Data-driven reduced-order modeling (ROM) methods are widely used in aerodynamic flow modeling. They are mainly used to predict distributed quantities obtained from high-fidelity simulations such as surface pressure distributions and flow fields. However, such models only consider the distributed quantities and do not consider the data of the aerodynamic coefficients which are also available from high-fidelity simulations. This work proposes a novel supervised ROM architecture that learns from both distributed quantities and aerodynamic coefficients. The proposed model is characterized using a transonic airfoil modeling problem and compared with a standard ROM and neural network model. The results demonstrate that the supervised ROM architecture can outperform a standard neural network by 12% in predicting aerodynamic coefficients when only using 25 training samples. Even with 300 training samples, the supervised ROM can outperform the neural network by 1 or 2%. This can be achieved while maintaining the same level of accuracy as a standard ROM in predicting airfoil surface pressure distributions. This demonstrates that the proposed model can lead to sample-efficient aerodynamic modeling by reducing computational cost and enhancing model accuracy.