使用生物醫學知識圖譜與深度學習方法的腫瘤生物標記辨識

Abstract

生物標記（Biomarkers）近年來已成為生物醫學領域的關鍵研究焦點。除了在癌症相關研究中獲得了顯著的關注外，生物標記也被視為實現精準醫學的重要途徑。其應用包括特定疾病的檢測、選擇個性化的治療方案，以及協助藥物開發。利用龐大的生物醫學文獻資料庫和資訊技術來進行生物標記的辨識，對生物醫學研究將帶來極大的助益。 SemMedDB 是一個從文獻中提取的、大型的生物醫學知識圖譜，使用基於規則的自動化技術來建構。儘管它涵蓋了眾多的醫學實體和關係，但它仍然面臨著無法完全捕捉到未知新知識（如生物標記）的挑戰。Wu（2021）提出了一種基於文獻分析的方法，通過識別知識圖譜中與生物標記高度相關的語義類型和關係，來構建潛在的生物標記範圍。該研究更強調了生物標記與知識圖譜中既存關係之間的高度相關性。在此基礎上，我們旨在建立標準化的生物標記標註程序，並開發自動化的生物標記識別模型，提供一個系統化的方式來發現新的生物醫學知識。 本研究提出了一個兩步驟的模型，將預訓練的知識圖譜嵌入與關於生物標記和既存關係的假設結合，構建了一個自動化的腫瘤生物標記分類模型（TBC）。在模型的第二階段，成功地運用了投影層和KGE集成加權總和技術，將預訓練的知識圖譜嵌入進行轉換並結合，以識別先前不存在的關係。在資料標註方面，對語義類型和關係的選擇、同義詞合併以及標準化的生物標記標註工作流程進行改良。實驗結果顯示，即使在資料有限的情況下，本研究的方法仍然有穩定的成效。後續的消融實驗，對模型的每個部分進行了分析和討論，期望能夠啟發未來對新生物醫學知識和關係的研究。

Biomarkers have recently become a key research focus in the biomedical field. In addition to receiving significant attention in cancer-related research, they are considered a crucial pathway to achieving precision medicine. Their applications include detecting specific diseases, selecting personalized treatment plans, and aiding drug development. Leveraging vast biomedical literature databases and information technology for biomarker identification could greatly benefit biomedical research. SemMedDB is a large-scale biomedical knowledge graph extracted from literature using rule-based automated techniques. While it encompasses numerous medical entities and relations, it still struggles to fully capture previously unknown knowledge, such as biomarkers. Wu (2021) proposed a literature-based approach to identifying highly relevant semantic types and relations within a knowledge graph to construct a potential biomarker scope. The study also highlighted the strong correlation between biomarkers and existing relations within the knowledge graph. Building on this work, we aim to establish a standardized biomarker annotation procedure and develop an automated biomarker identification model, offering a systematic approach for discovering new biomedical knowledge. Our study presents a two-step model that integrates pre-trained knowledge graph embeddings with assumptions about biomarkers and existing relations to construct an automated tumor biomarker classification model (TBC). In the second stage, we successfully employ projection layers and KGE-integrated weighted sum techniques to transform and combine pre-trained knowledge graph embeddings to identify previously non-existent relations. For data annotation, we refine the processes of selecting semantic types and relations, merging synonyms, and establishing a standardized biomarker annotation workflow. Our experimental results demonstrate stable performance even with limited data. Through further ablation studies, we analyze and discuss each model component, hoping to inspire future research on new biomedical knowledge and relations.