Gene expression profiling of tumours is an important source of information for cancer patient stratification. Detecting subtle alterations of gene expression remains a challenge, however. Here, we propose a novel tool for high-sensitivity detection of differential pathway activity in tumours. For a pathway defined by a collection of genes, the samples are projected onto a low-dimensional manifold in the subspace spanned by those genes. For each sample, a score is next found by calculating the distance between each projected sample and the projection of a subgroup of reference samples. Depending on the aim of the analysis and the available data, the reference samples may represent e.g. normal tissue or tumour samples with a particular genotype or phenotype. The proposed tool, PathTracer, is demonstrated on gene expression data from 1952 invasive breast cancer samples, 10 DCIS, 9 benign samples and 144 tumour adjacent normal breast tissue samples. PathTracer scores are shown to predict survival, clinical subtypes, cellular proliferation and genomic instability. Furthermore, predictions are shown to outperform those obtained with other comparable methods.