scOncoNet：用于单细胞癌症数据中程序感知型计算基因扰动的双图注意力网络

Identifying genes that functionally drive malignant programs in single-cell tumors remains difficult, particularly when CRISPR or Perturb-seq data are unavailable. Co-expression alone cannot separate correlation from regulatory influence, and existing analyses lack a framework that connects global molecular networks with patient-specific cellular programs. Current single-cell foundation models largely rely on gene-gene relationships while making limited use of the cell-cell structure intrinsic to tumor ecosystems. To address these gaps, we developed scOncoNet, a lightweight dual-graph attention network that integrates gene-gene and cell-cell structure, malignant meta-program supervision, and gene-to-program attribution to enable program-aware estimation of gene influence directly from patient scRNA-seq data. The model incorporates PPI networks to provide global molecular context for the gene encoder and constructs a program-aware cell graph blending local transcriptomic similarity with malignant program similarity. A two-layer architecture jointly learns gene and cell embeddings, with an attention mechanism that adaptively reweights cell-cell edges to highlight biologically meaningful transitions such as EMT or hypoxia. A program-prediction module anchors the cell embedding space to malignant programs, and gene impact scores combine PPI-informed gene embeddings with gradients that quantify each gene's influence on these programs, generating a mechanistic estimate of its potential effect on malignant states. Using an esophageal squamous-cell carcinoma (ESCC) single-cell dataset restricted to epithelial cells, we defined ESCC-specific malignant programs, including proliferation ( MKI67 ), EMT ( VIM ), hypoxia ( CA9 ), and immune evasion, and applied scOncoNet to prioritize candidate targets. Top-ranked genes were dominated by ESCRT and proteasome components, including multiple CHMP family members along with VPS25, TSG101, PSMD14 , and PSMC2 , reflecting core survival and vesicle-trafficking functions consistent with their pan-cancer essentiality. Compared with existing scRNA-seq-based approaches, scOncoNet shows stronger alignment between gene embeddings and malignant programs, enabling precise recovery of functional gene-program relationships. Across additional single-cell datasets, the model consistently preserved malignant program structure and generated stable target rankings. In summary, scOncoNet integrates global molecular networks with patient-specific cellular programs to identify functional gene drivers directly from tumor single-cell data. Its ability to reveal core cancer progression machinery and prioritize candidate CRISPR targets without perturbation experiments suggests utility for mechanistic studies and therapeutic target discovery in tumor types lacking perturbation datasets.