Textural features for pathway-level representation of omics data in biological networks

Abstract

More than 50 years ago, Haralick and co-authors proposed a family of gray-level co-occurrence statistics that became known as textural features. These features are widely used in image analysis, but their application to biological networks has remained limited because cellular networks are sparse, irregular graphs rather than regular pixel grids. This work presents a network-adapted version of Haralick texture analysis for generating pathway-level features from gene-level omics profiles. The resulting profiles reduce dimensionality and can be used as candidate predictors of anti-cancer drug response. Performance of these features is compared with original gene expression variables and with pathway features from network enrichment analysis (NEA), whose robustness has been demonstrated previously. Although technically simpler than NEA, Haralick features showed comparable sensitivity. More importantly, selected Haralick features were preserved between in vitro drug screens and clinical treatment-associated survival analyses, supporting their potential use for prioritizing robust pathway-level drug-response correlates.