基於生物醫學文獻及效應改善路徑重要性模型預測藥物新適應症

Abstract

開發新藥物的過程耗時、耗費成本高，且失敗率極高。為解決這個困難，研究人員提出「舊藥新用」的方法，利用原先已通過美國食品藥品監督管理局（FDA）核准的藥物，進行研究並尋找新的適應症，如此作法使藥物開發成本、開發時程大幅減低。 1986年，Swanson 透過尋找醫學文獻之間隱藏的關係來實現舊藥新用，之後的學者大多跟隨 Swanson 所提出的模型來改進研究。先前的研究提出各種不同的方法，例如有些研究使用文字在同一篇文章的出現次數計算關係的權重，另有其他研究在生醫概念網絡當中加入語義關係，而我們的前一篇研究（Lee, 2017）則使用監督式學習的方式，提出路徑重要性的分類模型。在本研究中，我們依循著前一篇研究的方法，將分類模型所使用的特徵加入生醫效應的考量，增加更多有生物醫學意義的特徵值，並且針對某焦點藥物，依照分類的結果計算分數，對其候選疾病進行排序，排序越前面，越可能是該藥物的新適應症。 最後，我們的實驗也證明加入生醫效應相關的特徵能有效提升原先的路徑重要性分類模型，且表現也比傳統的方法佳，可以更有效地找出某焦點藥物的潛在新適應症。

The drug development is an expensive and time-consuming process. Few drugs can be retained from the R&D, clinical trials, to the phase of FDA approval. Researchers begin to find alternative research methods. One of the well-known methods is drug repurposing, the application for approved drugs to new disease indications, which can cut down the cost and reduce the time from the R&D to FDA approval. In 1986, for drug repurposing, Swanson (1986) proposes a literature-based discovery (LBD) approach for discovering hidden relationships between drugs and diseases through some intermediate biomedical concepts. Afterwards, most of research studies follow Swanson’s model and propose various methods. Some prior studies weight links/paths by term co-occurrence; some add semantic relationships into the biomedical concept network; in our previous study, Lee (2017) proposes the supervised path importance classification method. In this research, we follow the classification method in Lee's study. We take biomedical effects of predicates into account and add additional features, called the effect-based features. Then for a focal drug, the scores for each disease are calculated through the results of path importance classification. According to the scores, we rank the disease candidates. Finally, for finding potential new indications effectively, we prove that by introducing the effect-based features, we improve the performance of the supervised path importance classification model of Lee’s study (2017), and our drug repurposing method also outperforms the traditional methods.