Realistic representation of protein flexibility in biomolecular simulations remains an unsolved fundamental problem and is an active area of research. The high flexibility of the cytochrome P450 2D6 (CYP2D6) active site represents a challenge for accurate prediction of the preferred binding mode and site of metabolism (SOM) for compounds metabolized by this important enzyme. To account for this flexibility, we generated a large ensemble of unbiased CYP2D6 conformations, to which small molecule substrates were docked to predict their experimentally observed SOM. SOM predictivity was investigated as a function of the number of protein structures, the scoring function, the SOM-heme cutoff distance used to distinguish metabolic sites, and intrinsic reactivity. Good SOM predictions for CYP2D6 require information from the protein. A critical parameter is the distance between the heme iron and the candidate site of metabolism. The best predictions were achieved with cutoff distances consistent with the chemistry relevant to CYP2D6 metabolism. Combination of the new ensemble-based docking method with estimated intrinsic reactivities of substrate sites considerably improved the predictivity of the model. Testing on an independent set of substrates yielded area under curve values as high as 0.93, validating our new approach.