Clinical resistance to crenolanib in acute myeloid leukemia due to diverse molecular mechanisms

FLT3 mutations are prevalent in AML patients and confer poor prognosis. Crenolanib, a potent type I pan-FLT3 inhibitor, is effective against both internal tandem duplications and resistance-conferring tyrosine kinase domain mutations. While crenolanib monotherapy has demonstrated clinical benefit in heavily pretreated relapsed/refractory AML patients, responses are transient and relapse eventually occurs. Here, to investigate the mechanisms of crenolanib resistance, we perform whole exome sequencing of AML patient samples before and after crenolanib treatment. Unlike other FLT3 inhibitors, crenolanib does not induce FLT3 secondary mutations, and mutations of the FLT3 gatekeeper residue are infrequent. Instead, mutations of NRAS and IDH2 arise, mostly as FLT3-independent subclones, while TET2 and IDH1 predominantly co-occur with FLT3-mutant clones and are enriched in crenolanib poor-responders. The remaining patients exhibit post-crenolanib expansion of mutations associated with epigenetic regulators, transcription factors, and cohesion factors, suggesting diverse genetic/epigenetic mechanisms of crenolanib resistance. Drug combinations in experimental models restore crenolanib sensitivity. FLT3 is commonly mutated in acute myeloid leukaemia and treatment with FLT3 inhibitors often ends with relapse. Here, the authors perform exome sequencing of samples from patients treated with the FLT3 inhibitor, crenolanib, to show that resistance occurs due to diverse molecular mechanisms, not primarily due to secondary FLT3 mutations.

A cute myeloid leukemia (AML) is a group of heterogeneous hematologic malignancies characterized by numerous cytogenetic and molecular alterations. Activating mutations in the FMS-like tyrosine kinase 3 (FLT3) gene represents the most frequent molecular abnormality in AML 1,2 . The majority of the mutations in FLT3 are internal tandem duplications (ITD), which are identified in approximately 30% of AML patients and are associated with a higher propensity for disease relapse and a shorter overall survival 3,4 , even after stem cell transplantation 5 . FLT3 point mutations in the activation loop of the tyrosine kinase domain (TKD), predominantly at residue D835, are found in an additional 7% of patients with uncharacterized prognosis 6,7 .
A growing number of small-molecule FLT3 tyrosine kinase inhibitors (TKIs) have been evaluated in preclinical experiments and clinical trials, but only one agent (midostaurin) has been recently approved for this specific use. Many of the firstgeneration FLT3 inhibitors including midostaurin, lestaurtinib, sunitinib and sorafenib have been limited by their suboptimal efficiency and sustainability as a single drug therapy 8,9 . However, recent clinical trials with some of these agents, notably midostaurin, have revealed durable improvements in patient outcomes when administered at diagnosis in combination with standard of care chemotherapy 10,11 . The second-generation inhibitors, including quizartinib, pexidartinib, gilteritinib and crenolanib, have demonstrated enhanced potency and selectivity when administered as single-agent therapies [12][13][14][15][16][17][18] . Compared to other FLT3 TKIs, crenolanib demonstrates several appealing characteristics to target FLT3 mutations in AML. As a potent type I pan-FLT3 inhibitor, crenolanib retains activity against FLT3 TKD mutations 19 , which have been shown to be the major resistance mechanisms for quizartinib and sorafenib [20][21][22][23][24] . Therefore, crenolanib is a candidate therapy for de novo AML patients with FLT3 TKD mutations as well as relapsed patients with TKD mutations acquired after treatment with other FLT3 TKIs 25 .
Crenolanib has been evaluated in two phase II clinical trials in chemotherapy or TKI refractory/relapsed AML patients with FLT3 mutations. Cumulatively, a high response rate (complete response with incomplete blood count recovery (CRi) of 37%, and partial response (PR) of 11% in prior TKI-naive group; 15% complete response (CR)/CRi and 13% PR in prior TKI group) was achieved with crenolanib single-agent therapy. 26 Details of the clinical trials are reported elsewhere 14,25,26 . However, similar to other FLT3 TKIs observed in early clinical trials, despite initial response, subsequent drug resistance and disease relapse occurred in the majority of patients 8,9,14,25,26 . We, therefore, performed whole exome sequencing (WES) and targeted deep sequencing on a series of samples from crenolanib-treated patients to investigate the relationship between drug resistance and genetic signatures (data can be explored and visualized in our Vizome, online data browser (www.vizome.org)). We were initially interested in investigating whether crenolanib resistance followed similar mechanisms as other FLT3 TKIs (quizartinib, gilteritinib and sorafenib) [27][28][29][30] , where secondary FLT3 mutations in the activation loop and/or gatekeeper residue play a major role. Given the nature of heterogeneous genetic alterations and selective pressure of chemotherapy and prior TKI treatment in relapsed/refractory AML patients on these trials, we also aimed to characterize the impact of co-occurring clones or subclones with other somatic mutations on crenolanib response and disease recurrence.
