This paper presents an extension of the Virtual Reference Feedback Tuning (VRFT) methodology dedicated to linear time-delay systems with known delay and unknown dynamics. The standard VRFT is not well suited for systems with dominant time-delay as it yields high order controllers. The proposed direct approach, relying on a Smith Predictor structure, guarantees the same level of performance as the standard VRFT but with lower order controllers. The joint direct data-driven design of the controller and the predictor is facilitated by the introduction of an ad-hoc optimization initialization. Effectiveness and robustness to uncertainty in the time-delay estimation are shown in a vehicle dynamics control problem.