Discovering the "recognition code" governing protein-DNA interaction has been an important topic for decades in bioinformatics. While other studies have focused on analyzing the frequency of amino acid-base contacts, this study here attempts to discover the structural and physicochemical features of proteins that determine the specificity of amino acid-base contacts. For each amino acid that contacts with DNA, we attempt to predict the type of bases (purines or pyrimidines) that it contacts. We extract 8 structural and physicochemical features from proteins and use a bottom-up approach to search for the combination of features that can be used to predict the specificity of amino acid-base contacts. In the end, 4 features are selected. Using these features, a support vector machine method can achieve 67.1% accuracy with 0.329 MCC in predicting the type of base (purines or pyrimidines) that an amino acid contacts. Analyzing the selected features will provide insights into the "recognition code" of protein-DNA interaction.