This Article Abstract Full Text (PDF) Alert me when this article is cited Alert me if a correction is posted Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Zhang, Y. Articles by Skolnick, J. Search for Related Content PubMed PubMed Citation Articles by Zhang, Y. Articles by Skolnick, J.

Tertiary Structure Predictions on a Comprehensive Benchmark of Medium to Large Size Proteins

Center of Excellence in Bioinformatics, University at Buffalo, Buffalo, New York 14203

Correspondence: Address reprint requests to Jeffrey Skolnick, E-mail: skolnick{at}buffalo.edu.

	ABSTRACT

TOP ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUDING REMARKS ACKNOWLEDGEMENTS REFERENCES

 
We evaluate tertiary structure predictions on medium to large size proteins by TASSER, a new algorithm that assembles protein structures through rearranging the rigid fragments from threading templates guided by a reduced C and side-chain based potential consistent with threading based tertiary restraints. Predictions were generated for 745 proteins 201–300 residues in length that cover the Protein Data Bank (PDB) at the level of 35% sequence identity. With homologous proteins excluded, in 365 cases, the templates identified by our threading program, PROSPECTOR_3, have a root-mean-square deviation (RMSD) to native < 6.5 Å, with >70% alignment coverage. After TASSER assembly, in 408 cases the best of the top five full-length models has a RMSD < 6.5 Å. Among the 745 targets are 18 membrane proteins, with one-third having a predicted RMSD < 5.5 Å. For all representative proteins less than or equal to 300 residues that have corresponding multiple NMR structures in the Protein Data Bank, 20% of the models generated by TASSER are closer to the NMR structure centroid than the farthest individual NMR model. These results suggest that reasonable structure predictions for nonhomologous large size proteins can be automatically generated on a proteomic scale, and the application of this approach to structural as well as functional genomics represent promising applications of TASSER.

	INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUDING REMARKS ACKNOWLEDGEMENTS REFERENCES

 
The protein structure prediction problem, that is, deducing the tertiary structure of a protein from its primary amino acid sequence, has attracted considerable interest in this postgenomic era (Baker and Sali, 2001; Skolnick et al., 2000a). At present, the success rate of structure prediction is dictated by two factors: First, the structure of smaller proteins is easier to predict than those of larger proteins. Given secondary structure assignments (that can be deduced from sequence alone with more than 80% accuracy using state-of-the-art predictors; Jones, 1999; Karplus et al., 1998), the number of ways to assemble the secondary structure blocks into tertiary models increases exponentially with the increasing number of such blocks. Second, since in principle similar sequences have similar folds (Holm and Sander, 1996), solved homologous protein structures can be exploited to greatly increase the accuracy of the predicted models (Marti-Renom et al., 2000). Therefore, in benchmarking tests to establish the applicability of an approach to weakly/nonhomologous proteins, such homologous structures should be carefully excluded.

Until now, most benchmark tests of protein structure prediction algorithms focused on small to medium size proteins. For example, based on an ab initio approach designed to globally optimize their potential energy function, Scheraga et al. could build models of root-mean-square deviation (RMSD) to native below 6 Å for protein fragments of up to 61 residues (Liwo et al., 1999). Using ROSETTA, Baker et al. report 73 successful structure predictions out of 172 target proteins with lengths below 150 residues, with a RMSD < 7 Å in the top five models (Simons et al., 2001). In recent works, we developed a threading template assembly/refinement approach, TASSER, and benchmarked TASSER on a comprehensive benchmark set of 1489 single-domain proteins in the Protein Data Bank (PDB) with length below 200 residues. We find that 990 targets can be folded by the approach; i.e., they have a RMSD < 6.5 Å in at least one of the top five models (Zhang and Skolnick, 2004a). Despite these important efforts, structure prediction on larger proteins with length greater than 200 residues, which is the range of protein lengths adopted by many enzymes and other functionally important proteins, has not previously been systematically explored. Although the Critical Assessment of Techniques for Protein Structure Prediction (CASP) provides a periodic and critical test of all size ranges of proteins (Moult et al., 2001, 2003), because of the relatively small number of targets in various specific categories, a comprehensive general trend still remains to be established.

