent intrinsic between-subjects variability or simply sampling noise.

In turn, Figure 3 provides a quantitative characterization of the spatial consistency, within and between participants, of the main contrasts from each task. Consistency is measured by a similarity metric, namely Pearson correlation. The mean of the correlations between participants stands always above zero, revealing some level of spatial consistency among individual maps. Nevertheless, they are always below 0.5, which quantifies the level of noise of these maps. Importantly, the contrasts show different levels of overlap because of either the low fSNR or the large variability present in each contrast. For instance, contrasts from both ARCHI Social and HCP Social tasks, that is, false-belief story versus mechanistic story, false-belief tale versus mechanistic tale and social-interaction motion versus random motion, as well as punishment versus reward from the HCP Gambling task exhibit lower consistency than tasks involved in language or perception.

On the other hand, spatial consistency is higher at the individual level in 27/33 contrasts displayed in Figure 3. Note that, since per-session data are used to compute the individual-maps consistency, the fSNR in every map is lower for the intrasubject consistency analysis than for the intersubject consistency. These correlations also reflect imperfectly corrected PA/AP spatial distortions. Overall, the fact that maps are generally more consistent within subjects than across subjects hints at subject-specific topographies.

We have also compared intersubject consistency of the main-contrast maps (except those from the RSVP Language task) estimated from the IBC dataset and the original datasets, that is, ARCHI and HCP datasets (see Figures A1 and A2 in Appendix C). Results show that, in the large majority of the cases, considerably higher correlations were achieved using the IBC data than the ARCHI data as well as the HCP data. Intrasubject consistency was also evaluated between the IBC and HCP data and correlations are also overall higher among contrasts from the IBC dataset (see Appendix C). They thus confirm that the IBC dataset hints at good levels of fSNR—in agreement with the results stated in Pinho et al. (2018)—as well as an overall high data quality, including an adequate parametrization of the MRI acquisitions.

3.3 A dictionary of cognitive components

We then analyzed the topographies of 20 cognitive components obtained by decomposing, per subject, the original data of all tasks—that is, 51 main-contrast maps—using sparse dictionary learning. These individual topographies are displayed as spatial maps (see top section of Figure 4), wherein each map represents the functional profile of a component in each subject and hemisphere (i.e., maps of the top-left and top-right panels refer to the left and right hemispheres, respectively). According to Smith et al. (2009) and given the total number of IBC contrasts, k = 20 represents a consensual number of components to feature this study. A larger number of components were deliberately not chosen for the sake of readability of the present results. These components synthesize the cognitive information across all contrasts, making between-subject comparison easier. They were named according to the main conditions of the contrasts that obtained the largest value in their functional fingerprint. Such labeling was only possible due to the use of task data. The functional fingerprints of the components are provided as Supplementary Material.

Since most of the activations depicted in the original contrast maps are located in the cortex, surface-based analysis was herein employed in order to enable a fine functional parcellation of the cortical structures from the individual topographies (see Section 2.7.1 for details).

The two red-framed brain maps, displayed at the right-top corner of each panel in Figure 4, show the median map of the components across individuals, that is, each vertex has the label corresponding to the strongest component observed in—at least—half of the participants. These median maps thus provide a robust consensus model of the topographies obtained across subjects. Overall, tasks consistently map brain networks across participants. Nevertheless, one can still observe differences between subjects that are worthy of notice. For instance, components tagged as math, saccades and consonant strings are observed in similar brain regions across subjects, but one typically dominates the others. In other words, their spatial representations compete across subjects. While their outline is broadly similar, they reveal in detail notable differences at high resolution.

To assess the stability of the dictionary, the dataset was split into two halves, according to the phase-encoding direction parameter. The split-half reproducibility of the 20 components was measured as correlations of the individual topographies both within and between subjects. The topographies were, in this case, obtained through the contrast z-maps estimated from the “PA” and “AP” runs of each task. Intrasubject stability was evaluated through the distribution of the correlations between PA- and AP-related topographies in every subject. On the other hand, intersubject stability was estimated through the distribution of correlations between PA- and AP-related topographies within pairs of subjects. The results, displayed at the bottom section of Figure 4, show higher correlations estimated from intrasubject data than intersubject data, highlighting that the variability of such spatial representations is linked to individual differences.

