Background Acute myeloid leukemia (AML), a biologically heterogeneous malignancy, requires improved prognostic models, particularly for patients with intermediate-risk profiles and lacking definitive genetic markers. Therefore, this study aims to identify biologically coherent and clinically informative gene signatures using a novel prognostic modeling approach integrating gene expression profiles with protein–protein interaction networks. Methods We applied network propagation using Personalized PageRank with seed genes from a literature-based six-gene signature (LBS6) and two recurrent AML mutations (IDH1 and IDH2). Network-informed modules were derived and optimized using LASSO–Cox regression models trained on the TCGA–LAML cohort (n = 132, adult AML) and externally validated in the BeatAML 1.0 (n = 308, adult AML) and TARGET–AML (n = 1,889, pediatric AML) cohorts. Cox proportional hazards models were used to evaluate associations with overall survival. Functional enrichment analyses were conducted using KEGG and Gene Ontology databases. Results From LBS6 propagation, a ten-gene signature (LBS6-Derived Network Gene Signature [LBSnet]: PTP4A3, HS3ST3B1, ECHS1, PLA2G4A, ETFB, NDST3, CSK, ARL6IP5, PLD1, and NDUFS8) was derived, stratifying patients in the TCGA–LAML based on overall survival (HR = 3.84, p < 0.0001) and was validated in the BeatAML 1.0 (HR = 1.94, p < 0.0001) and TARGET–AML (HR = 1.57, p < 0.0001) cohorts. Joint network propagation using IDH1 and IDH2 seed genes produced a five-gene signature metabolic and chromatin-modifying functions (G6PD, ENO1, SDHA, H3-3 A, IL4I1), demonstrating prognostic significance in the TCGA–LAML cohort (HR = 2.91, p < 0.0001), BeatAML 1.0 (HR = 1.33, p = 0.07), and TARGET–AML (HR = 1.34, p < 0.001). These network-derived risk scores remained independent predictors of overall survival in multivariate Cox models adjusted for age and key genetic covariates, including FLT3-ITD, NPM1, and CEBPA mutations. Functional enrichment analyses revealed significant involvement in fatty-acid oxidation, mitochondrial respiration, and platelet activation pathways. Conclusion This study presents a novel network-based framework for prognostic modeling in AML, generating biologically interpretable gene signatures with validated predictive power across adult and pediatric cohorts. Integrating transcriptomic data with molecular interaction networks provides a scalable strategy for biomarker discovery, enhancing risk stratification and offering insight into potential metabolic vulnerabilities in AML.

Integrated network propagation identifies prognostic metabolic signatures in acute myeloid leukemia

期刊论文
出版模式：

开放获取
发表日期：

2025年12月3日

查看全文

Jong Keon Song,
Hyery Kim,
Sang-Hyun Hwang

摘要

Background

Acute myeloid leukemia (AML), a biologically heterogeneous malignancy, requires improved prognostic models, particularly for patients with intermediate-risk profiles and lacking definitive genetic markers. Therefore, this study aims to identify biologically coherent and clinically informative gene signatures using a novel prognostic modeling approach integrating gene expression profiles with protein–protein interaction networks.

Methods

We applied network propagation using Personalized PageRank with seed genes from a literature-based six-gene signature (LBS6) and two recurrent AML mutations (IDH1 and IDH2). Network-informed modules were derived and optimized using LASSO–Cox regression models trained on the TCGA–LAML cohort (n = 132, adult AML) and externally validated in the BeatAML 1.0 (n = 308, adult AML) and TARGET–AML (n = 1,889, pediatric AML) cohorts. Cox proportional hazards models were used to evaluate associations with overall survival. Functional enrichment analyses were conducted using KEGG and Gene Ontology databases.

Results

From LBS6 propagation, a ten-gene signature (LBS6-Derived Network Gene Signature [LBSnet]: PTP4A3, HS3ST3B1, ECHS1, PLA2G4A, ETFB, NDST3, CSK, ARL6IP5, PLD1, and NDUFS8) was derived, stratifying patients in the TCGA–LAML based on overall survival (HR = 3.84, p < 0.0001) and was validated in the BeatAML 1.0 (HR = 1.94, p < 0.0001) and TARGET–AML (HR = 1.57, p < 0.0001) cohorts. Joint network propagation using IDH1 and IDH2 seed genes produced a five-gene signature metabolic and chromatin-modifying functions (G6PD, ENO1, SDHA, H3-3 A, IL4I1), demonstrating prognostic significance in the TCGA–LAML cohort (HR = 2.91, p < 0.0001), BeatAML 1.0 (HR = 1.33, p = 0.07), and TARGET–AML (HR = 1.34, p < 0.001). These network-derived risk scores remained independent predictors of overall survival in multivariate Cox models adjusted for age and key genetic covariates, including FLT3-ITD, NPM1, and CEBPA mutations. Functional enrichment analyses revealed significant involvement in fatty-acid oxidation, mitochondrial respiration, and platelet activation pathways.

Conclusion

This study presents a novel network-based framework for prognostic modeling in AML, generating biologically interpretable gene signatures with validated predictive power across adult and pediatric cohorts. Integrating transcriptomic data with molecular interaction networks provides a scalable strategy for biomarker discovery, enhancing risk stratification and offering insight into potential metabolic vulnerabilities in AML.