ChatBiobank links a 3M framework with source-linked, tiered AI guidance for multi-omics biobanking

High-quality biobanks are essential for modern biomedical research and precision medicine, but translating broad standards into practical, end-to-end workflows remains difficult. Different omics applications impose assay-dependent requirements for specimen type, input amount, timing, processing, storage, aliquoting, and metadata. Multimodal sampling adds further complexity because tissue, blood, urine, feces, and saliva may need to be coordinated across clinical departments, laboratories, ethics teams, data managers, and biobank staff.

To address this implementation gap, researchers at the First Affiliated Hospital of Zhengzhou University organized key requirements into an integrated “3M” framework: multi-omics integration, multimodal biospecimens, and multi-departmental coordination. The framework draws on ISO 20387:2018, ISBER Best Practices, the NCI Best Practices for Biospecimen Resources, peer-reviewed literature, and assay- or platform-specific implementation materials. It organizes existing requirements rather than establishing a new consensus standard.

The team then developed ChatBiobank, a dialogue-based web tool that makes the curated 3M resource accessible through natural-language questions. ChatBiobank Version 1 uses retrieval-augmented generation with source-linked retrieval and reranking. DeepSeek-V4-pro serves as the primary generation and adjudication model, while Kimi K2.5 provides an independent cross-check for in-domain questions outside direct knowledge-base coverage. Retrieved evidence is routed through four author-defined evidence-support and routing tiers. Tier A represents source-grounded responses; Tier B represents cross-model concordance; Tier C flags cross-model divergence for closer review; and Tier D refuses or redirects insufficiently supported or out-of-domain questions. These tiers describe evidence support and routing status, not calibrated probabilities of correctness.

The frozen Version 1 system was evaluated using a predefined 230-question benchmark: 100 knowledge-base-supported questions, 100 in-domain questions outside direct knowledge-base coverage, and 30 out-of-domain questions. Two experts independently reviewed de-identified outputs while benchmark-stratum and response-tier labels were concealed; disagreements were resolved with a third expert. Overall, 227 of 230 outputs (98.7%) met the prespecified professional-acceptability criterion. All 100 Tier A outputs were directly acceptable. Among the 100 in-domain questions outside direct knowledge-base coverage, 98 met the criterion. All 77 Tier B outputs met the criterion, including 75 directly acceptable and 2 broadly acceptable responses. Tier D appropriately refused 29 of 30 out-of-domain questions, while the only unacceptable substantive response occurred in Tier C.

In a separate single-center descriptive implementation audit, RNA quality was assessed in 153 randomly selected fresh-tissue biospecimens managed under a workflow informed by ChatBiobank guidance. Tissue evaluation, standardized processing, snap-freezing in liquid nitrogen, liquid-nitrogen storage, RNA extraction, and quality control were included. Of the 153 biospecimens, 150 met locally predefined quality criteria, a pass rate of 98.04% (Wilson 95% CI, 94.39%–99.33%). One failed sample showed severe RNA degradation in extensively necrotic tissue; the causes of the other two failures were undetermined. The audit was descriptive and was not designed to estimate a causal effect of ChatBiobank.

ChatBiobank is freely available at www.chatbiobank.com, and the 3M framework is available at www.chatbiobank.com/3m/. The complete Version 1 benchmark questions, responses, internal tier assignments, and final expert classifications are available at https://github.com/FuLab-ZhaoSun/3M_Framework.

Expert Contact: Yang Fu, Department of Gastrointestinal Surgery, The First Affiliated Hospital of Zhengzhou University; E-mail: fuyang@zzu.edu.cn

Custom Tweet: ChatBiobank organizes multi-omics biobanking guidance around a 3M framework with source-linked, tiered AI routing. In a 230-question benchmark, 98.7% of outputs met the professional-acceptability criterion. DOI: 10.1002/mdr2.70094