Many breast cancer patients are treated with chemotherapy before surgery for the removal of the tumour, which is known as neoadjuvant chemotherapy (NAT). It improves outcome for many patients, and their response to the treatment is prognostic for their overall outcome. However, not all patients who receive NAT respond to the treatment, and, as a result, many suffer unnecessarily from side effects and delays to surgery. In this thesis, I designed and evaluated two potentially independent and complementary strategies for predicting an individual patient’s response to treatment based on gene expression and computational pathology. Firstly, I attempted to develop a clinically practical approach for classifying breast tumours using RNA from routine formalin fixed paraffin embedded (FFPE) histopathological samples. Breast cancers can be classified into distinct groups, 'integrative clusters' (IntClusts), with different outcomes and potentially different response to NAT. The published classification is based on RNA expression from fresh frozen tissue, which is impractical in a clinical setting. Initially, I attempted to build an accurate classifier for IntClusts based on RNA expression from widely available FFPE tissue, using a user-friendly NanoString technique. Unfortunately, it was not possible to achieve reliable classification with this method using the gene probe set pre-selected based on fresh frozen tissue. Next, I sought to identify the genes whose expression was not affected by the fixative process by comparing paired FFPE and fresh frozen tissue with RNA sequencing, a method that allows the quantification of all genes. There was poor agreement in measured gene expression between the two types of tissue sample, FFPE versus frozen, and between the assessment methods on the same tissue type, RNA-sequencing versus Illumina microarray, resulting in unreliable classification of tumours into integrative clusters. These findings represent a challenge to the adoption of the integrative clusters in real-world precision medicine. Secondly, I developed quantitative computational methods for the assessment of digitized H&E slides, which are routinely produced clinically. I developed and validated two machine learning methods for cell classification, where cells on an image can be automatically detected and identified as tumour cells, lymphocytes, or stromal cells. I also show that my method can be effectively generalised to immunohistochemistry slides. In a novel dataset, using a new method, I replicate the previous finding that the presence of immune infiltrate in pre-treatment biopsies is predictive of NAT response. I then used this cell classification to demonstrate that the spatial profiles of tumour clusters and their relationship to immune cells are associated with treatment outcome. Specifically, I found that larger tumour clusters, more heterogeneous tumour clusters, and more lymphocytes in the region immediately bordering tumour clusters are all correlated with pathological complete response to NAT. The spatial features of the peri-tumour region were more predictive than the features of the tumour itself, suggesting a particularly important role for the interface between tumour clusters and their immune microenvironment. To conclude, I review the large number of further studies spawned by the work I present in this thesis, and explain how they might improve our ability to understand tumour biology, and translate this understanding to the clinical setting. In summary, I have explored the real-world applicability of gene expression profiling and computational pathology methods in widely available clinical samples. These approaches have the potential for translation into adjuncts to existing stratification methods, offering patients better care. This thesis provides a step towards the translation of the molecular classification of breast cancer, and computational methods for pathology image analysis, into real-world precision medicine, predicting response to neoadjuvant chemotherapy.