Finally, it is also worthy of note that these aggregated components are more stable than the original contrast maps (see Figure A1 of Appendix C). Overall, means of the distributions for both intrasubject and intersubject correlations of dictionary components (see bottom section of Figure 4) are higher than means of the distributions of z-maps correlations (see Figure 3).

3.4 Reconstructing functional contrasts from different tasks

Because different conditions across tasks share the same cognitive properties, the contrasts are expected to capture commonalities resulting from the corresponding functional activations. Such commonalities can be learnt and used to predict other contrasts from tasks supposedly sharing the same cognitive properties (see, e.g., Tavor et al. (2016)). Thus, contrast maps from a target task (one contrast map at a time) are predicted from contrast maps of the training tasks. The R²−statistic was used to quantify the discrepancy between the predicted and actual maps, and the maximum of this statistic was computed across tasks.

Figure 5—top shows that a good accuracy prediction was attained throughout all regions in the brain, except in the hippocampus, superior temporal asymmetrical pit (Glasel et al., 2011; Leroy et al., 2015), precuneus, and inferior temporal gyrus. The few exceptions might be explained by either the lack of tasks targeting these regions or a reduced fSNR at these regions.

Several tasks were well predicted by other tasks (see Figure 5—bottom). Indeed, tasks whose contrasts share cognitive properties with many contrasts from other tasks—like ARCHI Standard, RSVP Language, and HCP Working Memory—received higher scores. Contrariwise, a task like HCP Motor, that yields not only on hand movements (common to other tasks) but also on foot and tongue movements—which are unique to this task—could not be as well predicted by the other tasks. At last, this analysis was repeated, but with subjects permuted between train and test, potentially breaking the fine-grained structure of brain responses. Results show a clear decrease of the proportion of the predicted voxels, confirming that topographies are subject-specific. Interestingly, such decrease was not very evident for the HCP Motor task, highlighting that low-level cognitive processes are less sensitive to spatial variability across individuals. Chance level was computed for all tasks with and without permutations of subjects between train and test. In all cases, its value was zero.

Overall, results from Figure 5 highlight two important findings. First, brain responses in different contrasts underlie a latent structure, allowing for the prediction of a response to a given task in a new subject; this structure is, in turn, conditional to other brain responses. Second, the topographic organization of these components is subject-specific.

3.5 Functional mapping of the language network

As several contrasts cover the same psychological domains, the IBC dataset provides the appropriate framework to derive a fine and unambiguous characterization of brain regions with respect to cognition. To showcase the benefits resulting from contrast-maps accumulation, we thus investigated the cognitive profile of a set of brain regions known to be linked to language mechanisms. These regions-of-interest refer to: (a) the left inferior frontal gyrus (IFG) pars orbitalis and (b) pars triangularis, (c) the left temporoparietal junction, (d) the left temporal pole, (e) the left anterior, and (f) posterior superior temporal sulcus (STS). They were subsequently made subject-specific using the dual-regression approach described in Section 2.7.8. The maps of their projections onto individual templates are provided in Figure A3 of Appendix D.1. To illustrate the specific function of these regions, Figure 6 shows the activations elicited by—linearly independent—contrasts selected from the ARCHI Standard and RSVP Language tasks, whose effects-of-interest are related to language.

The two portions of the left IFG reveal greater responses to read sentence than read words, as highlighted in the selected contrast from the RSVP Language task, although effects are significant in both cases. Worthy of notice are the effects obtained in all regions for the contrast sentence⁵ versus mental subtraction (reverse contrast of mental subtraction vs. sentence in Figure 6), except in the left IFG pars triangularis. These results indicate higher level of activation in these regions, but in the left IFG pars triangularis for the general-domain semantic condition (i.e., sentence) than the math-specific semantic condition (i.e., mental subtraction). Indeed, Amalric and Dehaene (2016) report several regions of the language network activating less for math statements than nonmath statements. On the other hand, the left IFG pars triangularis was shown to be specifically implicated in arithmetic processes (Andin, Fransson, Rönnberg, & Rudner, 2015).

