Influence of Co-Mutational Patterns in Disease Phenotype and Clinical Outcomes of Chronic Myelomonocytic Leukemia

Abstract

Introduction: Chronic myelomonocytic leukemia (CMML) is characterized by recurrent somatic mutations in a conserved number of genes and functional pathways. Although specific genomic events are associated with disease phenotype and prognosis, identification of genomically-defined disease subgroups based on common recurrent cooperative genomic patterns is needed to refine disease classification and identify disease subsets with unique clinical behavior. Methods: To evaluate if CMML is defined by recurrent co-mutational profiles defining unique clinicopathologic subgroups we evaluated a cohort of 296 patients (pts) with newly diagnosed CMML. Whole bone marrow DNA was subject to 81 gene targeted next-generation sequencing analysis. Akaike information criterion (AIC) method was applied to evaluate mixture model components defining clusters based on mutational profile. Expectation-maximization (EM) algorithm was used for mixture model fitting to multivariate binary data. Unsupervised cluster membership was chosen as a function of the component with the highest weight for any individual, and we examined means of cluster centers. Cox proportional hazards models, regressing overall survival (OS) and leukemia-free survival (LFS), were fit on cluster in some cases and normalized cluster weights in other cases, with and without adjustment for age and sex. Two degree of freedom likelihood ratio tests was performed on cluster and cluster weight in all models regardless of covariate adjustment. Results: Median age of the population was 70 years [range 30-71]; 160 (54%) pts had myeloproliferative CMML (MP-CMML) and 152 (51%) had CMML-2 by WHO classification. Most frequent mutations included TET2 (51%), ASXL1 (48%), SRSF2 (40%), RUNX1 (22%), NRAS (18%), CBL (14%) and KRAS (14%) all present in >10% of pts. Based on frequency and co-occurrence of mutations, AIC optimization identified 3 unique clusters. Clusters based on mutation landscape are shown in Figure 1A. Unsupervised clustering allowed attribution of individual patients to specific clusters. Cluster 1 (C1, n=107), was characterized by near universal presence of ASXL1 mutations (86%), with lower frequency of TET2 mutations, and enrichment of RUNX1 (35%), NRAS (28%) and U2AF1 (18%) mutations. SETBP1 and EZH2 mutations were nearly exclusively observed in C1 (16% and 20% of C1 pts, respectively). All pts with ETV6 mutations belonged to this cluster. Cluster 2 (C2, n=135) was defined by universal presence of TET2 mutations (93%) with frequent SRSF2 co-mutation (56%) and lower frequency of ASXL1 (34%), RUNX1 (15%) or NRAS (12%) mutations. Despite their overall low frequency, BRAF mutations were restricted to C2, while CBL mutations were observed at similar frequencies in C1 and C2. This suggests that co-mutation profiles and genomic context differs among genes involved in RAS pathway signaling. Cluster 3 (C3, n=54) included a minor and more genomically heterogenous group enriched for SF3B1 (25%), KRAS (24%), DNMT3A (22%) and TP53 (15%) mutations, with absence of ASXL1 mutations, and low frequency of TET2 (12%) mutations. All BCORL1 and WT1 mutations and most ASXL2 and CEBPA mutations were in C3. RUNX1 mutations not only were more frequent among C1 but appeared at higher median VAF compared to other clusters (C1: 40%, C2: 11%, C3: 25%, p=0.003). Mutations in U2AF1 had higher median VAF in C3 compared to cluster 1 (49% vs 43%, p=0.017). C1 was enriched for MP-CMML (C1: 64%; C2: 48%; C3: 52%, p<0.001) and had higher frequency of del(7q)/-7 (C1: 11%; C2 0%, C3: 6%, p<0.001) and trisomy 8 (C1: 15%, C2: 7%, C3: 6%, p=0.05). Complex karyotype (C1: 1%, C2: 1%, C3: 22%, p<0.001) was nearly exclusive of C3, while C2 was more likely to be associated with a normal karyotype (C1: 59%; C1: 73%; C3: 41%, p<0.001) and higher median hemoglobin (C1: 9.8g/dL; C2: 12.1g/dL, C3: 9.7g/dL, p<0.001). In total 66% of pts received therapy with hypomethylating agents with no differences in response, median time to or duration of response between clusters. Use of Cox proportional hazards model identified that, when corrected by age and sex, patients on C1 and C3, based on attributed cluster weights, had significantly worse LFS (p=0.02, Figurer 1B) and OS (p=0.05) than those in C2. Conclusions: These data suggest that somatic mutations in CMML have unique patterns of clustering that define disease phenotype and influence disease outcomes.