Identification of key genes in signaling pathways using machine learning exemplified on acute myeloid leukemia data 9214

Description The identification of differentially expressed genes (DEGs) is crucial for studying diseases. Classical approaches rely on statistical methods (e.g., LIMMA, edgeR). Here, we introduce a machine learning (ML) approach to identify key genes in signaling pathways using gene expression data. The GEO dataset GSE13159, containing data from acute myeloid leukemia (AML) patients and healthy controls (HC), was analyzed. Seven different algorithms were used: Linear Regression, Ridge Regression (RR), Random Forest, Gradient Boosting, k-Nearest Neighbors, Support Vector Machines, and Bayesian Ridge (BR). Models were trained and (cross-) validated on HC data using genes from KEGG pathways (here: hsa05221); interacting genes (features) were collected from STRING DB. Predictions were performed on AML data, and crucial genes were identified based on the prediction error. Best results were achieved using BR (ØR2=0.23) and RR (ØR2=0.20), specifically for the genes IKBKB (R2=0.89), CCNA2 (R2=0.84), and RUNX1 (R2=0.81). Poor predictions were observed for genes like JUP (R2=-1,38) or CCNA1 (R2=-0.81). Comparing to statistical approaches, 43 out of 57 DEGs identified with edgeR were also significant using ML. Notably, genes known to be associated with AML (e.g., MYC, NFKB1) were identified using ML but not edgeR. The results show the feasibility of using ML for gene expression prediction as well as the identification of crucial key genes, offering an innovative alternative to statistical methods. Funding Sources Basic Funding Upper Austria (Land OÖ Basisfinanzierung) Topic Categories Computational and Systems Immunology (COMP)