Deep learning models to improve Variant calling in next generation sequencing data

Abstract

Next-generation sequencing (NGS) technologies have transformed genomic research by enabling high-throughput and cost-effective DNA analysis, unlocking new possibilities in understanding genetic variation. However, the accurate detection of genomic variantsespecially in clinical and cancer-related contextsremains a persistent challenge due to issues such as sequencing noise, low coverage, and complex variant patterns. Overcoming these limitations is critical to achieving the full potential of precision medicine. In recent years, artificial intelligenceparticularly deep learninghas emerged as a powerful approach for modeling and interpreting complex biological data. This thesis investigates deep learning-based solutions for improving variant calling from NGS data, with a focus on image-based representations. We propose two pipelines: the first reconstructs a complete variant calling workflow using public datasets (e.g., NCBI SRA), from raw sequencing reads to pileup image generation and classification with convolutional neural networks (CNNs). The second pipeline leverages clinical BAM and VCF files, integrating Oncomine annotations to create a multimodal model that combines visual and genomic features for variant classification and interpretation. We further incorporate multi-label and semi-supervised learning techniques to enhance model performance on noisy and partially labeled datasets. Experimental results demonstrate that our deep learning models outperform traditional variant calling approaches in both accuracy and interpretability. This work contributes to the field of computational genomics by demonstrating the value of integrating deep learning, image-based data, and structured clinical annotations in building scalable and clinically relevant variant detection tools.