- Email: [email protected]
Modelling antibody-antigen interactions: ferritin as a case study
Molecular
ImmunologyVol. 32, No. 13. pp. 1001-1010,
Pergamon 0161-5890(95)00027-5
1995 Copyright ‘0 1995 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0161-5890/95 $9.50 + 0.00
MODELLING ANTIBODY-ANTIGEN INTERACTIONS: FERRITIN AS A CASE STUDY MANUELA
HELMER-CITTERICH,*§
ALESSANDRA
ERMANNA
and
LUZZAGOf
ANNA
ROVIDA,?
TRAMONTANOS
*Department of Biology, University of Rome Tor Vergata, Via della Ricerca Scientifica, 00 133 Roma, Italy; TIstituto di Tecnologie Biomediche Avanzate XNR, Via Ampere 56, Milano, Italy; and $IRBM P. Angeletti, Via Pontina Km 30.600, 00040 Pomezia, Roma, Italy (First received 20 October 1994; accepted in revised.form 16 February 1995) Abstract-In this work, we propose a model for the structure of the antigen-antibody complex formed by human H-ferritin and an antibody that specifically recognizes it. We cloned and sequenced the antibody gene, predicted the antibody three-dimensional structure, and reconstructed the Hferritin-antibody complex using an automated docking procedure previously validated on known complexes. This procedure allowed us to identify one putative complex which we carefully analysed, in order both to evaluate its likelihood, in light of a set of experimental results described in the literature, and to predict precisely which are the sites of interaction between the two molecules. Our model is compatible with the experimentally determined characteristics of the complex. Some of the residues that form the predicted antigenic site of ferritin can be found in the amino acid sequence of peptides selected from a random peptide library because of their affinity for the ferritin monoclonal antibody. Furthermore, the structural difference between the antigenic site in human H-ferritin and the corresponding region in other species permits us to rationalize the inability of the antibody to recognize human L-ferritin and rat, chicken and mouse H-ferritin. Through the analysis of our model complex, we identify a number of other residues putatively involved in the interaction. This multidisciplinary approach shows that synergy between computational and experimental methods may bring further insight into the understanding of antibody-antigen recognition rules. Key words: antigen recognition,
antibody prediction. protein docking, ferritin.
INTRODUCTION
The problem of mapping the interaction surface between two macromolecules that bind each other is of outstanding interest in molecular immunology. The problem can be stated as follows: given the structure (or a model of the structure) of an antigen and the structure (or a model of the structure) of an antibody raised against the antigen, can we predict the site of interaction between the two molecules? The solution to this problem, even if approximate, would help in designing experiments to precisely map the residues involved in the interface and could be instrumental both in designing peptides able to mimic the interacting surface of the antigen and in understanding where important regions of an antigen are located in its three-dimensional structure. We have previously shown that an algorithm based on surface complementarity can be used to automatically dock two proteins of known structures or one protein of known structure and a model of the other (HelmerCitterich and Tramontano, 1994). The procedure allows
§Author to whom correspondence Abbreviations: mAb, monoclonal complementarity.
should be addressed. antibody; lsc, low surface
one to obtain a limited number of relative orientations of the two proteins (between one and four) among which the correct orientation, when known from crystallographic studies, is always found. We have also shown that it is possible to build a reliable three-dimensional model of immunoglobulins taking advantage of the “canonical structure model” of hypervariable loops and of the high conservation of the framework (Chothia et al., 1989).
Here we apply our docking procedure to a complex of unknown structure, where a wealth of experimental data is available. Furthermore, the availability of the reagents should facilitate the experimental test of the predictions of our model. The chosen system is that of human H-ferritin and a monoclonal antibody, mAb H107, raised against the molecule. Ferritin is a protein involved in iron storage and is expressed as a 24-mer consisting of a mixture of two homologous subunits, H and L (Arosio et al., 1978). The two subunits share 55% sequence homology and fold into a similar structure of four tightly packed alphahelices (Ford et al., 1984; Lawson et al., 1991). The H chain of human ferritin has been overproduced in E. coli (Levi et al., 1987), and the crystallographic structure of
1001
1002
M. HELMER-CITTERICH
H-ferritin 24-mer has been obtained at high resolution (Lawson et al.. 1991). A monoclonal antibody against human ferritin (H 107) has been raised in mouse and here we report the cloning of its gene and the determination of its nucleotide sequence. This antibody reacts with homo multimers of human H-ferritin synthesized and assembled in E. coli, does not react with human L-ferritin, mouse and rat H-ferritin (Luzzago et al., 1986). The results that we obtained by docking a model of the antibody to the H-ferritin structure are in excellent agreement with experimental results coming from phage library techniques and site-directed mutagenesis and furthermore, adds a substantial amount of interesting structural information on the putative complex formed.