In this work, we employ a representative benchmark set of all structures in the PDB ranging from 201 to 300 residues in length and present the results of the large-scale testing of TASSER for tertiary structure prediction on these medium to large size proteins. For the first time, folding simulations of multiple-domain proteins and membrane proteins are examined in a series of systematic tests. Finally, a direct comparison of the accuracy of the TASSER predicted models to the spatial uncertainty of NMR experimental structures is made.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUDING REMARKS ACKNOWLEDGEMENTS REFERENCES

 
The threading template assembly/refinement procedure, TASSER, consists of threading template identification, fragment assembly, and final model combination (Zhang and Skolnick, 2004a). A flowchart is presented in Fig. 1.

View larger version (34K): [in this window] [in a new window]   FIGURE 1  Flowchart of the TASSER structure prediction methodology that consists of template identification by threading, fragment assembly, and fold selection.

To select the final template alignments for TASSER assembly, we establish two sets of Z-score cutoffs based on benchmarking statistics (Skolnick et al., 2004): Z_struct, above which 95% of templates have their best structure alignment with a RMSD to native < 6.5 Å over the aligned regions, and Z_good, above which 80% of threading-predicted alignments have a RMSD < 6.5 Å in aligned regions. If a target has a template with a Z-score > Z_good or two templates with consensus alignments both having a Z-score > Z_struct, the target is assigned to the Easy set (note that Easy does not imply the results are trivially found; in a benchmark test, approximately half the proteins in the Easy set are not identified by PSI-BLAST; Frishman et al., 2003; Kawabata et al., 2002; Zhang and Skolnick, 2004a); if a target has a single template (or has multiple templates lacking a consensus structure) where Z_good > Z-score > Z_struct, the target is assigned as to the Medium set; all others are Hard targets.

TASSER force field
A protein's conformation is described by its C atoms and side-chain centers of mass (SG), called the CAS model. The force field employed in TASSER modeling consists of three classes of terms: 1), statistical potentials from the PDB database (Kolinski and Skolnick, 1998; Zhang et al., 2003), including long-range SG-pair interactions, local C correlations, hydrogen-bond, and hydrophobic burial interactions; 2), propensities for predicted secondary structures from PSIPRED (Jones, 1999); and 3), protein specific SG-pair potentials and tertiary contact restraints extracted from the threading templates by PROSPECTOR_3 (Skolnick et al., 2004).

The combination of all the energy terms was optimized by maximizing the correlation between the CAS energy and RMSD of decoy structures to native, on the basis of 100 training proteins outside the benchmark test set, each with 60,000 structure decoys (Zhang et al., 2003).

Compared with previous energy potential (Zhang et al., 2003; Zhang and Skolnick, 2004a), to increase the hydrogen-bond geometrical specificity, the hydrogen bond used in this work is constructed by including backbone N and CO groups rather than using just the C approximation. Protein specific pair interactions are derived on the basis of a freely jointed chain model simulation (our unpublished results) rather than using the quasichemical approximation (Skolnick et al., 2000b). These changes increase the correlation of the energy with RMSD (the average correlation coefficient between energy and RMSD in the decoy structures of 100 training proteins increases from 0.69 to 0.75); this also improved the performance of TASSER simulations.

On- and off-lattice model and structure assembly
A protein chain in TASSER modeling is divided into aligned and unaligned regions based on its PROSPECTOR_3 alignments, where the aligned regions are modeled off lattice for maximal accuracy and the unaligned regions are simulated on a cubic lattice system for computational efficiency (Fig. 2).

View larger version (58K): [in this window] [in a new window]   FIGURE 2  Schematic representation of a piece of polypeptide chain in the combined on- and off-lattice CAS model. Each residue is described by its C and side-chain center of mass (SG). Although C's of unaligned residues (white) are confined to the underlying cubic lattice system with a lattice space of 0.87 Å, C's of aligned residues (yellow) are excised from threading templates and traced off lattice. SGs are always off lattice (red) and determined using a two-rotamer approximation (Zhang et al., 2003).

The initial full-length models are submitted to parallel hyperbolic Monte Carlo sampling (Zhang et al., 2002) for assembly/refinement. Two kinds of conformational updates are implemented: Off-lattice movements of the aligned regions involve rigid fragment translations and rotations that are controlled by the three Euler angles. The fragment length normalizes the movement amplitude so that the acceptance rate is approximately constant for different size fragments. The lattice confined residues are subjected to two to six bond movements and multibond sequence shifts (Zhang et al., 2003). Overall, the tertiary topology varies by the rearrangement of the continuously aligned substructures, where the local conformation of the off-lattice substructures remains unchanged during assembly. Both movement of the aligned and the gap regions are guided by the same CAS force field.

