Study on genetic trait prediction and explanatory model based on large-scale phenotype-gene association data

In this study, a framework of genetic trait prediction and explanatory model based on large-scale phenotype-gene association data is proposed to solve the shortcomings of traditional methods in prediction accuracy and biological interpretability of complex traits. The framework integrates the convolutional neural network (CNN) and the Transformer architecture, and constructs a "prediction-interpretation" dual-task model-KG-DTT (Knowledge-Guided Dual-Task Transformer). By introducing the annotation of gene ontology (GO) as prior knowledge, the model pays more attention to the function-related gene modules, and an interactive visualization tool is developed to reveal the gene-phenotype association network. The experimental results show that KG-DTT is superior to the traditional multi-gene risk score, XGBoost and DeepSEA models in the data sets of human metabolic diseases and crop traits, with an average AUC-ROC value of 0.86 and an R value of 0.41 in rice yield prediction. In addition, the model successfully identified the key genes related to type 2 diabetes mellitus (T2D) through attention mechanism, and verified the biological effects of the newly discovered gene-phenotype association through wet experiments, such as the knockdown of CERS6 gene leading to a 32% decrease in glucose intake in insulin resistance (P = 0.004). This study not only improves the accuracy of genetic trait prediction, but also significantly enhances the biological interpretability of the model, providing new tools and ideas for precision medicine and breeding research.