MATERIALS
AND METHODS
Clonirrg afzd sequefrciny qf‘HlO7 cDNA synthesis and VH and VK amplification and assembly were performed essentially as described in Clackson et al. (1991). The assembled product was reamplified with the primers VK4FOR-Not1 (Clackson et a/.. 1991) and VHlBACKSFIIS (Hoogenboom et al., 1991), for cloning into the expression vector pHEN 1. The resulting clone was sequenced with Sequenase kit (U.S. Biochemicals, Cleveland, OH), with primers: LMB3 (5’CAGGAAACAGCTATGAC-3’). MO-LINK-BACK (5’-GGGACCACGGTCACCGTCTCCTCA-3’) MOLINK-FOR (S-TGGAGACTGGGTGAGCTCAATGTC-3’) and 0L6 (5’-CCCTCATAGTTAGCGTAACG3’). HI07 single chain Fv fragment was expressed as soluble form in E. coli HB2 15 I cells (Hoogenboom et al.. 1991) and its specificity tested by dot blot (data not shown). The deduced amino acid sequence of the light and heavy chain variable regions of mAb HI 07 is reported in Fig. 1. Antiho&
modelliny
The model of the H 107 monoclonal antibody was constructed using the automatic procedure described by Lesk and Tramontano (1993). using the database of known immunoglobulin structures. We aligned the sequences of VL and VH chains of the mAb with the sequences of the corresponding domains of the known immunoglobulin structures. The alignments were performed using the program GAP from the GCG package (Genetics Computer Group, 1991). For each domain. we selected as parent structure the immunoglobulin domain with the highest sequence identity. namely HyHEL-5 (Data Bank code 2HFL, Sheriff et al., 1987) for Vi_ and R19.9 (Data Bank Code 2F19, Lascombe et al., 1989) for VH. The alignments used for model building are shown in Fig. I. We superimposed the selected models of the frameworks by the least-squares fit method of the structurally conserved regions identified by Chothia et al. ( 1985) and merged the V, domain of HvHEL-5 and the V,I. domain ,
et ml.
of R19.9 in the resulting relative position. This is justified by the observation that, if one superimposes the conserved regions of the VL domains of any pair of immunoglobulins of known structure, the R.M.S. distance between the conserved regions of their V,, domains is always lower than 1 A (data not shown). We next analysed the canonical structures for each hypervariable loop. The Ll, L2, Hl and H2 loop main chain conformations were left unchanged with respect to their framework templates since they have the same canonical structure. The canonical structure of the HI07 L3 loop is different from that of the parent HyHEL-5 structure. This loop in H 107 is six residues long and is characterized by the presence of a proline in position 95 (in the cis conformation in all known immunoglobulins with this canonical structure) and by a group in position 90 able to hydrogen bond to main chain atoms of the loop. NQlO (Alzdri et al., 1990) is the structure with the highest sequence identity and the same canonical structure which was then selected as a template (Fig. 1). No canonical structure has been identified for the H3 hypervariable loop so far. This loop is the most variable in length, sequence and conformation. It is 15 residues long in the parent R19.9 and 11 in H107. Long H3 loops differ most at their tip (A. T., unpublished observation) so we reasoned that modelling this loop using R19.9 as a template and deleting the central four residues would produce a reasonable structure at least in the initial and final region of the loop, while reducing the unavoidable errors introduced by splicing into a framework a loop from a different structure. Furthermore. database loop searching techniques have been shown not to be reliable in the case of antibody loops (Tramontano and Lesk, 1992) and combined database searching and conformational energy calculations (Bruccoleri et al., 1988; Martin et al., 1989) are very CPU intensive and not yet very reliable for loops longer than eight residues. In the case of L3. where the canonical structure selected for the framework is different from that selected for the loop. we used the weighted least-squares fit of the main chain atoms of the four residues preceding and following the loop described by Tramontano and Lesk (1992) and grafted the loop into the model. The conformations of the side chains were modelled as follows: where the parent structure and the model had the same residue, we retained the conformation of the parent structures. If the side chain differed. we took the side chain conformation (if possible) from another immunoglobulin having the same residue in the corresponding position. When more than one structure was a candidate for this step. that more similar in sequence to H107 was selected. In all other cases, we used the most common conformer for the amino acid side chain. The model was subjected to 100 cycles of energy refinement (with the default parameters of Discover. Biosym Technologies. San Diego, CA) to obtain a reasonable stereochemistry. Complex predictiofz The PUZZLE automated Citterich and Tramontano,
