Structural modeling of HLA-B*1502/peptide/carbamazepine/T-cell receptor complex architecture: implication for the molecular mechanism of carbamazepine-induced Stevens-Johnson syndrome/toxic epidermal necrolysis

Drug-induced adverse reactions are a significant problem in healthcare worldwide and are estimated to cost billions of dollars annually in the United States. A portion of such reactions is observed to strongly associate with certain human leukocyte antigen (HLA) alleles; one of the strongest associations is the HLA-B*1502 protein with carbamazepine (CBZ)-induced Stevens–Johnson syndrome/toxic epidermal necrolysis (SJS/TEN) – the odds ratio value can even be higher than one thousand. The particularly strong association in CBZ-induced SJS/TEN suggests that the HLA-B*1502 is not only a genetic marker but also a participant in the pathogenesis of the disease. In the current study, we attempt to computationally model the atomic-level structure of the complete HLA-B*1502/peptide/CBZ/T-cell receptor (TCR) complex architecture based on prior knowledge obtained from epidemiological investigations as well as in vitro and in vivo assays. The model tells a different story about the molecular mechanism of CBZ-induced SJS/TEN from that previously reported for abacavir (ABC)-induced hypersensitivity (HSR); the CBZ molecule is located at the interface between HLA-B*1502/peptide and TCR, directly contacts the P3–P6 residues of antigen peptide, and bound within a pocket region encompassed by two TCR CDR3 fingers. Molecular dynamics simulation and binding energy analysis further reveal that the CBZ shows considerably high affinity to TCR over HLA-B*1502/peptide, which can tightly interact with the former rather than the latter. From the model, two hypotheses are proposed that can well explain most previous observations and are expected to guide next wet-lab experiments. This study could help to promote our understanding of the molecular mechanism and pathological implication underlying CBZ-induced SJS/TEN.


Introduction
Adverse drug reactions (ADRs) cause significant morbidity and mortality for patients and are an expense to the healthcare system. It was reported that more than two million hospitalized patients/year who experience a serious ADR and 106,000/year who die from an ADR in the United States (Bond & Raehl, 2006); the relevant cost has been estimated at US$136 billion annually (Johnson & Bootman, 1995). The ADRs can be classified into types A and B. Type A reactions are associated with drug's pharmacological activity; they are dose-dependent and are therefore readily reversible on reducing the dose or withdrawing the drug. In contrast, type B adverse reactions are caused by immune response and cannot be predicted from the known pharmacology of drugs (Pirmohamed, Breckenridge, Kitteringham, & Park, 1998).
In spite of strong genetic associations, the molecular mechanism of drug interactions with specific HLA alleles is not yet well characterized. Several theories, including hapten (or prohapten), p:i (immune receptor) and danger models, have been introduced to explain the HLA-associated drug hypersensitivity reactions (HSR) Pirmohamed, Naisbitt, Gordon, & Park, 2002). However, it is hard to determine whether these theories can be used for a specific reaction and, if yes, which one is applicable to the reaction. Recently, the high-resolution crystal structures of self-peptides presented by abacavir-modified HLA-B*5701 have been solved Ostrov et al., 2012), which gave the first molecular-level insight into structural basis of HLA-associated drug hypersensitivity . The crystallographic study, however, is extremely time-consuming and expensive, thus limiting its application even for a very few reported cases. Instead, the structural bioinformatics approach, which has been rapidly progressed in recent years, provides a promising way to computationally elucidate the structural basis, dynamics behavior, and energetic property of HLA-drug association and their biological role in diverse ADRs. Nevertheless, although the strategy has been widely applied to investigating various biological phenomena, it, to the best of our knowledge, still remains unexploited in the ADR area.
In this study, we attempt to understand the molecular mechanism and pathological implication underlying the HLA-B*1502-associated, CBZ-induced SJS/TEN. The CBZ is an anticonvulsant and mood-stabilizing drug used primarily in the treatment of epilepsy and bipolar disorder as well as trigeminal neuralgia, which has also been observed to cause several serious adverse reactions, including maculopapular exanthema, HSS and the SJS/ TEN . The CBZ-induced SJS/ TEN is estimated to be responsible for up to 35% of drug-induced SJS/TEN cases in the Han Chinese population, but only 6% of cases in Caucasians, which occurs within 2 months of drug administration, with a median onset of 15 days (Bharadwaj, Illing, & Kostenko, 2010). The nearly 100% association of HLA-B*1502 with CBZ-induced SJS/TEN suggests that the HLA-B*1502 is a central participant in the pathogenesis of the disease (Yang et al., 2007). Here, we attempt to computationally model the atomic-level structure of the complete HLA-B*1502/peptide/CBZ/T-cell receptor (TCR) complex by integrating theoretical methods and empirical knowledge. We demonstrated that the modeled complex can be readily utilized to explain most observations in previous reports. In addition, based on the model, two hypotheses were also proposed that may help to guide next wet-lab experiments.

