通讯作者: 唐小利, tang.xiaoli@imicams.ac.cn

DOI：10.12201/bmr.202605.00070

Identifying Biomedical Research Frontiers via Knowledge Flow: A Self-Supervised Graph Clustering Framework

• 摘要：目的/意义 生物医学领域知识体量激增与载体异构化的复杂性,对传统前沿识别方法提出了挑战。本文旨在探讨构建一种量化“科学-技术”知识溢出路径的计算框架,以揭示从基础研究发现到技术应用之间的隐性关联与演化规律,从而为科研选题与研发布局提供数据驱动的决策支持。方法/过程 本文提出了一种融合隐性语义与异构拓扑关联的研究前沿识别框架。首先,整合论文与专利数据构建“科学-技术”二部知识流动网络,利用PubMedBERT抽取文本语义并基于K近邻方法构建语义增强边,形成稠密的语义-拓扑耦合结构。其次,设计了三模态门控编码器,以自适应融合节点的内容、时间与结构特征,生成统一的表征。最后,通过自监督联合优化策略协同学习节点表征与社区结构,并基于链路预测概率定义“前沿指数”,实现高转化潜力研究前沿的识别。 结果/结论 在乳腺癌领域的实证研究表明,本框架有效克服了传统二部引用网络的稀疏性问题,能够识别出三类典型的前沿社区：“理论爆发型”、“产业成熟型”与“科学-技术并行突破型”。可视化分析进一步揭示了从基础研究社区(如三阴性乳腺癌分子分型)向应用研究社区(如靶向联合用药、AI影像诊断)的隐性知识流动。研究结果表明,本框架在语义深度和前瞻性预警方面具有优势,可作为生物医学领域科技情报分析的有效手段之一。

  关键词： 前沿识别; 知识流动; 图聚类; 链路预测; 乳腺癌

   

  Abstract: Purpose/Significance Driven by the exponential growth of biomedical information and the increasing heterogeneity of knowledge sources, conventional frontier identification techniques are struggling to maintain efficacy. This study develops a computational framework designed to quantify knowledge spillover pathways between science and technology, aiming to uncover the latent associations and evolutionary dynamics bridging basic research discoveries and technological applications.This framework offers instrumental decision support for optimizing research topic selection and R&D strategic planning.Method/Process This study develops an identification framework fusing implicit semantics and heterogeneous topological linkages. The process begins with the construction of a bipartite knowledge flow network spanning papers and patents. Leveraging PubMedBERT for semantic encoding and K-nearest neighbor for edge augmentation, we establish a dense semantic-topological structure. Subsequently, a tri-modal gated encoder is introduced to adaptively integrate nodes textual, temporal, and structural attributes into a unified representation. The framework utilizes a self-supervised joint optimization strategy to concurrently optimize node representations and community detection. Ultimately, research frontiers with significant translational potential are identified using a custom ‘Frontier Index derived from link prediction probabilities. Result/Conclusion Empirical results in the field of breast cancer demonstrate that the proposed framework effectively mitigates the sparsity issues inherent in traditional bipartite citation networks. It successfully identifies three archetypal frontier communities: Theoretical Burst, Industrial Maturation, and Sci-Tech Resonance. Visualization analysis further elucidates the latent knowledge flows emanating from basic research clusters (e.g., molecular subtyping of triple-negative breast cancer) toward applied research sectors (e.g., targeted combination therapies and AI-assisted imaging diagnosis). These findings suggest that the framework excels in semantic depth and proactive early-warning capabilities, serving as a robust computational tool for science and technology (S&T) intelligence analysis in the biomedical domain.

  Key words: identifying frontiers; knowledge flow; Graph Clustering; link prediction; breast cancer

  提交时间：2026-05-20

  版权声明：作者本人独立拥有该论文的版权，预印本系统仅拥有论文的永久保存权利。任何人未经允许不得重复使用。
• html

• 图表

• 刘燕, 张潇潇, 侯丽. 面向知识服务系统的学术知识图谱构建与应用研究. 2024. doi: 10.12201/bmr.202402.00015

  张馨月, 宋泽宇, 田晓伟, 韩霜, 孔杨. 基于K-means聚类的老年人养老服务需求识别. 2024. doi: 10.12201/bmr.202408.00059

  艾俊宇, 杨德林. 前列腺癌关键基因诊疗策略的前沿进展. 2025. doi: 10.12201/bmr.202501.00014

  姚根武, 王彦伟, 郭琎祎, 李露露, 党文博, 宋爱琳. 单细胞测序技术在乳腺癌中的研究进展. 2024. doi: 10.12201/bmr.202407.00053

  王华琼, 俞定国, 钱归平. 基于医学社交媒体数据的多模态知识图谱构建. 2022. doi: 10.12201/bmr.202209.00005

  车美龄, 南嘉乐, 林建海, 高东平. 多模态营养知识图谱构建. 2025. doi: 10.12201/bmr.202505.00042

  朱韶华, 李艳红, 张迅, 黄海量. 基于高质量文献的医学肿瘤学领域研究前沿识别研究. 2023. doi: 10.12201/bmr.202312.00010

  陈婕卿, 竹志超, 张锋, 曾可, 姜会珍, 程振宁. 面向知识图谱构建的中文电子病历命名实体识别方法研究. 2023. doi: 10.12201/bmr.202312.00011

  程洁, 余国政. 异粘蛋白在乳腺癌中的研究进展. 2026. doi: 10.12201/bmr.202601.00039

  田文琪, 刘嘉祯, 王之义, 崔欣. 基于分阶段VaDE聚类的跨区就诊行为分析——以上海市A区为例. 2025. doi: 10.12201/bmr.202509.00039

• 序号	提交日期	编号	操作
  1	2026-03-18	10.12201/bmr.202605.00070V1	下载

• 公开评论  匿名评论  仅发给作者 提交评论