Prediction of Temperature Factors in Proteins: Effect of Data Pre-Processing and Experimental Conditions

The B-factor or temperature factor is one of the most important parameters in addition to the atomic coordinates, and which is refined during the determination of the protein structure and stored in the Protein Data Bank. It reflects the uncertainty of the atomic positions and is closely linked to atomic flexibility. By using graphlet degree vectors as feature descriptors in a linear model—together with appropriate data transformation and consideration of various experimental factors—the model provides better prediction results. For example, the inclusion of crystal contacts in the linear model significantly improves the prediction accuracy. Since the distributions of the B-factors typically follow an inverse gamma distribution, applying a logarithmic transformation further improves the performance of the model. It has also been shown that large ligands, such as those found in protein–DNA complexes, have a significant impact on the quality of the prediction. A linear model based on graphlet degree vectors proves to be effective not only for the prediction of B-factors and the validation of deposited protein structures but also for the qualitative estimation of root-mean-square fluctuations derived from molecular dynamics.