Innovative Iowa State University Scientist Advances Machine Learning as Tool for Crop Breeding

 

Using machine learning to develop and utilize plant breeding tools that can deliver improved genetics to farmers faster is a dream of Asheesh (Danny) Singh, associate professor of agronomy at Iowa State University.

 

Singh, the Monsanto Chair in Soybean Breeding at Iowa State, collaborates across disciplines with fellow innovators, combining artificial intelligence and genetics to speed selection of crop varieties finely tuned to the needs of farmers now and in the future.

 

Breeding for desired traits in crops has long been a time-consuming, labor-intensive and expensive process. Breeders study generations of plants, identifying and tweaking desired genetics along the way as they assess how traits are expressed in offspring. If all goes well and the results are promising, a new genetic variation may progress step-by-step to commercial development, testing and marketing of a new plant variety.

 

Singh and colleagues in ISU’s Department of Mechanical Engineering think they can streamline the process, shortening the plant breeding timeline to create higher-yielding, more resilient cultivars for specific management situations.

 

In an article, “Machine Learning Approach for Prescriptive Plant Breeding,” recently published in the peer-reviewed journal “Scientific Reports,” Singh and fellow researchers describe a new method that can enable computers to train themselves to develop better, smarter plant breeding models. Co-authors are Kyle A. Parmley and Race Higgins, Singh’s previously graduated PhD agronomy students, Baskar Ganapathysubramanian, professor of mechanical engineering, and Soumik Sarkar, associate professor of mechanical engineering.

 

A related study, described in an article published in the journal “Plant Phenomics,” written by the same authors along with mechanical engineering PhD student Koushik Nagasubramanian, demonstrated developing methods to fuse machine learning with techniques to identify optimal suites of observable characteristics (phenotypes) for more informed breeding decisions and outcomes.

 

The researchers used aerial drones and sensors placed on ground-based robots to gather voluminous information on the soybean varieties’ growth and productivity in situations with varied row spacings and seeding densities. The sensors also collected data on environmental conditions, including temperature, rainfall, light quality and humidity.

 

The team wanted to explore the potential of using a deep learning approach: Could a computer use this detailed data to create a framework, or prescriptive model, to help predict which plant genetics will perform best in different management scenarios, specifically in different row spacings and varied seeding densities?

 

They fed the computer extensive beginning data from the sensors about the plants, their genetics and growing conditions. They also entered ending data about the plants' yield. In between, they wanted the computer to learn how to find its own path between the sets of data, to “connect the dots” to statistically predict which cultivars would perform best in different growing conditions.  

 

Their effort to create a predictive framework largely succeeded. For a number of varieties, the computer was able to find clear statistical patterns to recommend the best fit between the genetics and growing conditions most likely to achieve maximum yield.

 

“For soybeans, this means allowing faster development of genetic lines that will work best in different situations common to Iowa farmers, such as decisions about plant spacing,” Singh said. “But it can be used for many related purposes, including predicting plant lines tailored to different climatic regions and soil conditions, or to breed plants that will be more climate resilient. This is important to get the best plants to farmers to meet their needs, not just for now, but for 10, 20 or 30 years from now.”

 

The team’s focus is on soybeans, but Singh said the approach can work for other crops and cropping situations.

 

This research is of considerable interest to the soybean industry. Ed Anderson is executive director of the North Central Soybean Research Program and senior director of research for the Iowa Soybean Association, both organizations that have supported Singh’s work.

 

“Danny is among the best research leaders I know in working to understand farmer needs,” Anderson said. “He has been uniquely successful in establishing productive interdisciplinary teams of scientists, engineers and computational experts to meaningfully address basic and applied challenges and opportunities, advancing science and the soybean industry.”

 

Support for Singh’s current research has also come from the Iowa Crop Improvement Association, the Monsanto Chair in Soybean Breeding, the United Soybean Board, the North Central Soybean Research Program, USDA’s National Institute of Food and Agriculture, and Iowa State University sources, including the R.F. Baker Center for Plant Breeding, the Iowa Soybean Research Center, the Plant Sciences Institute and the Presidential Interdisciplinary Research Initiative. This work is also supported by team members of Singh’s program