We observed that crenolanib-resistant FLT3 secondary mutations (one patient with K429E mutation and two patients with gatekeeper mutations) are infrequent. The majority of patients exhibited a diverse spectrum of mutations associated with chromatin modifiers, cohesion, spliceosomes and transcription factors, which mostly expanded during treatment, suggesting an elaborate genetic/epigenetic mechanism of resistance to crenolanib.

Results
FLT3 secondary mutations are infrequent. We first determined whether secondary FLT3 mutations were acquired during treatment by sequencing available patient samples obtained after at least 28 days of crenolanib treatment as well as baseline samples obtained before crenolanib treatment initiation. Consistent with previous reports, no de novo activation loop mutations were detected in FLT3-ITD patients after crenolanib treatment as determined from 18 FLT3-ITD patients sequenced by exome sequencing (mean ± standard error of the mean (SEM) coverage is 177 ± 24) and 6 FLT3-ITD patients sequenced by Miseq (coverage 232,809 ± 9388) 31 . Variant allele frequencies (VAFs) of FLT3 D835 or FLT3-ITD were eliminated or cleared in 11 out of 21 or 11 out of 39 patients, respectively; and maintained or expanded in the rest of patients (Supplementary Data 1-2). FLT3 F691 mutations were detected in two patients (Fig. 1a). Both patients were previously treated with quizartinib and one patient harbored 17% VAF of FLT3 F691L prior to crenolanib treatment. Consistent with previous studies 14,20,23,25,26 , low VAFs of non-D835 FLT3 secondary mutations A833S, D839Y/G, N841K and Y842C and a small insertion/deletion (R834D835I836 (RDI->RP) were detected at baseline and eliminated during the course of crenolanib treatment ( Fig. 1a  To characterize the leukemogenic and drug-resistant potential, FLT3 wild-type (WT) and point mutations were introduced into Ba/F3 cells and evaluated by interleukin-3 (IL-3) withdrawal assay and drug sensitivity profiling. As expected, FLT3 D835Y (positive control) and Y572C transformed Ba/F3 cells (Fig. 1b). FLT3 K429E was also observed to transform Ba/F3 cells, although with slower kinetics compared to FLT3 D835Y and Y572C (Fig. 1b). The other two mutations (D200N and L601F) and the gatekeeper mutation F691L did not transform Ba/F3 cells.
Differential mutation profiles in prior TKI-treated patients. Since prior FLT3 inhibitor treatment was also identified as an adverse prognostic factor of crenolanib clinical response 26 , we compared the mutational profiles of these patients (pre-TKI) to patients who had not received prior TKI therapy (TKI naive) (Supplementary Table 2-3 and Supplementary Data 3-4).
Available data from the targeted gene panel and/or WES of 50 patients before crenolanib treatment were analyzed and recurrent variants within 24 commonly mutated genes were detected (Fig. 2a, b and Supplementary Data 1). The average number of pathogenic variants was highest among the pre-TKI group, while similar frequencies were observed between the TKI-naive and the de novo AML cases reported in The Cancer Genome Atlas (TCGA; Fig. 2d and Supplementary Table 8). Consistent with previous studies 24 , higher frequencies of FLT3-ITD and FLT3 TKD combination mutations were observed in the pre-TKI group (Supplementary Figure 2). Furthermore, the frequency of NRAS, KRAS, RUNX1, IDH1, WT1, TET2 truncation and ASXL1 mutations were higher in the pre-TKI group (Fig. 2b).
Differential mutation profiles in crenolanib poor responders. We next compared the coexisting mutation profiles between different crenolanib response groups. We observed that the average number of mutations at trial entry was higher for crenolanib poor responders (hematological improvement (HI)+resistant disease (RD)) compared to crenolanib good responders (CR/CRi +PR) (Fig. 2c). Furthermore, higher frequencies of NRAS, TET2, IDH1, IDH2, U2AF1, STAG2, KRAS, CSF3R, TET2 truncation    Fig. 2 Differential mutation profiles in pre-tyrosine kinase inhibitor (TKI)-treated patients and crenolanib poor responders. a Mutation spectrum identified by whole exome sequencing and targeted deep sequencing before crenolanib treatment in TKI-naive and pre-TKI group. b Comparison of frequencies of specific mutated genes detected in TKI-naive, prior TKI treatment patients. c The graph depicts mean ± SEM of mutation numbers discovered in The Cancer Genome Atlas (TCGA) de novo acute myeloid leukemia (AML) patients with FMS-like tyrosine kinase 3 (FLT3) mutations (n = 56, mutation number range 1-4), TKI-naive (n = 13, mutation number range 1-6) and prior TKI (n = 24, mutation number range 2-8) treatment patients. d The graph depicts mean ± SEM of number of mutations in TCGA