An Overview of Computational Methods for Gene Prediction in Eukaryotes: Strengths, Limitations, and Future Directions

The advancements in Next-Generation Sequencing (NGS) technologies and machine learning techniques have contributed to recent improvements in gene prediction in eukaryotes. Nevertheless, the computational prediction of genes remains a crucial task for complementing empirical data and annotating newly sequenced genomes. This task is challenging due to large genomes and the complex structure of their gene sequences, which include exons, introns that vary in length and composition, as well as nested and overlapping genes. The recent emergence of deep learning methods has significantly improved gene prediction accuracy. Without relying on predefined gene models, these approaches can learn complex patterns of gene structures from genomic data, enabling robust predictions that generalize across species. This advancement motivates the development of a new classification system for gene prediction methods. Unlike traditional classifications, which typically group tools by the type of empirical input data, this review categorizes gene prediction techniques according to their use of gene model information: gene-model-based, gene-model-free, and hybrid approaches. It also offers a dedicated discussion on the types of data used by these tools. The review examines representative methods within each category, highlighting their algorithmic strategies, input data types, strengths, and limitations. It also presents an updated overview of challenges previously identified in earlier reviews and introduces new limitations posed by recent techniques.