Prediction of myeloid malignant cells in Fanconi anemia using machine learning

Fanconi anemia (FA) is an inherited bone marrow failure syndrome with cancer predisposition. Most FA patients develop aplastic anemia during childhood and have an extremely high cumulative risk to develop cancer during their lifespan. Myeloid malignancy is one of the main neoplastic risks for patients with FA, including high-risk myelodysplastic syndrome (MDS), recently renamed as myelodysplastic neoplasm, and acute myeloid leukemia (AML). Although bone marrow transplantation is the treatment of choice for FA patients that develop aplastic anemia, patients with a more stable bone marrow remain not transplanted and at a high risk of presenting MDS/AML, these patients therefore should be monitored for appearance of myeloid malignant clones. Markers for an as-early-as-possible identification of emerging myeloid malignant cells are needed for the monitoring of patients with FA, since quick medical action after detection of neoplastic transformation is needed. In this work we have developed a deep neural network (DNN) model that was trained with publicly available single cell RNA-seq (scRNA-seq) datasets of patients with AML and used to predict the presence of AML-like cells in scRNA-seq datasets obtained from bone marrow samples of patients with FA. The predictor displayed high sensitivity, specificity, and accuracy for the detection of single-cell resolution myeloid malignant transcriptional profiles. Functional analyses of the predicted-AML cells from FA patients showed enrichment of lympho-myeloid-primed progenitor (LMPP) and granulocyte-monocyte progenitor (GMP) populations, as well as transcriptional profiles associated with malignant transformation. Cues of immune evasion were also detected using single cell pathway analysis (SCPA) and cell-cell communication profiles.