仍上市及下市藥物與其蛋白標的之機器學習研究以預測藥物引起的不良反應

Abstract

藥物設計的研發成本和時間消耗是藥廠的重大投資，但是藥物不良反應可能導致藥物在臨床試驗期間或藥物上市後被撤銷核准，造成藥廠財務上的重大損失。現今沒有論文直接研究已下市的藥物與目前仍在市場上的藥物之間的化學性質差異程度。了解下市藥物(withdrawn drugs；WDs）的相關因素及其不良反應將有助於臨床前毒性試驗。深究WDs與市場藥物(on market drugs；MDs）之間的性質分佈可為科學家們提供新概念，以避免開發在上市後或臨床試驗期間才被終止使用的藥物。 在本文中，我們系統地分析了WDs和MDs的化學特徵，結果使我們能夠提出「WD -like」和「MD -like」基團的化學性質範圍。使用前人建立的藥物設計規則（例如Rule of Five, Ghose Filter, and MDDR-Like Rule）辨認出WDs的效果也在本研究的範疇，我們並提出了新的藥物設計規則「Rule of on market drugs」。我們也開發了一種通過機器學習來預測WD和MD的新方法，還研究了WDs和MDs的蛋白質標的。為了探索蛋白質標的與藥物的不良反應之間的關係，我們研究了不良反應相關蛋白(adverse related proteins；ARPs）和非不良反應相關蛋白(non adverse related proteins；NARPs）的許多特徵。我們檢驗了這些功能在ARP和NARP之間是否存在顯著差異，並應用機器學習來分類ARPs和NARPs，評估了每個特徵對所得模型的重要性。本論文對WDs、MDs、ARPs和NARPs的研究使我們能夠為未來的藥物開發研究制定一些新指南，可避免或盡量減少藥物使用不良反應的風險。

The financial cost and time consumption of drug design are tremendous, yet adverse effects can cause drugs to be withdrawn during clinical trials or after the drugs are on the market. There had been no report directly investigating to what extent the chemical properties between drugs that have been withdrawn and those that are currently on the market are different. Understanding the factors connecting withdrawn drugs (WDs) and their adverse effects would help in preclinical toxicity studies. Knowing the distribution of properties between WDs and on market drugs (MDs) may provide new concepts for scientists to avoid developing a drug that may be withdrawn from the market or during clinical trials. In the works presented in this thesis, we systemically analyzed the chemical features of WDs and MDs. The result allowed us to propose ranges of chemical properties for “WD -like” and “MD-like” groups. The performance of filtering out WDs using previously established rules for drugs design such as Rule of Five, Ghose Filter, and MDDR-Like Rule was investigated and a new rule of a drug design “Rule of on market drugs” was proposed. We then developed a novel method to predict WDs and MDs by machine learning. Protein targets of WDs and MDs were also investigated. To explore the relationship between protein targets and drug-induced adverse effects, many features of adverse effect related proteins (ARPs) and non adverse effect related proteins (NARPs) were studied. We examined whether these features were significantly different between ARPs and NARPs. Machine learning was applied to classify ARPs and NARPs. The importance of each feature for the resulting model was assessed. The present studies of WDs and MDs, and also of ARPs and NARPs, allow us to develop some new guidelines for future drug development research to avoid or minimize the risk of adverse effect.