A new genome-wide selection method based on deep learning was proposed by the Institute of Composition, Chinese Academy of Agricultural Sciences

Recently, the Institute of Crop Science, Chinese Academy of Agricultural Sciences/Sanya Nanfan Research Institute of Big Data Intelligent Design and Breeding Innovation Team jointly proposed a deep learning method (Deep Neural Network for Genomic Prediction (DNNGP) that uses massive multiomics data of plants for whole genome prediction, which can realize the efficient integration and utilization of breeding big data, which will help the application of deep learning in whole genome selection.
Provide effective tools
for intelligent design breeding and platform building.
The results were published
online in Molecular Plant.

 

As a new generation of breeding technology, genome-wide selection predicts and selects early individuals according to the estimated breeding value of the genome by constructing a prediction model, thereby shortening the breeding generation interval, accelerating the breeding process, saving costs, and promoting the development
of modern breeding in the direction of precision and efficiency.
As the core of genome-wide selection, statistical models greatly affect the accuracy and efficiency
of genome-wide prediction.
Traditional prediction methods are based on linear regression models, which make it difficult to capture complex relationships
between genotypes and phenotypes.
Compared with traditional models, nonlinear models (such as deep network nerves) have the ability to analyze complex non-additive effects, artificial intelligence and deep learning algorithms provide new opportunities to solve problems such as big data analysis and high-performance parallel computing, and the optimization of deep learning algorithms will improve the prediction ability
of genome-wide selection.

 

Using population data from four different dimensions of corn, wheat and tomato as test materials, the research team developed a new genome-wide selection method DNNGP
through an innovative deep learning algorithm framework 。 With five other mainstream forecasting methods (i.
e.
: GBLUP based on linear models; Machine learning-based LightGBM and SVR; Compared with deep learning-based DeepGS and DLGWAS), DNNGP has the following advantages: genome-wide prediction can be carried out using multiomics data; The algorithm design includes structures such as batch normalization layer, callback function and corrected linear activation function, which can effectively reduce the error rate of the model and improve the running speed.
The prediction accuracy is robust, and the performance on small datasets is comparable to the current mainstream forecasting models, and the prediction advantages on large-scale datasets are more obvious.
The calculation time is similar to that of traditional methods, and nearly 10 times faster than existing deep learning methods.
Hyperparameter tuning is more user-friendly
.
DNNGP performs efficient mathematical modeling for massive data with complex structures, realizes the efficient integration and utilization of breeding big data, will facilitate the application of deep learning in whole genome selection, and provide effective tools
for intelligent design breeding and platform construction.

 

Wang Kelin, a master's student graduated from the Institute of Science and Technology, is the first author of this paper, and researcher Li Huihui, the leader of the team, is the corresponding author
.
The research has been supported
by the National Key Research and Development Program of China, the National Natural Science Foundation of China, Hainan Yazhou Bay Seed Laboratory and the Science and Technology Innovation Project of the Chinese Academy of Agricultural Sciences.

 

Paper Link:

 

　　https://doi.
org/10.
1016/j.
molp.
2022.
11.
004