長久以來, 在分子生物學上, 利用蛋白質的氨基酸序列來預測它的二次結
構一直是一個很重要卻始終無法完全解決的問題. 雖然已經有許多預測方

法被提出討論 (大致可分為四類, 即統計資訊, 樣式比對, 類神經網路,

以及複合式系統) , 但是這些方法的預測準確率始終停留在百分之六十到

七十之間. 在本論文中, 我們對蛋白質二次結構的形成機制提出一套假

設 (NDA: Neighbors Determine All 鄰居決定一切) , 並且利用遞迴式

類神經網路實作出一個NDA系統. 在使用實際蛋白質資料庫對這個系統加

以訓練及測試後, 我們發現其實驗結果足以證實NDA和實際資料庫間有極

高的相容性.也就是說, NDA可以解釋實際蛋白質資料庫中所含資料的行

為. 然而因為NDA本身的限制, 這個系統無法使用於實際的預測工作中.為

了利用存在此NDA系統中的資訊, 我們另外提出了兩種方法: 單邊NDA和

NECN. 單邊NDA簡化了NDA的嚴格限制, 因此可以於特定的實際預測工作中

使用. 而NECN則是結合了NDA網路以及傳統網路架構. 先從NDA網路中抽取

出新的氨基酸代碼, 再將這代碼應用於傳統的網路架構上, 形成一個兩階

段式的架構. 我們使用實際蛋白質資料庫以及人造NDA資料庫 (根據NDA假

設造出一個人造資料庫) 來訓練及測試NECN後, 得到的結果證實了NECN的

確優於傳統使用的網路架構.

The prediction of the secondary structure of a protein from its

amino acids sequence has long been a not-yet-completely solved

problem despite its importance in molecular biology. Although

various prediction methods have been proposed, the prediction

accuracy still hovered around 60-70%. In this thesis we

proposed a hypothesis on the formation mechanism of secondary

structures called NDA ( Neighbors Determine All ). A NDA

implementation using recurrent neural network is tested on a

database consists of proteins with known secondary structures

sequence. The result confirms that NDA hypothesis is highly

compatible to the database. However, the NDA network cannot be

applied on practical prediction tasks due to certain reasons.

To use the information contained in NDA network we proposed two

approaches: the Single-Side NDA and NECN. The Single-Side NDA

relieves the strict restrictions of NDA and provides

possibility of practical application. The NECN ( NDA Extracted

Code Neural network ) combines a NDA neural network and a

neural network which is traditionally used in this problem to

form a two phase secondary structure prediction system. NECN is

tested with the protein database and another database

constructed using NDA hypothesis. Both results show that NECN

performs better than conventional neural network methods used

in this problem. Although the performance gain is not large,

the NECN can still be improved in many aspect. In addition, the

NDA hypothesis and its neural network implementation are also

useful tools in this task.