Transformer-Enabled Multi-Omics Interpretation: LLM Pipelines for Cross-Dataset Annotation and Uncertainty-Aware Integration

Multi-omics studies now span genomics, epigenomics, transcriptomics, proteomics, metabolomics, microbiomics, and spatial and single-cell modalities, often generated across different cohorts, platforms, and labs. While these datasets are increasingly rich, their consistent interpretation, harmonized annotation, and reliable integration remain challenging. Batch effects, missing modalities, inconsistent nomenclature, incomplete metadata, and conflicting knowledge sources can all distort biological conclusions. At the same time, transformer models, including large language models (LLMs) and multimodal transformers, offer new ways to connect heterogeneous evidence through representation learning, retrieval-augmented workflows, and structured reasoning over biological knowledge. However, interpretability remains a major open problem: to support trustworthy biological discovery, researchers must be able to understand how outputs are generated and judge whether model reasoning is biologically meaningful.This Research Topic aims to showcase methods and applications that use transformer-based pipelines to interpret and integrate multi-omics data across datasets, with a strong focus on cross-dataset annotation, interpretability, and uncertainty-aware integration. We welcome contributions that bridge machine learning innovation with biological validity and reproducibility, including approaches that combine LLMs with knowledge graphs, ontologies, and classical statistical frameworks.Key themes include (but are not limited to):• LLM/transformer pipelines for cross-study harmonization of sample metadata and biological labels• Retrieval-augmented generation (RAG) for grounded annotation using curated resources (ontologies, pathway databases, clinical coding systems)• Uncertainty estimation, calibration, and error propagation in multi-omics integration and downstream prediction• Handling domain shift, batch effects, missing modalities, and weak supervision across cohorts and platforms• Multimodal transformers linking omics with text (papers, clinical notes), images (histology), and spatial context• Benchmarks, evaluation protocols, and reproducible pipelines for cross-dataset annotation quality and robustness• Interpretable transformer models: attribution, concept-based explanations, and biologically consistent reasoning• Case studies in cancer, immunology, neuroscience, infectious disease, rare disease, and population-scale biobanksThis Research Topic welcomes Brief Research Report, Case Report, Clinical Trial, Community Case Study, Correction, Data Report, Editorial, FAIR² DATA Direct Submission, FAIR² Data, General Commentary, Hypothesis and Theory, Methods, Mini Review, Opinion, Original Research, Perspective, Policy Brief, Policy and Practice Reviews, Review, Systematic Review, and Technology and Code articles. Topic editors: Yuda Munarko, Toshinori Endo, Nanqing Dong Source: https://www.frontiersin.org/research-topics/78531/transformer-enabled-multi-omics-interpretation-llm-pipelines-for-cross-dataset-annotation-and-uncertainty-aware-integration