A Protocol for Computer-Based Protein Structure and Function Prediction

Method overview

Following the sequence-to-structure-to-function paradigm, the I-TASSER procedure^1-4 for structure and function modeling involves four consecutive steps of: (a) template identification by LOMETS⁵; (b) fragment structure reassembly by replica-exchange Monte Carlo simulations⁶; (c) atomic level structure refinement using REMO⁷ and FG-MD⁸; and (d) structure-based function interpretations using COFACTOR⁹.

Template identification: For a query sequence submitted by the user, the sequence is first threaded through a representative PDB structure library by a locally installed LOMETS meta-threading server. Threading is a sequence-structure alignment procedure used for identifying template proteins which may have similar structure or contain similar structural motif as the query protein. To increase the coverage of homologous template detections, LOMETS combines multiple state-of-the-art algorithms covering different threading methodologies. Since different threading programs have different scoring systems and alignment sensitivities, the quality of the generated threading alignments from each threading program is assessed by normalized Z-score, which is defined as:

where Z-score is the score in standard deviation units relative to the statistical mean of all alignments generated by the program; and Z₀ is a program-specific Z-score cutoff determined based on large-scale threading benchmark tests⁵ to differentiate ‘good’ and ‘bad’ templates. A template with a high Z-score means that the top templates have an alignment score significantly higher than most of other templates, which usually implies that the alignment corresponds to a good model. If most of the top threading templates have high normalized Z-scores, the accuracy of the final I-TASSER model is usually high. However, if the protein is large and the coverage of the threading alignments is confined to a small region of the query protein, a high normalized Z-score does not necessarily mean a high modeling accuracy for the full-length model. Top two threading alignments from each threading program are collected and used for the next step of structure assembly.

Iterative structure assembly simulation: Following the threading procedure, query sequence is split into threading aligned and unaligned regions. Continuous fragments in threading alignment are excised from templates and used directly for the structure assembly, while the unaligned loop regions are built by ab initio modeling. The structure assembly procedure is performed on a lattice system guided by the replica exchange Monte Carlo simulations⁶. The I-TASSER force field includes hydrogen-bonding interactions¹⁰, knowledge-based statistical energy terms derived from known protein structures in the PDB¹¹, sequence-based contact predictions from SVMSEQ¹², and spatial restraints collected from LOMETS⁵ threading templates. The conformational decoys generated in the low-temperature replicas during the simulations are clustered by SPICKER¹³ to identify structures of low free-energy states. Cluster centroids of the top clusters are obtained by averaging the 3D coordinates of all clustered structural decoys and used for the final model generation. The simulation and clustering procedure are repeated twice for removing steric clashes and further refining the global topology.

Atomic-level model construction and refinement: The cluster centroids obtained after SPICKER clustering are reduced protein models (each residue represented by its C_α and side-chain center of mass) and have limited biological application. The construction of full-atomic model from the reduced models is done in two steps. In the first step, REMO⁷ is used to construct full-atomic models from C-alpha traces by optimizing the H-bond networks. In the second step, REMO full-atomic models are further refined by FG-MD¹⁴, which improves the backbone torsion angles, bond lengths, and side-chain rotamer orientations, by molecular dynamic simulations, as guided by the structural fragments searched from the PDB structures by TM-align. The FG-MD refined models are used as the final models for tertiary structure predictions by I-TASSER.

The quality of the generated models are estimated based on a confidence score (C-score), which is defined based on the Z-score of LOMETS threading alignments and the convergence of I-TASSER simulations, mathematically formulated as:

where M is the multiplicity of structure decoys in the structural clusters identified by the SPICKER¹³; M_tot is the total number of decoys submitted to the clustering; is the average RMSD of the clustered decoys to the cluster centroids; Norm.Z-Score(i) is the normalized Z-score (Eq. 1) of the top threading alignment obtained from ith threading server in LOMETS⁵; N is the number of servers used in LOMETS.

