DockTope: a Web-based tool for automated pMHC-I modelling

The immune system is constantly challenged, being required to protect the organism against a wide variety of infectious pathogens and, at the same time, to avoid autoimmune disorders. One of the most important molecules involved in these events is the Major Histocompatibility Complex class I (MHC-I), responsible for binding and presenting small peptides from the intracellular environment to CD8+ T cells. The study of peptide:MHC-I (pMHC-I) molecules at a structural level is crucial to understand the molecular mechanisms underlying immunologic responses. Unfortunately, there are few pMHC-I structures in the Protein Data Bank (PDB) (especially considering the total number of complexes that could be formed combining different peptides), and pMHC-I modelling tools are scarce. Here, we present DockTope, a free and reliable web-based tool for pMHC-I modelling, based on crystal structures from the PDB. DockTope is fully automated and allows any researcher to construct a pMHC-I complex in an efficient way. We have reproduced a dataset of 135 non-redundant pMHC-I structures from the PDB (Cα RMSD below 1 Å). Modelling of pMHC-I complexes is remarkably important, contributing to the knowledge of important events such as cross-reactivity, autoimmunity, cancer therapy, transplantation and rational vaccine design.

Scientific RepoRts | 5:18413 | DOI: 10.1038/srep18413 are costly and time-consuming. To overcome this, pMHC-I modelling represents an interesting and creative solution. Despite the availability of homology modelling techniques, each allotype presents specific particularities that cannot be simply determined through regular approaches. Thus, molecular modelling requires a careful structural study based on a solid validation process, which should take the wide MHC-I allotype variability in consideration. Immunoinformatics programs devoted to pMHC-I modelling have been developed or are under development [8][9][10][11][12][13][14][15][16][17] , but there is currently few online programs available to the scientific community.
Here, we present DockTope, a fully automated web-server tool designed with the purpose of modelling pMHC-I complexes for two human (HLA-A*02:01 and HLA-B*27:05) and two murine (H-2-Db and H-2-Kb) MHC-I allotypes. We have validated this tool through the cross-docking reproduction of 135 non-redundant structurally resolved pMHC-I structures available in the Protein Data Bank (PDB), using Cα and all atom Root Mean Square Deviation (RMSD) values for evaluation. DockTope has been fully automated using different programming languages, such as python and shell scripting. In addition, we have designed a dedicated web server providing free and easy access to any user throughout the world.

Results
The rationale behind DockTope. The DockTope tool is based on the D1-EM-D2 approach 18 , a pMHC-I modelling technique published by our group in 2010. The D1-EM-D2 approach is based on a protocol that employs a molecular docking step (D1), followed by an energy minimization (EM) of the pMHC-I complex and a final molecular docking round (D2). To develop DockTope, the D1-EM-D2 approach has gone through several implementations, including a new data validation improving its accuracy and reliability, and the full automation of the process. Here, we briefly describe important highlights of the technique improvement. A flowchart of the whole process is provided in Fig. 1.
Before the D1 step, the user provides the linear epitope sequence to be modelled, which is transformed into a three-dimensional structure. This is made possible by the fact that each epitope carries a specific backbone structure, depending on the MHC-I allotype where it is presented 18 . The epitope to be modelled is superimposed on another epitope (named here the 'Epitope pattern'), which was already determined by X-ray crystallography in the context of the target MHC-I (Table 1). Since the three-dimensional epitope structure is obtained by modelling the side chains over a constrained backbone, a brief energy minimization step is performed, allowing for a mild global relaxation of the peptide.
To perform the D1 step, all the MHC-I residues and the epitope backbone are kept rigid. Only the epitope side chains are allowed to move. During the molecular docking, the epitope can also perform rotational and translational movements, and the program AutoDock Vina 19 is used to search for the best epitope conformations inside the MHC-I cleft region. One round of molecular docking provides the best epitope conformations based on a scoring function, returning a Binding Energy (BE) value in kcal/mol, which is used by the program to rank the best conformations. To improve chances to find a suitable conformation, our approach performs twenty independent docking runs, using different initial points, which ensures that the program will search a wider range of conformations. In the end, the best conformation of each docking run is retrieved, producing a total of twenty structures.