de novo AML patients, crenolanib good responders (n = 10, mutation number range 1-5) and poorresponders (n = 27, mutation number range [1][2][3][4][5][6][7][8]. For (c, d), FLT3-ITD and FLT3 tyrosine kinase domain (TKD) compound mutation is counted as 2. e Comparison of frequencies of specific mutated genes detected in crenolanib good responders and poor responders. f Graph depicts variant allele frequency (VAF) change of gene mutations during crenolanib treatment. Statistical significance was assessed using one-way analysis of variance (ANOVA) and Kruskal-Wallis test comparing each condition to the respective FLT3 D835Y and expressed as: **p < 0.01, ****p < 0.0001 and ASXL1 mutations were present in poor responders compared with crenolanib good responders prior to crenolanib treatment (Fig. 2e, Supplementary Table 4 Figure 3). We did not observe an association of FLT3 mutation status (ITD vs. TKD vs. ITD+TKD) with crenolanib response (Supplementary Table 4 and Supplementary Figure 3). Changes in VAF were then analyzed from available paired samples to identify new or expanded clones during crenolanib treatment. We observed that the majority of TET2, DNMT3A, RUNX1, U2AF1, SF3B1 and IDH1 mutations were present with approximately 50% VAF, and the VAFs remained unchanged during crenolanib treatment (Fig. 2f), indicating they were present in the founder clone and may not be sensitive to crenolanib treatment. VAFs of variants of NRAS, BCOR, STAG2, CEBPA and ASXL1 increased during crenolanib treatment, suggesting these mutations may contribute to a clonal selection mechanism of drug resistance. We observed both the expansion and reduction of NPM1 and WT1 mutations during crenolanib treatment, indicating these mutations alone may not be sufficient to render drug resistance. Interestingly, loss of heterogeneity and VAF expansion were observed in two DNMT3A, one WT1, and one RUNX1 mutation at crenolanib resistance, suggesting they might contribute to disease resistance. Surprisingly, VAF of CSF3R (5 out of 5) and JAK3 (1 out of 1) and PTPN11 (2 out of 4) mutations dropped to undetectable levels during crenolanib treatment (Fig. 2f). Of note, multiple other gene mutations were present in these samples. There might be two possible explanations for this phenomenon: crenolanib may be able to inhibit CSF3R, JAK3 and PTPN11 mutation clones due to targeting kinases other than FLT3 (e.g., Janus kinases (JAKs), mitogen-activated protein kinase (MAPK) pathway); or CSF3R, JAK3 or PTPN11 mutation clones were suppressed by other more aggressive outgrowth clones with differing mutational profiles. Since crenolanib targets platelet-derived growth factor receptor (PDGFR) and FLT3 with high specificity and there are two patients clearly showing PTPN11 mutations expansion at disease relapse (A14 and B27), the second model appears more likely. Nonetheless, these data indicate that CSF3R and JAK3 mutations do not harbor high growth advantage during crenolanib treatment in the presence of other clones, and mutations in these genes are unlikely to confer crenolanib resistance, whereas PTPN11 mutations could be repressed or acquired during crenolanib treatment and contribute to drug resistance.

and Supplementary
Subclonal RAS mutations contribute to crenolanib resistance. Since mutations in RAS signaling pathway genes (NRAS, PTPN11, KRAS and CBL) are enriched in crenolanib poor responders and/or pre-TKI group (Fig. 2b, e), and/or expanded or acquired during crenolanib treatment (Fig. 2f), it is likely they can confer crenolanib resistance.
To determine whether these mutations co-occur with FLT3 mutations within the same clone or occur in an independent FLT3 WT clone, we compared the absolute VAF and the direction of VAF change of these gene variants with respect to absolute VAF and VAF changes of the FLT3 mutations. For example, if the VAF of the RAS pathway gene and the FLT3 mutations are both greater than 50%, or if a lower baseline VAF of the RAS pathway mutation changes in the same trajectory as the FLT3 mutation after crenolanib therapy, we can assume that they have a high probability of occurring within the same leukemic clone. In this manner, we observed that the majority of the NRAS and KRAS pathway mutations were present in independent clones not harboring the FLT3 mutations (Fig. 3a). However, we did observe that one NF1 and two CBL mutations co-occurred with FLT3 mutations in the same clone. Three out of four PTPN11 mutations (A14, A16 and A15) appear to co-occur with FLT3 TKD or ITD mutations. We, therefore, analyzed the drug response profile of cells harboring FLT3 and RAS pathway compound mutations by introducing FLT3 D835Y and PTPN11 A72D into Ba/F3 cells.