Clustering
Forty replicas are employed in the Monte Carlo simulation. Structures generated in the 14 lowest temperature replicas are submitted to an iterative structural clustering program, SPICKER (Zhang and Skolnick, 2004b), for clustering. The final models are combined from the clustered structures and ranked by the structure density D, i.e., where M is the multiplicity of structures in a SPICKER cluster and denotes the average RMSD of the structures to the cluster centroid.

	RESULTS AND DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUDING REMARKS ACKNOWLEDGEMENTS REFERENCES

 
Benchmark protein set
To construct a benchmark set, we group all protein entries in the PDB having from 201 to 300 residues in length based on their sequence similarity with a pairwise sequence identity cutoff of 35% and randomly select one protein from each group. The resulting representative benchmark set includes 745 proteins: 112, 132, and 501 of them are -, ß-, and ß-proteins, respectively; 258 of them have more than one domain according to DomainPhaser (Guo et al., 2003); and 18 are transmembrane proteins. The list of target proteins can be found on our website at http://www.bioinformatics.buffalo.edu/abinitio/745.

Threading
After homologous template proteins with sequence identity >30% are excluded from the library, 593 target proteins are assigned by PROSPECTOR_3 as the Easy set. The average coverage of template alignments for these Easy targets is 83%, with an average RMSD to native of 5.9 Å on the aligned regions (see Table 1). There are 418 Easy targets that have a RMSD to native below 6.5 Å; 363 of them have alignment coverage higher than 70%.

There are only two Hard cases (1hq0A and 1k24A) where PROSPECTOR_3 cannot identify a suitable template. Nevertheless, we still include the alignments of the highest scoring templates in the folding refinement for the Hard set targets. Although the global topology of the threading alignments are often wrong in the Hard set targets, the local fragments are still close to native in most cases, a fact that can be profitably exploited by TASSER (Zhang and Skolnick, 2004a).

Summary of folding results
The threading templates and alignments by PROSPECTOR_3 are taken as the initial inputs in the TASSER reassembly procedure. For the Easy targets, we take the two templates of the highest Z-score as well as their consensus substructure as an independent template. The consensus template is calculated as an average of the commonly aligned residues when their distances are <5 Å after superposition. For Medium and Hard targets, the top five and top 20 templates are taken, respectively.

Table 1 presents a summary of the PROSPECTOR_3 threading results as well as the final models produced by TASSER. If we define a successful prediction as the one where at least one of the top five full-length models has a RMSD to native below 6.5 Å (a statistically significant cutoff (Reva et al., 1998), but other cutoffs could be used as well), there are 396 foldable cases among the 593 Easy set targets (67%) with an average RMSD to native of 3.6 Å. Of the 152 Medium/Hard targets, there are only 12 foldable cases. This unfortunately shows the strong correlation between the final outcome of TASSER modeling and PROSPECTOR_3 threading alignments. Among these 12 foldable cases, 10 of them (all are ß- or ß-proteins, i.e., 1b5tA, 1bwzA, 1cmxA, 1e2tA, 1fs0G, 1g61A, 1gs5A, 1h8vA, 1isfA, and 1jtdB) have initial templates with incorrect alignments (RMSD > 8 Å) or <70% alignment coverage, demonstrating TASSER's ability to assemble big protein models from rather poor and incomplete template alignments.

In Fig. 3, we show a histogram of the percentage of foldable targets at different RMSD cutoffs where we categorize the targets into single domain and multiple domain proteins. For the 487 single domain proteins, in 61% of targets (296), the best of the top five models has a RMSD to native below 6.5 Å. For the 258 multiple domain proteins, there are 112 targets having RMSD < 6.5 Å in the top five models. However, in 172 cases, there is at least one domain with a RMSD < 6.5 Å and whose average length is 114 residues. This highlights a weak point of TASSER for predicting the mutual orientation of the domains even when the individual domains have correct topology; the solution to this issue is the next major challenge facing TASSER. In the meantime, an enlarged template library including various domain orientations within the same homologous subfamily will certainly be of help for use in TASSER domain assembly (our unpublished results).

View larger version (32K): [in this window] [in a new window]   FIGURE 3  Histogram of the percent of foldable targets by TASSER for single-domain and multiple domain proteins.

