The Study On Modeling Protein Side-chain With HMM And Its Application

It has been a challenge of bioinformatics to predict the protein structure from itsamino acid sequence for a long period of time. Side-chain prediction is an importantsub-problem of protein structure prediction and design.The thesis proposes a strategy of modeling protein side-chain based on HiddenMarkov Model. We construct sequence-dependent model and sequence-backbone-dependent model for protein side-chain. The sequence-dependent model takes aminoacids sequence information as main observed data; while the sequence-backbone-dependentmodel considers not only amino acids sequence information but also backbone dihedralangles as observed data. The models are trained to learn the relationships among theobserved data and side-chain conformation. After being trained based on compara-tive training data, these two models can produce the specific rotamer library giventhe target amino acids sequence by sampling the two trained models. Comparing torotamer library produced by the other popular side-chain modeling methods, ours aremore close to native conformation.We apply the rotamer libraries generated by our models to the state-of-the-artsystem of side-chain prediction. We compare the accuracy of predicted protein side-chain based on di?erent rotamer libraries generated by di?erent side-chain models. Wefind that the prediction accuracy outperforms on all the test targets to a certain extentcomparing with that based on traditional backbone dependent rotamer libraries.This thesis verifies that the sequence-dependent model and sequence- and backbone-dependent model will enable fully taking into account the possible a?ection of sequenceamino acids to the specific amino acid side-chain. The rotamer libraries generated bysequence-dependent models provide solid support to protein side-chain prediction.