Before the pMHC-I EM, the best epitope conformation is chosen among the twenty structures generated by AutoDock Vina. This choice is based on two variables: the BE and the average RMSD of each conformation in relation to all other structures. Since the epitope conformation having the best interaction with the MHC-I presents a low BE (a basic relationship between entropy and free energy calculations of non-covalent binding 20 ), we have designed a specific shell script to calculate the BE average among the twenty epitope structures generated from D1. This way we obtain a cut-off (Co) value that is used as a first structure filter. This will exclude spurious conformations that can arise and that do not represent the binding mode of the epitope with its respective MHC-I. Additionally, the remaining epitope conformations are compared with each other, using the g_confrms program from the GROMACS package 21 , which returns a RMSD value for each conformation pair. The epitope conformation with the lowest RMSD mean among all outputted conformations (i.e. the average structure) is chosen as the best structure. We describe the equation for choosing the best conformation in the 'Material and Methods section' . After that, the pMHC-I complex is submitted to an EM protocol, where all the residues are kept flexible to accommodate and correct the interactions between epitope and MHC-I.
The second and final molecular docking round, D2, is performed in the same way as the D1. The D2 step is intended to refine the pMHC-I structure, which is possible because the EM has already accommodated the MHC-I side chains to the target epitope structure. The final structure with the best epitope conformation is also chosen as explained above.
Automation and validation of DockTope. Back in 2010, the D1-EM-D2 approach validation process was performed over 46 pMHC-I structures available in the PDB, including the HLA-A*02:01, H-2-Db and H-2-Kb allotypes 18 . Here, thanks to the DockTope automated process, a broader validation analysis is reported over 135 pMHC-I structures, encompassing the previous dataset and including the HLA-B*27:05 allotype (Table 2). This represents almost three times the number of structures previously analysed. Also, the automation process first presented here, is a crucial feature of the DockTope web-based tool; it was made possible by writing and concatenating of more than 20 shell and python scripts. After submitting a sequence, the program automatically generates the three-dimensional epitope structure, according to the MHC-I of interest, and performs the molecular docking and energy minimization steps.
Since our method uses a reference MHC-I structure to build every model of a given allotype (also referenced here as 'MHC donor'), the validation process occurred through a cross-docking scheme. Each pMHC-I structure was modelled using only the epitope linear sequence as input and, in the end, the generated pMHC-I complex (model) was compared to its respective crystal structure deposited in the PDB (target). The comparison was based on RMSD values for the epitope atoms (considering Cα or all atoms) following the MHC-I chains superposition Scientific RepoRts | 5:18413 | DOI: 10.1038/srep18413 of model and target. This way, it was possible to obtain a RMSD value taking into consideration not only conformational changes on the modelled epitope, but also translational and rotational differences inside the MHC-I cleft.
A cut-off of 2 Å or less was used to indicate the accuracy of the modelling approach. As observed in Table 1, the reproduction of all MHC-I allotypes produced average Cα RMSD values below 2 Å. There were only two outliers, with Cα RMSD values of 3.129 Å and 2.061 Å, respectively corresponding to the attempts of reproducing an HLA-A*02:01 (PDB ID: 2GTW) and an HLA-B*27:05 (PDB ID: 3BP4) peptide-loaded complex. When all atoms , which depends on the epitope length (8-mer, 9-mer or 10-mer). The program prepares the files for the first docking (D1), where the best suited conformations will be saved during 20 rounds of simulation. The program checks whether the epitope is inside the cleft. In case of error, a report is written and the program stops. Otherwise, the program proceeds to the next step, where the best conformation is chosen. The program checks whether the structure generated is coming from D1. If so, the structure is energy minimized (EM) and a second docking is performed. In the end, the pMHC-I structure is generated in the PDB format.