Virtual mutation of amino acid residues
The few focused amino acid residues in HLA, peptide, or TCR can be virtually mutated to other residue types using an integrative protocol: first, the side chains of the focused residues were removed manually from structure architecture and then new side chains were added automatically to these incomplete residues using the rotamerbased SCWRL program (Krivov, Shapovalov, & Dunbrack, 2009). The SCWRL algorithm predicted all possible combinations of side chain rotamers in a specific structural context using tree decomposition strategy and optimized topological parameters for the interaction graph between these combinations based on a backbonedependent rotamer library derived from high-quality protein crystal structures. Previous comparison studies demonstrated that the SCWRL performed much better than other available tools such as SCATD, SPDBV, and SCit in predicting the side chain conformations of HLA/ peptide complexes (Knapp et al., 2008). In addition, the SCWRL has recently been successfully utilized to reproduce the crystal structures of peptide side chains in receptor-bound state . We therefore believe that this rotamer-based method can be applied to the current system as well.

Molecular docking
The molecular docking described in our recent work (Yang, Wang, Zhang, Huang, & Zhou, 2015) was carried out using Autodock Vina (Trott & Olson, 2010). The CBZ molecule was constructed and optimized with the MMFF94 force field (Halgren, 1996), and Gasteiger partial charges (Gasteiger & Marsili, 1996) were assigned for its atoms. The polar hydrogen atoms and Kollman charges (Besler, Merz, & Kollman, 1990) were added to HLA protein and bound peptide using the AutoDock Tools (Morris et al., 2009), which was also utilized to prepare the input pdbqt file for the CBZ ligand and to set the size and center of a grid box covering the CBZbinding site on HLA-B*1502/peptide complex surface, with the dimensions of x, y, and z in 1.0 Å spacing. In the docking procedure, the Lamarckian genetic algorithm was employed to generate thousands of potential conformations for CBZ within the box region, which were finally clustered into few representatives.

MD simulation
Important biological functions like protein conformationl changes (Balasco, Barone, & Vitagliano, 2015), new molecule design (Arfeen, Patel, Khan, & Bharatam, in press), and structural features (Fan, Zheng, Cui, Li, & Zhang, 2015) can be uncovered by investigating them using molecular dynamics (MD) simulations. Here, MD simulations were performed on the constructed HLA/ peptide, HLA/peptide/CBZ, and HLA/peptide/CBZ/TCR systems using the AMBER10 force field implemented in Amber11 package (Case et al., 2005). The complex systems were solvated in a periodic TIP3P water box extended 10 Å away from any solute atom. Counter-ions of Na + were placed around the solute molecules based on Columbic potential to keep the whole systems neutral. The antechamber tool (Wang, Wang, Kollman, & Case, 2006) was used to treat CBZ molecule. In the procedure, the AM1-BCC charge (Jakalian, Jack, & Bayly, 2002) was assigned to each CBZ atom; the topology and connectivity were generated for CBZ molecular structure, and the general AMBER force field (Wang, Wolf, Caldwell, Kollman, & Case, 2004) was employed to describe the force field parameters that were not found in the standard AMBER.
The systems were minimized using a combination of steepest descent algorithm and conjugate gradient algorithm; first, all hydrogen atoms were minimized, followed by all water molecules and counter-ions, and finally the minimization was addressed on the whole systems without any constraint. The maximum number of minimization steps was set to 5000. After the minimizations, the systems were heated to 300 K in 300 ps followed by constant temperature equilibration at 300 K for 500 ps. Subsequently, 100-ns MD production simulations were carried out in an isothermal isobaric ensemble with periodic boundary conditions. An integration step of 2 fs was used for the MD simulations and the particle mesh Ewald method (Darden, York, & Pedersen, 1993) was employed to calculate the long-range electrostatic interactions. A cut-off distance of 10 Å was used to calculate the short-range electrostatics and van der Waals interactions. In order to restrain all covalent bonds involving hydrogen atoms, the SHAKE method (Ryckaert, Ciccotti, & Berendsen, 1977) was used. Each simulation was coupled to a 300 K thermal bath at 1.0 atm through the Langevin algorithm (Wu & Brooks, 2003).

