Decoding genetic basis of nitrogen use efficiency in maize using AI-generated deep phenotypes

Area of Science:

• Agricultural Science
• Genetics
• Artificial Intelligence

Background:

• Improving nitrogen use efficiency (NUE) is crucial for sustainable agriculture.
• Conventional plant phenotyping methods have limitations in identifying genetic variation for NUE.

Purpose of the Study:

• To leverage artificial intelligence (AI) to discover novel phenotypic variation associated with NUE.
• To identify genetic loci and candidate genes related to NUE that are missed by conventional approaches.

Main Methods:

• Trained a convolutional neural network (CNN) on 25,080 maize images to differentiate responses to low- and high-nitrogen conditions.
• Defined AI-learned features as 'deep phenotypes' and compared their variation and heritability with conventional phenotypes.
• Identified significant genetic loci and functionally characterized candidate genes, including Liguleless2 (LG2).

Main Results:

• Deep phenotypes derived from AI showed greater variation and heritability than conventional phenotypes.
• AI identified 523 significant loci for NUE, compared to 21 using conventional methods.
• The LG2 gene was validated as a contributor to enhanced root architecture and nitrogen uptake.
• Field trials showed that beneficial alleles identified by AI increased ear weight under low-nitrogen conditions.

Conclusions:

• AI can uncover previously unrecognized phenotypic and genetic variation for complex traits like NUE.
• Deep phenotypes provide a more powerful approach for genetic analysis and breeding strategies.
• Integrating AI with biological data enhances the interpretability of AI models and accelerates crop improvement for sustainable agriculture.