Classifying whole-slide images of prostate cancer resections to derive an accurate prognosis for tumor progression is a highly challenging problem. We here introduce a novel type of high-level features which operate on the gland level for the application in computer-aided prognosis using a Tissue Phenomics approach. Since tissue architecture, and in particular, the gland distribution possibly provide information on the aggressiveness of the individual tumor, our features exploit the spatial relationship of different gland types. Glands are classified into cancerous and healthy glands, and also, into morphological classes based on size and shape. Co-occurrences of the classified glands are quantified and Haralick-like features are computed based on the derived gland co-occurrence matrices. The resulting gland co-occurrence features are mined to automatically determine the best parametrization. In experiments on whole-slide images it turned out that our novel features allow accurate stratification of patients into the prognostic groups tumor progression and nonprogression outperforming clinical features. Our results indicate a strong correlation of tumor progression with invasion phenotypes.