High-Fidelity computational fluid dynamics simulations, such as large eddy simulation and direct numerical simulation (DNS), are computationally expensive. For this reason, the Reynolds-averaged Navier-Stokes (RANS) equations are more popular for industrial applications. They are obtained by deriving time-averaged properties, removing the need to resolve small-scale unsteady turbulent fluctuations. However, RANS requires a closure model to relate unknown Reynolds stresses to the time-averaged velocities. Traditional turbulence closure models are prone to inaccuracies, particularly for separated and transitional flows, due to the chaotic and anisotropic nature of turbulence. In this study, a data-driven approach is used to model turbulence leveraging high-fidelity data from DNS. A canonical turbulent channel case is considered at three different Reynolds numbers and compared with reference results. A deep neural network is trained on these three datasets, and then used to predict two intermediate Reynolds numbers. Comparison with parallel DNS simulations demonstrates agreement from 70% to 98% R2 score for predictions of turbulent kinetic energy production and dissipation, depending on the chosen training and testing datasets. A second study is then performed on a NACA 0012 airfoil at a Reynolds number of 50,000 at angles of attack ranging from 4 to 12 degrees. A deep neural network was trained on a limited set of these angles of attack, and used to predict turbulent kinetic energy production and dissipation. Comparison with the testing dataset showed good agreement at lower angles of attack, with limited agreement at high angles of attack beyond stall. Based on these results, we can conclude that deep neural networks can be trained to accurately predict turbulent kinetic energy production and dissipation for wall-bounded turbulent flows, but there are some limitations for massively separated flows. Future work will focus on directly incorporating these results in a RANS turbulence modelling framework.