Atlases of cognition with large-scale human brain mapping

To map the neural substrate of mental function, cognitive neuroimaging relies on controlled psychological manipulations that engage brain systems associated with specific cognitive processes. In order to build comprehensive atlases of cognitive function in the brain, it must assemble maps for many different cognitive processes, which often evoke overlapping patterns of activation. Such data aggregation faces contrasting goals: on the one hand finding correspondences across vastly different cognitive experiments, while on the other hand precisely describing the function of any given brain region. Here we introduce a new analysis framework that tackles these difficulties and thereby enables the generation of brain atlases for cognitive function. The approach leverages ontologies of cognitive concepts and multi-label brain decoding to map the neural substrate of these concepts. We demonstrate the approach by building an atlas of functional brain organization based on 30 diverse functional neuroimaging studies, totaling 196 different experimental conditions. Unlike conventional brain mapping, this functional atlas supports robust reverse inference: predicting the mental processes from brain activity in the regions delineated by the atlas. To establish that this reverse inference is indeed governed by the corresponding concepts, and not idiosyncrasies of experimental designs, we show that it can accurately decode the cognitive concepts recruited in new tasks. These results demonstrate that aggregating independent task-fMRI studies can provide a more precise global atlas of selective associations between brain and cognition.


Modeling brain response to cognitive-ontology concepts
In a standard GLM framework, we use a design matrix capturing the effect on brain activity of the presence of a term in the task description, followed by a set of contrasts to isolate contribution of the term of interest opposed to related terms in the ontology.
Term effect We assign a set of terms to each image, forming a one-hot-encoding of the database, i.e. representing the occurrence of terms by a binary design matrix. We follow the standard fMRI analysis framework and perform a General Linear Model (GLM). This gives the correlation of each separate voxel with the terms within a set of images, and enables to test for their significance. Using the GLM formulation: y corresponds to the activation maps, X to the design matrix modeling the presence of terms, and β to the term effects. The input activation maps are subject-level condition versus baseline maps. S4 Fig. shows the effect map for the places term. We will use this term in the following to illustrate the differences between the types of inference.
Correlations in the terms induce correlations in the design matrix: effects of terms that appear always together in tasks cannot be teased out. S5 Fig. shows this correlation matrix for our database. We can see that the "visual" and "auditory" terms are very anticorrelated (their correlation is -.9). Indeed, our tasks are exclusively either visual or auditory, aside from the ds114 study in which there is no explicit stimuli. For this reason, we remove the regressor "auditory". The auditory map can be defined as the negated map for the visual term. Other terms suffer from strong correlation, in particular the "voice" and "auditory" terms, as most auditory stimuli are voices. However, some tasks involved non-voice auditory stimuli, such as the muslang study (see S2 Fig). Using contrasts, as detailed below, can then separate the terms corresponding to multiple different types of auditory stimuli.
Term contrasts A GLM estimates responses for each voxel with respect to a combination of terms. This entails that maps corresponding to the individual term effects show a certain degree of specificity: the effect of that term is conditional to the other terms. However, there is shared variance between the terms. To better isolate cognitive processes, a standard analysis in individual studies relies on contrasts in the GLM, e.g., a "face versus place" and a "face versus scrambled picture" contrast for a face recognition study. To disentangle the experimental factors without a too strong a priori on the control conditions, the alternative is to contrast a β map against all others, e.g., "face versus place and scrambled picture". To define such contrasts in a systematic way for the wide array of cognitive concepts touched in our database, we use the categories of our ontology. We form groups of terms within the task categories described in S4 Table: these are used to define the conditions and their controls. Inside each group, we perform a GLM analysis with all the "one versus all" contrasts. We denote these ontology contrasts. Note that we do not perform a 3rd level analysis [1] in the sense that the term effects are estimated directly from the subject-level maps, jointly across all studies.
Other regression approaches As outlined by one reviewer another potential approach to drawing relationships between cognitive concepts and brain activity is to rely on Partial Least Squares of Canonical Correlation Analysis methods -or more precisely, their predictive variants, namely reduced rank regression. These methods typically find combinations of terms that are highly correlated with combination of regional activities. However, they tend to combine many terms to form their prediction, creating latent factors -"loadings"-distributed across labels. In the present work we prefer to rely on term-specific mappings that avoid the additional difficulty of studying the cognitive loadings of the obtained components. The combination across terms is then done explicitly through contrasts and discriminative models.