The C-score has a strong correlation with the quality of the I-TASSER models. By combining C-score and protein length, the accuracy of the first I-TASSER models can be estimated with an average error of 0.08 for the TM-score and 2 Å for the RMSD¹⁵. In general, models with C-score > – 1.5 are expected to have a correct fold. Here, RMSD and TM-score are both well known measures of topological similarity between the model and native structure. TM-score values range in [0, 1], where a higher score indicates a better structure match^16,17. However for lower-ranked models (i.e. 2^nd-5^th models), the correlation of C-score with TM-score and RMSD is much weaker (~ 0.5), and cannot be used for reliable estimation of absolute model quality.

Is first model always the best model in I-TASSER simulations? The answer to this question depends on the target type. For easy targets, the first model is usually the best model and its C-score is usually much higher than the rest of the models. However, for hard targets, where threading does not have significant template hits, the first model is not necessarily the best model and I-TASSER actually has difficulty in selecting the best template and models. It is therefore recommended to analyze all the 5 models for hard targets and select them based on the experimental information and biological knowledge.

Function predictions: In the last step, final 3D-models generated from FG-MD are used to predict three aspects of protein function, namely: a) Enzyme Commission (EC) numbers¹⁸ and (b) Gene Ontology (GO)¹⁹ terms and (c) binding sites for small molecule ligands. For all the three aspects, functional interpretations are generated using COFACTOR, which is a new approach to predict protein’s function based on global and local similarity to template proteins in PDB with known structure and functions. First, the global topology of the predicted models is matched against functional template libraries using structural alignment program TM-align²⁰. Next, a set of proteins most similar to the target models are selected from the library based on their global structure similarity, and an extensive local search is performed to identify structure and sequence similarity near the active/binding site region. The resultant global and local similarity scores are used to rank the template proteins (functional homologues) and transfer the annotation (EC numbers and Gene Ontology¹⁹ terms) based on the top scoring hits. Similarly, ligand binding site residues and the ligand binding mode are inferred based on the local alignment of query with known ligand binding site residues in the top scoring function templates⁹.

The quality of function (EC and GO term) prediction in I-TASSER is evaluated based on functional homology score (Fh-score) which is a measure of global and local similarity between the query and template, and is defined as:

where C-score is an estimate of the quality of predicted model as defined in Eq. (2); TM-score measures the global structural similarity between the model and template proteins; RMSD_ali is the RMSD between the model and the template structure in the structurally aligned region from TM-align²⁰; Cov represents the coverage of the structural alignment (i.e. the ratio of the structurally aligned residues divided by the query length); ID_ali is the sequence identity in the TM-align alignment. The estimated confidence score for EC number predictions also includes a term for evaluating active site match (AcM) between query and template within a defined local region, calculated as:

where N_t represents the number of template residues present within the local area, N_ali is the number of the aligned query-template residue pairs, d_ii is the C_α distance between ith pair of aligned residues, d₀ = 3.0 Å is the distance cutoff, M_ii is the BLOSUM scores between ith pair of aligned residues. In general, the Fh-score is in the range [0, 5] and AcM score is between [0, 2], where higher scores indicating more confident functional assignments. AcM score is also used for evaluating the local structure and sequence similarity near the ligand-binding sites, which is referred to as BS-score.

1. Submission of protein sequence

1. Visit the I-TASSER web page at http://zhanglab.ccmb.med.umich.edu/I-TASSER to start with structure and function modeling experiment.
2. Copy and paste the amino acid sequence into the provided form or directly upload it from your computer by clicking the “Browse” button. I-TASSER server currently accepts sequences with up to 1500 residues. Proteins longer than 1500 residues are usually multi-domain proteins, and are recommended to be split into individual domains before submitting to I-TASSER.
3. Provide your e-mail address (mandatory) and a name for the job (optional).
4. Users can optionally specify external inter-residue contact/distance restraints, add-in an additional template or exclude some template proteins during the structure modeling process. Learn more about using these options in the “Discussion” section.
5. To submit the sequence, click on the “Run I-TASSER” button. The browser will be directed to a confirmation page displaying user specified information, job identification (Job ID) number and a link to a webpage where the results will be deposited after completion of job. Users may bookmark this link or note down the job identification number for future reference.