Concomitant TET2 mutations confer crenolanib resistance. Five patients with TET2 frameshift/nonsense mutations demonstrated adverse prognosis, whereas four other patients with TET2 missense mutations did not show a higher incidence of unfavorable response to crenolanib ( Fig. 4a and Supplementary Table 6), suggesting that TET2 truncation mutations may contribute to crenolanib resistance. VAF analysis demonstrated cooccurrence and persistence of TET2 and FLT3 mutations, indicating up-front drug resistance for cases harboring FLT3/TET2 compound mutations at the start of therapy. These findings are consistent with recent work showing TET2 mutations contribute to quizartinib resistance in mouse models 33 . To further test this hypothesis, we performed a drug sensitivity assay with BM stem/ progenitor cells from Flt3-ITD knock-in (Flt3 ITD ) and Flt3-ITD knock-in/Tet2 knockout mice (Flt3 ITD ; Tet2 +/− ). We observed that Flt3 ITD ; Tet2 +/− progenitor cells did not respond to crenolanib at 10 or 100 nM, whereas Flt3 ITD cells responded well at these drug concentrations (Fig. 4b-d). These data confirmed that FLT3-ITD, in concert with TET2 loss of function, is resistant to the single-agent FLT3 inhibitor. Excitingly, Flt3 ITD ; Tet2 +/− cells demonstrated sensitivity to azacytidine, and cell differentiation analysis demonstrated a clearance of Sca1+ progenitor cells with azacytidine in both Flt3 ITD and Flt3 ITD ; Tet2 +/− mouse BM stem/progenitor cells (Supplementary Figure 3).
IDH1 and IDH2 mutations contribute to crenolanib resistance. Higher frequencies of IDH1 and IDH2 mutations were observed in the pre-TKI and crenolanib poor responder groups (Fig. 2b-f), indicating IDH1 and IDH2 mutations might be associated with TKI treatment resistance. IDH1 mutations co-occurred in the same clone with FLT3-ITD/TKD mutations in five patients (B01, B07, B34, B32 and B16), whereas the IDH2 mutations were highly likely to be in FLT3-ITD/TKD independent clones in two patients (A08 and B04) and possibly co-occurred with FLT3-ITD in one patient (B06) (Fig. 5a). To investigate whether co-occurring IDH1 mutations perturb crenolanib sensitivity, we performed drug profiling of cells harboring both mutations. We did not observe decreased crenolanib sensitivity in FLT3-ITD stem cells transduced with IDH1 R132H compared to FLT3-ITD cells transduced with IDH1 WT (Fig. 5b). We observed that cells harboring IDH1 R132H/ FLT3 D835Y demonstrated similar crenolanib sensitivity as IDH1 WT/D835Y cells (16.2 ± 8.0 nM vs. 18.9 ± 13.9 nM) (Fig. 5c). Consistent with previous studies, an IDH1 inhibitor as a single agent did not impact on cell viability. However, the IDH1 inhibitor (AG5198) did enhance crenolanib sensitivity ( Fig. 5d and Concomitant TP53 mutations confer crenolanib resistance. TP53 pathway mutations are mutually exclusive with FLT3 mutations in de novo AML and were shown to be related to FLT3 inhibitor resistance 34,35 . Interestingly, we observed one TP53 mutation and one PPM1D mutation that co-occurred with FLT3-ITD and/or FLT3 TKD in two individual patients with crenolanib poor response. To test whether TP53 loss of function confers crenolanib resistance, we performed a knockout of TP53 expression in Molm13 cells using the clustered regularly interspaced short palindromic repeats (CRISPR)/CRISPR-associated 9 (Cas9) system (Supplementary Figure 7). We observed significant reduced crenolanib sensitivity ( Fig. 5e-g) (Fig. 6a).
In general, the clonal evolution patterns observed from this series of patient samples could be classified into three distinct groups. The first pattern is of a primary drug refractory clone (Fig. 6b). In this pattern, FLT3 mutations co-occur with other drug-resistant gene mutations in the same clone and the VAF of both mutations persist or increase during drug treatment. The cooccurring mutations are normally present in the founder clone, such as TET2, IDH1 and TP53 pathway mutations. In line with a previous study 33 and the data shown here, the patient presented in Fig. 6b exhibited both FLT3-ITD and TET2 mutations that cooccurred, and this concomitant mutation pattern was completely resistant to crenolanib. These data support the notion that up-front drug combinations may be useful to target certain mutational patterns, such as FLT3 and TET2 co-occurring mutation. The second pattern is acquisition or expansion of additional mutations in the context of a FLT3 mutation with either an original, dominant FLT3-mutant clone that exhibits new mutations at the time of drug resistance or a minor FLT3-mutant clone with additional mutations that expands during crenolanib treatment (Fig. 6c). Deep sequencing of samples at the start of FLT3 inhibitor therapy could identify problematic coexisting mutations and apply drug combinations that pre-emptively target the expanding clone. The third pattern is the acquisition of leukemic clones independent of the clone with the FLT3 mutation (Fig. 6d). In this pattern, the VAF of FLT3 mutations decreased, whereas drug-resistant subclones within a FLT3 WT clone emerged or expanded and became  Notably, we did not detect