Abstract
The growing literature reporting results of cognitive-neural mappings has increased calls for an adequate organizing ontology, or taxonomy, of these mappings. This enterprise is non-trivial, as relevant dimensions that might contribute to such an ontology are not yet agreed upon. We propose that any candidate dimensions should be evaluated on their ability to explain observed differences in functional neuroimaging activation patterns. In this study, we use a large sample of task-based functional magnetic resonance imaging (task-fMRI) results and a data-driven strategy to identify these dimensions. First, using a data-driven dimension reduction approach and multivariate distance matrix regression (MDMR), we quantify the variance among activation maps that is explained by existing ontological dimensions. We find that ‘task paradigm’ categories explain the most variance among task-activation maps than other dimensions, including latent cognitive categories. Surprisingly, ‘study ID’, or the study from which each activation map was reported, explained close to 50% of the variance in activation patterns. Using a clustering approach that allows for overlapping clusters, we derived data-driven latent activation states, associated with re-occurring configurations of the canonical fronto-parietal/salience, sensory-motor, and default mode network activation patterns. Importantly, with only four data-driven latent dimensions, one can explain greater variance among activation maps than all conventional ontological dimensions combined. These latent dimensions may inform a data-driven cognitive ontology, and suggest that current descriptions of cognitive processes and the tasks used to elicit them do not accurately reflect activation patterns commonly observed in the human brain.
Introduction
The exponential growth of functional neuroimaging studies mapping constructs at the psychological level to patterns of brain activity (cognitive-neural mappings) has led to issues surrounding an adequate organizing ontology or taxonomy of these results 1–4. An ontology of a scientific domain serves both an organizational and descriptive function. In terms of organization, an ontology provides a standardized framework for managing, sharing and analyzing increasingly large databases of experimental analyses and data. Importantly, they also serve to describe experimental results in terms of natural divisions or dimensions of the phenomena under study. This function is particularly important for cognitive neuroscience, where a central goal is to map constructs at the psychological level to phenomena studied at a neurobiological level of analysis. The present goal of this study is to identify data-driven dimensions of a potential ontology based upon its ability to explain variability between patterns of brain activity.
Important progress in the development of an ontology for cognitive-neural mappings has been made using a combination of large-scale meta-analysis of task-based functional magnetic resonance imaging (task-fMRI) activation maps and topic modeling 5–7. At a general level, these studies map latent cognitive dimensions, derived from text documents 5,6 or task-paradigm descriptions 7, to patterns of brain activity. However, an important factor that needs to be addressed for an ontology of cognitive-neural mappings is what cognitive features of the task-fMRI environment drive the greatest differences in observed activation patterns. In other words, an adequate ontology of cognitive-neural mappings needs to identify those features of the task-fMRI experiment that explain the most variance among task-fMRI activation patterns.
To address this issue in a quantitative framework, we used a whole-brain representational similarity analysis (RSA) approach 8. Using this approach, we quantified the spatial similarity among whole-brain task-activation maps from the BrainMap database (Fox & Lancaster, 2002). Using multivariate distance matrix regression (MDMR; Zapala & Schork, 2006), we predict the variability between task-activation maps by existing ontological categories of the BrainMap database. The fundamental assumption of this approach is that the important features of the exiting ontology are identified in terms of the amount of variance they explain in task-activation map similarities.
Using the spatial similarity estimates among whole-brain task-activation maps, we also explore the possibility of estimating latent ontological dimensions directly from these similarities using a graph clustering approach. We take the opposite strategy of recent topic modeling approaches 5–7. In this approach, latent cognitive dimensions are estimated from experimental-descriptions and then projected onto patterns of brain activity 5, or co-occurring latent cognitive dimensions and patterns of brain activity are simultaneously estimated 6,7. In our approach, we reverse this chain of analysis. Commonly re-occurring whole-brain activation patterns, or activation states, are estimated directly from task-fMRI activation maps, and then projected onto experiment descriptions for interpretation. This approach assumes that the quality of a latent dimension is determined by how well these dimensions explain observable differences in task-fMRI activation patterns. Thus, latent cognitive dimensions are determined not in terms of their co-occurrence in text, but their co-occurrence with dominant whole-brain activation states. Importantly, these dimensions may or may not map cleanly onto existing latent cognitive dimensions.
