MULTI-LEVEL FEATURE INTEGRATION IDENTIFIES A CONSENSUS TRANSCRIPTOMIC SIGNATURE FOR ACUTE MYELOID LEUKEMIA WITH NEAR-PERFECT DIAGNOSTIC DISCRIMINATION

Background: Acute myeloid leukemia (AML) is a heterogeneous hematologic malignancy where the need is for strong markers of diagnosis with high generalizability across independent cohorts. The limited number of studies performed so far have used only one data set, a limited sample of healthy controls, and have not been externally validated.

 Methods: We combined three independent RNA-seq datasets—Dryad, GSE201492, GSE327386—with 303 samples (285 AML, 18 healthy bone marrow) into a common expression matrix of 12,446 common genes that was batch-corrected. The differential expression analysis (limma, adj.Using the criterion (P<0.05, |logFC|>1), 1,661 DEGs were identified and ranked based on the composite score generated by incorporating consistency, fold separation, coefficient of variation and healthy-donor penalty. Random forest (5,000 trees) and linear support vector machine (SVM) feature selection algorithms were applied to the top 100 genes. The resulting 19-gene consensus panel was internally validated using a stratified 5-fold cross validation with bootstrap optimism correction, further validated on GSE317967 (3 AML, 5 healthy BM), cross validated with already known AML literature, COSMIC Cancer Gene Census, known leukemia stem cell markers, and assessed for drug-target potential.

Results: The 19-gene consensus list contained the genes: PCGF1, RNF220, SQSTM1, IRF2BP2, LAPTM4A, RPS27, VPREB3, BAIAP2, RAG1, GNA15, CD99, UBE2E1, HOMER3, MME, HBG2, IL2RB, KLRD1, PRG2, and HBG1. Near perfect discrimination was obtained with internal stratified 5-fold cross validation (RF AUC=1.0, SVM AUC=1.0, NN AUC=1.0) and with minimal use of bootstrap optimism (0.001). Subsets of the top 5 genes kept AUC>0.997. Three genes (RNF220, CD99, GNA15; AUC=1.0) and moderate panel level discrimination (combined AUC=0.667) was found in external validation with GSE317967, which is an independent cohort. Cross-referencing literature revealed 18 of 19 consensus genes are novel candidate AML biomarkers whereas only MME (CD10) has previously been associated with AML. Two genes in this wider set of top 100 (MECOM, WT1) are in the COSMIC AML census, and five others (MME, PBX1, IL1RAP, WT1, CD44) in AML literature. There is no consensus gene that is a known LSC marker. Ribosomal biogenesis, immune activation (TNF, NF-kB), hematopoietic cell lineage, and B-cell receptor signaling pathways were identified as the top four GO and KEGG pathways. The hub genes PLOD2, MAGI2-AS3, CDC42BPA and PLEKHG1 were identified through co-expression network analysis. 20 curated consensus genes were identified from drug-target literature mining, of which 18 were found to be related to a therapeutic intervention. High priority targets included those targeted by CD10-directed ADC, CD99-directed monoclonal antibody and IL1RAP CAR-T therapy.

 Conclusions: The multi-cohort integrative pipeline yielded a 19-gene consensus biomarker panel with good internal discrimination and moderate external validation. The top down-regulated candidates (VPREB3, RAG1, PAX5, EBF1, VPREB1) were predominantly of B-cell lineage and were probably not AML-specific but were more likely to be associated with myeloid replacement of B-lymphopoiesis in AML bone marrow. Three genes, RNF220, CD99 and GNA15, replicated perfectly across independent cohorts and deserve to be prioritised for validation. It is important to note that there was only limited representation of healthy controls (n=18) across all training cohorts, and that the performance of the model was lower on external testing (AUC=0.667) than on internal testing (AUC≈1.0), reflecting the importance of larger, well-powered external validation studies with adequate normal bone marrow reference samples.