Motivation: Acute Myeloid Leukemia is a highly heterogeneous disease. Although current classifications are well-known and widely adopted, many patients experience drug resistance and disease relapse. New biomarkers are needed to make classifications more reliable and propose personalized treatment. Results: We performed tests on a large scale in 3 AML cohorts, 1112 RNAseq samples. The accuracy to distinguish NPM1 mutant and non-mutant patients using machine learning models achieved more than 95% in three different scenarios. Using our approach, we found already described genes associated with NPM1 mutations and new genes to be investigated. Furthermore, we provide a new view to search for signatures/biomarkers and explore diagnosis/prognosis, at the k-mer level. Availability: Code available at https://github.com/railorena/npm1aml and https://osf.io/4s9tc/. The cohorts used in this article were authorized for use.