Results
Outline of Approach
Our approach proceeds in three general steps (Figure 1). First, sparse non-negative matrix factorization (sparse-NMF) is applied to task-fMRI activation maps (N = 8919) reconstructed from experiment-reported peak-activation coordinates to derive a sparse ‘dictionary’ of brain networks. Spatial similarity estimates between activation maps could be computed at the voxel level, but this ignores spatial dependence between voxels. As opposed to other techniques (e.g., independent component analysis) for estimation of brain networks using the BrainMap database 11,12, sparse-NMF provides a data-driven network estimation technique uniquely suited for the data reported in the BrainMap database: all data are positive (i.e. positive activation values) and sparse (i.e. a small number of reported coordinates for each experiment). Sparse-NMF naturally leads to a parts-based parcellation of the brain given its enforcement of positive weights 13, even under low signal-to-noise conditions 14.
Second, we compute the spatial dissimilarity between the network activation estimates to a construct an activation map*activation map dissimilarity matrix. This can be seen as a large-scale whole-brain representational similarity analysis 8, where dissimilarity estimates are computed between large-scale activity patterns across the entire brain. We then predict the variability among activation maps from the experimenter-labeled ontological categories of the BrainMap database using multiple distance matrix regression (MDMR) 10. MDMR regresses a distance matrix onto a set of continuous or categorical predictors, and has extensive use in behavioral genomics 15 and studies of individual differences in resting-state functional connectivity (known as connectome-wide association studies) 16. The parameter of interest from this analysis is the explained variance estimate (R2) for five sets of ontological categories provided by the BrainMap database: analysis decisions, response type, stimulus type, task-paradigm class, and behavioral domain. These estimates provide information regarding what features of the existing ontology drive the greatest differences in task-activation maps.
Third, we convert the dissimilarity matrix into a similarity matrix for graph-based clustering. We use a graph clustering algorithm, known as symmetric non-negative matrix factorization (Sym-NMF), that has been shown to outperform commonly-used spectral graph-clustering algorithms 17. Importantly, Sym-NMF allows for overlapping communities of activation maps. As noted above, the goal of this clustering analysis was to estimate potentially overlapping whole-brain activation states. These states may prove useful as candidates for latent dimensions of a new data-driven ontology of cognitive-neural mappings. All code used in this manuscript are provided on the following webpage: https://github.com/tsb46/BrainMap-Cognitive-Ontology (currently a work in progress).
NMF Results
In order to estimate the latent network activation estimates underlying each activation map, we applied a sparse-NMF algorithm to the full sample of 8919 task-fMRI activation maps. We derived a high-resolution network parcellation of 70 networks, the same high-model order estimated in previous analyses of the BrainMap database 11,18. The spatial similarity of activation maps replicated across alternative network solution sizes, as observed by the correlation between dissimilarity estimates of a 70-network solution size with smaller and larger solutions: N = 55, r = 0.91; N = 60, r = 0.93; N = 65, r = 0.96; N = 75, r = 0.95; N = 80, r = 0.94; N = 85, r = 0.92. Visual inspection of the 70-network solution (Figure 2) revealed a sparse parcellation of cortical and sub-cortical regions of the brain, corresponding to functionally relevant brain areas.
Prediction of Activation Map Dissimilarity from Existing Ontological Categories
Using the latent network activation estimates of the 70 networks we computed the activation map × activation map dissimilarity matrix with the cosine distance metric, representing the dissimilarity in latent network activation estimates between all activation maps. We used MDMR to regresses this dissimilarity matrix onto a set of ontological categories associated with each task-activation map. Categories provided by the BrainMap database fell into five domains: analysis decisions (subtraction vs. baseline contrast, and reported de-activation vs. activation), stimulus type (e.g. visual words, auditory stories, fixation cross, etc.), 3) response type (button press, speech, finger tapping, etc.), 4) paradigm class (encoding, task-switching, counting, etc.), and 5) behavioral domain (action execution, language semantics, working memory).
The BrainMap database contains nested data such that multiple task-activation maps are reported for a single study. To correct for this nesting in the BrainMap database, individual explained variance estimates for each categorical domain were corrected using a permutation test that respected study ID during permuting of task-activation map labels. Computation of the explained variance associated with each of the five categorical domains was carried out in a leave-one-out fashion. Each categorical domain was added to the remaining set of domains to compute the unique explained variance associated with each categorical domain.
