Determining sensitivity to FLT3 inhibitors prior to therapy in FLT3 mutant acute myelogenous leukemia

Abstract INTRODUCTION: FLT3 mutations occur in ~30% of Acute Myelogenous Leukemia (AML) cases and while FLT3 inhibitors (FLT3i) have improved prognosis, many patients are primary resistant. Previously, using Reverse Phase Protein Arrays (RPPA), we identified three proteomic clusters with differing FLT3i responses: 29% improved (C1), 21% unaffected (C2), and in 50% therapy was detrimental (C3). We also showed that a six-protein protein classifier predicted cluster membership with >97% accuracy. Here, we provide orthogonal validation of these findings using Mass Spectrometry (MS)-based proteomics and demonstrate differential FLT3i sensitivity in ex-vivo drug assays. METHODS The levels of 429 proteins (338 total and 91 post-translational modified (PTM)) were measured in 806 newly diagnosed, fresh, pre-treatment, >95% blast enriched AML samples using RPPA. Protein expression was normalized to non-G-CSF treated, normal bone marrow-derived CD34+ cells. FLT3 mutation status, therapy and outcome data were known for 636 patients, of which 127 were FLT3 mutant (MUT), including 54 treated with FLT3i. MS-based proteomics quantified 10,102 total proteins in 57 AML samples, of which Pearson's and Spearman's correlation were used for protein-protein correlations (p<0.01; R>0.5); Proprietary machine learning (ML) model coupled with Consensus Clustering was used for identifying FLT3i sensitivity sample subgroups based on MS proteomic measurements and ex-vivo sensitivity assays. For ex-vivo sensitivity assays, AML cells were treated with several dilutions of a FLT3i (Sorafenib, Midostaurin, Crenolanib, Quizartinib, Gilteritinib, Tuspetinib and MAX-40279 (a dual FLT3/FGFR inhibitor)) and cell viability was assessed with CellTiter-Glo® (Promega). RESULTS Using our MS-based proteomics measurements and sensitivity results with our proprietary ML model, we processed both AML patient samples (N=57) and patient-derived cell lines (N=59), and we were able to identify 7 distinct clusters. Most samples predicted to be sensitive to FLT3i by RRPA C1 (N=5) were clustered in the same MS-based category which was significantly enriched with ex-vivo FLT3i responders. In contrast, 4 of the samples that were predicted to be resistant to FLT3i (C2=3, C3=1) were clustered in a different MS-based category, which was enriched with ex-vivo insensitive FLT3i assays. The rest of the samples with a RRPA-based prediction were falling into 3 different categories with mixed or partial sensitivity enrichments to FLT3i (C1=1, C3=2 / C2=2, C3=2 / C1=1). 258 out of 338 RPPA proteins (76%) were detected by both RPPA and MS-based proteomics analyses, with a median correlation of R=0.4 and significant correlation (R>0.5) present in 85 proteins (25%). Importantly, 3 out of the 6 proteins from our RPPA protein classifier (SMARCA2, PXN, DPF2) showed significantly high correlation with MS-based proteomics (R=0.74, 0.68, 0.57) and passed an 80th percentile threshold on their intensities. These three markers exhibited concordant expression patterns across MS-based categories and their corresponding RPPA-based clusters, consistently showing up- or down-regulation in both measurement platforms. Of note, cases with higher overall resistance to FLT3i in the ex-vivo assays showed moderate sensitivity to a SMARCA2/4 inhibitor, corroborating findings from our previous work with RPPA. Finally, we identified cell lines in which the expressions of PXN, SMARCA2, and DPF2 match primary AML samples from each FLT3i group (sensitive vs resistant). CONCLUSIONS Previouslywe demonstrated that RPPA proteomics could discriminate FLT3i sensitive and resistant cases and here we present orthogonal confirmation by MS proteomics and ex-vivo sensitivity assays. Moreover, we show that the expression of only three proteins forms a robust biomarker to predict FLT3i sensitivity of FLT3-MUT AML patients prior to therapy. We also identified AML cell lines that mimic both FLT3i resistance and sensitivity, and combined with deep proteomics data, these provide an ideal tool to identify/optimize combination treatments for the resistant patient subsets. We aim to further validate our global proteomics findings retrospectively using samples from a separate cohort of FLT3-MUT and FLT3i treated AML patients. Our work could provide means to triage and recommend therapy selection for FLT3-MUT AML patients in the future.