The left temporoparietal Junction and left temporal pole show greater responses in the read sentence versus read words and read sentence versus read jabberwocky contrasts, revealing their involvement in the modulation of combinatorial semantics. In fact, the absence of activations in the left Temporoparietal Junction for the remaining contrasts highlights that this region is specifically responsive to sentence comprehension. Opposite effects in the read words versus read pseudowords, read words versus consonant strings and read pseudowords versus consonant strings contrasts indicate low responsiveness to lists of words. Additionally, the left Temporal Pole also exhibits a high level of activation in the listen sentence versus read sentence contrast, which reflects some sensitivity to auditory stimuli.

At last, the left STS exhibits overall strong effects to different levels of specificity in language mechanisms, given its high level of activation in most of the contrasts. Yet, its posterior portion is more responsive to list of words than the anterior, as shown in the contrasts read words versus read pseudowords, read words versus consonant strings and read pseudowords versus consonant strings. Contrariwise, its anterior portion is more responsive to combinatorial semantics than the posterior, as shown by the contrasts read sentence versus read words and read sentence versus read jabberwocky. Similarly to the left Temporal Pole, both portions of the left STS are responsive to auditory stimuli.

All regions were equally sensitive to the read sentence versus checkerboard contrast, highlighting the low level of specificity of this contrast for language-related mechanisms.

Overall, the variety of contrasts referring to language highlights differences in the functional specialization of the aforementioned regions-of-interest. To assess the statistical significance of these differences, we tested whether voxels could be reliably classified as belonging to one of these regions, based on their functional activation in the given contrasts. Figure 7 presents the prediction scores obtained between pairs of regions against their baselines, that is, the corresponding chance levels of such predictions. All scores outperformed the chance level, although the classifier between IPS pars orbitalis/pars triangularis was close to chance. In order to generalize our conclusions to other cognitive domains and demonstrate the benefits of accumulated task data, the same experiment was also performed with data extracted from all main-contrast maps obtained from the entire dataset. The decoding performance was slightly higher (see Figure A4 in Appendix D.2, which shows that including a wider range of cognitive domains improved the functional characterization of the regions-of-interest.

To further evaluate the generalizability of these results, we conducted the same analysis in another set of regions-of-interest, also pertaining to the language network but extracted from functional signatures obtained from a larger cohort. Concretely, the group-level contrast story versus math of the HCP Language task (for more information about this task, consult Section 2.4.2), referring to a subset of the HCP900 dataset (n = 786), was used to define functional regions participating in language-related mechanisms. A complete description of the procedures concerning regions' extraction can be found in Appendix D.3. This set of regions-of-interest partially overlaps with the previous one. The left Temporoparietal Junction was clearly isolated. Besides, it was possible to extract the left anterior STS as well as the left Temporal Pole but we were not able to dissociate both into two distinct regions; similarly, the results allowed the extraction of the IFG, yet comprising both pars orbitalis and pars triangularis. We further identified the Ventromedial Prefrontal Cortex, usually reported to be involved in semantic integration. In summary, the cognitive profiles obtained are consistent with the previous ones (see Figure A5 in Appendix D.3) and differences between them are statistically significant, since all prediction scores between pairs of regions were higher than chance (see Figure A6 in Appendix D.3).

The taskwise structure of the IBC dataset thus opens a new type of analysis, whereby brain-region profiles can be compared in order to highlight their functional specificity.

4.1 A small-n, yet reliable functional mapping

Fifty-one main-contrast maps were investigated in this cross-task analysis, yielding a unique high-coverage cognitive mapping. Despite the low number of subjects, group-average results are consistent with those from previous larger-scale datasets, as highlighted by the results presented on Section 3.1.