the IDH1 mutation in B32 before crenolanib treatment by exome sequencing. However, a targeted gene panel detected this IDH1 mutation before crenolanib treatment. b Graph depicts mean ± SEM of colony numbers of FLT3-ITD/IDH1 WT and FLT3-ITD/IDH1 R132H expressing mouse stem cells treated with gradient concentrations of crenolanib, AG5198 or two drugs in combination. c Graph depicts cell viabilities of FLT3 D835Y, IDH1 WT/ FLT3 D835Y and IDH1 R132H/ FLT3 D835Y expressing Ba/F3 cells treated with crenolanib, or crenolanib in combination with IDH1 Inhibitor (AG5198) determined by MTS assay. d Graph depicts decreased half-maximal inhibitory concentration (IC 50 ) of crenolanib and AG5198 in combination compared to crenolanib alone. Statistical significance was determined using two-tailed nonparametric Student's t-tests (Mann-Whitney test) comparing each group to the non-treated group. TP53 mutations co-occur with FLT3 mutations and confer crenolanib resistance. e Graph depicts VAFs of FLT3-ITD/TKD and TP53 or PPM1D mutation during crenolanib treatment. f Representative graph depicts higher mean ± SEM of cell viability of crenolanibtreated Molm13 cells expressing CRISPR/Cas9 and single-guide RNAs (sgRNAs) targeting TP53 compared to cells expressing CRISPR/Cas9 and a nonspecific targeting sgRNA (NS) control. TP53_1: sgRNA1 targeting TP53; TP53_2: sgRNA2 targeting TP53. g Graph depicts mean ± SEM of crenolanib IC 50 in (b). Data shown are from three or five biological replicates. Statistical significance was assessed using one-way analysis of variance (ANOVA) together with Dunn's multiple comparisons tests and expressed as: *p < 0.05   Fig. 6 Mutation spectrum and clonal patterns of patients treated with crenolanib. a Graph depicts mutations identified by exome sequencing, gene panel and/or targeted sequencing. Each column displays a patient; each row denotes a specific gene. Recurrently mutated genes are color-coded for only present before crenolanib treatment, only present in after crenolanib treatment samples, persistence before and after crenolanib treatment, with no samples available before or after crenolanib treatment. b Graph depicts primary drug-resistant clone. In this case, FLT3 mutation clones co-occur with a TET2 mutation clone and the variant allele frequencies (VAFs) of the combined mutation persist during drug treatment. c The second pattern is the acquisition or expansion of additional mutations in the context of a FLT3 mutation. In this case, the original, dominant FLT3 tyrosine kinase domain (TKD) clone was inhibited by crenolanib after three cycle treatments. However, CEBPA mutation was acquired and expanded at the time of drug resistance in FLT3 TKD clone. d The third pattern is the acquisition of subclones independent of FLT3 mutation clones. In this case, FLT3 mutation clones were eliminated by crenolanib; however, a PTPN11 mutation clone emerged during crenolanib treatment. Of note, the clonal patterns shown do not include mutation information before crenolanib treatment. C: crenolanib treatment cycle number; Chemo: chemotherapy. The y-axis stands for VAF Previous studies suggest that a common drug-resistant mechanism of TKI treatment is secondary mutations of the targeted kinase itself which could be classified into two major groups 21 . The first group is TKD mutations, which are specifically resistant to type II FLT3 TKIs including quizartinib, sorafenib, ponatinib and pexidartinib [36][37][38][39][40][41] . FLT3 N676K was first characterized to confer resistance to midostaurin and more recently also shown to contribute to resistance to quizartinib 22,42 . Point mutations at residues D835, I836 and Y842 have been shown to be predominantly associated with quizartinib and sorafenib resistance [20][21][22]43 . The second group is mutations on the gatekeeper residue of the targeted kinases. Gatekeeper mutations enhance the binding affinity of adenosine triphosphate (ATP) which competes with TKIs to bind to the ATP binding pocket, thereby preventing TKI binding to the kinase and rendering resistance to the TKI treatment 44 . FLT3 F691L was shown to be resistant to the majority of FLT3 TKIs including crenolanib, but not ponatinib and pexidartinib 36,39,45 . One study showed that 3 out of 8 quizartinib-treated patients 27 and 3 out of 16 and 4 out of 20 giltertinib-treated patients acquired a FLT3 gatekeeper mutation 29,30 . In vitro saturation assays have been widely used to screen and identify secondary drug resistance mutations. Previous studies showed that only two recurring mutant clones out of 300 × 10 6 initial clones screened were retained in the presence of crenolanib (100 nM) 25 . Consistent with the in vitro saturation assay, as a type I TKI, crenolanib did not induce novel FLT3 secondary TKD mutations, and only two crenolanib-treated patients demonstrate gatekeeper mutations in our study cohort. In addition, a novel FLT3 extracellular mutation at K429E was detected in one patient with high VAF, which showed increased crenolanib IC 50 . The structural basis for the drug resistance of FLT3 K429E requires further investigation.
