We use logic-based machine learning to distinguish DNAbinding proteins from non-binding proteins. We combine previously suggested coarse-grained features (such as the dipole moment) with automatically constructed structural (spatial) features. Prediction based only on structural features already improves on the state-of-the-art predictive accuracies achieved in previous work with coarse-grained features. Accuracies are further improved when the combination of both feature categories is used. An important factor contributing to accurate prediction is that structural features are not Boolean but rather interpreted by counting the number of their occurences in a learning example.