4.2 Spatial variability of functional signatures

High-resolution, individual contrast maps provide evidence of the intersubject spatial variability inherent to functional signatures associated with cognitive processes. In Figure 2, one can see, for example, the differences across participants in the cluster extent placed on a subsection of the transversal occipital sulcus, present in the contrast place image versus any images. This variability could be mitigated by normalization techniques relying on macroanatomical landmarks (Frost & Goebel, 2012) rather than MNI coordinates; this matter is subject of further investigation. Importantly, the number and diversity of contrasts allow us to quantify such variability within and between participants (see Figure 3). For instance, it shows which cognitive domains are more stable between participants. Contrasts isolating vision processes, motor performance and low-level of semantic comprehension in language (e.g., manipulation of words) are more consistent across individuals than contrasts pertaining to social cognition or theory-of-mind. Besides, spatial consistency can also be studied within subjects for every task, showing which brain networks display a low fSNR, due to for example, geometric distortions inherently associated with MRI acquisition parameters. Overall, intrasubject consistency is higher than intersubject consistency, which hints at subject-specific topographies.

4.3 Topographies of cognitive components

The taskwise structure of the IBC dataset can support the investigation of common functional profiles or fingerprints and link them to experimental tasks. Here, we use dictionary learning to extract these functional profiles in an automatic fashion. Note that, other approaches would also be possible, such as independent components analysis or clustering. Our data-driven analysis, exploiting individual topographies across tasks, highlights the intersubject variability inherent to cognitive mapping (see Figure 4). Similarly to the results in Section 3.2, intrasubject consistency was higher than intersubject consistency. Interestingly, the mean of intrasubject correlation for all components (see bottom section of Figure 4) is about twice the intrasubject consistency registered for the contrast maps themselves (see Figure 3). Obviously, more contrasts lead to less noisy, hence more reproducible components.

Note that we chose here to arbitrarily estimate 20 components to represent the data. There is no reason to believe that this is an optimal number. We defer a more thorough discussion on this topic to future work, bearing in mind that the level of description always depends on the scientific question posed. We also note some alternative modeling choices that have been proposed in the literature, for example, relying on group structure (Varoquaux et al., 2013) or spatial regularization (Abraham, Dohmatob, Thirion, Samaras, & Varoquaux, 2013). These algorithms contain some hyperparameters that are hard to set in practice, and their convergence is computationally challenging. Hence, we have decided to rely on a simpler, data-driven approach, expecting that the accumulation of contrasts obviates the need for regularization terms.

At last, we opted to map the individual topographies on the surface, since most of the activations in the original volume-based contrasts are located in the cortex. The importance of surface-based mapping comes to the forefront when one aims at individual-specific localization in the upper cortical areas of the brain and, thus, surface-based cortical registration becomes relevant toward setting a common system of topographical representation on the surface. Recent methods based on geometric features derived from functional and diffusion imaging, such as spherical daemons (Yeo et al., 2010) or multimodal surface matching (Robinson et al., 2014), are promising in regards to deliver a better surface alignment because of a greater sensitivity to a wider variety of surface descriptors. Therefore, we acknowledge the possibility to integrate, in the future, such steps in the preprocessing pipeline.

4.4 Prediction of contrasts within the dataset

Since many contrasts share the same cognitive components, they display overlapping functional patterns. This shared variability can be learnt and used to predict other contrasts. Such a successful prediction reveals the existence of a latent structure underlying them (see Figure 5). Additionally, predictions are higher when subject identity is not randomly shuffled, highlighting once again that functional anatomy varies across subjects.

Interestingly, this experiment, together with the reproduction of contrast maps from the ARCHI and HCP datasets (see Figure 1), implies that some brain maps could be transferred from IBC to these aforementioned datasets. For instance, we hypothesize the possibility of predicting the topographies of the HCP participants from the individual contrast maps of the IBC-ARCHI or RSVP-language tasks.

4.5 Cognitive regional profiles of the language network

The taskwise organization of the dataset also opens the possibility to investigate the functional profiles of brain regions linked to cognitive domains covered by the contrasts (see Section 3.5). For instance, the results shown in Section 3.5 can identify which regions elicit effects modulated by word-level (e.g., left posterior STS) and sentence-level semantics (e.g., left Temporoparietal Junction and left Temporal Pole). On the other hand, effects modulated by syntax are difficult to isolate in the context of brain mapping. Some studies have suggested that syntactic processing is distributed throughout an ensemble of brain regions that support high-level linguistic processing. Therefore, they cannot be separated from other aspects regarding language comprehension, like lexico-semantic processing (Blank, Balewski, Mahowald, & Fedorenko, 2016), overall suggesting that the language network might be more strongly concerned with meaning than structure (Siegelman, Blank, Mineroff, & Fedorenko, 2019). Yet, as an attempt to address this issue, we plan to integrate, in a future release, a naturalistic language-comprehension paradigm, dedicated to syntactic-composition modulation (Bhattasali et al., 2019). The more tasks are included in the dataset, the greater the richness of the contrasts that can be used to not only disambiguate the cognitive role of functional regions but also delineate finer demarcations of their anatomical boundaries.

