Variational inference for pattern extraction and recognition in genome sequences using state space models for cancer detection

Abstract This study presents VIPER (Variational Inference for Pattern Extraction and Recognition), a novel deep-learning framework for cancer-causing mutation detection from genomic data. VIPER uniquely combines 1D Convolutional Neural Networks (Conv1D) with Mamba blocks, a structured state-space model architecture, to capture local and long-range dependencies in genomic sequences. Unlike traditional methods or state-of-the-art models like Transformers, VIPER offers enhanced scalability and computational efficiency, making it well-suited for large-scale genomic analysis. The model was evaluated on the Genome Screen Mutants VCF dataset and achieved a training accuracy of 96.84%, a validation accuracy of 97.30%, and an F1 score of 97.13%. These results outperform conventional deep learning models on the same dataset, including RNNs, CNNs, and Transformers. VIPER also significantly reduced computational overhead while maintaining high precision (97.52%) and recall (96.51%), highlighting its utility in clinical workflows. By focusing on clinically significant mutations, such as driver mutations associated with oncogene activation, VIPER provides actionable insights for precision oncology. This work advances the state of the art in cancer genomics by delivering an accurate, efficient, and interpretable solution for large-scale mutation detection. Future extensions may integrate multiomics data to enhance diagnostic capabilities further.