Transformer Fusing Chromosome Conformation and Genomic Information for Soybean Trait Prediction

Genomic information is increasingly leveraged for the precise prediction of crop traits, with the adoption of advanced genomic prediction techniques resulting in substantial improvements in both crop yield and quality. However, traditional genomic prediction methods exhibit notable limitations in capturing long-range dependencies and fully utilizing prior information from chromosome structure. In this work, two novel Transformer models fusing chromosome conformation and genomic information have been proposed. One is the chromosomal self-attention fusion model, which captures cross-chromosomal interactions more precisely by introducing chromosomal conformation information into the self-attention mechanism of the Transformer. The other is the chromatin interaction squeeze excitation model, which extracts global features of the chromosome from all single nucleotide polymorphism sites on each chromosome. It then employs the chromatin interaction matrix to perform a weighted fusion of these global features, enabling the effective utilization of inter-chromosomal information. In addition, two novel metrics are introduced to comprehensively assess the effectiveness of the internal self-attention mechanism. They quantify the concentration of attention while measuring the alignment between the attention distribution and the chromosomal interaction priors. Experiments show that the two proposed models exhibit significant advantages in predicting soybean oil content and protein.