4.6 Limitations

According to the structure of the IBC dataset, the analysis is limited to 13 subjects. This is not enough if one were to report populations effects, yet it leaves the possibility of conducting parallel analysis on 13 unique brains and highlight commonalities among them. The question of deriving a sensible group-level model for this dataset remains thus open.

4.7 Toward more individualized models of brain function

One possible extension of the present work relates to the development and testing of feature-based alignment approaches (Haxby et al., 2011; Sabuncu et al., 2010) that can match specific properties of functional brain regions, such as cognitive responses or connectivity. They can thus be used to improve the estimation of functional templates in the presence of conspicuous between-subject variability, as demonstrated in the recent work by Bazeille, Richard, Janati, and Thirion (2019). Additionally, Bijsterbosch et al. (2018) have shown that most of the relevant cross-subject differences, identified through brain imaging, are rooted in the topography of their individual maps. These findings suggest that future investigations on the latent structure of individual contrast maps may contribute to improving the analysis of between-subject characteristics in brain imaging. Nevertheless, an important question still remains: what is the most appropriate method to achieving these correspondences?

4.8 Ongoing taskwise development of the Individual Brain Charting dataset

The collection of new data in IBC is ongoing until year 2022 and more releases are planned over the upcoming years. The second and third releases have recently been made publicly available—source data are available in OpenNeuro under the data accession ds002685⁶ and derivatives in NeuroVault with the id collection 6618 ⁷— and they comprise a variety of cognitive tasks. Like the first release, the second release covers tasks on both lower-order and higher-order cognitive functions, such as mental-time travel (Gauthier & van Wassenhove, 2016a, 2016b), positive-incentive value (Lebreton, Abitbol, Daunizeau, & Pessiglione, 2015), theory-of-mind and pain matrices (Dodell-Feder, Koster-Hale, Bedny, & Saxe, 2011; Jacoby, Bruneau, Koster-Hale, & Saxe, 2016; Richardson, Lisandrelli, Riobueno-Naylor, & Saxe, 2018), numerosity (Knops, Piazza, Sengupta, Eger, & Melcher, 2014), self-reference effect (Genon et al., 2014), and speech recognition (Campbell et al., 2015). The third release is dedicated instead to an extensive data collection tackling the visual system; it is concerned with tasks on visualization of naturalistic scenes (Huth, Nishimoto, Vu, & Gallant, 2012), classic retinotopy, and movie watching (Haxby et al., 2011). The fourth release is currently in preparation and it includes a tonotopy task (Santoro et al., 2017) as well as two batteries of localizers (Aron, Behrens, Smith, Frank, & Poldrack, 2007; Bissett & Logan, 2011; Crump, McDonnell, & Gureckis, 2013; Eriksen & Eriksen, 1974; Fan, McCandliss, Sommer, Raz, & Posner, 2002; Figner, Mackinlay, Wilkening, & Weber, 2009; Hamame et al., 2012; Kaller, Rahm, Spreer, Weiller, & Unterrainer, 2011; Kirby & Maraković, 1996; Ossandon et al., 2012; Otto, Skatova, Madlon-Kay, & Daw, 2014; Perrone-Bertolotti et al., 2012; Saignavongs et al., 2017; Schneider & Logan, 2011; Shallice, 1982; Stroop, 1935; Vidal et al., 2010; Ward & Allport, 1997) addressing many other cognitive modules, such as: stimulus salience, working memory, visual object categorization, audio perception, risk-associated decision making, motor inhibition, planning and vigilance. Our ultimate goal is to achieve a comprehensive brain coverage of functional signatures, associated with a large variety of mental functions, by the end of the project.