BackgroundGenomic mutations drive the pathogenesis of myelodysplastic syndromes (MDS) and acute myeloid leukemia (AML). While morphological and clinical features, complemented by cytogenetics, have dominated the classical criteria for diagnosis and classi cation, incorporation of molecular mutational data can illuminate functional pathobiology.
MethodsWe combined cytogenetic and molecular features from a multicenter cohort of 3588 MDS, MDS/ myeloproliferative neoplasm (including chronic myelomonocytic leukemia [CMML]), and secondary AML patients to generate a molecular-based scheme using machine learning methods and then externally validated the model on 412 patients. Molecular signatures driving each cluster were identi ed and used for genomic subclassi cation.
FindingsUnsupervised analyses identi ed 14 distinctive and clinically heterogenous molecular clusters (MCs) with unique pathobiological associations, treatment responses, and prognosis. Normal karyotype (NK) was enriched in MC2, MC4, MC6, MC9, MC10, and MC12 with different distributions of TET2, SF3B1, ASXL1, DNMT3A, and RAS mutations. Complex karyotype and trisomy 8 were enriched in MC13 and MC1, respectively. We then identi ed ve risk groups to re ect the biological differences between clusters. Our clustering model was able to highlight the signi cant survival differences among patients assigned to the similar IPSS-R risk group but with heterogenous molecular con gurations. Different response rates to hypomethylating agents (e.g., MC9 and MC13 [OR: 2.2 and 0.6, respectively]) re ected the biological differences between the clusters. Interestingly, our clusters continued to show survival differences regardless of the bone marrow blast percentage.
InterpretationDespite the complexity of the molecular alterations in myeloid neoplasia, our model recognized functional objective clusters, irrespective of anamnestic clinico-morphological features, that re ected disease evolution and informed classi cation, prognostication, and molecular interactions. Our subclassi cation model is available via a web-based open-access resource as well (https://drmz.shinyapps.io/mds_latent).