In recent years, causal insights have been used to improve machine learning methods by introducing assumptions. Following this trend, we propose a new method to improve the estimation of Conditional Average Treatment Effect (CATE) based on ML methods. In CATE estimation, a common approach is to use metalearners, which are able to estimate CATE using conventional machine learning models if certain identification properties (the backdoor criterion) are satisfied. In this approach, the causal knowledge of the problem (e.g., the causal graph) is used only for the identification of the estimand, not for the estimation itself. We describe a new approach that exploits causal knowledge during the estimation phase by adding constraints during model training. These constraints are based on the conditional independence structure encoded in the causal graph. We apply the constraints to tree based algorithms and show that models trained with these constraints achieve higher performance and lower variability when used to estimate CATE. Our experiments also show that this approach can improve performance even in cases where the causal knowledge of the data is unknown and must be obtained by causal discovery algorithms.