COMPOSITE TRANSCRIPTOMIC SCORING AND DEEP LEARNING FRAMEWORK FOR ACUTE MYELOID LEUKEMIA DIAGNOSIS USING MULTI-COHORT RNA-SEQ INTEGRATION
Keywords:
Acute myeloid leukemia; RNA-Seq; batch correction; machine learning; deep learning; SHAP; biomarker discovery; ribosome biogenesis; UBE2N; FLT3Abstract
Background: Acute myeloid leukemia (AML) is a highly aggressive hematological malignancy with heterogeneous molecular features, which present a diagnostic challenge in terms of accuracy. The ability to study gene expression changes in AML is unprecedented using transcriptomic approaches based on RNA sequencing (RNA-Seq).
Methods: Two independent RNA-Seq datasets were included in a combined expression matrix of 12,675 common genes: a Dryad dataset (n = 20; 10 AML, 10 healthy controls) and GSE201492 (n = 157 AML). ComBat was used to correct for batch effects. Using differential expression analysis (limma) identified 8,325 differentially expressed genes (DEGs). A new six-biological-metric composite scoring system identified the top 100 candidate genes. A consensus panel of three genes (CYCS, POLR1C and ZPR1) was identified by Random Forest and Linear Support Vector Machine classifiers. The top 100 features were used to train a deep learning model called a multilayer perceptron (MLP), and then interpret the model using the SHAP and LIME explainability frameworks.
Results: The MLP classifier achieved perfect discrimination (AUC = 1.000, accuracy = 100%) on the held-out test set. Each consensus gene panel (CYCS, POLR1C, ZPR1) had an AUC = 1.000. Using the SHAP analysis, the most predictive feature was identified as UBE2N (ubiquitin-conjugating E2 N enzyme) and was validated by concordant LIME analysis. Pathway enrichment revealed that the genes were mainly associated with ribosomal biogenesis and cytoplasmic translation, which is in line with the current models of AML pathobiology.
Conclusions: This integrative approach shows that a short, biologically meaningful transcriptomic biomarker can accurately distinguish AML from normal bone marrow with high accuracy. The biomarkers identified, especially UBE2N, CYCS and the ribosomal gene cluster, are promising targets for prospective clinical use for diagnostics or therapeutics.
