Write a PREreview

The rapid expansion of high-throughput sequencing technologies and the completion of telomere-to-telomere (T2T) assemblies have transformed genomics into a data-driven discipline, shifting the research focus from data generation to large-scale computational discovery. This literature review synthesizes foundational and emergent pathways in bioinformatics, genomics, and their integration into health applications. We examine the critical role of genomic reproducibility and benchmarking in establishing clinical trust, alongside mathematical models for comparative genomics, such as the Double-Cut-and-Join (DCJ) distance. A significant portion of this review is dedicated to methodological shifts in representation learning, specifically evaluating the impact of Byte-Pair Encoding (BPE) tokenization on genomic language models and the dominance of repetitive elements in sequence vocabularies. Furthermore, we explore the evolution of deep learning architectures, contrasting traditional convolutional and recurrent neural networks with recent advancements in State Space Models (SSMs). These emergent architectures, such as Caduceus and Mamba, demonstrate linear-time complexity and superior performance in capturing long-range regulatory dependencies across ultra-long genomic sequences. Finally, we discuss how these computational innovations converge to support the goals of precision medicine. By mapping these trajectories, this review provides a comprehensive overview of the technical and theoretical challenges inherent in modeling the complexity of the human genome for clinical and biological insights.