Quantitative biomarkers of depression are critical for development of rational therapeutics, but limitations of current low-resolution, indirect brain assays may impede their discovery. We applied graph theory and machine learning to a large unique dataset of intracranial electrophysiological recordings to generate a four-dimensional whole-brain model of neural activity. Using this model, we found patterns of network activity that correctly classified depression in over 80% of individuals. These complex patterns were especially evident in alpha and beta spectral power across frontal and occipital brain regions, respectively. Our findings reveal a widespread network of abnormal activity that may inform advanced personalized treatment.