Interaction free energy analysis
The mm-pbsa program in Amber11 package (Case et al., 2005) was used to analyze the interaction free energies between different parts of a complex system. The analysis was performed over the thousands of snapshots evenly extracted from the last 50 ns of MD trajectory. The total free energy change DG ttl upon the binding of two components consists of direct nonbonded interaction DG nbd and indirect solvent effect DG slv (Zhou, Huang, & Tian, 2012), i.e. DG ttl ¼ DG nbd þ DG slv ; the former was calculated using the molecular mechanics (MM) approach based on AMBER10 force field, and the latter was expressed as solvation free energy, which can be decomposed into electrostatic and nonpolar contributions. The electrostatic contribution to solvation free energy was calculated by finite difference solution of nonlinearized Poisson-Boltzmann (PB) equation using the delphi program (Rocchia, Alexov, & Honig, 2001). In the calculation, the solute and solvent were assigned with dielectric constants 2 and 78, respectively. The atomic radii and partial charges were taken from the parse parameter set (Sitkoff, Sharp, & Honig, 1994) and the AMBER10 force field, respectively; the nonpolar solvation contribution was derived from the surface area (SA) model b þ c A, where A is solvent accessible surface area computed with a solvent-probe radius of 1.4 Å, and b and c are coefficient terms that were defined to be .92 kcal/mol and .00542 kcal/mol/Å 2 , respectively, as suggested by Kollman and co-workers (Kollman et al., 2000).

