مقاله انگلیسی رایگان در مورد محلی سازی زیرسلولی های پروتئین انسانی با یادگیری عمیق – الزویر 2018

Knowledge of the subcellular localization of proteins is critical for the understanding of their functions and biological processes in cells. Protein subcellular location information is of highly importance in various areas of research, such as drug design, therapeutic target discovery, and biological research, etc [1]. Accurate prediction of protein subcellular localization is the prequiste to help in-depth understanding and analysis of various protein functions. As the wide applications of next sequencing techniques, protein sequences have accumulated rapidly during the last decades [2, 3]. Facing with such large-scale sequences, experimental determination of their protein subcellular locations is extremely inefficient and expensive in this post-genomic era. Therefore, effective and efficient computational methods are desired to assist biologists to address these experimental problems. During the past few years, many computational efforts have been done for predicting protein subcellular locations, thus generating a serious of computational methods. Most of high-performance computational methods use machine learning algorithms together with diverse feature representations to make predictions [4-8]. These machine learning based methods can be roughly divided into two classes: (1) sequence-based, and (2) annotation-based. Sequence-based methods use sequential information from primary sequences of proteins. For instance, Park et al. [9] trained a set of SVMs based on multiple sequence-based feature descriptors, such as amino acid composition, amino acid pair, and gapped amino acid pair composition, and proposed a voting scheme using the trained SVMs to predict protein subcellular locations. The upcoming problem is that features based on amino acid pair composition would lost sequence order effect. In order to address this problem, Chou et al. [10] proposed a modified feature encoding method, namely Pseudo Amino Acid  Composition (PseAAC), sufficiently taking the sequence order information for the remarkable improvement of the predictive performance. Most recently, Rahman et al. [11] proposed to fuse PseAAC, physiochemical property model (PPM), and amino acid index distribution (AAID) to improve the prediction accuracy. Moreover, other sequential information such as the sequence homology and sorting signals are often used to train machine learning models [12-15].