Another drug resistance mechanism of TKIs is expansion or activation of alternative signaling pathways 33,35,[46][47][48] . Comparing the mutation profiles of chronic myeloid leukemia (CML) and AML patients may provide clues as to the mechanisms allowing BCR-ABL TKIs to demonstrate a low incidence of relapse and a long relapse-free latency. BCR-ABL was so far identified as the exclusive driver mutation in CML. Recent studies have not identified other recurrent mutations in a majority of CML patients by exome sequencing 49,50 . However, AML is a highly heterogeneous group of diseases exhibiting diverse coexisting mutations with great potential to generate multiple subclones during treatment 1,2,51,52 . In contrast to the mutual exclusivity of FLT3-ITD and D835 mutations seen in de novo AML (TCGA dataset), 35% of patients in TKI-naive postchemotherapy group and 51.4% of pre-TKI patients in the current cohort harbored both FLT3-ITD and D835 mutations 26 , suggesting that both chemotherapy and type II FLT3 inhibitor treatment could select drug-resistant FLT3 TKD clones. Furthermore, pre-TKI-treated crenolanib refractory and resistant patients exhibited more coexisting driver mutations compared to TKI-naive and crenolanib responders, respectively, suggesting a drug-resistant clone selection or induction pressure of TKI, including crenolanib treatment. In addition, previous studies have shown that FLT3-ITD demonstrated mutual exclusivity with NRAS, KRAS, RUNX1, SF3B1, TP53, SRSF2 and ASXL1 mutation in de novo AML 51 , whereas in our current cohort, we observed high frequencies of coexisting NRAS, KRAS, RUNX1, SF3B1, TP53, SRSF2 and ASXL1 mutations with FLT3 mutations respectively, suggesting a drugresistant clonal selection or induction pressure of previous chemotherapy and/or previous TKI treatment.
Deciphering whether the drug treatment leads to the acquisition of resistance mutations, or whether resistance is indicative of the expansion of pre-existing sub-clones under selective pressure, is important to understanding drug resistance mechanisms that may inform the design of better treatment regimens. New somatic mutations were detected in 10/30 samples with available paired samples, including mutations in NRAS (n = 1), PTPN11 (n = 2), ABL1 (n = 1), FLT3 F691L (n = 1), ASXL1 (n = 1), BCORL1 (n = 1), CEBPA (n = 1), WT1 (n = 1) and IDH1 (n = 1) (Fig. 6, dark  red). Notably, we did not detect the IDH1 mutation in B32 and CEBPA mutation in A10 before crenolanib treatment by exome sequencing. However, targeted gene panels detected the IDH1 and the CEBPA mutation before crenolanib treatment. The rest of the NRAS mutations as well as the majority of the mutations of IDH2, RUNX1 and STAG2 were present before crenolanib treatment with low VAF, and they expanded independently of the FLT3 mutations. In contrast, mutations in TET2, U2AF1, SF3B1 and IDH1 were present in the same clone of FLT3 mutations before crenolanib treatment and maintained during drug treatment. These data highlight the heterogeneity of AML clones and the expansion of pre-existing sub-clones under TKI selective pressure. Several alternative pathway mutation clones (Fig. 6, dark red) were not detected prior to treatment, which might be due to the limitation of current deep sequencing approaches to detect clones at very low levels. A higher prevalence of NRAS and KRAS mutations (26%) in de novo AML patients was detected in a recent study using deep targeted sequencing methods in comparison to 12% detected by WES in the TCGA AML cohort, indicating RAS mutations are frequently present in minor clones in de novo AML 51 . However, prolonged drug could also induce alternative pathway mutations shown by previous studies 47,53,54 and two PTPN11 mutations in the current study (A14 and B27). Nevertheless, future high coverage sequencing methods before and during treatment may be needed to uncover drug-resistant and low-level clones.
In some instances, previous studies have neglected to analyze the co-occurrence or subclonal patterns of coexisting mutations. It is expected that mutations with different oncogenic pathways co-occurring in a distinct clone from the FLT3 mutations can escape drug treatment and expand to become dominant clones. However, when they occur in the same clone as the FLT3 mutation, the capacity of the concomitant clone to confer drug resistance needs further validation. Determination of the mutation co-occurrence pattern is important for understanding drug resistance and developing alternative treatment strategies. By analyzing the VAF dynamics, we could distinguish between cooccurrence and independent sub-clone patterns. Consistent with previous studies, we observed that TET2 mutations co-occurred with FLT3 mutations and induced primary crenolanib resistance. We also identified two patients with TP53 pathway mutations on the same clone with FLT3 mutations. Previous studies have shown that the loss of TP53 led to midostaurin resistance. 34 In line with this study, we observed that knockout of TP53 conferred crenolanib resistance. In addition, we observed high cooccurrence of IDH1 and FLT3 mutations in crenolanib poor responders. Although we did not observe significant reduced crenolanib sensitivity of double mutation comparing to single mutation in ex vivo assays, this may be due to the models we used not accurately reflecting the patient actual biological condition. Furthermore, biologically relevant models are needed to assess the drug sensitivity of the combination mutations. In contrast, IDH2, BCOR, STAG2 and signaling pathway mutations (NRAS and PTPN11) were present in subclones independent of FLT3 mutations and escaped during crenolanib treatment, resulting in disease relapse. However, whether these mutations cause disease relapse alone or whether they cooperate with additional uncharacterized gene mutations and how exactly these mutations enable bypass of crenolanib inhibition are as-yet unknown.
