Biomedical Entity Mining from Multi-Source Evidence (Manuscript Under Review)

Problem and Research Question

Biomedical research generates vast quantities of unstructured text — journal articles, clinical reports, patent filings, and public databases — that contain valuable information about virus-host interactions, drug targets, and disease mechanisms. Manually extracting structured entities (viruses, host organisms, interaction types) from this evidence is time-consuming, error-prone, and does not scale to the volume of available literature. The central question: how can instruction-tuned language models be adapted to reliably extract virus, host, and interaction entities from multi-source biomedical evidence, and how can evidence from heterogeneous sources be reconciled into a consistent structured representation?

Existing biomedical named-entity recognition (NER) systems are typically trained on single-source corpora and struggle with the diversity of writing styles, terminology, and evidence quality across different source types (scientific articles, databases, clinical notes). A model that can handle multi-source evidence while maintaining high precision and recall would significantly accelerate biomedical knowledge discovery.

Method

This manuscript describes an instruction-tuned language model for mining virus, host, and interaction entities from multi-source biomedical evidence. The approach combines:

• Instruction tuning — fine-tuning a pre-trained language model on instruction-formatted examples that specify the entity types to extract (virus, host organism, interaction type) and the source context, enabling the model to follow extraction instructions rather than relying on a fixed label set.
• Multi-source evidence integration — processing and reconciling entity mentions from heterogeneous sources (scientific literature, public databases, clinical records) with different writing styles, terminologies, and evidence quality levels.
• Structured entity representation — producing a consistent structured output (virus–host–interaction triples) that can be used for downstream analysis, database construction, or hypothesis generation.

The work is part of a broader collaboration applying natural language processing to biomedical discovery, related to the closed-loop discovery platform for antibacterial combinations.

Personal Contribution

Frederik Banis is a co-author on this manuscript. His contribution focuses on the methodological framing and the connection to the broader uncertainty-aware decision and inference framework. The first author (Zheng Zhang) led the model development and experimental design. Full author list and individual contribution details will be disclosed upon publication.

Validation and Key Results

• The manuscript is under review (submitted to Genomics, Proteomics & Bioinformatics). It is not counted as a published output.
• Specific quantitative results (precision, recall, F1 scores on benchmark datasets) are pending peer review and are not reported here to avoid presenting proposed outcomes as completed work.
• Detailed results, model architecture, and evaluation methodology will be available upon publication.

Related Outputs

• Related manuscript (under review): AntiSyn-AI for Accelerating the Discovery of Synergistic Antibacterial Combinations against MRSA and Their Hydrogel Translation — submitted to npj Antimicrobials and Resistance. See the publication page.
• Related project: Closed-Loop Discovery for Synergistic Antibacterial Combinations — the broader platform context for this work.

Collaborators and Institutions

• Zheng Zhang (first author) — model development and experimental design
• Weichao Li, Jiajun Ren, Yufei Chen, Ben He, Jinhe Li — co-authors
• Qiang Sun, Haibo Wang — senior authors
• Hangzhou Institute of Technology, Xidian University — host institution

Status and Next Steps

Status: Under review (submitted 2026). Details will be disclosed upon publication.

Transfer to current research: The instruction-tuning approach for structured entity extraction from heterogeneous evidence sources shares the same methodological principle as the broader research program: using data-driven models with structured priors to extract meaningful information from noisy, multi-source data. The shift is from biomedical text to physiological signals, but the inferential architecture — combining learned representations with domain structure — is shared.

See the related project page for the broader research context and the Applications page for the biomedical sensing application domain.