Machine learning for time-series genomics: Advances, challenges, and future directions

Recent developments in single-cell and high-throughput sequencing technologies have enabled the collection of genomic data at multiple time points; therefore, the investigation of dynamic processes and the observation of disease prognosis, progression, and therapy effects are now possible. Yet, challenges, including biological variability, high dimensionality, noise, and temporal resolution in these datasets, need to be addressed. This review aims to categorize the machine learning (ML) methods used to create models that analyse the temporal dynamics of genomic data. In addition, the best way to apply either the classical or modern deep learning approaches to analyse temporal genomic and multi-omics datasets is being discussed. ML methods used in time-series genomics analysis have been organized by data type and modeling approach to compare their efficiency in handling sparsity and nonlinear temporal relationships, thereby helping to understand the biological mechanisms underlying them. Furthermore, we focus on the recent frameworks that combine temporal learning models with long-read and single-cell sequencing. While ML provides robust tools to reveal temporal dynamics, data standardization, scalability, interpretability, and benchmarking remain challenges. We summarize best practices for model evaluation and outline future directions, including multi-omics integration, interpretable artificial intelligence, and large-scale, reproducible benchmarks. Via combining computational ingenuity and biological insight, we enable a deep understanding of the complexity of biological processes and pave the way for applying precision medicine through ML-based time-series genomics.