Two major limitations of the current study must be noted. First, the use of VAF to analyze clonal architecture might be misinterpreted in cases where there is loss of heterozygosity and presence of biallelic mutations. More sensitive and accurate sequencing methods, e.g. single cell sequence, allele-specific PCR or digital PCR, are needed to validate the clonal architectural data. Second, while AML is a highly heterogeneous disease with various cytogenetic and genetic alterations, the sample size of the current study is small. Future larger cohort studies are needed to validate and identify more recurrent coexisting mutations that confer crenolanib resistance.
Overall, we identify three distinct patterns of mutation dynamics during crenolanib treatment. Each clonal evolutionary pattern represents a distinct prognosis and indicates different potential strategies to circumvent drug resistance. Our study also provides clinical implications: comprehensive sequencing should be carried out on patient samples at the start and during the treatment in order to identify and pre-emptively target problematic clones. Additionally, while single-agent therapy with FLT3 inhibitors may be of marginal clinical benefit to patients with high VAF of FLT3 mutations, it is imperative to combine a FLT3 inhibitor with chemotherapy or agents targeting cooperative lesions to achieve deep and durable remission. Assessment of outcomes. Complete response was adapted according to the International Working Group Criteria for AML and key criteria were defined as follows: CR was defined as ≤5% blast in the borrow marrow, no circulating blasts, with complete blood count recovery (neutrophil count ≥1000/μL and platelet count ≥100,000/μL). CRi required the same criteria as CR with incomplete count recovery (neutrophil count <1000/μL and platelet count <100,000/μL). PR required all the hematologic values of CR with a decrease of ≥50% bone marrow blast but still >5%. HI was defined as ≥50% blast reduction in peripheral blood or bone marrow.
Whole exome sequencing. WES was performed as previously described 55 and summarized in a reporting summary statement as a Supplementary Information file. Briefly, DNA was extracted from leukemia patient specimens using Qiagen DNeasy according to the manufacturer protocols. For exome sequencing we used the Illumina Nextera capture probes and protocol (12 samples per capture group with each sample run on 3, 5 or 6 lanes) with libraries run on a HiSeq 2500 using paired-end 100 cycle protocols. Initial data processing and alignments were performed using our in-house workflows that we describe here briefly. For each flowcell and each sample, the FASTQ files were aggregated into single files for reads 1 and 2. BWA MEM version 0.7.10-r789 56  (3) Read realignment around indels from the reads RealignerTargetCreator/ IndelRealigner; (4) Base Quality Score Recalibration. The resulting BAM files were then aggregated by sample and an additional round of MarkDuplicates and indel realignment was carried out at the sample level. For genotyping, single-nucleotide variations (SNVs) and small indels were called using the UnifiedGenotyper and VarScan2 58 . Additionally, SNVs were called by MuTect 59 . Each VCF file was annotated using the Variant Effect Predictor v83 60 against GRCh37.
Variant calling. Since no paired normal tissue controls were available, we compiled a list genes associated with human hematologic cancers according to these two papers 61,62 . In total, 170 genes were selected (Supplementary Table 2). We used the global filtering as previously described 63 . On top of that, the following filters were used: (1) excluding variants found in more than 0.1% of ExAC samples; and excluding variants found in normal samples from more than 20% Beat AML normal controls; (2)  Validation sequencing. DNA extraction was performed the same as the 'WES' procedure. Libraries were created and hybridized using custom designed Nimblegen (SeqCap EZ) probes covering genes of interest containing both known variants and novel recurrent variants seen in other AML samples sequenced in-house. Because of the smaller library size, we ran 12 samples per lane. Paired-end 100 base reads were generated, aligned and post-processed using the WES protocol described above. For each unique variant position observed in the WES data, the number of reads supporting each observed allele in the validation library was determined using bam-readcount 56 . A variant was considered to be 'validated' by sequencing if there were at least three reads supporting the called variant.
FLT3-ITD quantification with PCR. FLT3 exons 14 through 20 were amplified using forward primer 5'-GCAATTTAGGTATGAAAGCCAGC-3' and reverse primer 5'-CTTTCAGCATTTTGACGGCAACC-3'. The PCR products were then run on a 2% agarose gel stained with ethidium bromide and visualized and imaged under a Lummi imager (Roche Applied Science). The intensity of FLT3-ITD band and WT band were quantified by Image Lab software. The FLT3-ITD VAF was determined by calculating the ratio of FLT3-ITD to FLT3-ITD plus FLT3 WT band intensity.
Cell lines and reagents. HEK 293T/17 cells (provided by Dr. Richard Van Etten) were maintained in Dulbecco's modified Eagle's medium (Invitrogen). Ba/F3 cells (ATCC) were maintained in RPMI 1640 (Invitrogen) supplemented with 15% WEHI-conditioned medium. Molm13, Molm14 and MV4-11 cells (DSMZ) were maintained in RPMI 1640. All mediums were supplemented with 10-20% fetal bovine serum (Atlanta Biologicals), L-glutamine, penicillin/streptomycin (Invitrogen) and fungizone (Fisher). Mycoplasma contamination was routinely tested (once per month). Only mycoplasma-free cells were used in the experiments. The cell lines were authenticated by internal FLT3-ITD PCR and small inhibitor screening as well as short tandem repeat analysis.
