Large DNA and protein language models enhance discovery of deleterious mutations in maize.

backgroundThe maize inbred line Chang7-2 and derived lines are important male donors for hybrid production, contributing significantly to the development of high-yield and stress-tolerant hybrids. Additionally, Chang7-2 serves as a valuable model inbred line for genetic and genomic studies, facilitating the discovery of genes underlying hybrid vigor and other agronomic traits.

resultsHere, a reference genome assembly and a chemical-induced mutant population (N = 1,716) through ethyl methyl sulfonate (EMS) treatments are generated using Chang7-2. Each EMS line is whole genome sequenced and compared to the Chang7-2 genome, identifying 2,586,769 mutations with 4,939 mutations causing premature stop codons or altered splicing sites. The effect estimation of mutations using two large language artificial intelligence (AI) models, namely the protein language model ESM1b and the DNA language model PlantCaduceus, reveals 15,264 and 18,326 deleterious mutations, respectively. Mutation effects estimated with AI models accelerate revelation of four causal mutations underlying phenotypes of albino leaf, reduced cuticular wax, altered seed color, and male sterility. In addition, allelic expression quantification of genic mutations in 13 EMS M1 lines and their M2 heterozygous progeny, which contain both wildtype and mutant alleles, shows that mutant alleles are overall accumulated at a lower level compared to wildtype. Such allelic disparity is observed for some synonymous mutations, indicating they may not be biologically inconsequential.

conclusionsAI-based estimation of mutation effects offers cross-species evidence for functional impacts of mutations. Our study demonstrates its application in revealing deleterious EMS mutations and identifying causal mutations responsible for mutant phenotypes.