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

Improving nitrogen-use efficiency (NUE) is essential for sustainable agriculture, yet conventionally measured plant characteristics have limited value as NUE proxies. Here, we show that artificial intelligence (AI) can uncover previously unrecognized phenotypic variation associated with NUE, showing genetic variation that is largely missed by conventional phenotypes. We trained a convolutional neural network (CNN) on 25,080 maize images to learn features that distinguish how plants respond to low- and high-nitrogen (N) conditions, achieving 96.7% accuracy. The learned features were defined as deep phenotypes. Compared with conventional phenotypes, deep phenotypes showed greater phenotypic variation and higher heritability, enabling the identification of 523 significant loci compared with 21 for conventional phenotypes. We next investigated candidate genes underlying these loci and used these findings to interpret the learned features. Lower CNN layers largely reflected visual patterns overlapping with conventional phenotypes, whereas deeper layers encoded additional features associated with N-responsive genetic variation. To validate candidate genes identified by the AI framework, we functionally characterized Liguleless2 (LG2), a basic-leucine zipper transcription factor, and demonstrated that lg2 mutants exhibit enhanced root architecture and increased N-uptake efficiency. Field trials of 200 hybrids across diverse N environments further supported the AI findings, with each beneficial allele increasing ear weight by an average of 18 g per plot under low-N conditions. These results show how integrating AI and biology can uncover biologically relevant variation underlying complex traits such as NUE and enhance the interpretability of AI models.