docking procedure (Helmer1994) is based only on geo-
Antibody-antigen
1003
interaction
A) * * ** 1 DIELTQSPAIMSASLGEEITLTCSASSSVSYIHWYQQKSDTSPKLLIYST II 1lIIIIIIIII II.:I:IIIIIIIl.I:.IIIIII:IIlI :Il.I 1 DIVLTQSPAIMSASPGEKVTMTCSA~~QQKSGTSPKRWIY~ l
H107 hyHEL- 5
H107
* * * 51 SSLASGVPSRFSGSGSGTFYSLTISSVEAEDAADYYCHQWSFYPWTFGGG I.IIlIII
HyHEL - 5 nql0
IIIIIIIII
I
IIIIIII:I.IIII:III:II:
IIIII
51 SKLASGVPVRFSGSGSGTSYSLTISSMETEDAAEYYCQQFP,TFGGG
H107
101
HyHEL-5
100
TKLEIK Il1llI TKLEIK
B) H107 R19.9
* ** * 1 DVQLQESGTELVKPGASVKLSCKASGYTFTSYWMHWVKQRPGQGLEWIGE : IIII:II.III:: I.lII:IIIIIIIIIIII ::llllllllllllll 1 QVQLQQSGAELVRAGSSVKMSCKAS GYTFTSXGVNWVKQRPGQGLEWIGY
H107
* 51 IYPSNGATNYNEKFKSKATLTVDKSSNTASMQLSSLTSEDSAVFFCASSG
R19.9
51 INaYL&‘NEKFKGKTTLTVDRSSSTAYMQLRSLTSEDAAVYFCAw
I
I:.l
IIIIII:I.IIIII:II.II
IIl.IIIIII.II:III.I
H107
101
AYY.. ..GNPFTYWGQGTTVTVSS.
R19.9
101 &LAVYYA&WGQGTTLTVSS
IIIIII:llll
Fig. 1. Alignment of the V, (A) and VH (B) chain sequences of mAb H107 with the corresponding sequences of HyHEL-5 and RL9.9, respectively. The sequence ofthe L3 loop of NQIO is also reported. Asterisks indicate positions mainly responsible for the canonical structures of the hypervariable loops (underlined).
metric criteria applied on the solvent accessible surface of the two proteins (Connolly, 1983). Each one of the two proteins is projected onto a cylinder, cut in slices orthogonal to the cylinder axis and opened onto a grid.
The grid is equivalent to a matrix, in which each protein slice is described as a series of values in a row, Each protei! slice (2 A thick) is described as a polygon with sides 2 A in length. All vertices of the polygon coincide with a protein surface point. Each element of a matrix row is calculated as the difference between adjacent radial coordinates of the polygon vertices. Each position in the matrix row describes a fixed area of the protein surface with a number which is related to the protein surface shape in that slice. The search for protein surface complementarity becomes a search for submatrices which describe regions of surface complementarity. A complete search in the space of all possible protein/protein orientations can be made by rotating one of the two proteins, projecting its surface onto the un-
rotated cylinder and applying the comparison algorithm to the new matrix obtained. Firstly, a coarse sampling of the rotation space at 10” steps is performed; regions around the obtained solutions are subsequently scanned at 10 steps. Lacking any a priori knowledge about how many monomers are involved in the 24-met/antibody interaction, one should in principle, use the unique symmetry element of ferritin, that is the heptamer. However, this choice would substantially increase the CPU time required by the procedure. Consequently, we decided to try and use the ferritin monomer as a docking probe and look for relatively small regions of surface complementarity with the target antibody. Since, as we show later, this allowed us to detect contacts between the antibody and more than one monomer, we concluded that this faster approach is indeed adequate. We allowed the listing of areas of surface complementarity lower than the threshold used in our
M. HELMER-CITTERICH
1004
Table 1. List of solutions
test runs. The procedure then became able to construct more than just a few solutions, even with a relatively low area of surface complementarity (> - 100 AZ/protein). We will call these solutions low surface complementarity (kc) solutions. After the coarse run, the solutions were clustered in groups whenever they differed less than 5 A translation and 20” rotation and carefully analysed at the graphics. Solutions not discarded at this stage (see Results) were subjected to the fine run of the programme and finally subjected to energy refinement (1000 iterations of steepest descent algorithm with the default parameters of Discover). Structural