The explained variance accounted for by all categorical domains (Npredictors = 603) in the model was 19.67%. The explained variance accounted for by analysis decisions (subtraction vs. baseline contrasts, and reported activation vs. de-activations), controlling for other domains, was 0.49% (p = 0.001). The explained variance accounted for by stimulus type, controlling for other domains, was 3.1% (p = 0.001). The explained variance accounted for by response type, controlling for other domains, was 0.57% (p = 0.014). The explained variance accounted for by paradigm class, controlling for other domains, was 4.17% (p = 0.001). The explained variance accounted for by behavioral domain, controlling for other domains, was 3.71% (p = 0.001). In summary, all categorical domains, which correspond to traditional ontological categories commonly used to describe cognitive-neural mappings, accounted for less than 20% of explained variance in activation maps.
We also ran a second model using only study ID as a predictor and we found that study ID, the study from which the task-activation map was reported, accounts for 44.71% of the variance in task-activation map differences. The single predictor of study ID accounted for more than twice the variance of the five categorical domains provided from the BrainMap database.
Identification of Re-occurring Whole-Brain Activation States
As the full set of available ontological categories explain a limited amount of variance between whole-brain activation patterns, we next explore whether differences in whole-brain activation patterns can be explained by a smaller set of latent whole-brain activation states. We first apply a graph clustering approach, symmetric non-negative matrix factorization (Sym-NMF), to derive overlapping clusters of task-activation maps. Overlap allows for the possibility that a single task-activation map may represent a combination of multiple whole-brain activation states. We then project these latent dimensions onto the task-descriptive categories for behavioral and cognitive interpretation.
To choose the optimal number of clusters, we used a cross-validation approach to identify cluster sizes that exhibited stable cluster centroids across split-half samples. We repeated the two-fold cross-validation procedure for cluster sizes ranging from 2 to 20. Examination of average cluster centroid stability across cross-validation samples revealed two strong local minima of stability at a cluster sizes of four (Mstability= 0.043) and seven (Mstability= 0.055). Thus, further analyses were carried out on these cluster solutions.
In order to provide a functional interpretation of the clustering results, we computed the centroid activation maps and the prevalence of BrainMap ontological categories in each cluster. The centroid activation maps represent the weighted average activation pattern of the activation maps within the cluster. The association of a BrainMap category with a cluster was computed as the probability that BrainMap category appears given the cluster, normalized by the base probability of the cluster (see Method and Materials).
For the four-cluster solution, the centroid activation maps corresponded to functionally relevant brain areas (Figure 4). The clusters included a visual/dorsal-parietal activation pattern (Four Cluster 1; 4C1), fronto-parietal activation pattern (4C2), a default mode network (DMN) activation pattern (4C3), and an auditory/motor activation pattern (4C4). Consistent with previous reports of domain-general BOLD activation in anterior insula (AI) and dorsomedial prefrontal cortex (DMPFC) 19, activation in the AI and thalamus was observed across all centroid activation patterns. The BrainMap categories associated with the four clusters are consistent with previous functional descriptions of these brain areas. 4C1 was associated with paradigms requiring viewing and responding to or manipulating visual objects, including ‘mental rotation’, visual shape perception, and ‘spatial cognition’. 4C2 was associated with paradigms requiring a variety of complex high-level actions and behaviors, including ‘working-memory,’ ‘action inhibition’ and ‘language processing’. 4C3 was associated with paradigms requiring recall/recognition of previous information, inferences regarding others’ behavior (‘social cognition’), and various affective states (e.g. sadness, fear, reward, etc.). In addition, consistent with a DMN activation pattern, the activation maps within 4C3 were more likely to be reported as ‘de-activations’. 4C4 was associated with paradigms requiring overt/covert motor responses, along with listening or responding to auditory stimuli.
The seven-cluster solution presented a slightly more fine-grained partition of the four-cluster solution. While 4C2 remained stable from the four-cluster to seven-cluster solution (7C7), the other three clusters from the four-cluster solution were split into two separate clusters in the seven-cluster solution. 4C4 in the four-cluster solution was split into two clusters with activation predominantly in sensory/motor cortices (7C2) and auditory cortices (7C5). Consistent with previous functional descriptions of these areas, the BrainMap categories associated with 7C2 included ‘action execution’, flexion/extension and tactile stimulation, and passive listening and pitching monitoring/discrimination for 7C2.
4C3 in the four-cluster solution was split in the seven-cluster solution into two clusters with activation predominantly in the default-mode network (7C4) and sub-cortical structures (basal ganglia and amygdala; 7C6). Consistent with previous functional descriptions of these areas, the categories associated with 7C4 included ‘theory of mind’, ‘social cognition’, and autobiographical recall, and affective stimuli and long-term memory for 7C6.
