The research, an accepted early version of which was published in Communications Biology, introduced an artificial intelligence framework called HRAftu-LM-RAG to automatically extract features related to the Human Reference Atlas’s (HRA) functional tissue units from scientific publications. HRA aims to map the multiscale structure of the human body, from organ systems down to the single-cell level, by bringing together experts from more than 25 international consortia. So far, experts have produced 2D illustrations detailing functional tissue units in 10 organs.
The validated system combines Large Language Models for textual reasoning, Large Image Models for visual interpretation, and Retrieval Augmented Generation to ground information in sources. The study scanned 244,640 PubMed Central publications containing 1,389,168 figures related to 22 functional tissue units. Of these, 617,237 figures containing microscopy and schematic images were identified. From the images and associated text, 331,189 scale bars, 1,719,138 biological entity mentions, and donor information such as sex and age were automatically extracted. The resulting data is intended to support the design, review, and approval of future tissue unit illustrations within HRA.
HRA is being developed by HuBMAP, SenNet, KPMP, GUDMAP, and NIDDK, together with the HRA Editorial Board and experts from more than 20 other consortia. The authors reported no conflicts of interest, and the funders were stated to have had no role in the research or publication processes.
Background
The atlas is not a new name in the FikirPilot archive: in the past 90 days, we published one article mentioning this name; that article was dated September 2, 2026.