CMML2AML: Machine-learning discovery of co-mutations predictive of blast transformation in chronic myelomonocytic leukemia

Abstract Background Blast transformation (BT) occurs in approximately 15-30% of patients with chronic myelomonocytic leukemia (CMML) and remains a leading cause of death. Allogeneic stem cell transplantation (ASCT) is currently the only treatment modality with the potential to cure the disease or prolong survival. Optimal timing of ASCT is critical for maximizing benefit while minimizing risks. To that end, contemporary risk models have focused on the prognostic relevance of individual, as opposed to concurrent mutations. In the current study, we looked into the possibility of prognostic prominence from concurrent mutations in predicting BT in CMML. Methods CMML diagnostic criteria were according to the International Consensus Classification (Arber et al. Blood 2022; 140:1200). A machine-learning hierarchical clustering algorithm was developed and tailored for patient stratification using survival outcomes and co-expression of genomic alterations. To reduce complexity and improve interpretability, we generalized each cluster using existing rule induction algorithms, including the in Trees framework and JRip. The final output was a set of mutation-based cluster definitions, each representing a distinct patient subgroup with similar survival trajectories. Competing risk analysis and cumulative incidence functions were used in downstream evaluation to validate the clinical distinctiveness of the clusters. For all survival analysis, patients were censored at the time of ASCT. Time-to-BT was calculated from the date of diagnosis to the date of BT or ASCT or last contact. BT-free-survival was calculated from the date of diagnosis to the date of BT, ASCT, death, or last contact. Results The core study cohort included 605 patients from the Mayo Clinic, USA, and the external validation cohort 501 patients from Humanitas Cancer Center, Milan, Italy. Using the patient cohort from the Mayo Clinic, machine-learning algorithms identified five molecular clusters with 3-year blast transformation (BT) rates ranging from 0% to 100% (AUC at 3 years 0.78): the order of molecular signature assignment (probability of BT/death from another cause) was i) PHF6MUT/ASXL1WT (0%/17% at 3 years; N=32), ii) NPM1MUT OR BCORMUT/ASXL1MUT OR SETBP1MUT/NRASMUT (48%/30% at 1 year; N=24), iii) RUNX1MUT/ASXL1MUT OR SRSF2MUT/NRASMUT OR EZH2MUT/ASXL1MUT OR SETBP1MUT OR BCORMUT (31%/55% at 3 years; N=132), iv) ASXL1MUT/TET2MUT OR DNMT3AMUT OR JAK2MUT (24%/28% at 3 years; N=123), and v) all other permutations (10%/40% at 3 years; N=294) [Figure 1]. Additional analysis confirmed significant differences in survival between RUNX1MUT/ASXL1MUT vs. RUNX1MUT/ASXL1WT (p<0.01) OR RUNX1WT/ASXL1WT (p<0.01) OR RUNX1WT/ASXL1MUT (p=0.038) [Figure 2]. Similar patterns of differences in survival were also documented for NRAS/SETBP1, ASXL1/EZH2, and NRAS/SRSF2 mutation combinations (Figure 2). A subsequent Cox regression analysis confirmed independent prognostic contributions from “PHF6MUT/ASXL1WT” (HR 5.43e-10; p<0.01), NPM1MUT (HR 26.7; p<0.01), “SETBP1MUT/NRASMUT” (HR 12.7; p<0.01), BCORMUT(HR 5.8; p<0.01), “RUNX1MUT/ASXL1MUT” (HR 2.3, p<0.01), JAK2MUT (HR 2.1; p<0.01), and “ASXL1MUT/TET2MUT” (HR 1.7; p=0.02). The prognostic relevance of “SETBP1MUT/NRASMUT”, “RUNX1MUT/ASXL1MUT”, NPM1MUT, and BCORMUT was validated in the external cohort from Italy (N=501). In the Mayo Clinic cohort, presence of any of the latter mutations was associated with 1-, 3-, and 5-year BT (death from other cause) rates of 27% (21%), 44% (49%), and 44% (51%), respectively (Figure 3). The corresponding values in the absence of high risk mutations were 7% (15%), 15% (37%), and 18% (52%) [Gray's p value <0.01 for both BT and death from another cause; Figure 3]. Similarly, the 1-, 3-, and 5-year BT (death from another cause) rates in the Italian cohort were 21% (19%), 37% (47%), and 42% (54%) in the presence and 8% (14%), 20% (35%), and 24% (45%) in the absence of high-risk mutations (Gray's p value <0.01 for BT and 0.18 for death from another cause; Figure 4). Conclusions In the current study, machine-learning algorithms enabled the discovery of concurrent mutations in CMML that were shown to be prognostically more significant than their individual constituents. Such prognostic interaction might have contributed to some of the discrepancies noted in current literature regarding the prognostic relevance of certain mutations in CMML and should be accounted for in the development of future risk models.