4C1 in the four-cluster solution was split in the seven-cluster solution into visual (7C1) and parietal-frontal clusters (7C3). Activation in 7C1 was predominantly observed in the visual cortex, with slightly weaker activation in the dorsal-parietal and frontal cortices. Activation in 7C3 was predominantly observed dorsal-parietal and frontal cortices, with slightly weaker activation in the visual cortex. Categories associated with 7C1 included passive viewing paradigms, including ‘action observation’, and visual shape perception, and perception of moving objects. Task-descriptive categories associated with 7C3 included active viewing paradigms, including ‘mental rotation’, ‘spatial cognition’, and delayed match to sample tasks.
An important observation apparent in the clustering solution is that BrainMap categories ‘distant’ in semantic space (farther apart in the word cloud) load strongly onto single activation states. In other words, categories that are semantically ‘distant’, are often not so ‘distant’ in the latent BOLD activation space. For example, consider the following two task paradigms: the Stroop color word task, and an overt word generation task. The psychological processes generally inferred to underlie these tasks include action inhibition 20 for the Stroop color word task, and lexical processing 21 for a word generation task. However, the whole-brain activation patterns of these paradigms are similar, as they both load strongly onto the same latent whole-brain activation state: 4C4 in the four-cluster solution and 7C7 in the seven-cluster solution. This suggests the neural response measured by task-fMRI is similar for the Stroop color word task and word generation task. Thus, whole-brain activation states may be informative in terms of mapping the similarity in functional anatomy between different cognitive or behavioral processes.
Whole-Brain Activation State and BrainMap Category Comparison
To compare the explained variance from these latent whole-brain activation states to the existing ontological categories provided by the BrainMap database, we conducted an MDMR analysis regressing the weighted membership coefficients of the four- and seven-cluster separately onto the activation map * activation map dissimilarity matrix. The total explained variance for the four-cluster solution and seven-cluster solution were 20.95% (p = 0.001) and 34.37% (p = 0.001), respectively. Thus, both lower-order latent activation states (N =4 and N = 7) explained a greater amount of variance in task-activation map dissimilarity than the full model of existing ontological categories (N = 603). This difference in explained variance can be visualized with a re-ordered activation map*activation map similarity matrix according to either the latent activation states or two example categorical domains (Behavioral Domain and Paradigm Class; Figure 5).
Discussion
A driving goal of cognitive neuroscience is to generate and constrain theories and models of cognitive processes using functional neuroimaging data. Directed at this goal, the primary aim of this study was to identify the dominant cognitive dimensions upon which whole-brain blood-oxygen-level dependent (BOLD) activation maps vary. In other words, what dimensions of the task-fMRI experiment is the BOLD response most attuned do? Task-based fMRI research has conventionally used a fluid set of ontological categories to interpret the differences between activation maps. Using a novel analysis pipeline applied to the BrainMap database, we quantify the variance explained by these existing ontological categories, describing various features of the task-fMRI experiment. The full model of existing ontological categories (N = 603 predictors) explained a moderate proportion of variance (R2= 0.197) in the similarities between activation maps. We find that ‘stimulus type’, ‘paradigm class’ and ‘behavioral domain’ uniquely explain a small (R2< 0.05), but non-trivial amount of variance in similarity between activation maps.
It is difficult to assess these explained variance estimates in terms of absolute value, as the disparate quality of the activation maps (derived from peak-activation coordinates) place an unknown upper limit on the variance explainable by the task-descriptive categories. However, we can make relative comparisons. We find that the observable domain of ‘paradigm class’ (or task paradigm), explains the greatest amount of variance among the task-descriptive categories, controlling for other domains, (R2 = 0.0417). This suggests that differences in task experiments drive the most observed differences in fMRI task-activation patterns over and above differences in other domains (e.g. behavioral domain). Thus, researchers should be aware that the choice of experimental task may be of more crucial importance than whether they are thought to be eliciting the same purported cognitive or behavioral process. Surprisingly, a substantial portion of variance in the similarity between activation maps can be explained by the study ID (R2 = 0.447); over two times greater than the full set full model of task-descriptive categories (R2 = 0.197). This can be for multiple reasons: 1) differences in peak-coordinate reporting, 2) differences in pre-processing pipelines and scanner hardware, or 3) differences in significance thresholding procedures 22. These differences in study ID may be reduced with the sharing of unthreshold statistical maps 23, that avoids peak-coordinate reporting and thresholding procedures 19.