were considered, the RMSD values were slightly higher in comparison to Cα RMSD values. This was expected since epitope side chains could present high flexibility in the MHC-I cleft, especially the residues involved in the interaction with the T Cell Receptor (TCR) 22 . Considering all modelled epitopes, the overall RMSD average was 0.882 Å ± 0.437 Å (s.d.) and 1.964 Å ± 0.655 Å (s.d.) for Cα and all atoms, respectively. As observed in Fig. 2, which shows the data distribution around the median with interquartile range, most of the Cα RMSD measurements were grouped even below 1.5 Å, which strongly highlights the precision of the DockTope modelling approach. Of note, the median for Cα /all atoms RMSD values were 0.854 Å/1.723 Å, 0.764 Å/2.578 Å, 0.629 Å/1.856 Å, 0.355 Å/1.685 Å and 1.292 Å/2.153 Å for HLA-A*02:01, HLA-B*27:05, H-2-Db (9-mer epitope), H-2-Db (10-mer epitope) and H-2-Kb, respectively. All 135 RMSD values for the evaluated structures are available in Supplementary Table S1 online.
Web server. DockTope is a freely accessible tool available through the website dirac.cesup.ufrgs.br/bio/ home.php, or from the CrossTope platform (http://www.crosstope.com.br) under the 'Tools' tab 23 . First, the user should register an account. After that, the user receives an email with access data (login and password) to the site (Fig. 4a). To submit a new job, the user should provide a valid linear epitope sequence. The web server automatically recognizes the epitope sequence and provides a list with the possible MHC-I allotypes that can be used in the modelling (Fig. 4b). After submitting the job, the user can follow the process steps by clicking on the 'Processing Jobs' tab. A table is provided containing information about all jobs, such as Job ID, Job Name, Epitope sequence, MHC-I allele, Status, and Submission Date (Fig. 4c). After the job submission, a "Queued (qw)" flag is assigned. The time that the job stays with this flag will mainly depend on the demand. After that, the job proceeds to the "Running" state where the files are individually stored on our server. In case of error, the server stops the job. At the end of the process, a "Finalized" flag is assigned to the particular modelled epitope, and the pMHC-I structure file in the PDB format is sent to the registered email account provided by the user. The time spent on each job, after it enters the running state, will depend on the epitope sequence and allotype, though it should not exceed 6 hours. A 10-mer epitope of H-2-Db, for example, is expected to take longer, since the addition of one residue (in comparison with 9-mer epitopes) will increase dimensionality and require more computational time. It is also expected that epitopes with a large content of arginines, for example, take longer because of the increase in the number of side chain torsions. The performance of the DockTope web server was assayed through the modelling of 238 immunogenic epitopes obtained from Immune Epitope Database and Analysis Resource 24,25 . In the end, 226 epitopes were modelled by DockTope without any error and 12 were aborted along the process-six of MHC-I allotype H-2-Db (9-mer epitope) and six of H-2-Db (10-mer epitope). These results represent an accuracy of approximately 95%.  The Cα RMSD mean value of all MHC-I allotypes was significantly below the crystal resolution mean values (***p < 0.0001, **p = 0.049). For H-2-Db (9-mer epitope), the all atoms RMSD mean value was also significantly below (#p = 0.0037). On y-axis, RMSD stands for Root Mean Square Deviation; on the x-axis, the MHC types are represented. Discussion In this work we described a fully automated tool for the structural prediction of peptide:MHC-I (pMHC-I) complexes, DockTope, which was developed and validated for the MHC-I allotypes HLA-A*02:01, HLA-B*27:05, H-2-Kb and H-2-Db. DockTope was able to reproduce 135 crystal structures from the PDB with a RMSD mean value of 0.882 Å and 1.964 Å, for epitope Cα and all atoms, respectively. The final accomplishment of this tool is (i) the complete automation, first presented here, of the established approach D1-EM-D2 18 , (ii) the modelling validation of all non-redundant pMHC-I structures available in the PDB at this time, and (iii) the tool availability as a web server for any researcher or user interested in pMHC-I modelling.
