Driver genes are pivotal in different biological processes. Current methods generally identify driver genes by associative analysis. Leveraging on the development of current large language models (LLM) in single cell genomics, we proposed a Causal inference based approach to Identify Driver genes (called CID) from scRNA-seq data. Through experiments on three different datasets, we show that CID can (1) identify biologically meaningful driver genes that have not been captured by current associative-analysis based methods, and (2) accurately predict the direction of expression changes in downstream target genes if a driver gene is knocked out. This study presents a resource-efficient in silico framework for identifying key regulatory genes from learned cell embeddings and simulated perturbations, thereby streamlining follow-up experiments and deepening our understanding of how gene-level interventions influence cellular phenotypes and disease.