2. Availability of results

1. Check the status of your submitted job by visiting the I-TASSER queue page at http://zhanglab.ccmb.med.umich.edu/I-TASSER/queue.php. Click on the search tab and use the Job ID number or the query sequence to search your submitted job.
2. After the structure and function modeling is finished, a notification e-mail containing image of the predicted structures and a web-link will be sent to you. Click on this link or open the link bookmarked at Step 1.5 to view and download the results.

3. Secondary structure and solvent accessibility predictions

1. Check the FASTA formatted query sequence displayed on the top of the result page. If any additional restraint/template was specified during sequence submission, a link to the webpage displaying user-specified information can also be seen (Figure 1A).
2. Examine the secondary structure prediction displayed as: alpha helix (H), beta strand (S) or coil (C) and confidence score of prediction (0 = low, 9 = high) for each residue. Look for region with long stretches of regular secondary structure (H or S) predictions, to estimate the core-region in the protein. Structural class of protein can also be analyzed based on distribution of secondary structures elements. Also, long regions of coil elements in the protein usually indicate unstructured/disordered regions.
3. View the predicted solvent accessibility (Figure 1C) to ascertain buried and solvent exposed regions in the query. Values of predicted solvent accessibility range from 0 (buried residue) to 9 (exposed residue). Region containing mostly buried residues can be used to delineate the core region in the protein, while regions with solvent exposed and hydrophilic residues are potential hydration/functional sites.

4. Tertiary structure predictions

1. Scroll down to view the predicted tertiary structures of query protein, displayed in an interactive Jmol applet (Figure 2). Left click on the applet to change the appearance of displayed structure, zoom into specific region, select specific residue types in the predicted model or calculate inter-residue distances.
2. Analyze the models for the presence of long unstructured regions. These regions usually correspond to disordered regions in protein or indicate lack of template alignment. These regions generally have low modeling accuracy and removing these regions during modeling from N & C-terminus region will improve the modeling accuracy.
3. Download the PDB formatted structure files of the model by clicking on the “Download Model” links. You can open these files in any molecular visualization software (e.g. Pymol, Rasmol etc.) for further analysis of the structural features.
4. Analyze the confidence score (C-score) of structure modeling to estimate the quality of predicted structures. C-score (Eq. 2) values are typically in the range [-5, 2], wherein a higher score reflects a model of better quality. The estimated TM-score and RMSD of first model is shown as “Estimated accuracy of Model 1”. For long proteins, it is recommended to evaluate the model quality based on TM-score, as TM-score is more sensitive to the topological changes than RMSD.
5. Click on “more about C-score” link to analyze C-score, cluster size and cluster density of all the models. Estimated TM-score and RMSD are presented only for the first I-TASSER model, because C-score of lower ranked models is not strongly correlated with TM-score or RMSD. Quality of lower-ranked models can be partially assessed based on their cluster density and cluster size relative to the first model, wherein models from bigger cluster and higher density are on average closer to the native structure.
6. Low C-score predictions usually indicate a low accuracy prediction. In most such cases, the query protein lacks a good template in the library and has a size beyond the range of ab initio modeling (i.e. >120 residues). In these cases, users can seek for additional spatial restraints and use them to improve the I-TASSER modeling (see Discussion section). It is also encouraged to submit the sequences to our QUARK server (http://zhanglab.ccmb.med.umich.edu/QUARK/) for a pure ab initio modeling if the protein size is below 200 residues.