To automate and validate DockTope, a specific bash script was developed and executed in each step of the pMHC-I complex construction. The structure validation was performed using each pMHC-I structure in the PDB as a target to calculate the RMSD value with its respective model. Target and model were always fitted by MHC-I residues, which ensures that not only the difference between each epitope residue pair is considered, but also its displacement inside the MHC-I cleft after the molecular docking/energy minimization process. It should be noted that molecular docking programs can find unusual conformations after the searching process. To avoid this, DockTope performs a total of 20 rounds of molecular docking, generating up to 1000 conformations, which increases the probability of finding a proper epitope conformation. Still, unusual conformations can be generated (such as an inverted epitope inside the cleft or a protuberant C-terminal/N-terminal extremity pointing outside the MHC-I cleft). This phenomenon can be biologically explained, since some MHC-I allotypes do not have the capability to interact with determined epitope residues, but it can also be simply due to the fact that the docking algorithm was unable to find the correct solution. From the biological point of view, a work developed by Sidney et al. encompassing 945 HLA-A and HLA-B molecules reveals that some physicochemical specificities are not found in the evaluated MHC-I allotypes (considering B and F pocket residues), which in turn prevents the binding and presentation of peptides with such features 26 . In order to avoid a misleading result, DockTope also automatically checks the epitope position and orientation (but not the binding affinity) after molecular docking, confirming its position inside the MHC-I cleft before proceeding to the search for the best pMHC-I structure. Of note, this is one of the most common sources of error reported by DockTope.
Before the implementation and automation of DockTope, the best pMHC-I structure was chosen through visual inspection only, where the most frequent conformation among the twenty generated was selected. This way, a user intervention was required, which could bias the result. Here, a new and improved algorithm is used to choose the best structure, based on the mean RMSD value between each epitope pair and on the binding energy value generated by AutoDock Vina.
Since there is a lack of pMHC-I crystal structures available in the PDB, our analysis was restricted to MHC-I allotype H-2-Kb (8-mer epitope), HLA-A*02:01 (9-mer epitope), HLA-B*27:05 (9-mer epitope), and H-2-Db (9-mer epitope and 10-mer epitope). This ensures that only experimentally-resolved protein structures are used to identify the MHC-I allotype-specific epitope pattern, which reinforces the technique specificity. Also, the low number of MHC-I allotypes available for modelling by DockTope should not be seen as a weakness, since it opens the theoretical possibility to model roughly 1.2 × 1013 pMHC-I structures, which would be unfeasible through X-ray crystallography or any other method currently available. Moreover, the importance of each one of these allotypes should be highlighted. The HLA-A*02 molecule is expressed by approximately half of the human population, and the HLA-A*02:01 allele is found in a relatively high frequency all over the world 27 . For this reason, it is one of the most studied alleles. The HLA-B*27:05 has been associated with spondyloarthropathies disorders, such as ankylosing spondylitis [28][29][30] , vaccine response 31,32 , and HIV in elite controllers 33,34 . The H-2-Db and H-2-Kb are widely-studied murine alleles, and recent studies have demonstrated its importance in synapse pruning of developing brain in murines 35,36 .
