MESSM : a framework for protein threading by neural networks and support vector machines
Protein threading, which is also referred to as fold recognition, aligns a probe amino acid sequence onto a library of representative folds of known structure to identify a structural similarity. Following the threading technique of the structural profile approach, this research focused on developing and evaluating a new framework - Mixed Environment Specific Substitution Mapping (MESSM) - for protein threading by artificial neural networks (ANNs) and support vector machines (SVMs). The MESSM presents a new process to develop an efficient tool for protein fold recognition. It achieved better efficiency while retained the effectiveness on protein prediction. The MESSM has three key components, each of which is a step in the protein threading framework. First, building the fold profile library-given a protein structure with a residue level environmental description, Neural Networks are used to generate an environment-specific amino acid substitution (3D-1D) mapping. Second, mixed substitution mapping--a mixed environment-specific substitution mapping is developed by combing the structural-derived substitution score with sequence profile from well-developed amino acid substitution matrices. Third, confidence evaluation--a support vector machine is employed to measure the significance of the sequence-structure alignment. Four computational experiments are carried out to verify the performance of the MESSM. They are Fischer, ProSup, Lindahl and Wallner benchmarks. Tested on Fischer, Lindahl and Wallner benchmarks, MESSM achieved a comparable performance on fold recognition to those energy potential based threading models. For Fischer benchmark, MESSM correctly recognise 56 out of 68 pairs, which has the same performance as that of COBLATH and SPARKS. The computational experiments show that MESSM is a fast program. It could make an alignment between probe sequence (150 amino acids) and a profile of 4775 template proteins in 30 seconds on a PC with IG memory Pentium IV. Also, tested on ProSup benchmark, the MESSM achieved alignment accuracy of 59.7%, which is better than current models. The research work was extended to develop a threading score following the threading technique of the contact potential approach. A TES (Threading with Environment-specific Score) model is constructed by neural networks.