Developing transcriptomic signature for IDH1 and IDH2 acute leukemia and the demonstration of high prevalence of these signatures in mutation-negative leukemia

Abstract

Abstract Introduction: Despite multiple studies, the impact of IDH1 and IDH2 (IDH1/2) mutations on the overall biology and clinical outcome of acute myeloid leukemia (AML) remains controversial. This is likely due to the fact that AML with IDH1/2 mutations is not significantly biologically different from non-IDH1/2 mutated AML. However, current data suggest that adding IDH inhibitors to a combination therapy improves the outcome of patients with IDH1/2 mutations. We explored the potential of using artificial intelligence (AI) and transcriptomic data to define a specific transcriptomic signature for IDH1-positive (IDH1p) and IDH2-positive (IDH2p) AML. We then used this signature to screen IDH1/2-negative acute myeloid leukemia (AMLn). Methods: RNA was extracted from the bone marrow samples of 1186 cases of AML or advanced myelodysplastic syndrome with increased blasts. The RNA was sequenced by next generation sequencing (NGS) using a targeted RNA panel of 1600 genes. Hybrid capture sequencing library preparation was used and RNA was quantified using transcript per million (TPM). IDH1 mutation was detected in 83 cases (7%) and IDH2 was detected in 120 (10%). Four cases (0.3%) had mutations in both IDH1 and IDH2. These four cases were excluded from the study. A set including 83 cases with IDH1 mutation and 156 random AMLn was isolated to develop an IDH1 transcriptomic signature. A second set including the 120 cases with IDH2 mutation and 180 random AMLn cases was used for developing the IDH2 transcriptomic signature. The rest of the cases were used for testing these signatures. Bayesian statistics were used to rank the genes that distinguish between two groups, then random forest was used to establish the signatures. Two thirds of the sets used for developing the signatures were used for training and one third was used for testing the model. A score for the combination of relevant genes with a cut-off point was established that distinguishes each signature. The same Bayesian/random forest algorithm was used to test the rest of the AML cases that were not used in developing the models. Results: In developing an IDH1 signature, the AI algorithms showed that only 8 genes (TRAF3, TRAF2, HMGB1, CDK2, LRRC59, MEAF6, CREB1, RPN) were adequate to distinguish IDH1p cases from IDH1n cases with AUC of 0.873 (95% CI: 0.801-0.945). For developing an IDH2 signature, the AI algorithms required 35 genes to distinguish IDH2p group from IDH2n cases with AUC: 0.976 (95% CI: 0.947-10.00). Using the same AI modeling and genes, testing the 918 cases of AMLn cases that were not used for using the model shows that 739 cases (80%) had a transcriptomic signature similar to that of IDH1p cases. Testing the 887 AML cases that were IDH2n for the presence of IDH2 transcriptomic signature showed that 579 (65%) cases showed a signature similar to that seen in IDH2p cases. Conclusions : This data suggests that despite AML with IDH1/IDH2 mutations overlapping biologically with other types of AML, transcriptomic signatures can be developed for IDH1p and IDH2p cases. More importantly, this signature can be seen in a significant number of cases without IDH1/2 mutations. This also suggests that a significant number of AML cases may benefit from adding IDH1or IDH2 inhibitors to their combination therapy even if they do not have IDH1/IDH2 mutations but do have the signature. Clinical trials are needed to confirm the role of using IDH1/IDH2 signatures in selecting patients for treatment with ID1/IDH2 inhibitors.