Regarding the RMSD values for the DockTope validation (see Table 2 and Fig. 2), the overall RMSD average for all modelled epitopes, considering Cα and all atoms, remained below 2 Å; this is considered a reference cut-off value indicating a valid crystal reproduction obtained through a cross-docking approach 19,[37][38][39] . In fact, the validation values are reinforced after the comparison of the Cα and all atoms RMSD values by the average resolution value extracted from the PDB for all 135 structures analysed in this work (Fig. 3). The crystal resolution average value of the reproduced dataset, considering all allotypes, was 2.163 Å, which is higher than the RMSD value obtained for Cα and all atoms (0.839 Å and 2.012 Å, respectively). Analysing each pMHC-I individually, we observed that all Cα RMSD mean values were significantly below the crystal resolution mean. This indicates that subtle deviations between target and model are expected, especially because the epitope is not a rigid body inside the MHC-I cleft, and normal amino acid fluctuations can occur 22 . It came to our attention that only HLA-B*27:05-restricted complexes presented all atoms RMSD values greater than the respective crystal resolution average. This is attributed to the fact that HLA-B*27:05-restricted epitopes present a high proportion of arginine residues 40 , containing long side chains, which in turn accounts for most of the RMSD deviation observed.
Most of the Cα RMSD data was distributed below 1.5 Å, indicating a high precision of our technique (Fig. 2a). However, we observed an incoherent value of 3.129 Å for one of the epitopes bound to HLA-A*02:01. This value corresponds to the epitope LAGIGILTV derived from the MART-1/Melan-A protein (PDB ID: 2GTW). This epitope represents a variant of the 10-mer epitope ELAGIGILTV, which is recognized by MART-1-reactive T cells 41 . The interesting fact is that this 10-mer epitope presents a bulged conformation comprising the residues Gly-Ile-Gly-Ile, which is replicated by the 9-mer epitope LAGIGILTV, comprising the same residues. To adopt this conformation, the P1 leucine residue of the 9-mer epitope is inserted into the P2 pocket, exactly as it occurs with the 10-mer epitope 42 . This bulged conformation accounts for the major deviation values observed between the model and the crystal structure (Fig. 5). As discussed by Borbulevych et al., this bulged conformation differs from other HLA-A*02:01 bound 9-mer epitopes from MART-1/Melan-A protein, such as ALGIGILTV (PDB ID: 2GTZ) and AAGIGILTV (PDB ID: 3QFD), which present a common extended conformation and incidentally produced better Cα RMSD values here (1.435 Å and 1.377 Å, respectively). It is important to note that DockTope Scientific RepoRts | 5:18413 | DOI: 10.1038/srep18413 is based on a technique that uses epitope backbone patterns inside the MHC-I cleft; thus it is possible that unusual or aberrant epitope conformations will not be properly assessed.
Structures generated using DockTope can be used in several immunology fields, such as cancer research, transplantation, in silico stabilization assays and cross-reactivity assessment, expanding the range of possibilities to study these topics. In fact, our tool have already proven to be useful when studying cross-reactivity among different pMHC-I complexes. In a previous work, Principal Component Analysis (PCA) and Hierarchical Clustering Analysis (HCA) were employed to compare electrostatic potential data of TCR-interacting residues presented on the pMHC-I surface 43 . A total of 28 known HCV targets (epitopes from NS3 protein) were modelled and analysed. The differences observed in PCA and HCA were evidences for structure-dependent immunogenic patterns and were in accordance with in vitro data of IFN-γ releasing assays 44 . After that, 55 pMHC-I complexes including epitopes from different viral proteins were also modelled; this allowed us to infer other potentially cross-reactive targets with HCV-NS3 1073 , such as LMP2 329 from Epstein-Barr virus (EBV), Gag 77 from Human Immunodeficiency virus (HIV), and NA 231 from Influenza virus (IV). Of note, cross-reactive responses of NS3 1073 -specific CD8 + T cells against all of these targets were later confirmed through in vitro assays 45 . Intriguingly, the linear sequence of the confirmed cross-reactive epitope EBV-LMP2 329 presents no similarities in amino acid sequence with the reference HCV-NS3 1073 epitope, and shares only 33% of biochemical properties. Sequence-based analysis would most likely be unable to predict such cross-reactivity. However, an incredible resemblance is observed in a higher level of complexity, through the analysis of the TCR-interacting surface of the pMHC-I. Such analysis was made possible by modelling these pMHC-I complexes through D1-EM-D2, the approach behind DockTope (Fig. 6).
