Acute leukemia is the most common cancer of childhood and one of the most deadly due to relapsed disease. Predicting patients at high risk of relapse is an imperfect science and currently combines clinical features, somatic alterations and early response to therapy. Yet, despite this, for the 20% of children who will relapse, at least half will die of their leukemia. Using single-cell, high-parameter analysis of primary leukemia specimens from diagnosis, we demonstrate the ability to more accurately predict patients at high risk for relapse utilizing novel developmental classification and machine learning tools.  This approach also enables identification the cells that are responsible for relapse, thereby providing insight into potential therapeutic strategies.