Transcript-level quantification is often measured across two groups of patients to aid the discovery of biomarkers and detection of biological mechanisms involving these biomarkers. Statistical tests lack power and false discovery rate is high when sample size is small. Yet, many experiments have very few samples (≤ 5). This creates the impetus for a method to discover biomarkers and mechanisms under very small sample sizes. We present a powerful method, ESSNet, that is able to identify subnetworks consistently across independent datasets of the same disease phenotypes even under very small sample sizes. The key idea of ESSNet is to fragment large pathways into smaller subnetworks and compute a statistic that discriminates the subnetworks in two phenotypes. We do not greedily select genes to be included based on differential expression but rely on gene-expression-level ranking within a phenotype, which is shown to be stable even under extremely small sample sizes. We test our subnetworks on null distributions obtained by array rotation; this preserves the gene-gene correlation structure and is suitable for datasets with small sample size allowing us to consistently predict relevant subnetworks even when sample size is small. For most other methods, this consistency drops to less than 10% when we test them on datasets with only two samples from each phenotype, whereas ESSNet is able to achieve an average consistency of 58% (72% when we consider genes within the subnetworks) and continues to be superior when sample size is large. We further show that the subnetworks identified by ESSNet are highly correlated to many references in the biological literature. ESSNet and supplementary material are available at: http://compbio.ddns.comp.nus.edu.sg:8080/essnet .