Retroviral vector production and transduction. PTPN11 and IDH1 mutations were generated using the QuikChange II XL site-directed mutagenesis kit (Agilent Technologies) on the respective pENTR vectors (GeneCopoeia GC-Z2134 and Invitrogen clone IOH11942) and cloned into a gateway compatible MSCV-IRESpuromycin retroviral vector or a Tet-inducible lentiviral vector, pInducer20 (Addgene, #444012) via Gateway Cloning System (Invitrogen). Retrovirus was produced by transfecting HEK 293T/17 cells together with an EcoPac helper plasmid). Lentivirus was produced by transfecting HEK 293T/17 cells together with psPAX2 (psPAX2 was a gift from Didier Trono (Addgene plasmid # 12260) and pLP/VSVG (Invitrogen). After 2 days, the virus containing supernatants were filtered, and infected to cells followed by flow cytometry (fluorescence-activated cell sorting (FACS)) sorting or puromycin selection.
Ba/F3 IL-3 withdrawal assay. Stably transduced Ba/F3 cells (1 × 10 6 ) were washed three times and cultured in cytokine-free media. Viable cell number was determined on a Guava Personal Cell Analysis System (Millipore) every 1-2 days.
Colony-forming unit (CFU) assay. Bone marrow cells were harvested from BALB/ c mice (The Jackson Lab, #000651), Flt3 ITD and Flt3 ITD ;Tet2 +/− transgenic BALB/ c mice 33 (a kind gift from Evan Lind's lab, Oregon Health & Science University, Portland, OR). All mouse work was performed with approval from the Oregon Health & Science University Institutional Animal Care and Use Committee. For the BM transduction experiment, BM lineage-negative cells were enriched using Lineage Cell Depletion Kit (#130-090-858, Miltenyi Biotec), cultured overnight in medium containing IL-3, IL-6, and stem cell factor, and infected with retrovirus expressing FLT3 or IDH1 WT and mutants. Cytokines are purchased from PepreTech. A total of 2000 lineage-negative cells per well were seeded into 6-well plate with 1.1 mL of Methocult M3534 methylcellulose medium (StemCell Technologies) for 10 days. Images were taken and colonies (>50 cells) were counted via STEMvision™ colony counting software (StemCell Technologies). Colony cells were harvested, stained with cell surface markers and analyzed by FACS.
FACS. Cells were stained with antibodies (Biolegend) for 20 min at room temperature and washed twice with PBS. Membrane expression of CD11b, Sca1, c-KIT, Ter119 and CD71 were analyzed by FACS. All antibodies were used at 1:1000 dilution.
Inhibitor assay. Transformed Ba/F3 cells were seeded in 384-well plates (1250 cells per well) and exposed to increasing concentrations of crenolanib, trametinib or twodrug combination for 72 h. Cell viability was measured using a methanethiosulfonate (MTS)-based assay (CellTiter96 Aqueous One Solution; Promega), and read at 490 nm after 1-3 h using a BioTek Synergy 2 plate reader (BioTek). Cell viability was determined by comparing the absorbance of drug-treated cells to that of untreated controls (four replicates for each condition) set at 100%. The IC 50 values were calculated by a regression curve fit analysis using GraphPad Prism software.
Evaluation of combinatorial effect of combination drugs. We used Excess over Bliss (EOB) independence model 64 to quantify the synergy for crenolanib/trametinib combination at each drug concentration. EOB evaluates if the combined effect of two compounds is significantly greater or smaller than the combination of their individual (independent) effects and is measured by calculating the difference between the observed and predicted inhibition of the drug combination. For two single compounds with inhibition effects A and B, the predicted inhibition for the drug combination is calculated as C = A + B − A × B. The two-agent combination inhibition is defined as AB. EOB can be calculated by Z = AB − C. Z Plus score (>0) indicates a synergistic effect, and Z minus score (<0) indicates an antagonistic effect. We define EOB Z-score ≥0.2 as strongly synergistic and ≤ −0.2 as strongly antagonistic. The predicted combination viability of drug A and B combination is defined as (A + B − A × B) %. Highest single-agent (HSA) model 64 was also used to evaluate if the combined effect of two drugs compounds is significantly greater or smaller than the higher individual drug effect.
Statistical analysis. Statistical analyses were performed on GraphPad Prism software. The data were expressed as the mean ± SEM. Statistical significance was determined using two-tailed nonparametric Student's t-tests (Mann-Whitney test) or one-way analysis of variance (ANOVA) and expressed as p values (*p < 0.05, **p < 0.01, ***p < 0.001 and ****p < 0.0001).
Reporting summary. Further information on experimental design is available in the Nature Research Reporting Summary linked to this article.

Data availability
All sequence data have been deposited at dbGaP and Genomic Data Commons. The study ID is 29125 and the accession number is phs001628. In addition, all data can be accessed and queried through our online, interactive user interface, Vizome, at www.vizome.org.