利用深度學習預測蛋白質活性位點－－以ACE及ZMR配體為例

蛋白質複雜的結構和交互作用使其有許多不同的功能，藥物設計也是靠著研究蛋白質與藥物配體的接合來進行，其接合點，我們稱之為蛋白質的活性位點，本研究就是要設計一套輔助藥物設計的預測活性位點系統。我們所使用的預測架構為深度學習中循環神經網路（Recurrent Neural Networks，RNN）搭配長短期記憶（Long Sort Terim Memory，LSTM），利用此架構分別對包含ZMR配體；ACE配體的蛋白質進行訓練，包含ZMR配體的蛋白質測試資料中，有66%的蛋白質活性位點被成功預測，但包含ACE配體的蛋白質測試資料中，卻只有55%的蛋白質活性位點成功預測，而這樣的結果本研究推測可能與配體的結構複雜度相關。 Understanding protein is the direction that biologists and pharmacologists have been working on for many years. Protein is made up of many macromolecules and small molecules. Due to the complex structure and interaction of proteins, it has many different functions. Drug design is also based on the studying of proteins and drug ligands binding. The docking position(active site) is special structure of the protein that ligand binding to. This thesis is to design a system for assisted drug design that due to predict active site. Since the rise of deep learning in the field of artificial intelligence, many studies have also begun to experiment with deep learning architectures, which include many predictive studies. The architecture of neural networks used in this paper is recurrent neural networks(RNN) with long-term and short-term memory(LSTM) which is one method of deep learning. This architecture improves the shortcomings of previous recurrent neural networks, allowing the recurrent neural network to have deeper training. The final experimental results will be used to evaluate the accuracy of the active sites predicted by the system