Analysis of epidemiological investigations and experimental evidences
The CBZ-induced SJS/TEN has been previously reported to strongly associate with HLA-B*1502 allele in Han Chinese populations (Chung et al., 2004). The nearly 100% association implied that the HLA-B*1502 is not only a genetic marker but also a participant in the pathogenesis of the disease (Yang et al., 2007). Later, from gene fine mapping, Hung et al. narrowed the susceptibility locus for the CBZ-induced SJS/TEN to within 86 kb region where HLA-B is the only known gene, and confirmed a very strong association of the adverse reaction with HLA-B*1502 (OR = 1357, 95% CI: 193 to 8838) (Hung et al., 2006).
Based on the epidemiological investigations, Chen and co-workers have performed a series of experimental analyses to give molecular insights into the strong association (Yang et al., 2007;Ko et al., 2011), and they found that (i) the antigen peptides of HLA-B*1502 prefer use of serine residues at their nonanchoring positions P5, P6, P7, and P8, (ii) CBZ does not covalently modify these peptides, (iii) CBZ does not alter the peptide repertoire of HLA-B*1502, (iv) HLA-B*1502 directly presents CBZ without intracellular metabolism and antigen processing, (v) the presentation of CBZ requires endogenous peptides loaded in the antigen-binding groove of HLA-B*1502, (vi) the cytotoxicity elicited by CBZ or its metabolites can be abolished by a washing procedure, (vii) the 5-carboxamide moiety is functionally required for CBZ and its reactive metabolites, (viii) CBZ can directly interact with intact HLA-B*1502/peptide complex, (ix) the N63, I95, and L156 of HLA-B*1502 are the key residues for CBZ binding, and (x) restricted TCRs participate in CBZ-induced SJS/TEN. From these observations, it is suggested that CBZ non-covalently locates on the surface of HLA-B*1502/ peptide complex and is directly touched by TCR, supporting the p:i mechanism involved in CBZ-induced SJS/TEN. This is different from abacavir (ABC)-induced HSRanother HLA participating ADR which was observed to strongly associate with HLA-B*5701 allele (OR = 117, 95% CI: 29-481) (Mallal et al., 2002). Recently, the high-resolution crystal structures of ABC in complex with HLA-B*5701 and self-peptides were solved Ostrov et al., 2012), from which it is revealed that the ABC molecule is rooted on the bottom of HLA-B*5701's peptide-binding groove, reshaping the geometrical property and chemical environment of the antigen-binding cleft, thereby altering the peptide repertoire presented by HLA-B*57:01. The significant difference in the interaction manner of CBZ with HLA-B*1502 and ABC with HLA-B*5701 imparts distinct molecular mechanisms underlying CBZ-induced SJS/TEN and ABC-induced HSR, which is schematically illustrated in Figure 1. As can be seen, the ABC locates beneath antigen peptide and is separated from TCR by the peptide, while CBZ bound at the interface between HLA-B*1502/peptide and TCR. The model can readily explain many experimental observations. For example, the CBZ-specific TCR types are restricted (Ko et al., 2011), whereas broadly polyclonal TCR usage was observed in response to ABC ; ABC alters HLA peptide repertoire (Norcross et al., 2012), but CBZ does not (Yang et al., 2007); the washing procedure can abolish the cytotoxicity elicited by CBZ (Wei, Chung, Huang, Chen, & Hung, 2012) rather than by ABC (Chessman et al., 2008).
3.2. Setup of HLA-B*1502/peptide/CBZ complex structure model First, we tried to construct the structure model of CBZ interaction with HLA-B*1502/peptide complex, where TCR was not present. Because the HLA-B*1502 crystal structure is currently unavailable in the Protein Database Bank (PDB) (Berman et al., 2000), we employed a computational protocol to model its complex structure with cognate peptide. The high-resolution crystal structure (solved at 1.79 Å) of HLA-B*1501 bound with a nonapeptide ILGPPGSVY was retrieved from the PDB database under the accession ID: 1XR9, which was used as template to generate HLA-B*1502/peptide complex structure models.
The HLA-B*1501 and HLA-B*1502 are highly conserved in their primary sequence and advanced structure. Considering that there are only five residue differences between the two alleles, i.e., E63 N, T94I, L95I, H113Y, and W156L from the former to the latter (the sequence alignment between HLA-B*1501 and HLA-B*1502 is shown in Supplementary Figure S1), instead of the sophisticated homologous modeling method that are widely used to predict protein structure from homologous crystal template (Martí-Renom et al., 2000), we herein employed virtual site-directed mutagenesis strategy to model the HLA-B*1502 structure based on the HLA-B*1501 template. In the procedure, the mutations E63 N, T94I, L95I, H113Y, and W156L were addressed directly on the crystal structure of HLA-B*1501 to obtain the preliminary structure of HLA-B*1502 complexed with its non-cognate peptide ILGPPGSVY using the SCWRL method (Krivov et al., 2009). Then, the peptide was automatically mutated to several HLA-B*1502 binders separately, including SLFVSNHAY, ELWKNPTAF, NVIRDAVTY, QLGPVGGVF, FLFDGSPTY, and SVKPASSSF; these HLA-B*1502-cognate peptides were used previously by Wei et al. (2012) to investigate CBZ-elicited cytotoxic activity. The modeled HLA-B*1502/peptide systems were relaxed by MM minimization and MD simulation to eliminate potential atomic collisions and bond distortions involved in the artifacts.
Subsequently, computational solvent mapping algorithm (Ngan et al., 2012) was employed to detect CBZ-binding hot spots on the surface of HLA-B*1502/ peptide complex. The mapping rendered 9 hot spots; most of them located on the region nearby conserved α3  domain and β2 microglobulin that cannot contribute to the HLA-B*1502-CBZ association specificity, and only three hot spots are close to the B and C pockets of the variable antigen-binding cleft (Figure 2). Therefore, we set a box that fully covers the three effective hot spots and carried out molecular docking calculations to predict the binding modes of CBZ in the box. Consequently, eight clusters of docking-generated modes were obtained, and we performed 100-ns MD simulations for each of the clusters in order to reach at equilibrated state for the HLA-B*1502/peptide/CBZ system. After the MD simulations, two clusters achieved stable form while other six were considerably unstable that the CBZ molecule cannot maintain complex with the HLA-B*1502/peptide. One of the two stable complex structures is shown in Figure 3; it is seen that the CBZ occupies on the surface of the complex, exhibiting a weak binding behavior toward the HLA-B*1502/peptide. MM-PB/SA analysis unraveled that the CBZ interaction energies with HLA-B*1502/peptide in the two stable complexes are separately −13.4 and −15.1 kcal/mol, which are significantly lower than those of small-molecule drugs bound into the narrow cleft or pocket in their cognate targets. Further, we changed the peptide and repeated above procedure, and consequently, a similar result was obtained, that is, most docking clusters cannot maintain in a tightly bound state during MD equilibration, and in few cases the CBZ only showed a low affinity to HLA-B*1502/peptide, suggesting that the direct binding of CBZ to HLA-B*1502/ peptide is unstable that may need additional component such as TCR to stabilize the three-body system.

