In Acute Myeloid Leukemia (AML), leukemic stem cells (LSCs) play a critical role in disease initiation and in sustaining the leukemic bulk of immature cells, which infiltrate the bone marrow and other tissues, leading to clinical symptoms. Due to its marked heterogeneity, AML comprises distinct cellular subtypes with varying degrees of immaturity (stemness), contributing to treatment resistance through the persistence of a quiescent population. Although stemness pathways have been explored using molecular signatures such as LSC17, current scoring systems often fail to uncover therapeutic vulnerabilities. Furthermore, the large volume of data associated with cellular immaturity may include irrelevant variables that compromise precision.

This project aims to integrate and refine stemness-related gene signatures using machine learning to identify key cellular dependency processes that could support the development of more effective risk-adjusted therapies.

For model training, transcriptomic and clinical data from 121 AML patients in The Cancer Genome Atlas (TCGA) were used. A total of 137 functionally validated stemness-associated genes were included and modeled using LASSO regression (via the glmnet package in R), with overall survival as the response variable. The model’s robustness and generalization capability were evaluated in independent cohorts (BEAT-AML, HOVON, GSE71014, GSE37642). Patients were stratified into high- and low-risk groups using the cutpointr package. To confirm the biological relevance of the index in terms of cellular immaturity, immune cell deconvolution was performed using CIBERSORTx. Additionally, differential expression analyses employing Gene Ontology (GO) and Gene Set Enrichment Analysis (GSEA) were used to identify pathways differentially enriched between groups.

The final LASSO model yielded a six-gene prognostic panel—SPI6 (Stemness Prognostic Index 6): TFPT, CALCRL, IL2RA, PLXNC1, SPINK2, and IL3RA. Model validation (Harrell’s C-index, λ = 0.06) demonstrated strong predictive performance, with SPI6 outperforming (AUC = 0.742) the established LSC17 signature (AUC = 0.612). After dichotomize patients into high and low SPI6 groups, Kaplan-Meier analysis revealed significantly poorer overall survival among patients in the high SPI6 group (p < 0.0001), a finding consistently validated across all external cohorts (BEAT-AML, HOVON, GSE71014, GSE37642). In multivariate Cox regression analyses, adjusted for confounders such as white blood cell count (WBC), gender, age and ELN risk stratification, SPI6 remained an independent and significant predictor of prognosis in all datasets (p < 0.001). CIBERSORTx analysis showed a higher proportion of primitive (immature) cells in the high SPI6 group. GO-based differential expression analysis revealed reduced dependence on mitochondrial metabolism in high-risk patients. This was supported by GSEA, which identified negative enrichment of the oxidative phosphorylation pathway (NES = –2.07, FDR = 0.003) and positive enrichment of the cholesterol metabolism pathway (NES = 1.5, FDR = 0.01) in the SPI6-high group.

In summary, this study developed and validated a novel six-gene stemness-based prognostic index for AML. SPI6 outperformed existing markers (LSC17) and was strongly associated with poor survival and distinct metabolic profiles, supporting future drug repurposing strategies for high-risk patients.