5. LOMETS target template alignment

1. Scroll down to analyze the top ten threading templates of the query protein, as identified by LOMETS threading programs (Figure 3). View the normalized Z-score (Eq. 1), shown in ‘Norm. Z-score’ column, to analyze the quality of threading alignments. Alignments with a normalized Z-score > 1 reflects a confident alignment and most likely have the same fold as the query protein.
2. Analyze the sequence identity in the threading-aligned region (column ‘Iden. 1’) and for the whole chain (column ‘Iden. 2’) to assess the homology between the query and the template proteins. High sequence identity is an indicator of evolutionary relatedness between the query and template proteins.
3. View the threading aligned residues shown in colored to visually identify conserved residues/motifs in the query and the template proteins. A higher sequence identity in threading-aligned region, compared to whole-chain alignment also indicates presence of conserved structural motif/domains in the query.
4. Assess the coverage of the threading alignment by viewing the ‘Cov.’ column and inspecting the alignment. If the coverage of the top alignments is low and confined to only a small region of the query protein or absent for a long segment of query sequence, then the query protein usually contains more than one domain and it is recommended to split the sequence and model the domains individually (Figure 3).
5. Download the PDB formatted sequence-structure alignment files by clicking on the “Download Align” links. These alignment file can be opened in any molecular visualization program listed in the Materials section, and can also be used for adding additional restraints during the structure modeling (Step 1.4).

6. Structural analogs in PDB

1. View the next table (Figure 4) of the result page to determine the top ten structural analogs of the first predicted model, as identified by the structural alignment program TM-align²⁰. A TM-score >0.5 indicates that the detected analog and model have a similar topology and can be used to determine the structural class/protein family of the query protein¹⁶, while those with TM-score <0.3 means a random structure similarity.
2. Analyze the sequence identity and RMSD in the structurally aligned region shown in ‘IDEN^a‘ and ‘RMSD^a‘ columns to assess the conservation of spatial motifs in the model and the structural analog. Visually inspect the colored and aligned residue pairs in the alignment to identify these structurally conserved residues and motifs.
3. Click on the PDB code shown in the ‘PDB Hit’ column to visit RCSB website and learn more about their structural classification (SCOP, CATH and PFAM) and functional information (EC number, associated GO terms and bound ligand).

7. Function prediction using COFACTOR

1. Scroll down in the result page to analyze functional interpretations for the query protein. Protein functions are enumerated in three context tables, displaying: Enzyme Commission (EC) numbers, Gene Ontology (GO) terms, and ligand binding sites.
2. View the ‘TM-score’, ‘RMSD^a‘, ‘IDEN^a‘ and ‘Cov.’ columns in each table to analyze parameters of global structure similarity and conservation of spatial patterns between model and identified functional homologues (templates).

8. Enzyme Commission number prediction

1. View the top five potential enzyme homologues of query protein shown in the “Predicted EC numbers” table (Figure 5). The confidence level of EC number prediction using these templates is shown in ‘EC-Score’ column. Based on benchmarking analysis²³, functional similarity (first 3 digits of EC number) between the query and template protein can be reliably interpreted using EC-score >1.1.
2. Look for consensus of function (EC numbers) amongst the templates, which have the similar fold (i.e. TM-score >0.5) as the query protein. If multiple templates have same EC number and EC-score > 1.1, the confidence level of prediction is very high. However, if the EC-Score is high but there is a lack of consensus among the identified hits, then the prediction becomes less reliable and the users are recommended to consult the GO term predictions.
3. Click on the link provided on the EC numbers to visit the ExPASy Enzyme database and analyze the function, including the reaction catalyzed, co-factor requirements and the metabolic pathway, of the template protein in detail.

9. Gene Ontology (GO) term predictions

1. View the “Predicted GO terms” table (Figure 6) to identify top ten homologues of query protein in the PDB library, annotated with Gene Ontology (GO) terms. Each protein is usually associated with multiple GO terms, describing its molecular functions (MF), biological processes (BP) and cellular component (CC). Click on each term to visit the Amigo website and analyze its definition and lineage.
2. Analyze the Fh-score (Functional homology score) column to access the functional similarity between the query and template proteins and estimate the confidence level of transferring functional annotation from these proteins. In our benchmarking study²³, 50% of the native GO terms could be correctly identified from the first identified template using an Fh-score cutoff of 0.8, with an overall accuracy of 56%.
3. View the “Consensus prediction of GO terms” table to analyze the concurrence of function between the templates. These common functions are used for predicting the GO terms (MF, BP and CC) of the query protein and assess the confidence level (GO-score) of GO term predictions. Based on the benchmarking test²³, the best false positive and false negative rates are obtained for predictions with GO-score cutoff=0.5, with decreasing coverage of prediction at deeper ontology levels.