We assume that the quality of latent cognitive or behavioral descriptors for a task-fMRI ontology is contingent upon how well these descriptors explain differences in BOLD activation patterns. Thus, we attempt to derive a set of latent activation states and associated cognitive and behavioral descriptors that maximize the explained variance between BOLD activation patterns. With just a set of four (or seven) latent activation states with sensible cognitive interpretations, we can explain a greater amount of variance in BOLD activation maps than the full set of ontological categories of the BrainMap database (N = 603). One might argue that this is trivial: one would expect that a solution built to maximize explained variance between activation patterns will of course explain more variance than the a priori ontological categories. However, we do not believe this is trivial for two reasons: 1) though it is trivial fact that all clustering algorithms, at a general level, estimate clusters that maximize explained variance, it is not trivial that a low-dimensional clustering solution (N= 4 or N = 7) can explain greater variance than the high-dimensional task-descriptive categories (N = 603). Relatedly, 2) explained variance of activation map dissimilarities should be our criterion for what constitutes an adequate ontology of cognitive-neural mappings. For example, if a four-cluster solution explained 99% of the variance in task-activation map dissimilarity, adding more ontological categories amounts to an unnecessary overparameterization of this space.
The BrainMap categories associated with the latent activation states from the 4-cluster and 7-cluster map onto plausible neurocognitive systems. The four latent activation states can be generally described as object-viewing (4C1), higher-order cognition (inhibition/control/language) (4C2), self-memory/affective/social (4C3), and auditory-stimulus/motor-action (4C4) states. The seven latent activation states can be generally described as passive object-viewing (7C1), sensory-motor (7C2), active object-viewing (7C3), social/theory of mind (7C4), auditory (7C5), affective (7C6) and higher-order cognition (7C7). Examination of the four-cluster and seven-cluster solutions centroid activation patterns revealed a set of regions commonly activated across the BrainMap database. Consistent activation was observed across regions that make up the traditionally described ‘task-positive’ and ‘task-negative’ activation/de-activation pattern 19,24,25. The ‘task-positive’ activation pattern traditionally includes elements of the fronto-parietal (DLPFC, SPC and DMPFC) and salience networks (dorsal anterior cingulate cortex and AI). Consistency of activation was particularly predominant in the AI across all centroid activation maps, in accord with previous studies demonstrating the ubiquity of AI activations across task paradigms 26. The ‘task-negative’ de-activation pattern, generally restricted to the DMN 25, was represented in both the four- and seven-cluster solution. Consistent with the ‘task-negative’ ascription, the activation maps of these clusters were more likely to be reported as ‘de-activations’.
Limitations
One objection to the current approach is that it is circular in some sense. According to this objection, the use of conventional task-descriptive categories of cognitive functioning and task paradigms for interpretation in the behavioral decoding analysis essentially reifies these categories in our neural-driven categorization. Thus, the analysis relies on a circular inference. In response, we agree that the mix of ordinary and technical terms codified in the BrainMap experimental descriptors is relied upon in our descriptions of the data-driven categories, but reliance on this terminology does not constitute circularity. The behavioral decoding analysis maps clusters to their most representative experimental descriptors, but does not identify the clusters with those experimental descriptors. For example, descriptions of complex physics topics often rely on ordinary linguistic terminology, but this description does not then identify the physical systems themselves with these ordinary terms. In the same manner, describing the data-derived categories in terms of their association with ordinary or current technical terms of cognitive science does not constitute circularity. By examining the pattern of experimental descriptors associated with each cluster, we may find pointers to an underlying neural process that unifies these descriptors.
Due to the disparate representation of activation maps in the BrainMap database in terms of peak-activation coordinates, the precision of the distinctions between possible clusters is limited. However, the substantial sample size of reported task-fMRI experiments provided by the BrainMap database is unmatched in terms of power and detailed experimental metadata. The increasing size of databases containing unthresholded activation maps, such as NeuroVault 23, offers the potential for a more precise or fine-grained categorization using the approach applied in this study. In addition, the task-activation maps in the BrainMap database were computed using a variety of analytic approaches and experimental designs, which introduces an extra source of variability among the activation maps that cannot be accounted for in our analysis. Nevertheless, we hope the results presented here provide a starting point for a well-developed ontology of cognitive-neural mappings.
Acknowledgements
Peter Fox/Brainmap. This work was supported by award R01MH107549 from the National Institute of Mental Health to LQU and award 1631325 from the National Science Foundation to ARL.