Addition of TCR to the HLA-B*1502/peptide/CBZ system
From above computational modeling analysis, we found that the CBZ cannot bind tightly to the HLA-B*1502/ peptide. It is known that small-molecule drugs are commonly bound within the narrow cleft and deep cave of their protein targets, such as enzyme active site. However, the HLA-B*1502/peptide complex has only a large, flat surface that may not be effective to hold CBZ molecule steadily. In fact, some experimental observations also support this point. For example, no Figure 4. The HLA-B*1502/peptide/CBZ complexes have not yet been successfully crystallized to date, and a washing procedure can readily abolish CBZ-elicited cytotoxic effect. Thus, it is implied that the presence of TCR might be the prerequisite for forming stable reactive system. HLA-B*1502/peptide/CBZ complexes have been successfully crystallized to date and a simple washing procedure can abolish CBZ-elicited cytotoxic effect. Thus, it is implied that TCR may play a critical role in stabilizing the system (Figure 4).
Further, we attempted to model the complete HLA-B*1502/peptide/CBZ/TCR complex structure. A computational procedure shown in Figure 5 was used to fulfill the modeling, that is, the HLA-B*1502/peptide/ CBZ complex modeled above was superposed onto the HLA-B*3501/peptide/TCR crystal structure (PDB ID: 3MV7) by means of least-squares fitting algorithm (Martin, 1982). The peptide bound to HLA-B*3501 is a 11-mer sequence that forms a protrudent loop that  occupies at the pocket region between the two chains of TCR. Instead, HLA-B*1502 is complexed with a 9-mer peptide that exhibits an extended conformation within the peptide-binding cleft of HLA-B*1502. It is evident in Figure 6 that the CBZ-binding site on the surface of HLA-B*1502/9-mer peptide complex corresponds to the protrudent loop of the 11-mer peptide in complex with HLA-B*3501, suggesting that the CBZ molecule in HLA-B*1502/peptide/CBZ system plays a role similar to that of the peptide protrudent loop in HLA-B*3501/peptide/TCR system, which can facilitate the proper recognition of HLA-B*1502/peptide by TCR. Next, the HLA-B*3501 and peptide in original crystal structure were manually removed from the superposition, resulting in an artificial HLA-B*1502/peptide/CBZ/TCR system. Recently, Ko et al. observed a shared and restricted TCR usage in the T-cell clones expanded from CBZ-induced SJS/TEN patients; sequencing of TCR CDR3 regions found that few particular sequence patterns such as VA-22-FISGTY and VB-11-ISGSY are frequently used by CBZ-specific T cells (Ko et al., 2011). In this respect, we adopted different combinations of the CBZ-specific CDR3 sequences to correct TCR CDR3 in the artificial HLA-B*1502/peptide/CBZ/TCR system obtained above. The sequence combinations are extracted from Ko et al. (2011) and tabulated in Table 1. The SCWRL algorithm was employed to mutate the original TCR CDR3 regions, and then the mutated HLA-B*1502/peptide/CBZ/TCR system was equilibrated with 100-ns MD simulation.  The resulting HLA-B*1502/peptide/CBZ/TCR complex architecture is shown in Figure 7. The system is stable and can maintain in bound state over the MD simulation procedure. According to the model, CBZ is located at the interface between the HLA-B*1502/peptide complex and TCR (Figure 7(A)), directly contacts the P3-P6 residues of antigen peptide (Figure 7(C)), and bound tightly into the pocket formed by two TCR CDR3 fingers (Figure 7(B)). It is worth noting that the surface of HLA-B*1502/peptide complex has no typical cleft to 'catch' CBZ, whereas the deep pocket in TCR appears to well accommodate the CBZ, suggesting a preferred interaction of the ligand with TCR relative to HLA. To gain a quantitative picture, we performed MM-PB/SA analysis to calculate the binding free energies (affinities) ΔG ttl of CBZ separately to HLA-B*1502/peptide and to TCR over the thousands of snapshots extracted from the MD trajectory of complete system. The obtained results are shown in Figure 8(A), where the peptide is SLFVSNHAY and the two TCR CDR3 fingers adopt different combinations of CBZ-specific sequence patterns. It is evident that the ΔG ttl value upon the binding of CBZ to TCR is much higher than those to HLA-B*1502/peptide; the former is larger than −25 kcal/mol, while the latter is only about −10 to −15 kcal/mol. We demonstrated that peptide categories have only a modest influence on CBZ affinity to HLA-B*1502/peptide, varying from −9.7 kcal/mol (QLGPVGGVF) to −14.1 kcal/mol (SVKPASSSF) (Figure 8(B)). In addition, we also investigated the interaction between HLA-B*1502/peptide complex and TCR in the presence or absence of CBZ. It is seen from Figure 8(C) that in four of five examined cases, presence of CBZ can enhance TCR interaction potency with HLA-B*1502/peptide, indicating that the CBZ is an effective adjustor to mediate the recognition of HLA-B*1502/peptide complex by its specific TCR.