Comparison with the initial template
In Fig. 4, for the threading aligned regions, we show a detailed comparison between the final models and initial template alignments. As expected, the template alignments in the Easy set have much better quality than those in the Medium/Hard set. The majority of the template alignments in the Easy set have a RMSD to native below 6 Å (Fig. 4 a), whereas the alignments in Medium/Hard set have a quite broad RMSD distribution ranging from 1 Å to more than 20 Å. This highlights the threading alignment problems shown by PROSPECTOR_3 on the Medium/Hard targets, even though the majority of templates (90%) in this category can have good structure alignments (Skolnick et al., 2004).

View larger version (20K): [in this window] [in a new window]   FIGURE 4  RMSD to native of the best models in top five by TASSER versus the RMSD to native of the best initial template by PROSPECTOR_3; both RMSD calculated over the same aligned regions. (a) Easy set targets; (b) Medium/Hard set targets.

Unaligned loops/tails modeling
In Fig. 5, we show the results of TASSER modeling for the unaligned loop and tail regions. Here, an unaligned loop (tail) region is defined as a piece of continuous sequence that has no coordinate assignments in the middle (terminus) of a target protein from the PROSPECTOR_3 threading alignments.

View larger version (41K): [in this window] [in a new window]   FIGURE 5  (a and b) Size distribution of the unaligned loops and tails, with the last points including all loops (tails) of length above 25 (50) residues. The solid lines connect the data points denoting all loops and tails. The dashed lines signify those loops with good stem backbones having a RMSD to native below 4 Å. (c and d) Average RMSD to native of the unaligned loops and tails by TASSER modeling as a function of the size of the modeled regions. RMSDlocal () denotes the root-mean-square deviation with direct superposition of native and the modeled regions; RMSDglobal () is the root-mean-square deviation after the superposition of up to five neighboring stem residues in both sides of the loops or in a single side of the tails. The dashed-dotted line signifies a RMSD cutoff of 6.5 Å. The solid lines connect the data points denoting the results for all modeled loops/tails; the dashed lines denotes the results for the loops with good stem backbones.

The accuracy of loop modeling is obviously influenced by the accuracy of the neighboring stem backbone. For example, if the distance between the two stem backbones is much larger than that in the native structure, the loop conformation will tend to be extended because of the constraint of geometric connectivity even if the force field favors a compact loop conformation and vise versa. If we only count those loops where the RMSD of the residues in the stem backbones is below 4 Å, there are 3821 loop regions in total. The modeling results of these loop regions are shown in Fig. 5 c marked as stem filtered. They clearly have higher accuracy than if we count all loops regions, especially for the big loops because the bigger loops have more opportunities to be embedded between distorted stem backbones (see Fig. 5 a). After removing those loops of distorted stem backbones, TASSER can have the average RMSD_global < 6.5Å for the unaligned loop regions up to 20 residues.

There are 785 unaligned regions at the N- or C-termini in the PROSPECTOR_3 alignments, with lengths ranging from 1 to 173 residues. Some of the big tails include an entire individual domain in multidomain proteins. The size distribution of the unaligned tails is presented in Fig. 5 b, with the last point including all tails greater than or equal to 50 residues. In comparison with loop modeling, because of a lack of a second spatial constraint on the free end of the tails, the orientation of the unaligned tails can be seriously misplaced, even though the local conformation can be correct. As shown in Fig. 5 d, TASSER has an average RMSD_global below 6.5 Å for tails under seven residues long. Here, the RMSD_global for tails is defined as the root-mean-square deviation between the modeling region and native after a superposition of five neighboring residues on the stem of the tail. For local tail conformations, TASSER can generate RMSD_local < 6.5 Å for tails up to 35 residues long (Fig. 5 d).

Most of the unaligned loop and tail regions in PROSPECTOR_3 alignments are of small size (see Fig. 5, a and b), which are relatively easier to model because of the limited configuration entropy. If we only focus on the loop/tail regions greater than or equal to four residues in size, there are in total 1345 unaligned loops with an average length of 8.1 residues; there are 464 unaligned tails with an average length of 52.6 residues.

We summarize in Table 2 the RMSD distribution for loops and tails with length greater than or equal to four residues. In 42% (561/1345) of the cases, TASSER loop modeling has acceptable accuracy with RMSD_global < 4 Å. The average RMSD_global and RMSD_local for these 1345 loops are, respectively, 5.03 and 1.82 Å. If we consider the loops with good stem backbones, 59% (497/837) of loops have a RMSD_global < 4 Å; and the average RMSD_global and RMSD_local for these 837 loops are 4.03 Å and 1.47 Å, respectively. For the tails, 53% (246/464) of the cases have a RMSD_local below 4 Å. Considering the global conformation of tails, only 11% (50/464) of the tails have RMSD_global below 4 Å. Again, the data show much better control of local conformation than the global orientation for tails in TASSER modeling. This is reminiscent of the problem TASSER has with predicting domain-domain orientations.

