Single-cell RNA sequencing (scRNA-seq) was used to analyze the GSE161277 dataset to identify candidate NK cell–associated genes. A prognostic model based on NK cell-associated gene signatures was constructed via LASSO regression. Deep learning using the Clustering-constrained Attention Multiple Instance Learning model extracted pathomic features from 458 TCGA-COAD whole-slide images. A multimodal prognostic framework was developed by integrating scRNA-seq, transcriptomic, pathomic, and clinical data. scRNA-seq analysis of the GSE161277 dataset revealed diverse immune and stromal cell populations within the microenvironment, highlighting its cellular heterogeneity. Based on these data, NK cell-associated candidate genes were identified for subsequent prognostic modeling. The NK cell-associated gene signature prognostic model showed high accuracy in predicting overall survival in TCGA COAD cohort, with a concordance index (C-index) of 0.835. The pathology-based prognostic model achieved a C-index of 0.871. The multimodal prognostic framework achieved a C-index of 0.889, outperforming single-modality approaches. This study proposes a multimodal prognostic framework integrating NK cell-associated molecular features, pathomic features, and clinical variables.
Yang et al. (Wed,) studied this question.