KAN and DeepInsight: Advanced Deep Learning Methods for Pan-Cancer Gene Expression Classification

Gene expression profiles constitute a quantitative source of molecular information on the transcriptional activity of malignant tumors. In contrast to the limitations and subjectivity inherent to microscopic histopathologic observation, mRNA expression profiling offers a more objective, reproducible, and detailed classification method. This study proposes a sophisticated pan-cancer approach that integrates preprocessing techniques based on feature selection and dimensionality reduction. These techniques are critical in managing high redundancy and noise that is characteristic of gene expression data. This management is essential for extracting pertinent disease information. The performance of several machine learning models is then evaluated, including Kolmogorov-Arnold networks (KAN). Furthermore, the efficacy of an improved DeepInsight algorithm, which facilitates the conversion of data into image format, is investigated. Subsequently, the transformed data is analyzed using a neural architecture designed for visual data processing. The approach was evaluated using several datasets containing expression profiles of different cancer types to assess its robustness and generalizability. The findings indicate that this framework notably enhances the precision of pan-cancer classification, providing novel insights.