Foundation models in omics research: a comprehensive survey

The rapid expansion of high-throughput omics has created molecular datasets of unprecedented scale and complexity. These data are rich in biological information yet inherently sparse and high-dimensional, often limiting the effectiveness of conventional machine learning techniques. Foundation models (FMs), built on large-scale self-supervised pretraining, offer a robust alternative by learning generalizable representations directly from raw biological data. This review systematically analyzes the emerging landscape of FMs in omics research, spanning sequence modeling, cell state characterization, and multimodal integration. We organize the current literature into three distinct paradigms-sequence-centric, cell-centric, and multi-omics-to clarify a field currently fragmented by diverse tokenization strategies and architectural choices. Beyond methodology, we evaluate the practical utility of these models in tasks ranging from biomarker discovery to perturbation response prediction. We also identify critical barriers to adoption, including high computational costs, interpretability challenges, and the lack of standardized benchmarks. To support reproducible research, we provide a curated catalog of essential datasets and evaluation frameworks. Finally, we propose a roadmap for the next generation of FMs, advocating for architectures that move beyond statistical correlation to incorporate causal reasoning, temporal dynamics, and autonomous experimental validation.