Cross-Scale Biomedical Data Construction and Representation in the Foundation Model Era: A Survey

Foundation models (FMs) pre-trained on biological, biomedical, and clinical corpora are reshaping life-science research across protein structure prediction, single-cell analysis, computational pathology, and clinical decision support. Surveys are model-centric, cataloging architectures while treating data as secondary and single-scale. Biological FMs integrate sequences, structures, molecular interactions, omics profiles, images, and clinical records across scales and time. These corpora are sparsely paired, unevenly curated, and dependent on ontologies and measurement protocols, so data construction and alignment determine generalizability. We focus on data construction, representation, and cross-scale alignment. We introduce a three-dimensional tensor index spanning biological scale (molecular to population/cohort, plus cross-scale alignment), data operation (construction to service), and technical era (curated knowledge to foundation models). We characterize modalities, curation pipelines, foundation models, and evaluation resources, and synthesize three alignment paradigms: knowledge-graph bridging, multimodal co-training, and simulation-mediated grounding. We argue that cross-scale alignment is an organizing problem for biomedical AI, with knowledge-graph scaffolds enabling links across scales. We organize open challenges into four groups: data quality and standardization, cross-scale integration, evaluation, and governance. Preprint status: This manuscript has not yet undergone peer review. A version of this manuscript is currently under journal review.