|本期目录/Table of Contents|

[1]顾林林,王志勇,方铭.一种基于人工智能的基因组选择方法[J].集美大学学报(自然科学版),2023,28(3):205-213.
　GU Linlin,WANG Zhiyong,FANG Ming.A Genomic Selection Method Based on Artificial Intelligence[J].Journal of Jimei University,2023,28(3):205-213.
点击复制

一种基于人工智能的基因组选择方法(PDF)
分享到：

《集美大学学报（自然科学版）》[ISSN:1007-7405/CN:35-1186/N]

卷:
第28卷
期数:
2023年第3期
页码:
205-213
栏目:
水产、食品与生物工程
出版日期:
2023-05-28

文章信息/Info

Title:
A Genomic Selection Method Based on Artificial Intelligence
作者:
顾林林1; 2; 王志勇1; 2; 方铭1; 2
(1.集美大学水产学院，福建 厦门 361021；2.农业农村部东海海水健康养殖重点实验室，福建 厦门361021)
Author(s):
GU Linlin1; 2;  WANG Zhiyong1; 2;  FANG Ming1; 2
(1.Fisheries College, Jimei University, Xiamen 361021, China;2.Key Laboratory of Healthy Mariculture for the East China Sea of Ministry of Agriculture, Xiamen 361021, China)
关键词:
基因组选择; 基因组估计育种值; 神经网络; 深度学习
Keywords:
genomic selection; genomic estimation breeding value; neural network; deep learning
分类号:
-
DOI:
-
文献标志码:
A
摘要:
基因组选择 (genomic selection, GS) 是一种利用全基因组的标记信息进行选择育种的新方法。它利用参考群体的基因型和表型信息进行训练，对只含基因型信息的候选群体预测基因组估计育种值 (genomic estimated breeding value, GEBV)。开发了一种新的深度学习基因组选择 (genomic selection, GS) 方法，并命名为DRNGS (deep residual network genomic selection)。新方法的特点有：1) 以深度残差网络来预测基因组估计育种值 (genomic estimated breeding value, GEBV)，可捕获基因型内部的复杂关系，提高预测准确性；2) 采用卷积和池化策略来降低高维基因型数据的复杂性，加快计算速度；3) 方法中引入批量归一化层，加快了收敛速度。将新方法应用于CIMMYT小麦数据集，实验结果表明：DRNGS的效果比前馈神经网络 (feedforward neural network, FNN)提高了101.59%~130.83%；在对大部分性状的表型预测中，DRNGS比GBLUP (genomic best linear unbiased prediction) 提高了2.24%~20.19%；在计算耗时方面，DRNGS仅次于GBLUP，比DeepGS快了大约18~22倍，比FNN快了24~26倍。为进一步比较DRNGS和DeepGS，用伊朗面包小麦 (Triticum aestivum) 数据集进行测试，结果表明:DRNGS收敛速度优于DeepGS;在对所有性状的表型预测过程中，DRNGS的计算耗时始终较DeepGS短;而且DRNGS在预测准确性方面优于DeepGS,在8个性状中，DRNGS较DeepGS提高0.12%~1.59%。并将DRNGS开发成R包，可通过https://github.com/GuLinLin.JMU/DRNGS访问。
Abstract:
Genomic selection (GS) is a new approach to selective breeding by using genome-wide markers. It uses genotypic and phenotypic information from a reference population for training, and then predicts This paper has improved and developed a new deep learning Genomic selection (GS) method named deep residual network genomic selectio (DRNGS).The features of the new method were: 1) a deep residual network was used to predict genomic estimated breeding value (GEBV), which could capture the complex relationships within genotypes and improve the prediction accuracy; 2) convolution and pooling strategies were used to reduce the complexity of high-dimensional genotype data and speed up the computation; 3) a batch normalization layer was introduced in the method to speed up the convergence speed.The new method applied the CIMMYT wheat dataset,and the experimental results showed that DRNGS outperformed Feedforward Neural Network (FNN) with a relative improvement of 101.59%-13083%. DRNGS outperformed Genomic Best Linear Unbiased Prediction (GBLUP) by 2.24% to 20.19% for phenotypic prediction of most traits. It was the second only to GBLUP in terms of computational time consumption, and was approximately 18.22 times faster than DeepGS and 24.26 times faster than FNN. To further compare DRNGS with DeepGS, we applied the Iranian bread wheat (Triticum aestivum) dataset for testing and showed that DRNGS converged faster than DeepGS, consistently took less time to compute than DeepGS in predicting phenotypes for all traits, and that DRNGS outperformed DeepGS in terms of prediction accuracy. For eight traits, DRNGS improved 0.12%-1.59% over DeepGS.DRNGS has been developed as an R package, which can be accessed at https://github.com/GuLinLin.JMU/DRNGS.

参考文献/References:

相似文献/References:

备注/Memo

备注/Memo:
更新日期/Last Update: 2023-09-12