Loopholes and missing links in protein modeling

Abstract

This paper provides an unbiased comparison of four commercially available programs for loop sampling, Prime, Modeler, ICM, and Sybyl, each of which uses a different modeling protocol. The study assesses the quality of results and examines the relative strengths and weaknesses of each method. The set of loops to be modeled varied in length from 4-12 amino acids. The approaches used for loop modeling can be classified into two methodologies: ab initio loop generation (Modeler and Prime) and database searches (Sybyl and ICM). Comparison of the modeled loops to the native structures was used to determine the accuracy of each method. All of the protocols returned similar results for short loop lengths (four to six residues), but as loop length increased, the quality of the results varied among the programs. Prime generated loops with RMSDs <2.5 Å for loops up to 10 residues, while the other three methods met the 2.5 Å criteria at seven-residue loops. Additionally, the ability of the software to utilize disulfide bonds and X-ray crystal packing influenced the quality of the results. In the final analysis, the top-ranking loop from each program was rarely the loop with the lowest RMSD with respect to the native template, revealing a weakness in all programs to correctly rank the modeled loops.

Closing the side-chain gap in protein loop modeling

“…Since the RMSDs of the starting side-chains were initially quite small it is possible that the software algorithms were simply unable to improve upon such good results. The current results support the published data, including our own, which suggest that modeling software can accurately model short loops in proteins [12][13][14]20]. Table 1 shows that while the individual results from each method demonstrated little improvement, examination of the percent improved value tells a different story.…”

“…As mentioned above, the initial conformations of the loops originated from our previous study [12] where loops with the lowest backbone RMSD relative to the native form were selected regardless of which algorithm was used to calculate the loop. The previous study concluded that Prime generated the most accurate loops in the test set, therefore, it is not surprising that the majority of starting loop and side-chain orientations (126 out of 197) originated from Prime calculations.…”

“…Therefore, calculated loops from our previous study were used as the starting point for side-chain sampling. We used our previous dataset containing a total of 197 loops with lengths ranging from 4 to 12 residues [12]. The native protein structures selected for this dataset are a subset of those used by both Sali [13] and Friesner [14], that contain high quality X-ray structures with resolutions better than 2.0 Å and average temperature factors of below 35.…”

“…Target model refinement In our previous work, we addressed issues related to protein sequence alignment and model backbone generation [1]. Those initial studies prompted us to explore the accuracy of protein loop modeling, and results of a subsequent publication showed that loops up to 12 residues in length could be modeled accurately with a backbone RMSD of \2.5 Å [12]. However, only the RMSD of the backbone atoms were scrutinized.…”

GPU-accelerated multi-scoring functions protein loop structure sampling

“…The main reason is that the 1xyz(813:824) loop is deeply buried into the protein, which leads to strong interactions with the other atoms in 1xyz and thereby relatively high score values in all three scoring functions we used. The difficulty of modeling 1xyz(813:824) is also discussed in [27]. …”