INTRODUCTION: Newer AML therapies with Hypomethylating Agents & Venetoclax (HV) provide an alternative to Conventional Chemotherapy (CC), with anthracycline and cytosine arabinoside, but criteria to select which to use are lacking. We investigated whether proteomic classification could provide guidance. METHODS: Using Reverse Phase Protein Array (RPPA) methodology, the level of 405 proteins (327 total; 78 post-translational modified) were measured in 818 samples from newly diagnosed adult AML patients and normalized to normal CD34+ cells. Patients treated with HV (N=85) or CC (N=335) were analyzed. Overall Survival (OS) and Complete Remission Duration (RD) analysis was performed with pairwise LogRank tests (p < 0.05), adjusted for False Discovery Rate (FDR), to identify individual proteins predictive of outcome. Patients were clustered based on protein expression. Categorical clinical variables were compared with Fisher's Exact test with simulated p-values (10000 replicates), and continuous ones with Kruskal-Wallis test. Univariate (UV) and Multivariate (MV) models were built with Cox proportional hazards model (CoxPH). Deep neural network analysis using random forest was performed to create a Protein Classifier for treatment recommendation. RESULTS: Individual protein analysis found 55 prognostic proteins (Protein Selector 1 - PS1), which clustered patients into 3 groups. A preferred therapy for cluster 2 was not identified with PS1, so a second Protein Selector 2 (PS2) was generated, identifying 45 predictive proteins in this subset. From the PS1/2 analysis, the 95 proteins formed 4 distinct expression profiles (C1=red, C2=gold, C3=magenta, C4=green) which were stratified by treatment (HV or CC), as shown in Fig. 1. Patients in C2 were enriched for secondary AML (57% vs C1=41%, C3=33%, C4=32%; p<0.001) and unfavorable cytogenetics (41% vs C1=32%, C3=39%, C4=35%; p<0.001), and were older (60yo vs C1=61yo, C3=54yo, C4=56yo; p<0.001). Mutations in CEBPA, GATA2, FLT3 and NPM1 were more common in C4 (22%, 13%, 35%, 35% vs C1=12%, 1.7%, 5.7%, 6%; C2=7.8%, 6.2%, 18%, 15%; C3=9.3%; 1.4%, 29%, 21%; p=0.037, p=0.013, p<0.001, p<0.001, respectively), highlighting the heterogeneity between groups. We evaluated OS and RD (Fig. 2) in each cluster (C1=red, C2=gold, C3=magenta, C4=green), considering each treatment (HV= solid line, CC=dashed line). Patients in C1 and C2 treated with HV had a diametrically different and superior response compared to those treated with CC (C1-HV: Median OS (MS)>65mo, Median RD (MRD)>60mo; C2-HV: MS=15mo, MRD=24mo vs C1-CC: MS=20mo, MRD=39mo; C2-CC: MS=10mo, MRD=17mo). C3 showed the opposite, with CC outperforming HV (MS=19mo, MRD>102mo vs MS=7mo, MRD=9mo). Similarly, in C4, CC patients did dramatically better, (MS=40mo, MRD>96mo vs MS=9mo, MRD=7mo). In the UV CoxPH model, we included variables for PS cluster and treatment and all groups predicted OS, except for C3-CC and C4-CC, as did age, secondary AML, complex karyotype, and CEBPA, FLT3, KIT, MLL, NPM1, PTPN11 and TP53 mutations. In the MV model, all clusters were predictive for OS, together with age, secondary AML, complex karyotype and NPM1 mutation. Regarding RD, in the UV model, only C2-HV, C2-CC, and C4-HV predicted RD, as did age, secondary AML, complex karyotype, unfavorable cytogenetics, and FLT3, RUNX1 and TP53 mutations. In the MV model of RD, all clusters, except for C3-HV were predictors, together with age. Importantly, optimized treatment triaging based on PS1/2 would have reassigned 52% (218/420) of cases, and is estimated to increase the 3yrs OS of the whole cohort from 35.6% to 48.8%, a 37.2% increase. Moreover, the 2yrs Relapse Free Survival would increase from 57,9% to 63,9%, a 10.39% increment. Furthermore, our Protein Classifier identified a set of 3 proteins (SPI1, NOTCH1.cle, and TGM2) whose expression can be used to rapidly classify patients and recommend either HV or CC. This classifier has a very high c-index (a measure of individual patient discriminatory power) of 0.93 (>0.7 is considered predictive, 1 would indicate perfection) demonstrating that it robustly predicts optimal therapy choice. CONCLUSION: Protein profiling identified protein expression patterns which had diametrically different responses according to treatment (HV or CC). Given the US annual incidence of 20000 newly diagnosed AML cases, proteomic triaging could result in ~2600 more cures using existing therapy. Figure 1View largeDownload PPTFigure 1View largeDownload PPT Close modal
Read full abstract