基于多示例多标记迁移学习的蛋白质功能预测

发布时间：2018-11-15 22:20

【摘要】：随着各种基因组测序计划的推出,不断有很多物种被新测序完成,需要对这些物种的蛋白质功能进行注释.这些物种中已知功能的蛋白质数量少,可以考虑使用亲缘关系近、已知功能蛋白质数量多的物种来帮助这些物种进行蛋白质功能预测.本文把这个任务抽象为多示例多标记迁移学习问题,并提出了第一个多示例多标记迁移学习框架TR-MIML来解决此任务.TR-MIML通过最小化投影空间上加权源域样本中心点与目标域样本中心点的距离,给源域样本赋予不同权值,并基于目标域和源域样本训练多示例多标记学习模型.在两个新完成测序物种上,实验结果证明了迁移学习有助于它们的蛋白质功能预测.另外,亲缘关系越近的物种作为源域进行迁移学习帮助越大.
[Abstract]:With the introduction of various genome sequencing projects, many species have been sequenced, and the protein function of these species needs to be annotated. Since the number of known proteins in these species is small, we can consider the use of species with close kinship and a large number of known functional proteins to help these species predict their protein function. This article abstracts this task as a multi-example, multi-tag migration learning problem. The first multi-example multi-label migration learning framework (TR-MIML) is proposed to solve this task. TR-MIML assigns different weights to the source domain samples by minimizing the distance between the center points of the weighted source domain and the center points of the target domain samples in the projection space. And the multi-example multi-label learning model is trained based on target domain and source domain samples. In two newly sequenced species, the results show that migration learning is helpful to predict their protein function. In addition, the more closely related species are used as source areas for migration learning, the greater the help.
【作者单位】：南京邮电大学通信与信息工程学院;南京信息工程大学江苏省优势学科江苏省大气环境与装备技术协同创新中心;河海大学计算机与信息学院;南京邮电大学地理与生物信息学院;
【基金】：国家自然科学基金(批准号:61571233,61271082) 国家重点基础研究发展计划(973)(批准号:2011CB302903) 江苏省高校自然科学研究重大项目(批准号:14KJA510003) 江苏省重点研发计划(批准号:BE2015700) 南京信息工程大学PAPD与CICAEET资助项目
【分类号】：Q51;TP181

【相似文献】