Abstract Background Protein carbonylation, an irreversible and non-enzymatic post-translational modification (PTM), is often used as a marker of oxidative stress. When reactive oxygen species (ROS) oxidized the amino acid side chains, carbonyl (CO) groups are produced especially on Lysine (K), Arginine (R), Threonine (T), and Proline (P). Nevertheless, due to the lack of information about the carbonylated substrate specificity, we were encouraged to develop a systematic method for a comprehensive investigation of protein carbonylation sites. Results After the removal of redundant data from multipe carbonylation-related articles, totally 226 carbonylated proteins in human are regarded as training dataset, which consisted of 307, 126, 128, and 129 carbonylation sites for K, R, T and P residues, respectively. To identify the useful features in predicting carbonylation sites, the linear amino acid sequence was adopted not only to build up the predictive model from training dataset, but also to compare the effectiveness of prediction with other types of features including amino acid composition (AAC), amino acid pair composition (AAPC), position-specific scoring matrix (PSSM), positional weighted matrix (PWM), solvent-accessible surface area (ASA), and physicochemical properties. The investigation of position-specific amino acid composition revealed that the positively charged amino acids (K and R) are remarkably enriched surrounding the carbonylated sites, which may play a functional role in discriminating between carbonylation and non-carbonylation sites. A variety of predictive models were built using various features and three different machine learning methods. Based on the evaluation by five-fold cross-validation, the models trained with PWM feature could provide better sensitivity in the positive training dataset, while the models trained with AAindex feature achieved higher specificity in the negative training dataset. Additionally, the model trained using hybrid features, including PWM, AAC and AAindex, obtained best MCC values of 0.432, 0.472, 0.443 and 0.467 on K, R, T and P residues, respectively. Conclusion When comparing to an existing prediction tool, the selected models trained with hybrid features provided a promising accuracy on an independent testing dataset. In short, this work not only characterized the carbonylated substrate preference, but also demonstrated that the proposed method could provide a feasible means for accelerating preliminary discovery of protein carbonylation.