Rank Cl
RESULTS In order to address the problem of precisely identifying the surface of interaction between an antibody and its antigen we selected a well studied system as a case study, namely that of human H-ferritin and of the monoclonal antibody H107. We cloned and sequenced the anti-ferritin H 107 monoclonal antibody gene. modelled its three-dimensional structure (Tramontano and Lesk, 1992) and used the automatic docking method described above (HelmerCitterich and Trdmontano, 1994) to predict the antibody-antigen complex structure. The search for complementarity between the ferritin H monomer and the monoclonal antibody H 107 ended up with 20 Isc solutions, which were carefully analysed at the graphics in order to discard the matches proposed if they were not compatible with the overall 24-mer structure or with an antibody-antigen interaction involving the antigen binding site of the antibody. The results are reported in Table 1. Only eight solutions out of 20 correctly positioned the antibody outside the 24-mer cavity and were compatible with the overall 24-mer structure. These solutions were compared, clustered in three groups (Table 1) and ranked by PUZZLE according to a scoring system based on the dimension of the matching submatrices. The member of the cluster with the largest buried area then became representative of its cluster and underwent the fine run analysis, which resulted in three solutions listed in Table 2 where the value of buried surface area upon complex formation is also shown. Solutions 1 and 2 are relatively similar to each other
Group ‘
Notes ”
I
I
2 3 4 5 6
A
7
A B
II D I
I
8 9 IO II 12 13 I4 15 I6 17 18 19 20
analysis qf’the putative compleses
The proposed complexes were analysed by measuring their buried surface area upon complex formation, the number of contacts and the number of hydrogen bonds between the antibody and the antigen. Buried surface area upon complex formation was measured using the program Insight (Dayringer et al., 1986) and the values reported always refer to the H-ferritin 24mer buried surface area upon complex formation. All other analyses were performed using the program PINQ (Lesk, 1986).
et (11.
I III D 1 B A II C B D D A
“The solutions are ranked here on the basis of the approximate complementary surface evaluation performed automatically by the program. *Discarded solutions are indicated by a letter in this column: A. antibody positioned inside the 24-mer cavity; B, antibody partially positioned inside the 24-mer cavity; C. part of the antibody colliding with a monomer other than the one used as probe protein; D. contacts not involving the hypervariable loops of the antibody. ‘Solutions are subsequently clustered in groups of similarity (indicated by roman numerals).
(3.7 A and 14.4’ of translation and rotation, respectively), while solution 3 substantially differs from them (Table 2). Interestingly, all three of them involve the interaction with another ferritin monomer (no experimental data are available on this point). It is worth noting that the surface of contact of each of these solutions with the second monomer is below the threshold used in the docking procedure, and for this reason they do not appear in the final list of solutions. The three solutions have been analysed for antibody-antigen contacts and potential hydrogen bonding (Table 3). Solution 1 shows the more extended buried surface area upon formation of the dimer-antibody complex (554 Table 2. List of accepted
Solution Solution Solution
1 2 3
solutions
Group
T (A) il
R ( ‘) ”
A (A’) *
II 1 III
3.7 22.0
14.4 34.0
554 283 224
“For each solution, T and R are the translations and rotations calculated with respect to solution 1. ‘A is the H-ferritin 24-mer buried surface area upon complex formation.
Antibody-antigen
interaction
1005
006
M. HELMER-CITTERICH
er rrl
Fig. 2. (a) Ribbon representation of the H-ferritin 24-mer complexed with the V,_ and Vn variable portion of the mAb H107. The ferritin dimer directly involved in binding according to our solution 1 is depicted in yellow (monomer 1) and purple (monomer 2). In blue and white, the remaining portion of H-ferritin and in cyan, the antibody moiety. (b) Solution 1 is shown here in more detail with the same colours as in (a). The side chains of the amino acids directly involved in binding are shown in ball and stick representation.
