Malignant Cancer Cell Classification Modelingusing Single-Cell RNA-Sequence from Glioblastoma Tumours

Glioblastoma (GBM) is the most aggressive primary brain tumour (WHO Grade IV), and its significant molecular heterogeneity poses challenges for diagnosis and treatment decisions. Although single-cell RNA sequencing (scRNA-seq) has greatly advanced our understanding of this diversity , reliance on manual cell annotation remains a key bottleneck. This study proposed a machine learning pipeline integrating feature selection (SelectKBest with ANOVA F-test) and a four-layer neural network to classify malignant cells in a dataset of 40,026 single cells (17,403 malignant, 22,623 non-malignant).The analysis employed ANOVA-based feature selection (SelectKBest with f_classif) to identify the top 30 most discriminative genes, including established markers (EGFR, MALAT1) and novel candidates. A four-layer neural network (128-64-32-1 architecture) was implemented using Keras, with standardized inputs and early stopping to prevent overfitting. The model demonstrated strong performance with 89.4% sensitivity for malignant cells and 88.8% specificity for non-malignant cells (overall accuracy=89.1%, ROC-AUC=0.93). Notably, the pipeline reduced manual annotation time by 92% while maintaining clinical-grade precision, though 11.9% of non-malignant cells were conservatively misclassified as malignant - likely reflecting transitional states in the tumor microenvironment.