DISGRO-H3 also performs well in identifying native-like conformations using an empirical potential function. Therefore, prediction of protein loop structures is an important problem, and has received considerable attention.423 Computational prediction of loops in antibodies has the promise to determine the detailed structural basis of specific antibody-antigen interactions. Antibodies are a class of Y-shaped proteins produced by Btk inhibitor 2 the immune system that identify and neutralize foreign pathogens. They can recognize and bind to antigens with remarkable affinity and specificity,24,25and can be used for preventing and treating various diseases, e.g. cancer,26,27arthritis,28and infectious diseases.29The remarkable binding specificity and affinity are decided mainly by the six hypervariable loops, also referred to as the complementary determining regions (CDR). Three of the loops (L1, L2, and L3) belong to the variable domain name of the light chain, and Btk inhibitor 2 the other three (H1, H2, and H3) belong to the variable domain name of the heavy chain. Among the six loops, the H3 loop locates at the center of the binding site and plays important roles in determining the specificity of antibody-antigen interactions.30,31H3 loop is also the most diverse in terms of sequence identity and length, resulting in a larger and more complex structure space. The five non-H3 loops are usually short, with length ranges between 2 and 8 residues. In contrast, the H3 loop may have length up to 26 residues.32The structures of the five non-H3 loops can be predicted with high accuracy using canonical classes of non-H3 CDR loops,32,33which describe the relationships between the sequences and structures of those loops. However, the prediction of H3 loops has not been very satisfactory due to the increased sequence length and much larger conformational space. A number of methods have been developed for modeling antibody H3 loops. They are either based onab initioconformational search,34or rely on knowledge-based databases of structural templates of loop conformations.35,36One of the widely used methods is RosettaAntibody,34which is anab initiomethod based on fragment assembly. It achieves an average RMSD of 2.18 for the lowest energy conformations on a dataset of 53 loops. The 53 average minimum RMSD of the top 10 lowest energy conformations is usually 1.51 . Casp3 FREAD predicts antibody H3 loops based on local sequence and geometric similarities.35H3Loopred, a recently published method uses Random Forest method to select structural templates for H3 loops from Btk inhibitor 2 a set of candidates. After modeling the other CDR loops using canonical structure model implemented in PIGS,37it achieves an average of 2.4 for the same RossetaAntibody data set of 53 loops, and 2.5 for H3 loops using a set of 50 recently solved structures.36However, the strategy of using a knowledge-based database does not apply if there are no appropriate templates available for the loops to be modeled. As many loops may not be adequately described by a single conformation, generating an ensemble of near-native loop conformations can provide better characteristics of the dynamics of these loops, and can help to assess their entropic effects.38,39As current knowledge-based methods mostly focus on generating a single conformation of low energy, they are not well suited for generating large ensembles of near-native loop conformations. Despite extensive efforts with significant progress made in recent years, conformational sampling and structural Btk inhibitor 2 prediction of H3 loops remains a challenging problem when structural templates are not available, especially when loops are long (e.g.,l 17). In this study, we report a novel method for conformational sampling and structural prediction of H3 loops. Our method is called Distance-guided Sequential chain-Growth Monte Carlo forH3loops (DISGRO-H3). It is based on chain growth with importance sampling.21,4044The strategy of sequential chain-growth sampling has been used in various of biological studies,4249for example, analysis of protein packing and void formation,43HP sequence folding problem,44RNA loop entropy calculation,45reconstruction of transition state ensemble of protein folding,47and reaction network sampling.48Prior to conformation sampling, we employ the H3-rules to classify CDR-H3 structure into two conformation types (kink or extended) using sequence information,50,51as successfully used in RosettaAntibody.34The H3-rules can predict a loop in either kink or extended conformation type with 85% accuracy, which substantially reduces the conformational space for the subsequent sampling step. Overall, DISGRO-H3 has.