Antibody-antigen
interaction
1007
Table 3. List of close contacts and hydrogen bonds between the H-ferritin 24-mer and the antibody for each accepted solution Solution number Solution 1
CDR
FW
Ll L2 L3
H2
Monomer I”
Monomer II”
Total number of H-bonds
Y40, R43, D45, D91, D92 KS7 Y40, C90, D91, D92, W93, E94. N98 N109
R79, 180, F81
15
T5, R9. E17,
S66, G67, T68
Solution 2
R79 D89, C90, N98, C102, H105 D84, KS7
H3 Ll
F81, L82, Q83
F81, L82,
17
Q83,DW 486 Y29,486, K87, P88, D89, C90. D91 El16 K87
L3
H2
Solution 3
Ll
Y29, Q86, K87, D89
L3 Dl
Q83, D84
F81, 483,
6
D84, Q86 483, D84
D89, C90, D91
“Two residues are considered in close contact when the distance between any pair of their atoms is less than their van der Waals radii plus 0.6 A; amino acids involved in putative hydrogen bonding are in bold.
A2), 40 amino acid contacts have been detected between the antibody and the ferritin dimer (27 with monomer I and 13 with monomer II). Fifteen hydrogen bonds can be formed (10 with monomer I and five with monomer II). Five CDRs out of six are involved in the binding and each of them always forms two or more hydrogen bonds. Also, three residues of the framework contact amino acids of the antigen and possibly form hydrogen bonds with them. Visual inspection of the complex shows a good shape complementarity between the protein surfaces and no major steric hindrance is detected. Solution 2 has a considerably less extended buried surface area (283 A’); nevertheless, 23 contacts between amino acids of the complex are detected (12 with monomer I and 11 with monomer II). Seventeen hydrogen bonds can be formed to stabilize complex formation. Only three out of six hypervariable loops of the antibody are involved in antigen binding and this is in contrast with what is generally seen in other known antibodyantigen complexes (Chothia and Lesk, 1987). Solution 3 shows a less extended buried surface area upon formation of the putative complex (224 A’) and the analysis of all the parameters which can be measured suggests that it is not a good candidate as a stable complex. Eighteen amino acid contacts can be detected (11 with monomer I and seven with monomer II); the formation of only six hydrogen bonds can be proposed and only two CDRs out of six are involved in the for-
mation of the putative complex. Shape complementarity is very good, even if much less extended than what is observed in known antibody-antigen complexes. Taken together, the results of our analysis show that the solution to be preferred on the basis of the extension of surface complementarity, number of contacts, number of hydrogen bonds and number of antigen hypervariable loops involved in the interaction is indeed the first solution proposed by the automatic procedure. The structure of the proposed complex is shown in Fig. 2. According to the structure of the resulting complex, we propose that the complex epitope recognized by the H107 monoclonal antibody comprises residue 29, residues 40 through 46, 87 through 95,98 through 103, 105 through 109 in one monomer, and residues 5 through 9, 17 through 21 and 77 through 82 in the second monomer. By screening a nonapeptide random library inserted in the major coat protein of the filamentous phage fl (Felici et al., 1991), Luzzago et al. (1993) affinity-purified 11 independent clones with the H107 monoclonal antibody (Luzzago et al., 1986) and the deduced amino acid sequences were then arranged into two groups on the basis of their similarity. The first group was formed by three sequences, and characterized by the consensus Y,D,N,-X-X-X-X-W, the second cornprized eight sequences and had the consensus GSA2SI-S8-X-F,Y, (subscripts indicate the number of occurrences of the amino acid at that position, and dashes separate positions in the
M. HELMER-CITTERICH
1008
et ul. ++
1 Frih_Human
0
Frih_Mouse Frih_Rat Frih_Chick Fril_Human
TTASTSQVRQ NYHQDSEAAI NRQINLELYA SYVYLSMSYY FDRDDVALKN .. ..P ........ ..A .................... ..C. .......... .. ..P ............................... ..C. .......... .ATPP ........ ..C .................................. .. ..S..I .. ..ST.V ...V .SLV..Y.Q. ..T...LGF. ...... ..EG
Frih_Human Frih_Mouse Frih_Rat Frih_Chick Fril_Human
51 FAKYFLHQSH EEREHAEKLM KLQNQRGGRI FLQDIKKPDC ....................................... . ....................................... . ....................................... . VSHF.RELAE .K..GY.R .L .M..... ..A LF......A E
Frih_Human Frih_Mouse Frih_Rat Frih_Chick Fril_Human
101 ECALHLEKNV NQSLLELHKL ATDKNDPHLC DFIETHYLNE QVKAIKELGD ...... ..S. ............................... ..S ...... ..S ...... R ..... ..S. ............................... E ................... ..E ............. ..D. .... ..Q ... KA.MA...K L ..A..D..A. GSART ..... ..L...F.D. E..L..K.M..
