Do ontological concepts stand up to a meta-analysis of neuroimaging data?

Background. Large scale sharing of cognitive neuroscience data and results requires common semantic frameworks for describing the experimental methods, as well as the data analyses and results.  We have developed and released version 1 of the Cognitive Paradigm Ontology (CogPO, www.cogpo.org), which is an ontological form of part of the BrainMap schema (www.brainmap.org) for describing neuroimaging experiments.  The key components of CogPO include the representation of experimental conditions, focused on the stimuli presented, the instructions given, and the responses requested.  The aim of the ontology is to focus on experimental methods to identify when patterns in neuroimaging results cross pre-defined cognitive domains. We used the ontological categorizations in conjunction with meta-analysis of the activation patterns archived in the BrainMap repository to identify where the ontology provided meaningful categorizations of neuroimaging data, and to highlight concepts and relationships that need refining.

Methods. Peak coordinates from the published experiments in the BrainMap repository were smoothed using a Gaussian distribution (FWHM=12 mm) to accommodate spatial uncertainty; these smoothed maps provided experiment images as the initial data.  Independent component analyses was applied to this 4D data (space x experiment-ID) using MELODIC (Beckmann et al., 2005) in FSL (Smith et al., 2004; Woolrich et al., 2009) to decompose the experiment images into spatially independent components, which represent the major modes of co-activation across the experiments represented in the BrainMap database. Along with each spatial map, a corresponding experiment-ID vector was generated that describes how strongly a given component relates to each of the original 8,637 experiment images. Each of the original experiment images were annotated with CogPO terms; the strength of the relationships between different CogPO concepts and the spatial maps can be explored through clustering analyses on the weighting coefficients across the experiment images.

Results. Different stimulus, response, and instruction categories showed some expected relationships to certain patterns. For example, tactile stimuli and fine motor responses were strongly weighted on the spatial components that included subcortical regions (e.g., basal ganglia and thalamus) and primary motor cortex. Visual stimuli were linked to the primary visual areas, as well as inferior frontal networks. However, other stimulus or response types were related to multiple components, and some components were equally strongly related to a large cluster of seemingly unrelated stimulus or response types.  Stimulus x response and stimulus x instruction combinations were also explored.

Conclusions. Our findings indicate that the distinctions in some branches of the ontology are not as important as others for mapping cognitive constructs onto neurophysiological circuits. Our results serve as a resource for functional interpretations of brain networks in resting state studies and future investigations into mental operations and the tasks that drive them.