Hypotheses and implications
From the structure model as well as dynamics simulations and energetic analysis, we herein proposed two hypotheses and discussed their biological implications.
Hypothesis 1. The CBZ is first bound to TCR to form stable TCR/CBZ complex, and then the complex  . The CBZ is first bound to TCR to form TCR/CBZ complex, and then the complex binds HLA/peptide, resulting in the complete HLA/peptide/CBZ/TCR system. recognizes and binds HLA/peptide presented on cell surface, finally giving rise to the complete HLA/peptide/ CBZ/TCR system (Figure 9). According to MM-PB/SA analysis, the binding energy of CBZ to TCR is considerably larger than that of CBZ to HLA-B*1502/peptide (Figure 8(A)), which means that in solvent condition, CBZ can only weakly interact with HLA-B*1502/peptide, maintaining a dynamic balance between bound and unbound states; the presence of TCR can effectively promote the balance shifting into bound state. A direct experimental evidence for this hypothesis is that a simple washing procedure before performing T-cell cytotoxicity assay can completely abolish the cytotoxicity (Wei et al., 2012). Another indirect support is no report for successfully crystallizing the HLA-B*1502/peptide/CBZ threebody system up to now, and we therefore suggest that the crystallization efforts would be attempted in the presence of appropriate TCR subtypes.
Hypothesis 2. Binding of CBZ to TCR shifts TCR specificity; the CBZ-modified TCR may recognize selfpeptides presented by HLA-B*1502 and then cause autoimmune reaction (Figure 10). The T-cell thymocytes mature in thymus, where they undergo negative selection that removes portion of the thymocytes with particular TCR subtypes capable of strongly binding with self-peptides presented by medullary thymic epithelial cells (mTECs). In the procedure, only those incapable of recognizing and binding self-peptides can survive and be released to lymph circulation. However, some survived T-cells present particular TCR subtypes with high affinity for CBZ, resulting in CBZ-modified TCR that retrieve the capability of recognizing self-peptides, thereby activating self-reactive cytotoxic T lymphocyte (CTL).

Concluding remarks
Although intensive efforts have been addressed on epidemiological investigation of ADRs over the past decades, the molecular mechanism and biological implication underlying ADRs still remain largely unexplored to date. In this study, we attempted to computationally model the first three-dimensional structure of complete HLA-B*1502/peptide/CBZ/TCR complex, regarding a deep understanding of the molecular mechanism and structural basis of CBZ-induced SJS/TEN. The model can be used to explain most observations in previous experiments, according to which the CBZ is located at the interface between the HLA-B*1502/peptide and TCR, directly contacts the antigen peptide, and bound within TCR pocket. We demonstrated that the CBZ can bind TCR more tightly than HLA-B*1502/peptide, suggesting a crucial role of TCR in stabilizing the complex architecture. We also proposed two hypotheses that could be used to guide next wet-lab experiments. This study would help to establish a different molecular framework for CBZ-induced SJS/TEN from that recently published for ABC-induced HSR Ostrov et al., 2012).

Supplementary material
The supplementary material for this paper is available online at http://dx.doi.org/10.1080/07391102.2015.1092476.

Disclosure statement
No potential conflict of interest was reported by the authors. Figure 10. The binding of CBZ to TCR shifts TCR specificity; the CBZ-modified TCR may recognize self-peptides presented by HLA-B*1502, thus leading to autoimmune reaction.