10. Protein-ligand binding site predictions

1. Scroll down to the bottom of the page to view top ten ligand binding site predictions for the query protein. Predicted binding sites are ranked based on the number of predicted ligand conformations that share common binding pocket. The best identified binding site is already displayed in the Jmol applet. Click on the radio buttons to analyze other predictions and visualize the ligand interacting residues.
2. Analyze the BS-score column to evaluate local similarity between the model and template’s binding site. Based on the benchmark⁹, BS-score > 1.1 indicates high sequence and structure similarity near the predicted binding site in model and known binding site in the template.
3. Download the PDB formatted structure file of the complex by clicking on the “Download” link. Users can open these files in any molecular visualization program and interactively view the predicted binding site and ligand-protein interactions on their local computer.

11. Representative results

Figure 1. An excerpt of I-TASSER result page showing (A) FASTA formatted query sequence; (B) predicted secondary structure and associated confidence scores; and (C) predicted solvent accessibility of the residues. Analyzed core region and potential hydration site in the query are highlighted in cyan and red rectangles, respectively.

Figure 2. An example of I-TASSER result page showing tertiary structure predictions for the query proteins. The predicted models are displayed in an interactive Jmol applet, allowing the user to change the display of the molecule. The models can also be downloaded by clicking on the “Download” links. The confidence score to estimate the quality of the model is reported as C-score.

Figure 3. An example of I-TASSER result page showing top ten identified threading templates and alignments by LOMETS⁵ threading programs. The quality of the threading alignments is evaluated based on normalized Z-score (highlighted in green), where a value >1 reflects a confident alignment. Aligned residues in the template that are identical to the corresponding query residues are highlighted in color to indicate presence of conserved residue/motif, while a lack of alignment in most of the top templates indicates presence of multiple domains in the query protein and the unaligned residues correspond to domain linker regions. Click here to view the full-sized version of figure 3.

Figure 4. An example of result page showing top ten identified structural analogs and structural alignments, identified by TM-align²⁰ structural alignment program. The ranking of the analogs shown in is based on the TM-score (highlighted in blue) of the structural alignment. A TM-score >0.5 indicates that the two compared structures have a similar topology, while a TM-score <0.3 means a similarity between two random structures. Structurally aligned residue pairs are highlighted in color based on their amino-acid property, while the unaligned regions are indicated by “-“. Click here to view the full-sized version of figure 4.

Figure 5. An example of I-TASSER result page showing identified enzyme homologues of the query protein in the PDB library. The confidence level of EC number prediction is analyzed based on EC-score (highlighted in green), where EC-score > 1.1 indicates functional similarity (same first 3 digits of EC number) between query and template protein.

Figure 6. An example of I-TASSER result page showing GO term predictions for the query protein. Functional homologues for the query protein in the Gene Ontology template library are ranked based on their Fh-score (in orange rectangle). Common functional features from these top-scoring hits are derived to generate the final GO term predictions for the query protein. The quality of the predicted GO terms is estimated based on GO-score (shown in green), where a GO-score >0.5 indicates a reliable prediction. Click here to view the full-sized version of figure 6.

Figure 7. An example of I-TASSER result page showing top ten protein ligand binding site predictions using the COFACTOR⁹ algorithm. The ranking of the predicted binding sites is based on the number of predicted ligand conformations that share common binding pocket in the query. BS-score (highlighted in red) is a measure of local sequence and structure similarity between the predicted and template’s binding site; and is useful for analyzing the conservation of binding site pockets.

Figure 8. An example of external restraint files used to for specifying residue-residue contact/distance restraints.

Figure 9. Example of restraint files used for specifying a template protein to the I-TASSER server. User can specify the query-template alignment either in (A) FASTA format; or (B) 3D format.

Figure 10. An example file used for excluding template during the I-TASSER structure modeling procedure. The first column contains the PDB ID of the template proteins to be excluded. The second column is used to specify the sequence identity cutoff which will be used for other similar templates in the template library.