Icare 1: A Prospective Clinical Trial to Predict Treatment Response Based on Mutanome-Informed Computational Biology in Patients with AML and MDS

Background: Hypomethylating agents (HMAs) (e.g., azacitidine (aza), decitabine (dec)) and lenalidomide (len) are approved agents and used in the treatment of patients with myelodysplastic syndromes (MDS) or acute myeloid leukemia (AML). Despite their widespread use, HMAs fail in the majority of MDS and AML patients, and len fails in 75% of non-del(5q) MDS. Unfortunately, no method exists to predict disease response, thus the management of MDS and AML patients is challenging. Predicting treatment response would improve treatment effectiveness, restrict treatment-related adverse events to those who would benefit, and reduce health care costs. Ideally, patient prediction would be based on disease biology.

Aim: To determine the biological and clinical predictive values of a genomics-informed computational biology method in patients with AML and MDS who are treated with aza, dec or len.

Methods: Patients with AML or MDS were recruited in a prospective clinical trial (NCT02435550) designed to assess predictive values by comparing computer predictions of treatment response to actual clinical response. Genomic profiling was conducted by conventional cytogenetics, whole exome sequencing (SureSelectXT Clinical Research Exome, Agilent), and array CGH (Agilent). These genomic results were inputted into computational biology software (Cellworks Group), which generates disease-specific protein network maps using PubMed and other online resources. Digital drug simulations were conducted by quantitatively measuring drug effect on a cell growth score, which is a composite of cell proliferation, viability and apoptosis. Each patient-specific protein network map was digitally screened for the extent by which aza, dec or len reduced simulated disease growth in a dose-respondent manner. Treatment was physician's choice based on SOC. Before initiating treatment, treating physicians were masked to the results of whole exome sequencing and computational predictions. Clinical outcomes were prospectively recorded. To be eligible for efficacy assessment, patients must have had at least 4 cycles of HMA treatment or 2 cycles of len treatment. For AML, CR+PR was used to define response (IWG 2003). For MDS, CR+PR+HI was used to define response (IWG 2006). To validate the predicted protein network perturbations, Western blot assays were performed on pertinent pathway proteins. Comparisons of computer-predicted versus actual responses were performed using 2x2 tables, from which prediction values were calculated. Fisher's exact test was used to compare prediction values of the genomics-informed computer method versus empiric drug administration.

Results: Between June 2015 and June 2016, 80 patients were recruited. 40/80 (50%) had AML and 40/80 had MDS (50%). The median age was 66 (range 24-91). 44/80 (55%) were treatment-naïve and 36/80 (45%) were treatment-refractory. 99% completed all planned molecular tests and computational analyses. Laboratory validation study of computer-predicted, activated protein networks in 19 samples from 13 different patients showed correct prediction of 5 activated networks (Akt2, Akt3, PIK3CA, p38, Erk1/2) in 17 samples, exhibiting 89% accuracy. At the time of this report, 20/80 patients were eligible for efficacy evaluation. 6/20 patients showed clinical response to SOC therapy, while 14/20 did not achieve clinical response. 18 patients' outcome predictions were correctly matched to their actual clinical outcomes, and 2/20 were incorrectly matched, resulting in 90% prediction accuracy, 75% positive predictive value (PPV), 100% negative predictive value (NPV), 100% sensitivity, and 86% specificity. The accuracy of the genomics-informed computer method was significantly greater than empiric drug administration (p=1.664e-05). New genomic signature rules were discovered to correlate with clinical response after aza, dec or len.

Conclusions: A computational method that models multiple genomic abnormalities simultaneously showed high predictive value of protein network perturbations and clinical outcomes after standard of care treatments. The network method uncovered molecular reasons for drug failure and highlighted resistance pathways that could be targeted to recover chemosensitivity. This technology could also be used to establish eligibility criteria for precision enrollment in drug development trials.