A Comparative Evaluation of Batch-Normalized Multi-Layer Perceptron Models for Multi-Class Cancer Classification Using TCGA RNA-Seq Data

The classification of cancer types using gene expression data presents significant challenges due to high dimensionality, limited sample sizes, and class imbalance inherent in genomic datasets. The Cancer Genome Atlas (TCGA) has enabled large-scale pancancer studies by providing standardized RNA-Seq gene expression profiles across multiple cancer types. In this work, a comparative evaluation of Multi-Layer Perceptron (MLP) models incorporating Batch Normalization is presented for multi-class cancer classification using TCGA-derived RNA-Seq data. The dataset comprises 801 tumor samples with 20,531 gene expression features spanning five cancer types: BRCA, KIRC, LUAD, PRAD, and COAD. To ensure robust and unbiased performance estimation, stratified five-fold cross-validation is employed, and class imbalance is addressed through the use of class-weighted training and macro-averaged evaluation metrics. Model performance is assessed using accuracy, macro-averaged precision, recall, and F1-score. Experimental results indicate that the inclusion of Batch Normalization improves training stability and generalization compared to standard MLP configurations, while deeper or more complex architectures offer no additional benefit for this high-dimensional setting. The findings highlight the importance of careful architectural design and rigorous evaluation protocols when applying deep learning models to genomic cancer classification tasks. and cancer classification based on transcriptomic signatures (Figure 1). RNA-Seq, a next-generation sequencing method, has emerged as the preferred approach to gene expression profiling due to its high sensitivity, wide dynamic range, and ability to detect novel transcripts. Unlike earlier microarray technology, RNA-Seq does not rely on predefined probes, allowing more accurate and unbiased measurement of transcript abundance across the genome. This technology has facilitated large-scale projects such as The Cancer Genome Atlas (TCGA), which has generated RNA-Seq data encompassing dozens of cancer types with matched clinical annotations. These public resources have become benchmarks for developing and evaluating computational methods for cancer classification. Machine learning (ML) methods have been widely applied to gene expression data for cancer classification. Conventional algorithms such as Support Vector Machines, Random Forests, and k-Nearest Neighbors have demonstrated utility in discriminating among cancer types but often require manual feature selection or dimensionality reduction to handle the high dimensionality of transcriptomic datasets. Deep learning (DL) techniques, including neural networks with multiple hidden layers, have shown promise in automatically learning complex nonlinear relationships in high-dimensional data without extensive feature engineering. Among these, multi-layer perceptrons (MLPs), a type of fully connected neural network, have been used in prior studies for multi-class classification tasks in cancer genomics. Despite the potential of deep learning models for cancer classification, several challenges remain. Gene expression datasets typically contain far more features (genes) than samples, which can lead to model overfitting if not addressed with appropriate architectural design and validation protocols. Moreover, class imbalance — where some cancer types are represented by many more samples than others — can bias learning and evaluation metrics unless handled explicitly. Proper normalization, model regularization, and rigorous evaluation such as