STRUCTURAL MODELING OF PROTEIN-PROTEIN INTERACTIONS USING MULTIPLE-CHAIN THREADING AND FRAGMENT ASSEMBLY

Abstract

Since its birth, the study of protein structures has made progress with leaps and bounds. However, owing to the expenses and difficulties involved, the number of protein structures has not been able to catch up with the number of protein sequences and in fact has steadily lost ground. This necessitated the development of high-throughput but accurate computational algorithms capable of predicting the three dimensional structure of proteins from its amino acid sequence. While progress has been made in the realm of protein tertiary structure prediction, the advancement in protein quaternary structure prediction has been limited by the fact that the degree of freedom for protein complexes is even larger and even fewer number of protein complex structures are present in the PDB library. In fact, protein complex structure prediction till date has largely remained a docking problem where automated algorithms aim to predict the protein complex structure starting from the unbound crystal structure of its component subunits and thus has remained largely limited in terms of scope. Secondly, since docking essentially treats the unbound subunits as "rigid-bodies" it has limited accuracy when conformational change accompanies protein-protein interaction. In one of the first of its kind effort, this study aims for the development of protein complex structure algorithms which require only the amino acid sequence of the interacting subunits as input. The study aimed to adapt the best features of protein tertiary structure prediction including template detection and ab initio loop modeling and extend it for protein-protein complexes thus requiring simultaneous modeling of the three dimensional structure of the component subunits as well as ensuring the correct orientation of the chains at the protein-protein interface. Essentially, the algorithms are dependent on knowledge-based statistical potentials for both fold recognition and structure modeling. First, as a way to compare known structure of protein-protein complexes, a complex structure alignment program MM-align was developed. MM-align joins the chains of the complex structures to be aligned to form artificial monomers in every possible order. It then aligns them using a heuristic dynamic programming based approach using TM-score as the objective function. However, the traditional NW dynamic programming was redesigned to prevent the cross alignment of chains during the structure alignment process. Driven by the knowledge obtained from MM-align that protein complex structures share evolutionary relationships and the current protein complex structure library already contains homologous/structurally analogous protein quaternary structure families, a dimeric threading approach, COTH was designed. The new threading-recombination approach boosts the protein complex structure library by combining tertiary structure templates with complex alignments. The query sequences are first aligned to complex templates using the modified dynamic programming algorithm, guided by a number of predicted structural features including ab initio binding-site predictions. Finally, a template-based complex structure prediction approach, TACOS, was designed to build full-length protein complex structures starting from the initial templates identified by COTH. TACOS, fragments the templates aligned regions of templates and reassembles them while building the structure of the threading unaligned region ab inito using a replica-exchange monte-carlo simulation procedure. Simultaneously, TACOS also searches for the best orientation match of the component structures driven by a number of knowledge-based potential terms. Overall, TACOS presents the one of the first approach capable of predicting full length protein complex structures from sequence alone and introduces a new paradigm in the field of protein complex structure modeling.