Compared with patients with common diseases, cancer patients usually have a more fragile cellular microenvironment and more complex or varied complications. Therefore, to meet treatment needs while also keeping the body in a state of balance between resistance and protection, there is an urgent need for the large-scale accurate identification of effective cancer drug combinations. Inspired by many approaches to drug molecule representation learning and cancer-related gene recognition, we propose a feature fusion model in this paper. The model can combine the molecular representation of drugs with the representation of target genes containing cancer information. It is a means to infuse specific disease information into drugs to guide drug combination prediction for specific cancers. Experimental results show that our method has achieved outstanding results in terms of predictive performance across seven common cancers, such as colorectal adenocarcinoma. For instance, in the case of osteosarcoma, the model achieved an AUROC of 0.9246. Additionally, excellent results were obtained in terms of running time and model complexity. New pairs of effective drug combinations can be mined for each of them, and out-of-sample predictions can be verified against relevant databases. This demonstrates that the feature fusion model can make full and convenient use of the existing advanced feature extraction technology and effectively promote the realization of precision therapy. In addition, we integrated a cancer drug combination dataset constructed from anticancer drugs. All entries in this dataset are standardized and uniformly identified, containing a wealth of information for the search of experts in related fields.