The folding results of 18 large membrane proteins in current benchmark set are summarized in column 3 of Table 3. In one-third of the cases, TASSER generates at least one model in the top five that has a RMSD to native below 5.5 Å. In column 2 of the table, we also show the TASSER folding results of 20 membrane proteins from the representative benchmark set of smaller proteins (41 200 residues), which has a slightly better success rate of 45% because of their smaller size. The overall folding rate in the combined membrane benchmark is 40% (15/38).

In Fig. 6, we show three typical examples of membrane proteins predictions, 1jgjA, 1fqyA, and 1bh3_, with the well-known GPCR rhodopsin protein 1jgjA having the highest accuracy. The best template hits by PROSPECTOR_3 for 1jgjA, 1fqyA, and 1bh3 are 1ap9_ (1.47 Å over 96% coverage), 1fx8A (5.20 Å over 92% coverage), and 2por_ (13.44 Å over 88% coverage). The final models in these three cases have a RMSD to native of 1.1/0.89 Å, 3.3/3.1 Å, and 5.3/5.2 Å with full-length/aligned regions, respectively. These data show that TASSER has the potential to draw the stem fragments closer to native with respect to the threading templates and build reasonable loops for the membrane proteins as well.

View larger version (40K): [in this window] [in a new window]   FIGURE 6  Three representative foldable examples of transmembrane proteins by TASSER. The thin lines denote the C-backbone of experimental structures, and the thick lines are the predicted models. Blue to red runs from the N- to C-terminus. Below the structures are the PDB code, RMSD between the model and native structure, and the protein size.

There is no significant difference between the loop modeling for membrane proteins and that for nonmembrane proteins. For the 18 big membrane proteins of size from 216 to 299 residues, there are 31 unaligned loops with length greater than or equal to four residues. The average values of RMSD_global and RMSD_local for these loops are 5.36 Å and 2.04 Å, respectively, which are comparable with the values of 5.03 Å and 1.82 Å for all loops including both membrane and nonmembrane proteins (see Table 2). The small difference between the membrane and nonmembrane proteins may be due to the fact that the membrane proteins on average have a worse global quality in comparison with nonmembrane proteins. For example, if we only count the 19 loops with RMSD of stem residues below 4 Å, the average values of RMSD_global and RMSD_local are 4.01 Å and 1.49 Å, respectively, which are almost the same as the corresponding values for all loops in Table 2.

Comparison of TASSER predictions with structures determined by NMR
The structures determined by either x-ray crystallography or NMR experiments are almost always nearly identical (Branden and Tooze, 1999). To compare the accuracy of the models predicted by TASSER to that of protein structures determined by NMR, we calculate the centroid of the set of structures provided by NMR experiments and compare the deviation of the TASSER models to the NMR structure centroid with that of the individual NMR structure located at the maximal distance from the centroid. Since the set of NMR models equally well satisfy the experimental data, the maximal distance between the centroid and the individual structures represents the inherent uncertainty and resolution of the NMR structure. The reason we compare the TASSER models to NMR data rather than x-ray structural data is that three-dimensional structures in NMR are usually derived from distance/contact constraints from their nuclear Overhauser effect (NOE) spectrum, and a collection of models consistent with the experimental restraints is often provided. Thus, there is an envelop of experimental structures to which we can readily compare the quality of our predictions to assess whether or not they are distinguishable.

In our complete benchmark set of 2234 proteins (including 1489 targets between 41 to 200 residues and 745 targets between 201 to 300 residues), there are 503 targets whose experimental structures are determined by NMR; 92% (463) are below 150 residues, a fact due to the difficulty of applying NMR spectroscopy to the structure determination of larger proteins (Branden and Tooze, 1999). Among these 503 NMR targets, there are 363 proteins with 5 56 individual models that simultaneously satisfy the NMR spectra (for the other 140, the authors just provide the minimized average structure).

