Rapid and Accurate Protein Side Chain Prediction Using Local Backbone Information Only

Abstract. High-accuracy protein structure modeling demands on accurate and very fast side chain prediction since such a procedure must be repeatedly called at each step of structure refinement. Many known side chain prediction programs, such as SCWRL and TreePack, depend on the philosophy that global information and pairwise energy function must be used to achieve high accuracy. These programs are too slow to be used in the case when side chain packing has to be used thousands of times, such as protein structure refinement and protein design. We present an unexpected study that local backbone information can determine side chain conformations accurately. LocalPack, our side chain packing program which is based on only local information, achieves equal accuracy as SCWRL and TreePack, yet runs many times faster, hence providing a key missing piece in our efforts to high-accuracy protein structure modeling. keyword: side chain prediction, local backbone features, multiclass Support Vector Machines.