A MULTI-METHOD MACHINE LEARNING FRAMEWORK FOR IDENTIFICATION OF A CONSENSUS TRANSCRIPTOMIC SIGNATURE IN ACUTE MYELOID LEUKEMIA

Abstract

Acute myeloid leukemia (AML) is an aggressive form of hematological malignancy characterized by the abnormal proliferation of immature myeloid precursor cells, and it is highly heterogeneous at the molecular level. Although there have been great strides in the development of genomic profiles and targeted drugs, early diagnosis and molecular classification remain significant clinical challenges. The potential of transcriptomic profiling to identify disease-associated biomarkers, enhance diagnostic accuracy and provide insight into disease biology is considerable. We created and tested a comprehensive multi-method machine learning (ML) framework to uncover a strong consensus transcriptomic signature that distinguishes AML patients from healthy individuals. Transcriptomic data (10 AML samples and 10 healthy controls) was retrieved from a publicly available repository and processed using a strict pipeline of data preprocessing, normalisation and quality control (QC). After filtering low expressed genes, 13,329 expressed genes were retained, and the top 1,000 most variable genes were selected for downstream analyses. Four supervised ML algorithms were trained and tested using Leave-One-Out Cross-Validation (LOOCV): Least Absolute Shrinkage and Selection Operator (LASSO), Elastic Net, Random Forest (RF) and Support Vector Machine (SVM). Five different feature-selection methods were used to increase robustness and decrease algorithm-specific bias: LASSO, Elastic Net, RF importance ranking, SVM-Recursive Feature Elimination, and Boruta. All four classifiers had an area under the receiver operating characteristic curve (AUC) of 1.0, indicating that they could perfectly separate AML and healthy samples. A combination of several feature-selection methods resulted in an 11-gene consensus signature that included ARRB1, CD44, FAM178B, HDHD5, HOMER3, MRPS23, NRROS, NTPCR, SNHG32, TCL1A and TCN2. Nine of these genes showed 100% individual diagnostic accuracy; the multivariable logistic regression model had an AUC of 1.0, with 100% sensitivity, and 100% specificity. Biological processes and pathways important for AML pathogenesis, such as MAPK signaling, regulation of apoptosis, cytokine receptor activity, HIF-1 signaling, intracellular signal transduction, and differentiation of hematopoietic cells, were observed to be significantly involved in functional enrichment analysis. Together, these results demonstrate the potential of multi-method ML algorithms for identifying robust transcriptomic biomarkers in AML. The proposed 11-gene signature represents a promising approach for molecular-based diagnostic and may indicate biologically relevant targets that should be further validated in large independent cohorts and functional experimental studies.