Parallel workflows for data-driven structural equation modeling in functional neuroimaging

We present a computational framework suitable for a data-driven approach to structural equation modeling (SEM) and describe several workflows for modeling functional magnetic resonance imaging (fMRI) data within this framework. The Computational Neuroscience Applications Research Infrastructure (CNARI) employs a high-level scripting language called Swift, which is capable of spawning hundreds of thousands of simultaneous R processes (R Development Core Team, 2008), consisting of self-contained SEMs, on a high performance computing system (HPC). These self-contained R processing jobs are data objects generated by OpenMx, a plug-in for R, which can generate a single model object containing the matrices and algebraic information necessary to estimate parameters of the model. With such an infrastructure in place a structural modeler may begin to investigate exhaustive searches of the model space. Specifi c applications of the infrastructure, statistics related to model fit, and limitations are discussed in relation to exhaustive SEM. In particular, we discuss how workfl ow management techniques can help to solve large computational problems in neuroimaging.