Azacitidine Response Prediction in MDS Patients with NGS Data Using a Computational Biology Modeling (CBM) Based Clinical Decision Support System

BACKGROUND: Azacitidine (AZA) is currently a drug of choice for most of high-risk MDS patients. However, only 40-50% of MDS patients achieve clinical improvement with AZA. There is a need for a predictive clinical decision support tool that can identify MDS patients with higher or lower likelihood of AZA response. Ideally, patients with no chance of response would be spared of life-threatening toxicities and expense; while patients with high chance for response would receive maximized treatment.

AIM: The objective of this study was to predict response to AZA in a cohort of intermediate- and high-risk MDS patients using CBM approach in retrospective blinded manner.

METHODS: We analyzed the clinical and genomic (NGS, cytogenetics and FISH) data for a cohort of 48 Int-2 and high risk MDS patients treated with AZA for median of 12 (4-34) cycles. Median age was 70.4 years, M:F ratio 1:1. MDS subtypes: EB-2 (35.4%, N=17), EB-1 (39.6%, N=19), MDS/AML (18.6%, N=9), CMML1/2 (4.2%, N=2), and RARS (2.1%, N=1). Median IPSS-R was 5 (3-10), cytogenetic score was 1. Median BM blasts were 10% (0.88-43.12), Hemoglobin 91g/L (62-145), ANC 1.24x10⁹/mL (0.09-11.64), Platelets 84x10⁹/mL (2-576). Patients were treated by AZA until progression to AML. Median AZA cycles was 14. Median overall survival (OS) on therapy was 24.2 months (4.4-61.6). Median progression free survival (PFS) was 15.9 months (4.0-61.6). Clinical responders were defined by IWG2006 criteria. Overall response rate (ORR) was 60.4% (CR/PR: 18 of 48 patients, stable disease with hematology improvement (SD-HI): 11/48). While 29 of 48 (60.4%) patients progressed to AML following the AZA therapy; 5 patients (10.4%) were primarily AZA-resistant.

CBM for each MDS patient was created utilizing genomic data to create a predictive workflow (Cellworks Group) complemented with digital mechanistic model of AZA and other FDA-approved drugs. Drugs were modeled by programming their mechanism of action on pathways and simulated individually and in combination. A disease inhibition score (DIS) characterized the drug impact to which malignant phenotypes was inhibited. For AZA non-responder profiles, unique combinations were selected that reduced DIS.

RESULTS: CBM cohort analysis performed on 37 out of 48 patients predicted 20 clinical responders (54.05%) and 17 clinical non-responders (45.95%). CBM accurately predicted the clinical outcomes of 14 out of 20 responders and 17 out of 17 non-responders with overall accuracy 83.78%. Sensitivity of identifying a responder is 70% while non-responders are called with 100% specificity. The CBM identified AZA based combination in 17 patients who did not respond to AZA monotherapy: AZA+Lenalidomide (N=6), AZA+Dasatinib (1), AZA+Ruxolitinib (2), AZA+Sorafenib (1), and AZA+Venetoclax (7). In the patient AZA014 (EB1 with isolated 5q-) who underwent AZA for 17 cycles and responded with mCR (MLD) without HI, the CBM predicted Lenalidomide sensitivity. Following the treatment with Lenalidomide the patient entered a long lasting CR (> 3 years). Digital droplet ddPCR technology during patient follow-up confirmed loss of the clone characterized by TP53(p.Thre377Pro) mutation upon Lenalidomide therapy. In contrast, ddPCR of AZA048 utilized SF3B1 (p.Lys700Glu) mutation tracking during patient follow-up that predicted the relapse by 8 months.

CONCLUSION: The CBM prediction of AZA clinical response in newly diagnosed, higher risk MDS patients showed high predictive accuracy of AZA resistant patients. The study validates the approach to a priori predict response and identify the right therapy option for the patient and could be used to establish criteria for precision enrollment in drug development trials. In addition, analysis uncovered possible mechanisms for AZA resistance that could be targeted to induce response. Clonal architecture is trackable using ddPCR technology providing time for additional NGS analysis to target the progressing clone. Ultimately, improving patient selection and mutation tracking may avoid unnecessary chemotherapy toxicity and reduce health care expenses.