Template Based Modeling and Structural Refinement of Protein-Protein Interactions.

Abstract

Determining protein structures from sequence is a fundamental problem in molecular biology, as protein structure is essential to understanding protein function. In this study, I developed one of the first fully automated pipelines for template based quaternary structure prediction starting from sequence. Two critical steps for template based modeling are identifying the correct homologous structures by threading which generates sequence to structure alignments and refining the initial threading template coordinates closer to the native conformation. I developed SPRING (single-chain-based prediction of interactions and geometries), a monomer threading to dimer template mapping program, which was compared to the dimer co-threading program, COTH, using 1838 non homologous target complex structures. SPRING’s similarity score outperformed COTH in the first place ranking of templates, correctly identifying 798 and 527 interfaces respectively. More importantly the results were found to be complementary and the programs could be combined in a consensus based threading program showing a 5.1% improvement compared to SPRING. Template based modeling requires a structural analog being present in the PDB. A full search of the PDB, using threading and structural alignment, revealed that only 48.7% of the PDB has a suitable template whereas only 39.4% of the PDB has templates that can be identified by threading. In order to circumvent this, I included intramolecular domain-domain interfaces into the PDB library to boost template recognition of protein dimers; the merging of the two classes of interfaces improved recognition of heterodimers by 40% using benchmark settings. Next the template based assembly of protein complexes pipeline, TACOS, was created. The pipeline combines threading templates and domain knowledge from the PDB into a knowledge based energy score. The energy score is integrated into a Monte Carlo sampling simulation that drives the initial template closer to the native topology. The full pipeline was benchmarked using 350 non homologous structures and compared to two state of the art programs for dimeric structure prediction: ZDOCK and MODELLER. On average, TACOS models global and interface structure have a better quality than the models generated by MODELLER and ZDOCK.