Abstract 1: Novel pediatric AML patient risk stratification by inferred protein activity through integrative network analysis and machine learning

Abstract

Abstract Despite recent advances, overall survival for children with acute myeloid leukemia (AML) remains relatively low at 70 to 80%. Further refinement of risk stratification is paramount to guiding optimal therapy selection and improving outcomes. In this study, we leverage machine learning and network-based systems biology approaches that infer tumor protein activity and define key disease drivers - master regulator (MR) proteins - in pediatric AML to further refine existing patient risk stratification and identify potential therapeutic targets. Using the ARACNe and VIPER algorithms1, we reverse-engineered an AML gene regulatory network and inferred protein activity from enrichment of the proteins' transcriptional targets in gene expression signatures from a novel 1080-patient cohort treated on COG AAML1031. We integrated ensemble clustering with feed-forward feature selection on VIPER-inferred protein activity to identify groups of patients with unique protein activity signatures and correlated them to survival by Cox proportional-hazards regression analysis. We identified features predictive of each group by Monte Carlo cross-validation and binary classification. We found that the activity of 2-4 MR proteins was sufficient to predict patient stratification for each group with areas under the receiver-operator characteristic curve above 0.9. We validated our findings in three external test datasets: St-Jude Children's Research Hospital (n=249), TARGET (n=682) and Erasmus MC Sophia's Children's Hospital (n=237). We identified 10 groups with significant survival differences (P < 0.0001, overall survival range 42-85%). This includes a new high-risk group that is not associated with any known cytogenomic features, but shows high activity of NOTCH signaling pathway regulators HES1 and HEY2. Other groups show aberrant activity for known transcription regulators; for example RUNX1, KDM5B and POU4F1. In conclusion, we propose novel patient risk stratification based on VIPER-inferred MR protein activity in pediatric AML. We define 10 groups with differential survival, including refinement of a classically-characterized intermediate-risk cytogenomically normal group as high-risk. We classify patients into therapeutically-relevant risk groups that can guide treatment decision-making. We define the MR proteins driving these groups to offer mechanistic insights into AML tumorigenesis as well as suggest potential molecular targets for therapies. Reference: 1. Califano A, Alvarez M, The recurrent architecture of tumour initiation, progression and drug sensitivity, Nat Rev Cancer (2017), 17(2): 116-130 Citation Format: Alexandre P. Alloy, Jovana Pavisic, Rhonda E. Ries, Tanja Gruber, C Michel Zwaan, Monique L. den Boer, James R. Downing, Adolfo Ferrando, Soheil Meshinchi, Andrea Califano. Novel pediatric AML patient risk stratification by inferred protein activity through integrative network analysis and machine learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 1.