Deep learning reveals a microprotein atlas in maize and uncovers novel regulators of seed amino acid metabolism

Abstract

Microproteins represent a vast and functionally important class of genes that remain largely unexplored in plant genomes. Here, we developed DeepMp, a deep learning framework that integrated multi-omics evidence and built the most comprehensive plant microprotein atlas to date, identifying 18,338 high-confidence candidates in maize. The majority of these appear to have originated de novo from regions previously annotated as noncoding, and they show hallmarks of rapid, lineage-specific evolution and pronounced tissue specificity. These novel microproteins were found integrated into core regulatory networks, particularly in organ development and nutrient storage. Focusing on the maize kernel, population-scale analyses linked microprotein expression to natural variation in amino acid content. We functionally validated three grain-filling-specific candidates originating from noncoding regions by CRISPR-Cas9 knockouts, which confirmed their roles as precise modulators of arginine, aspartate, and methionine levels, without pleiotropic effects on kernel morphology. Our study provides a foundational resource and analytical framework, establishes microproteins as a new and functionally important coding layer in maize, and uncovers a previously untapped source of targets for crop improvement.