Preliminary Lasso Regression Analysis Identifies DNA-Damage Gene Expression Signature Predictive of Clinical Outcomes in Patients Using Gemtuzumab Ozogamicin

Abstract

Background: In September 2017, the FDA announced re-approval of gemtuzumab ozogamicin (GO), a CD33-directed antibody-drug conjugate, for treatment of newly diagnosed and relapsed/refractory AML setting the pace for a new era of personalized therapeutic options for AML. While the future of GO as a therapy in AML is bright, studies have shown that clinical response to GO is subject to interpatient variability (Gbadamosi 2018). Hitherto, efforts to understand the interpatient variation have focused on profiling CD33 expression levels (Pollard et al. 2016) and more recently on genetic variation in CD33 and ABCB1 predicative of GO response (Chauhan et al. 2019, Rafiee et al. 2019). However, efforts to understand interpatient variability in regards to calicheamicin, the DNA-damaging cytoxin linked to the antibody portion of GO, have been limited. Thus, we hypothesized that interpatient differences in expression levels of genes involved in pharmacodynamic effects of calicheamicin may impact GO response/resistance. Methods: Herein, our group has used the least absolute shrinkage and selection operator (LASSO) regression analysis method to identify critical genes with expression levels predictive of clinical outcomes in response to GO. Using data from the TARGET database, the expression levels of 18 genes, selected for their role in calicheamicin induced DNA-damage response, were extracted for AML patients treated with GO in the AAML03P1 and AAML0531 clinical trials (N=128; Table 1). Using a penalized LASSO regression algorithm (glmnet R-package), a Cox regression model was fit on to the expression levels of the selected gene. A thousand bootstraps of LASSO regression were performed and genes included in greater than 95% of the models were further selected. Average coefficients from the LASSO model were used to generate a gene signature equation designated as the DNA-damage response score (DDRS) given the nature of the genes included. Patients were classified into high (DDRS High) or low (DDRS Low) values for DDRS according to the median value and evaluated using Cox-proportional hazard model for survival data analysis, while the Chi-square test was used to examine the differences between categories, and the Wilcoxon rank-sum test was used to assess the differences in averages where appropriate. Results: The DDRS equation was defined as DDRS = (AKT1*-0.070) + (CASP9*-0.091) + (H2AFX*-0.160) + (XRCC4*0.373) where gene expression levels are multiplied by the coefficients obtained from the LASSO regression (Figure 1). Patients in the DDRS High group had significantly worse event free survival (EFS; HR = 2.22, P < 0.001; Figure 2A), lower complete remission rates (67.7% vs 94.6%, P < 0.001; Figure 3A), and a trend towards worse overall survival (OS; HR = 1.70, P = 0.058) as compared to patients within DDRS Low group. DDRS score was also significantly different amongst patients that were MRD+ve vs. MRD-ve after induction 1 therapy (21% difference, P = 0.038; Figure 3B). Consistent results were seen within standard risk group patients where patients in the DDRS High group had significantly worse EFS (HR = 2.29, P = 0.01; Figure 2B) as compared to those in the DDRS Low group. In preliminary multivariate Cox regression analysis, the DDRS remained a significant predictor of EFS amongst age, risk group, FLT3 status, and WBC (HR 2.42, P < 0.001; Table 2). Conclusion: Our preliminary investigation using LASSO regression model defined DDRS, a gene signature predictive of clinical response in patients treated with GO. The model included four genes well known for their involvement in DNA-damage response: AKT1, a kinase that regulates cell growth and division; CASP9, The initiator caspase involved in apoptosis; H2AFX, a DNA-damage marker that recruits other DNA-damage response proteins to damaged loci; and XRCC4, a ligase for DNA damage repair. Our future work focuses on expanding this investigation in bigger cohorts of patients representing different risk groups of AML as well as in vitro mechanistic studies. Once validated for its sensitivity and specificity to calicheamicin response, our results hold promise towards developing strategies for understanding interpatient variability in GO response and personalizing GO therapy based on diagnostic gene expression signatures. Disclosures No relevant conflicts of interest to declare.