Predicting gene mutation–phenotype associations is a core task of precision medicine—providing molecular evidence for disease stratified diagnosis, individualized treatment, and prognostic assessment based on the correspondence between individual genetic variants and clinical phenotypes. However, existing methods face three bottlenecks: traditional graph models lack biological prior constraints, multi-omics fusion models struggle with sparsely labeled data (including association sparsity with extremely low verified association ratios and sample sparsity with ≤ 50 labels for some phenotypes), and predictions lack biological interpretability. To address these, we introduce a novel ComLeaCor-Net model for the first time, adopting a closed-loop architecture of “sparse association graph completion - global association feature learning - sparse sample error correction” On ClinVar and GWASdb datasets, it achieves AUC-ROC (overall ability to distinguish positive/negative associations) of 0.886 ± 0.012 and 0.857 ± 0.014, 5.5 percentage points higher than mainstream model MILTON. Performance degradation in sparse scenarios is only 9%, and SPA (sparse phenotype prediction accuracy) is improved by 6.7%–9.7%. GO/HPO enrichment analysis and SHAP interpretation confirm its biological validity, accurately identifying key pathogenic mutations like TP53 and BRCA1. This model establishes a new paradigm of “performance-sparseness adaptation-biological interpretability” collaborative optimization, offering an efficient computational tool for precision medicine.
Yao et al. (Thu,) studied this question.