Characterization of breast parenchyma on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is a challenging task owing to the complexity of underlying tissue structures. Current quantitative approaches, including radiomics and deep learning models, do not explicitly capture the complex and subtle parenchymal structures, such as fibroglandular tissue. In this paper, we propose a novel method to direct a neural network's attention to a dedicated set of voxels surrounding biologically relevant tissue structures. By extracting multi-dimensional topological structures with high saliency, we build a topology-derived biomarker, TopoTxR. We demonstrate the efficacy of TopoTxR in predicting response to neoadjuvant chemotherapy in breast cancer. Our qualitative and quantitative results suggest differential topological behavior of breast tissue on treatment-na\"ive imaging, in patients who respond favorably to therapy versus those who do not.