Three approaches stand out among previously published methodologies aiming at the pMHC structural prediction: (i) MHCsim 10 , (ii) pDOCK 8 , and (iii) a Biased-Probability Monte Carlo docking protocol published by Bordner and Abagyan 9 . MHCsim was the first automated server designed to model pMHC complexes. The server uses the input sequences (MHC and epitope) to perform a search in an internal database for the pMHC structure that is the most similar to the input sequences. Then, the template is modified at the positions where the residues differ to generate a new 3D structure. Some aspects not included in MHCsim are addressed by DockTope. The MHCsim methodology is based only on sequence similarity, which might not be sufficiently accurate to predict the 3D structure of a pMHC complex, especially when it comes to epitope conformation. This way, the provided pMHC structure is not final, but can be used for posterior refinement 46 . Second, the MHCsim server allows the pMHC construction for human allotypes only, and is restricted to 9-mer epitopes. The pDOCK methodology is based mainly on ICM docking, Monte Carlo sampling and local minimization. In its validation, the authors presented Cα RMSD values below 1 Å in a set of 186 pMHC-I and pMHC-II structures. However, contrary to DockTope, which used cross-docking to reproduce crystal structures, the validation process of pDOCK was performed through a re-docking approach and all atom RMSD values were not provided in the text. Also, pDOCK is currently not available as a web server, but only as an in-house protocol. The method published by Bordner and Abagyan is based on ICM docking, homology modelling and Support Vector Machine (SVM). They were able to reproduce through cross-docking a set of 14 HLA-A*02:01 epitopes and 9 H-2-Kb epitopes with epitope backbone RMSD values inferior to 1 Å. Like pDOCK, their method is not available as a web server.
DockTope emerges as a free, automated, well-validated (Cα RMSD mean values below 1 Å), and user-friendly web-server tool for modelling pMHC-I complexes in a reliable way. Its usefulness was already demonstrated by previously published work. The possibility to construct pMHC-I complexes will open new avenues for structural immunoinformatics, hopefully triggering new discoveries in basic immunology and health applied sciences. Methods DockTope Automation. DockTope was developed as an optimized tool based on the D1-EM-D2 pMHC-I modelling approach. In order to automate the process, we employed a series of 9 shell scripts, 13 python scripts, 7 C++ executables and 2 python executables to perform the following steps: (i) Epitope structure modelling, (ii) first molecular docking (D1), (iii) choosing the best structure from D1, (iv) second molecular docking (D2), (v) choosing the best structure from D2 and (vi) writing the output. These steps are represented in the flowchart of Fig. 1 and in Supplementary Fig. S1.
Epitope structure modelling. The epitope to be modelled is provided as a linear amino acid sequence (without three-dimensional coordinates). A python script, which launches a built-in PyMOL 47 plug-in, uses the backbone of the epitope pattern to give shape to the modelled epitope. This epitope undergoes energy minimization, allowing for a mild global relaxation of the peptide. First (D1) and Second (D2) Molecular Docking. Molecular docking is performed using the programs AutoDock Tools 48 and AutoDock Vina 19 ; it involves three main steps. In the first step, the MHC-I molecule is prepared according to the following protocol: (i) adding all hydrogens, (ii) adding Gasteiger charges and (iii) removing non-polar hydrogens. In the second step, the modelled epitope is prepared by repeating the same protocol used for the MHC-I, but including an additional level: the torsion tree is set in a manner to maintain the epitope backbone rigid during the molecular docking process, thus only allowing the movement of side chains. In the final step, a box grid is configured to allow the search for the best epitope conformations inside the MHC-I cleft (formed by the α 1 and α 2 domains). The search for the best conformation is performed according to specific algorithms 19 along twenty rounds (arbitrary value). In the end, the best structure is chosen according to an algorithm developed by our group.
