The protein structure prediction problem could be solved using the current PDB library

Abstract

For single-domain proteins, we examine the completeness of the structures in the current Protein Data Bank (PDB) library for use in full-length model construction of unknown sequences. To address this issue, we employ a comprehensive benchmark set of 1,489 medium-size proteins that cover the PDB at the level of 35% sequence identity and identify templates by structure alignment. With homologous proteins excluded, we can always find similar folds to native with an average rms deviation (RMSD) from native of 2.5 A with approximately 82% alignment coverage. These template structures often contain a significant number of insertions/deletions. The tasser algorithm was applied to build full-length models, where continuous fragments are excised from the top-scoring templates and reassembled under the guide of an optimized force field, which includes consensus restraints taken from the templates and knowledge-based statistical potentials. For almost all targets (except for 2/1,489), the resultant full-length models have an RMSD to native below 6 A (97% of them below 4 A). On average, the RMSD of full-length models is 2.25 A, with aligned regions improved from 2.5 A to 1.88 A, comparable with the accuracy of low-resolution experimental structures. Furthermore, starting from state-of-the-art structural alignments, we demonstrate a methodology that can consistently bring template-based alignments closer to native. These results are highly suggestive that the protein-folding problem can in principle be solved based on the current PDB library by developing efficient fold recognition algorithms that can recover such initial alignments.

Fig. 1.

Overview of the tasser structure prediction methodology that consists of template identification (here by structure alignment), force field construction, structure assembly, and model selection.

Fig. 2.

RMSD to native of the templates identified by the structure alignment program sal (20) versus the alignment coverage.

Fig. 3.

Comparison of initial and final alignments to the target structure. (a) Scatter plot of RMSD from native of the final models built by tasser refinements versus RMSD from native of the initial template alignments identified by sal. The same aligned regions are used in both RMSD calculations. (b) Similar data as in a, but the models are from modeller refinements. (c) Fraction of targets with an RMSD improvement “d” by tasser greater than some threshold value. Here d = (RMSD of template) – (RMSD of final model), where both RMSDs are calculated over aligned regions. Each point in c is calculated with a bin width of 1 Å. (d) Similar data as in c, but the models are from modeller.

Fig. 4.

RMSD_local (a) and RMSD_global (b) of unaligned/loop regions as a function of loop length (L). tasser and modeller models are denoted by triangles and circles, respectively. The lines connecting the points serve to guide the eye. The dashed line in b denotes an RMSD_global cutoff of 7 Å.

Fig. 5.

Representative examples showing the improvement of final models with respect to their initial template alignments. The left column is the superposition of the template alignments and native, and the right column is that for refined models and native. The thin lines are native structures; the thick lines signify templates or final models. Blue to red runs from N to C terminus. The numbers in parentheses are the coverage of the templates or models on which the denoted RMSD has been calculated.

Fig. 6.

A representative example of targets where the final model has an RMSD to native >6 Å. The native structure of 1k5dB is shown in white with thin backbones; the predicted model of the highest cluster density in blue with thick backbones. The red wire-frame in a denotes the partner chains in the Ran-RanBP1–RanGAP complex. (a) 1k5dB in the entire complex. (b) Native model superposition. (c) As in b but with the tail cut off.

Fig. 7.

Side-chain contact probability for random chains as a function of chain length. (a) Contact probability q₀(i,j,N/L) for the ith and jth residues at given scaled contact-numbers, N/L, versus the distance |i – j| along the protein chain, from simulations of an FJC of 50 linkers with excluded volume. The Kuhn-length of the FJC is taken to be 11.4 Å, according to experimental single protein molecule stretching data (43). The definition of excluded volume is introduced on the basis of the minimal observed C_α distance in the PDB, i.e., no pair of linkers of the FJC could go closer than 4.5 Å. Contacts are defined based on a weighted average of C_α distances for contacting residues in the PDB, i.e., a contact occurs if two linkers are closer than 7.27 Å. The curves are the least square fits to the power-law at each given N/L.(b) The power γ versus the scaled contact-number N/L. Data are from the FJC of different lengths, i.e., 50-, 100-, 200-, and 300-residue linkers, which correspond to the protein lengths of 150, 300, 600, and 900 residues, respectively, because one Kuhn-length here corresponds approximately to three C_α–C_α backbone lengths. The error bars denote the power-law fitting errors of a. The curve denotes the least square fit equation: γ = –2.8 + 2.5√N/L.