A narrative review of the evolving role of language models in biomedicine: bridging text, omics, and multiscale integration

Background and Objective: The rapid evolution of language models (LMs) is facilitating their expansion from general-purpose domains into specialized biomedical fields. This transition brings forth unprecedented opportunities alongside unique challenges, encompassing the management of complex terminologies, virtual-actual combination, and the assurance of reliability in clinical contexts. The purpose of this review is to investigate and synthesize genuine LMs in the biomedical field, explore the applications of three key areas inspired by LM architectures (protein structure prediction, cell annotation, and foundation models), and discuss the integration of text and omics data. Methods: This review adopted a non-systematic search strategy to analyze relevant literature from the PubMed database. This search mainly focused on publications from 2017 to February 2025, with keywords including “language models”, “entity recognition”, “relation extraction”, “ontology”, “knowledge graph”, “protein structure prediction”, “and “cell annotation”. Key Content and Findings: This review examines the applications of LMs in three key areas: the application of LMs in biomedical text mining, omics researches inspired by LMs architectures, and the application in virtual-actual data combination. A key finding is that while LMs exhibit exceptional capabilities, their effectiveness is frequently constrained by inherent challenges, including model hallucination, poor reproducibility, and a fundamental reliance on the quality of their training data. We detail how integrating knowledge graphs (KGs) and ontologies can significantly enhance the interpretability of LMs and mitigate these risks. Furthermore, we analyze the paradigm of leveraging LMs to fuse multi-source data, which addresses data silos between textual and omics datasets. This fusion substantially deepens biomedical research capabilities and supports the visualization of complex biological associations. Conclusions: The advancement of LMs in biomedicine depends on establishing an intelligent framework that synergizes knowledge-guided and data-driven approaches. This synergy is critical for enabling the transition from simplistic “data splicing” to meaningful “knowledge fusion”, thereby enhancing model interpretability and decision-making transparency. By delineating this pathway, this review informs future research by promoting the development of robust, reliable, and clinically translatable artificial intelligence (AI) models.