Energy Minimization (EM). The energy minimization process is performed using the GROMACS package 21 . This process is used twice: on the modelled epitope and on the best pMHC-I complex produced by the first molecular docking, with the final goal of removing possible steric clashes and correcting distances between atoms in the system. The EM protocol is performed using a virtual cubic box filled with the protein and water (Simple Point Charge water model). Ions Na + and Cl − are included to neutralize the system, maintaining a final concentration of 0.15 M/L. The GROMOS53a5 force field 49 is used to compute inter-and intramolecular interactions. The cut-off distances for the Coulomb (electrostatic and long-range attraction) and Lennard-Jones (repulsion and short-range attraction) forces are set to 1 nm. Molecular dynamics parameters include the steepest descent method of integration, with no constraints and a total of 10,000 steps, with an initial time step of 0.001 nm. The minimization converges after the maximum force is smaller than 2,000 kJ mol −1 nm −1 ; the lowest energy coordinates are then written to a file.
Choosing the best structure. The output of the first and second molecular docking process is composed of the best 20 epitope conformations in a set that could contain up to 1000 conformations. The best conformation, among these twenty, is chosen according to the following equations: Where, BE represents the binding energy value provided by AutoDock Vina for each epitope conformation (i = 1, i = 2,…, i = 20); Co represents a cut-off value based on the average of the twenty BE; "n" represents the structure conformation chosen based on equation (2) and that will be used in the next steps of the calculation; "M" represents a matrix used to combine the selected data from equation (2); "k ie " is the resultant file from the union ("U") of all data from equation (3); "RMSD j " represents the file containing the RMSD average from the data contained in the k ie files; and "Best" is the final structure containing the three-dimensional coordinates of the chosen epitope.

DockTope validation.
For each pMHC-I structure downloaded from the PDB, the epitope and MHC-I parts were separated. Next, the epitope linear sequence and the MHC-I allele name (according to Table 1) were used as input for a cross-docking process, using DockTope. At the end of the process, the modelled pMHC-I structure was compared to its respective structure available in the PDB; quantitative data was obtained through the RMSD analysis of the two structures, considering the Cα and the all atoms RMSD displacement of the epitope. For the analysis, the structures available in the PDB were refined to contain only the pMHC-I structure, without TCR and possible ligands interfering with the peptide:MHC-I interaction. In the end, a total of 135 pMHC-I structures encompassing the MHC-I allotypes HLA-A*02:01, HLA-B*27:05, H-2-Db and H-2-Kb were evaluated ( Table 2). The performance of DockTope was also evaluated through the modelling of 238 epitopes downloaded from the Immune Epitope Database and Analysis Resource (IEDB). The IEDB parameters for epitope search were set to contain only linear epitopes, from any disease, and confirmed by T cell assays (positive). Web server. DockTope can be accessed through the CrossTope website (http://www.crosstope.com.br), in the "Tools" tab, or directly from dirac.cesup.ufrgs.br/bio/home.php. To use this tool, the user should sign in providing basic information such as name, email address, institution and academic degree. After logging in, the user will find the following tabs: Home, Submit, Processing Jobs, About DockTope, Collaborators, Contact Us and Frequent Asked Questions (FAQ). The web server includes two interfaces: user-tool and tool-server. The user-tool interface uses more than one programming language to better integrate all the modules. The visual module (web interface) was developed in PHP and jQuery Ajax, which are based on an HTML structure. All internal actions of the web interface are controlled and executed through JavaScript, especially processes validation, such as login validation, for example. The interface management and integration service available to the user, as well as the non-visible part (such as the execution of .js files) were obtained through the XAMPP server, which includes the APACHE, MySQL and PHP packages. The tool-server interface works exclusively through JavaScript (connection, submission and receipt of submitted jobs). A verification module works constantly over each created page to ensure the database connection, allowing access to the user. All the jobs, after being submitted, enter in a queue until the server checks and allows them to run.