Predicting Phase 3 clinical trial results is a critical step of Go/No-Go decision making and optimization of Phase 3 clinical trial designs. Traditionally, because of limited historical efficacy and safety data about the investigational product, a homogeneity assumption is often made in such predictions that Phase 3 patients are similar to Phase 2 patients in terms of distributions of baseline characteristics. However, Phase 3 clinical trials are larger than Phase 2 clinical trials and such homogeneity assumption is often violated. Thanks to availability of electronic health records, we can better simulate the Phase 3 patients with a random sample of the real world data. Thus, we propose a deep learning approach, where the Phase 2 clinical trial data and the real world data are coupled together to predict Phase 3 clinical trial results without the above homogeneity assumption. We demonstrated that the proposed method can yield much more accurate and robust predictions.