To calculate the structure centroid, for each of the 363 proteins, we superimpose all the NMR models to the first model in the PDB record and average the coordinates of the corresponding residues after superposition. Then, we calculate the maximal root-mean-square deviation of the individual models from the structure centroid, RMSD_NMR. For the 363 NMR targets, the average value of RMSD_NMR = 2.64 Å. For the models predicted by TASSER, we also calculate the RMSD of the theoretical models from the NMR structure centroid. The resultant average value for the TASSER models is 4.84 Å, which is considerably higher than that of NMR experiments. In 72 cases (27 -proteins, 23 ß-proteins, and 22 ß-proteins), however, the RMSD of the theoretical models from the NMR centroid is smaller than RMSD_NMR. Among the 72 cases, seven cases (i.e., 1cw5A, 1g9pA, 1h9fA, 1i5jA, 1imuA, 2prp_, and 3lriA) are classified by PROSPECTOR_3 as Medium targets, with the best templates having an average RMSD to native 6.9 Å and an average alignment coverage of 60.6%. All other 65 cases are Easy targets, with PROSPECTOR_3 templates having average RMSD to native 4.1 Å and 83.9% alignment coverage.

In Fig. 7, we present three typical examples of the superposition of TASSER models on the NMR structures for 1adr_ (an -protein), 2fnbA (a ß-protein), and 1dbyA (an ß-protein). The maximal RMSD of NMR models from their centroid for the 1adr_, 2fnbA, and 1dbyA are 3.6 Å, 2.3Å, and 1.3 Å, respectively, whereas the RMSD of the TASSER models to the centroids are 1.6 Å, 1.9 Å, and 1.1 Å, respectively. These results show that, in 20% of cases, TASSER generates models of accuracy comparable to the NMR experimental methods where the predicted structures are closer to the NMR centroid structure than that of the farthest NMR structure.

View larger version (41K): [in this window] [in a new window]   FIGURE 7  Three representative examples of TASSER predicted models that are structurally closer to the NMR structure centroid than some of individual NMR structures. The thick backbone shows the rank-one models predicted by TASSER; the wire frame presents the structures satisfying the NMR distance constraints equally well. Blue to red runs from the N- to C-terminus. The RMSD of TASSER models to the NMR centroid for 1adr_ (-protein), 2fnbA (ß-protein), and 1dbyA (ß-protein) are 1.6 Å, 1.9 Å, and 1.1 Å, respectively; the maximal RMSD of NMR models to the centroid are 3.6 Å, 2.3 Å, and 1.3 Å, respectively.

	CONCLUDING REMARKS

TOP ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUDING REMARKS ACKNOWLEDGEMENTS REFERENCES

For approximately three-fifths of larger single domain proteins, TASSER can generate models ranked in the top five that have a RMSD to native below 6.5 Å. For multidomain proteins, the success rate drops to approximately two-fifths; although in two-thirds of the multiple domain proteins, the individual domains of the complex are correctly predicted. Further development of the TASSER force field to control the interdomain orientation is required to improve the quality of the predictions for multiple domain complexes. Keeping in mind that many multidomain proteins with high sequence identity have different domain orientations, an immediate follow-up project will be the construction of an extended multidomain orientation library, which allows TASSER simulations to select suitable domain orientations among the homologous templates. For membrane proteins that have a very limited number of solved template structures, 40% of such transmembrane proteins from 41 to 300 residues can be folded to a RMSD below 6 Å. Finally, when TASSER models are compared with the set of experimental structures determined by NMR, 20% of the TASSER models are closer to the centroid of the set of NMR structures than the farthest NMR structure consistent with experimental data.

For all the categories of the target proteins, TASSER models show obvious improvement with respect to the initial threading templates from PROSPECTOR_3 (Skolnick et al., 2004). Over the same aligned regions, the average RMSD of all 745 proteins is reduced by TASSER modeling from an initial average RMSD of 7.2 to 5.6 Å, and the number of cases with a RMSD to native < 6.5 Å increases from 463 to 552.

For the unaligned loop regions with good stem backbone conformations (where the RMSD of the stem residues is below 4 Å), the TASSER ab initio modeling approach can generate reasonable loop models with an average RMSD_global < 6.5 Å for loops up to 20 residues long. For the loops of size greater than or equal to four residues (8.1 residues on average), 59% (497/837) have a global RMSD below 4 Å. For the unaligned tail regions with an average length of 52.6 residues, although in most cases the correct global orientation of the tails are not reproduced, TASSER generates tails whose RMSD_local is below 4 Å in 53% (246/464) of the cases.

One purpose of this work is to focus on proteins greater than 200 residues in length, a size range where most enzymes and other functionally important proteins are often found. Although there is still considerable room for improvement of TASSER methodology, especially for multiple domain complexes and membrane proteins, the results of the large-scale benchmark test reported here suggest that reliable predicted structures by automated computational approaches is becoming a reality for at least a subset of non-/weakly homologous large size proteins.

	ACKNOWLEDGEMENTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUDING REMARKS ACKNOWLEDGEMENTS REFERENCES

 
This research was supported in part by grants GM-37408 and GM-48835 of the Division of General Sciences of the National Institutes of Health.

Submitted on May 3, 2004; accepted for publication July 1, 2004.

	REFERENCES

TOP ABSTRACT INTRODUCTION METHODS RESULTS AND DISCUSSION CONCLUDING REMARKS ACKNOWLEDGEMENTS REFERENCES

 
Baker, D., and A. Sali. 2001. Protein structure prediction and structural genomics. Science. 294:93–96.[Abstract/Free Full Text]

Baleja, J. D. 2001. Structure determination of membrane-associated proteins from nuclear magnetic resonance data. Anal. Biochem. 288:1–15.[CrossRef][Medline]

Berman, H. M., J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov, and P. E. Bourne. 2000. The Protein Data Bank. Nucleic Acids Res. 28:235–242.[Abstract/Free Full Text]

Branden, C., and J. Tooze. 1999. Introduction to Protein Structure. Garland Publishing, Inc., New York.

Fiser, A., R. K. Do, and A. Sali. 2000. Modeling of loops in protein structures. Protein Sci. 9:1753–1773.[Abstract]

Frishman, D., M. Mokrejs, D. Kosykh, G. Kastenmuller, G. Kolesov, I. Zubrzycki, C. Gruber, B. Geier, A. Kaps, K. Albermann, A. Volz, C. Wagner, M. Fellenberg, K. Heumann, and H. W. Mewes. 2003. The PEDANT genome database. Nucleic Acids Res. 31:207–211.[Abstract/Free Full Text]

Guo, J. T., D. Xu, D. Kim, and Y. Xu. 2003. Improving the performance of DomainParser for structural domain partition using neural network. Nucleic Acids Res. 31:944–952.[Abstract/Free Full Text]

Holm, L., and C. Sander. 1996. Mapping the protein universe. Science. 273:595–603.[Abstract/Free Full Text]

Ikeda, M., M. Arai, T. Okuno, and T. Shimizu. 2003. TMPDB: a database of experimentally-characterized transmembrane topologies. Nucleic Acids Res. 31:406–409.[Abstract/Free Full Text]

Jones, D. T. 1999. Protein secondary structure prediction based on position-specific scoring matrices. J. Mol. Biol. 292:195–202.[CrossRef][Medline]

Karplus, K., C. Barrett, and R. Hughey. 1998. Hidden Markov models for detecting remote protein homologies. Bioinformatics. 14:846–856.[Abstract/Free Full Text]

Kawabata, T., S. Fukuchi, K. Homma, M. Ota, J. Araki, T. Ito, N. Ichiyoshi, and K. Nishikawa. 2002. GTOP: a database of protein structures predicted from genome sequences. Nucleic Acids Res. 30:294–298.[Abstract/Free Full Text]

Kolinski, A., and J. Skolnick. 1998. Assembly of protein structure from sparse experimental data: an efficient Monte Carlo model. Proteins. 32:475–494.[CrossRef][Medline]

Levy, D., M. Chami, and J. L. Rigaud. 2001. Two-dimensional crystallization of membrane proteins: the lipid layer strategy. FEBS Lett. 504:187–193.[CrossRef][Medline]

Liwo, A., J. Lee, D. R. Ripoll, J. Pillardy, and H. A. Scheraga. 1999. Protein structure prediction by global optimization of a potential energy function. Proc. Natl. Acad. Sci. USA. 96:5482–5485.[Abstract/Free Full Text]

Marti-Renom, M. A., A. C. Stuart, A. Fiser, R. Sanchez, F. Melo, and A. Sali. 2000. Comparative protein structure modeling of genes and genomes. Annu. Rev. Biophys. Biomol. Struct. 29:291–325.[CrossRef][Medline]

Moult, J., K. Fidelis, A. Zemla, and T. Hubbard. 2001. Critical assessment of methods of protein structure prediction (CASP): round IV. Proteins. (Suppl. 5):2–7.[Medline]

Moult, J., K. Fidelis, A. Zemla, and T. Hubbard. 2003. Critical assessment of methods of protein structure prediction (CASP)-round V. Proteins. 53(Suppl. 6):334–339.[CrossRef][Medline]

