Predicting a DNA-binding protein using random forest with multiple mathematical features

Abstract

DNA-binding proteins are involved and play a crucial role in a lot of important biological processes. Hence, the identification of the DNA-binding proteins is a challenging and significant problem. In order to reveal the intrinsic information correlated to DNA-binding, nine classes of candidate features based on different mathematical fields are applied to construct the prediction model with random forest. They are fractal dimension, conjoint triad feature, Hilbert-Huang Transformation, amino acid composition, dipeptide composition, chaos game representation, and the corresponding information entropies. These mathematical expressions are evaluated with 5-fold cross validation test. The results of numerical simulations show that the mathematical features consisted of amino acid composition, fractal dimension and information entropies of amino acid and chaos game representation achieve the best performance. Its accuracy is 0.8157, and Matthew's correlation coefficient (MCC) achieves 0.5968 on the benchmark dataset from DNA-Prot. By analyzing the components of top combination of the nine candidate features, the concepts of fractal dimension and information entropy are the effective and vital features, which can provide complementary sequence-order information on the basis of amino acid composition.