Deconvolving Complex Neuronal Networks into Interpretable Task-Specific Connectomes

Yifan Wang, Ananth Grama, Vikram Ravindra

Keywords: functional connectomics, supervised dimension reduction, supervised nonnegative matrix factorization

Abstract: Neuronal responses associated with complex tasks are superpositions of several elementary physiological and functional responses. Important challenges in this context relate to identification of elementary responses (also known as basic functional neuronal networks), combinations of responses for given tasks, and their use in task and efficacy prediction, and physiological characterization. Task-specific functional MRI (fMRI) images provide excellent datasets for studying the neuronal basis of cognitive processes. In this work, we focus on the problem of deconvolving task-specific aggregate neuronal networks into elementary networks, to use these networks for functional characterization, and to ``explain'' these networks by mapping them to underlying physiological regions of the brain. This task poses a number of challenges due to very high dimensionality, small number of samples, acquisition variability, and noise. We propose a deconvolution method based on supervised non-negative matrix factorization (SupNMF) that identifies elementary networks as factors of a suitably constructed matrix. We show the following important results: (i) SupNMF reveals cognitive "building blocks" of task connectomes that are physiologically interpretable; (ii) SupNMF factors can be used to predict tasks with high accuracy; and (iii) SupNMF outperforms other supervised factoring techniques both in terms of prediction accuracy and interpretability. More broadly, our framework provides important insights into the physiological underpinnings of brain function and individual performance.

Submission Number: 13237