论文部分内容阅读
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.
With the introduction of various genome sequencing programs, there are many species that are continuously sequenced and need to be annotated on the protein functions of these species. The small number of proteins that are known to function in these species can be considered using close genetic relationships, known functions Protein number of species to help these species for protein function prediction.This task abstracts this task as a multi-sample multi-label migration learning problem and proposes the first multi-instance multi-label migration learning framework TR-MIML to solve this task.TR -MIML Weighted the source domain samples by minimizing the distance between the weighted source domain sample center and the target domain sample center in the projection space and trained the multi-instance multi-label learning model based on the target domain and the source domain samples. Newly completed sequencing of the species, the experimental results demonstrate that migration learning contributes to the prediction of their protein function.In addition, the more closely related species help migrate learning as a source domain.