以生物醫學文獻建構之語義網路預測藥物新適應症

Abstract

一般而言，藥物開發的過程耗時而且成本高昂，但失敗的機率卻非常高。「舊藥新用」提供了藥物開發的另一個可行方向。它利用既有藥物的作用機轉或與其他生物因子間（例如：基因、蛋白質…）的關係來找出可能的新適應症。而那些未能通過藥物開發流程的藥物也是「舊藥新用」方法實施的對象之一。雖然目前已經有許多藥商、醫藥研究人員和學者引入這個方法，但仍有許多改善的空間。因此，本研究提出SemDRLTR模型，其利用自然語言技術處理後的醫學文獻資料，加上多個醫藥知識庫來建立關係網路，並以Swanson ABC模型和Chen Path 模型中不同權重計算方法與排序方法的組合作為特徵值（Feature），利用這些特徵值來進行監督式排序學習（Supervised Learning to Rank），再以學習後的模型來預測藥物可能的新適應症。另外，本研究亦提出一個整合In-predication和Co-occurrence概念的方法以期能提高SemDRLTR模型的準確度。從實驗結果可知，SemDRLTR模型能夠比既有的方法表現更好，其中In-predication的表現較Co-occurrence佳，而整合式的方法能夠達到更好的效能提升，這代表將此整合式方法運用至我們的SemDRLTR模型中能夠更有效地找出潛在的藥物與疾病關係。

The process of drug development consumes much time and money but it often fails during clinical tests or other stages of drug development. Drug repositioning gives another way for drug discovery. Its goal is to find new indications for existing drugs or those drugs which ever failed at one of drug development stages. Although many researchers engage in this topic, there is still room for improvement. Hence, we propose a Semantic-based Drug Repositioning Learning-to-Rank (SemDRLTR) method, which considers the semantic relation between concepts. It is based on Swanson’s ABC model as well as Chen’s Path model and combines literature and ontologies to construct a comprehensive concept network. We further propose hybrid methods that combine the in-predication method and the co-occurrence method. From the results of our experiments, it is proven that in-predication, which keeps the meaning of relation between concepts, outperforms the co-occurrence method, which only considers whether concepts co-occur or not and the hybrid methods further perform better than either one.