Developing artificial intelligence-based transcriptomic signature for selecting patients with HOXA-MEIS1 pathway abnormalities for the treatment with menin inhibitors

Abstract

Abstract Introduction: The menin-KMT2A complex plays a major role in activating the HOXA-MEIS1 pathway. Abnormalities in this pathway lead to leukemogenesis. KMT2A gene rearrangement (KMT2Ar) and the expression of various KMT2A fusion genes is a common driving abnormality in pediatric acute myeloid leukemia (AML) and in subgroups of adult AML. In such leukemia, disrupting the menin-KMT2A complex using menin inhibitors has been established as an effective therapy in AML. However, multiple mechanisms other than KMT2Ar can activate the HOXA-MEIS1 pathway and these cases may respond to menin inhibitors. We hypothesized that KMT2Ar leads to generalized RNA signature in leukemic cells and this signature can be generated by mechanisms other than KMT2Ar. Using transcriptomic data from AML cases with KMT2Ar, we established a unique expression signature for KMT2Ar AML using an artificial intelligence (AI) model. Then we used this AI model for testing KMT2A-negative (KMT2An) AML cases for the presence or absence of such signature. Methods RNA was extracted from the bone marrow samples of 759 cases with AML. 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). Of the 759 AML cases, 52 were KMT2Ar positive and 707 were KMT2An. A set of 102 KMT2An cases and the 52 KMT2Ar (total 154) was used to establish the KMT2Ar signature and the rest of the KMT2Arn cases (N=657) were used for testing. Bayesian statistics were used to rank the genes that distinguish between KMT2Ar and KMT2An, then eXtreme Gradient Boosting (XGBoost) was used to establish the KMT2Ar signature. Two thirds of the 154 cases 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 distinguish TMT2Ar from KMT2An. The same Bayesian/XGBoost algorithm was used to test the rest of the KMT2An AML cases and to stratify as signature KMT2A positive vs. negative. Results Using 52 KMT2Ar cases and 102 KMT2An cases in the Bayesian/ XGBoost model described above, we show that in testing set KMT2Ar can be distinguished from KMT2An with AUC of 0.993 (95% CI: 0.973-1.00) using only 5 genes (TRAF2, TRAF5, TRAF3, CCND2, NEDD4). Using 65 genes increased the accuracy of distinguishing the two groups to AUC of 0.998 (95% CI: 0.988-1.00). To increase stringency, we used the 65-gene AI model and tested the remaining 657 KMT2An AML cases for the presence of the KMT2Ar signature. Of these 657 cases, 130 (20%) showed biologically KMT2Ar transcriptomic signature. These cases showed a significantly higher level of HOXA9 (P=0.004) and significantly (P<0.0001) different RNA levels in TRAF2, TRAF5, TRAF3, CCND2, NEDD4 expression as compared with KMT2An signature. The cases classified as biologically similar to KMT2Ar contained two cases (1.5%) with KMT2A-PTD (Partial tandem duplication) and 39 (30%) cases with mutation in NPM1 gene. In contrast, 21% of the KMT2An cases had NPM1 mutation. Conclusions: This data shows that AML with KMT2Ar has a unique transcriptomic signature that identifies activation of HOXA-MEIS1 pathway. This signature, when used in AI model, can identify a significant number of KMT2An AML cases with the same KMT2Ar signature that potentially could benefit from treatment with menin inhibitors. This model is robust based on using 65 genes and justifies initiating a clinical trial for selecting patients for the treatment with therapy that includes menin inhibitors.