應用類神經網路分析胜▇質譜以辨識其一級結構

Abstract

雙向關聯式記憶(Bidirectional Associative Memory)網路，及雷射攝 影(Holographic)網路均可以提供類神經網路雙向空間的內容定址記憶( Content Addressable Memory)，利用向量空間非齊次二項式 Lyapunov 能量的最佳線性組合，雙向關聯式記憶網路及雷射攝影網路可以有效組成 修正式內連結雙向關聯式記憶 (Modified Intraconnection aBidirectional Associative Memory)網路，此網路可以用於稀疏向量( Sparse Vector)的信號擷取。胜▇(Peptide)的快速原子撞擊質譜(Fast Atom Bombardment Mass Spectrum)，及聯結掃描質譜(Linked Scan Mass Spectrum)中各個胺基酸殘基的離子碎片組，可以視為一組稀疏向量 。本篇將組成胜▇的胺基酸殘基的理論離子碎片組，視為修正式內連結雙 向關聯式記憶網路的編碼向量(Encode Vector)，組成網路的長期記憶( Long Term Memory)。由質譜所測得的離子碎片組經過此網路運算後，就 可以辨識胺基酸殘基的種類。我們發現利用此方法比傳統辨識的方法方便 ，但也有其測量上的限制。 實驗的內容是利用 B/E 子離子聯結掃描質譜 測量胜▇，討論所測質譜經過我們設計的網路運算的結果及其限制。 Bidirectional Associative Memory (BAM) and Holography are two artificial neural networks that have the characteristic of granting content addressable memory of bidirectional space. Using the optimized linear combination of heterogeneous forms of Lyapunov energy, BAM and holography can be shaped into a so- called Modified Intraconnection Bidirectional Associative Memory (MIBAM) network. The network developed in this laboratory is applied in the analysis of mass spectra of peptides. The a mino acid sequence is an essential piece of information in the study of peptide molecules. Since the invention of Fast Atom Bombardment (FAB) ionization, the peptide sample can be directly introduced to the mass spectrometer without a lot of pretreatment. Accordingly, mass spectrometry has evolved as one of the sequencing techniques. In this study, the calculated amino acid mass fragments are input as the encode vector of the MIBAM network to obtain the weighting matrix of the long term memory. The ne twork then recalls (deduces) the residue of the peptides sequentially according to the mass spectra. The MIBAM network developed in this work proves to handle signals of sparse vector (such as mass signals in this study) extraordinarily well. Several known peptides are analyzed to test the feasibility of this technique. The advantages and limitations are discussed. Apparently this preliminary work has pointed the right direction for identification of the primary structure of peptides using our MIBAM network.