Needleman, S. B., and C. D. Wunsch. 1970. A general method applicable to the search for similarities in the amino acid sequence of two proteins. J. Mol. Biol. 48:443–453.[CrossRef][Medline]

Reva, B. A., A. V. Finkelstein, and J. Skolnick. 1998. What is the probability of a chance prediction of a protein structure with an rmsd of 6 A? Fold. Des. 3:141–147.[CrossRef][Medline]

Simons, K. T., C. Strauss, and D. Baker. 2001. Prospects for ab initio protein structural genomics. J. Mol. Biol. 306:1191–1199.[CrossRef][Medline]

Skolnick, J., J. S. Fetrow, and A. Kolinski. 2000a. Structural genomics and its importance for gene function analysis. Nat. Biotechnol. 18:283–287.[CrossRef][Medline]

Skolnick, J., and D. Kihara. 2001. Defrosting the frozen approximation: PROSPECTOR–a new approach to threading. Proteins. 42:319–331.[CrossRef][Medline]

Skolnick, J., D. Kihara, and Y. Zhang. 2004. Development and large scale benchmark testing of the PROSPECTOR 3:0 threading algorithm. Proteins. 56:502–518.[CrossRef][Medline]

Skolnick, J., A. Kolinski, and A. Ortiz. 2000b. Derivation of protein-specific pair potentials based on weak sequence fragment similarity. Proteins. 38:3–16.[CrossRef][Medline]

Stevens, T. J., and I. T. Arkin. 2000. Do more complex organisms have a greater proportion of membrane proteins in their genomes? Proteins. 39:417–420.[CrossRef][Medline]

White, S. H., and W. C. Wimley. 1999. Membrane protein folding and stability: physical principles. Annu. Rev. Biophys. Biomol. Struct. 28:319–365.[CrossRef][Medline]

Zhang, Y., D. Kihara, and J. Skolnick. 2002. Local energy landscape flattening: parallel hyperbolic Monte Carlo sampling of protein folding. Proteins. 48:192–201.[CrossRef][Medline]

Zhang, Y., A. Kolinski, and J. Skolnick. 2003. TOUCHTONE II: a new approach to ab initio protein structure prediction. Biophys. J. 85:1145–1164.[Abstract/Free Full Text]

Zhang, Y., and J. Skolnick. 2004a. Automated structure prediction of weakly homologous proteins on a genomic scale. Proc. Natl. Acad. Sci. USA. 101:7594–7599.[Abstract/Free Full Text]

Zhang, Y., and J. Skolnick. 2004b. SPICKER: a clustering approach to identify near-native protein folds. J. Comput. Chem. 25:865–871.[CrossRef][Medline]

This article has been cited by other articles:

				S. Y. Lee and J. Skolnick Benchmarking of TASSER_2.0: An Improved Protein Structure Prediction Algorithm with More Accurate Predicted Contact Restraints Biophys. J., August 15, 2008; 95(4): 1956 - 1964. [Abstract] [Full Text] [PDF]

				A. Jagielska, L. Wroblewska, and J. Skolnick Protein model refinement using an optimized physics-based all-atom force field PNAS, June 17, 2008; 105(24): 8268 - 8273. [Abstract] [Full Text] [PDF]

				H. Chen and J. Skolnick M-TASSER: An Algorithm for Protein Quaternary Structure Prediction Biophys. J., February 1, 2008; 94(3): 918 - 928. [Abstract] [Full Text] [PDF]

				M. Brylinski and J. Skolnick A threading-based method (FINDSITE) for ligand-binding site prediction and functional annotation PNAS, January 8, 2008; 105(1): 129 - 134. [Abstract] [Full Text] [PDF]

				S. B. Pandit, Y. Zhang, and J. Skolnick TASSER-Lite: An Automated Tool for Protein Comparative Modeling Biophys. J., December 1, 2006; 91(11): 4180 - 4190. [Abstract] [Full Text] [PDF]

				Y. Zhang and J. Skolnick TM-align: a protein structure alignment algorithm based on the TM-score Nucleic Acids Res., April 22, 2005; 33(7): 2302 - 2309. [Abstract] [Full Text] [PDF]

This Article Abstract Full Text (PDF) Alert me when this article is cited Alert me if a correction is posted Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Zhang, Y. Articles by Skolnick, J. Search for Related Content PubMed PubMed Citation Articles by Zhang, Y. Articles by Skolnick, J.