Machine Learning Allows the Identification of New Co-Mutational Patterns with Prognostic Implications in NPM1 Mutated AML - Results of the European Harmony Alliance

Background: Mutations of NPM1 are generally considered as a favorable prognostic marker in acute myeloid leukemia (AML), although patients present with distinct co-mutations that might influence outcomes. The recently updated European LeukemiaNet classification (ELN2022) does now group NPM1mut AML patients with FLT3-ITD (regardless of the allelic ratio) as intermediate prognosis, while NPM1mut without FLT3-ITD is part of the favorable group. However, the mutational landscape of NPM1mut AML is more complex and analysis of a large cohort of patients might allow further dissection to identify less frequent co-mutational patterns of prognostic and predictive importance. Aims: To identify clinically significant co-mutational patterns in NPM1mut AML to further refine risk stratification models for this subset of patients. Methods: From the HARMONY Alliance database, a total of 1093 intensively treated NPM1mut patients were selected. Patients treated with targeted therapies (e.g. FLT3 or IDH1/2 inhibitors) were not included and NPM1 MRD information was not available at this stage of the analysis. A machine learning algorithm was developed in order to identify combinations of up to 4 co-mutated genes with potential impact on overall survival (OS). A heuristic search algorithm was implemented and bootstrap sampling was applied to estimate the impact of all possible gene combinations on OS. In addition, a global dashboard was developed where it was possible to compare mutational combinations with Kaplan Meier and Cox regression models. Finally, clinically significant co-mutational patterns were summarized in a novel risk stratification model for NPM1mut AML, which was validated using a publicly available external patient cohort (Awada et al, Blood, 2021). Results: The study population of 1093 NPM1mut AML patients included 57% females and median age was 53 years. Regarding ELN2022 classification, 57% of patients were classified into the favorable, 42% into intermediate and only 1% into adverse risk groups. The most frequent co-mutations were DNMT3A (54%) and FLT3-ITD (42%), followed by NRAS, FLT3-TKD and TET2 (20% each). Mutations on IDH1 (13%) and IDH2 (15%) showed a similar behavior for all the analyses performed, so for simplification purposes we will refer to IDHmut when any of them was mutated and to IDHwt when both were wildtype. The triple combination of NPM1mut + FLT3-ITD + DNMT3Amut identified a subgroup of patients with adverse prognosis (2-year OS of 33%), similar to patients with TP53mut. Of note, not all FLT3-ITD patients carried an intermediate or adverse prognosis, as we were able to identify a subgroup (FLT3-ITD + IDHmut + DNMT3Awt) with excellent prognosis (2-year OS of 80%), which represented 4% of NPM1mut AML. However, in the absence of FLT3-ITD and the presence of DNMT3Amut, the addition of IDHmut decreased OS towards the intermediate risk (2-year OS 59%). Notably, not all DNMT3Amut patients carried an intermediate or adverse prognosis. In the absence of FLT3-ITD and with IDHwt, mutations on either NRAS, KRAS, PTPN11 or RAD21 revealed a subgroup with favorable prognosis even when DNMT3A was mutated (2-year OS 80%) and this group represented 11% of NPM1mut AML. This information is summarized in a 4-category risk stratification model (figure 1). The revised NPM1mut favorable group presented with a 3-year OS of 78%, which was 63%, 48% and 29% for intermediate-1, intermediate-2 and adverse risk groups respectively (p<0.001). Regarding 3-year relapse free survival (RFS), it was 71%, 59%, 39% and 26% accordingly (p<0.001). In the validation cohort, 3-year OS of was 73% for NPM1mut favorable group, being 59%, 40% and 22% for intermediate-1, intermediate-2 and adverse groups respectively (p<0.001). Multivariate OS analysis in NPM1mut AML identified the following independent prognostic factors: revised NPM1mut model (taking favorable group as reference, HR 1.77 for intermediate-1, HR 2.92 for intermediate-2 and HR 5.13 for adverse group, p<0.001); secondary or therapy-related AML (HR 1.93, p<0.001), age >60 years (HR 1.84, p<0.001), WBC at diagnosis >100x103/uL (HR 1.59, p<0.001), and male gender (HR 1.21, p=0.04). Conclusion/summary: Analysis of large NPM1mut AML cohorts allows the identification of clinically significant co-mutational patterns. We propose a new genetic stratification model for NPM1mut AML that identifies 4 groups with different OS and RFS. Figure 1View largeDownload PPTFigure 1View largeDownload PPT Close modal