Frih_Human Frih_Mouse Frih_Rat Frih_Chick Fril_Human
151 HVTNLRKMGA PESGLAEYLF DKHTLGDSDN ES ........................ ..HG .E S. ............ ..M ......... ..HG .E S. .......... .KY.M ......... ..E ..S .. .L...H.L .G ..A..G .... ERL..KHD ....
M
. .
DDWESGLNAM .......... .......... ..... ..T .. .E.GKTPD ..
Fig.3.Multiple alignment ofhuman.mouse.rat, chicken H-ferritin and human L-ferritin. Swissprot data bank codes identify the sequences. Only positions human H-ferritin are shown. Amino acids in bold indicate positions our modelled complex.
alignment). H-ferritin was able to compete the interaction of all the 11 selected phage clones with mAb H 107, indicating that binding occurred at the level of the antigenbinding site. The attempt of mapping the conformational epitope of ferritin was facilitated by the observation that a set of the selected clones contained a tryptophan and that a unique tryptophan was present in the linear sequence of H-ferritin. To definitely prove this point, the authors subsequently proceeded to the site-directed mutagenesis of residues surrounding Trp 93 in the threedimensional structure of H-ferritin. They purified and probed two groups of H-ferritin mutants with mAb H107. The first group involved mutations surrounding Trp 93 in the primary structure of H-ferritin (D91V. D92N, E94G. S95K), the second included residues neighbouring the tryptophan in the three-dimensional structure (F411, Y40F, F41 S, Y40G, Y4OC, D45A) (see Figs 3 and 4). Mutants in the first region abolished binding; significant reduction of the binding was observed for the mutants of the second group. These results strongly support the hypothesis that these two regions are part of the discontinuous epitope identified through phage library experiments and are compatible with our results (Table 4). Phenylalanine 41 was not directly involved in antigen-antibody contacts in our model. However, its
different from the sequence in contact with the antibody
The of in
mutation caused a significant reduction of binding. Since residues surrounding this position (Table 3) are in contact with the antibody, we propose that the F411 mutation can affect the conformation of the adjacent interface residues.
Fig. 4. Ribbon representation of the H-ferritin monomer. Only the side chains of residues F41 and W93 are shown to illustrate their proximity in the three-dimensional structure of the protein.
Antibody-antigen
Table 4. Comparison
interaction
1009
of experimental and theoretical results on the ferritin-H107
complex
Mutant ‘I
Source
Binding h
Interface contacts ‘
D91V, D92N E94G, S95K F411 Y40F. F41S Y40G Y4OC
Mut Mut Mut Mut Mut Mut Mut Mouse
No No No No No No No No
Yes Yes No Yes Yes Yes Yes Yes
Rat
No
Yes
Chicken
No
Yes
D45A
T5P, Y38C, C90R, N109S, A144S, D177H, S178G, N180E, E181S T5P, Y38C, C90R, ElOlR, N109S, A144S, L165M, D177H, S178G. N180E, E181S T2A, A3T. S4P, T5P, C90R, N98T, N139D, E147Q, El62K, S163Y, L165M. D177E, Nl80S
Both the sequence variations in site-specific and natural isoforms are included. Bold case indicates residues predicted to be in contact in our modelled complex. *Adapted from Luzzago et al. (1986) and Levi et al. (1987). ‘Interface residues have their surface within 2.5 A from the surface of at least one Ab atom.
DISCUSSION The experimental results reported in the literature (Luzzago et al., 1993) and summarized above are con-
sistent with our theoretical prediction of the antibodyantigen complex structure (solution l), in that they identify the very same amino acids that are involved in the formation of our predicted complex. Tyrosine 40 and Trp 93, identified as discontinuous epitopes in the phage library experiment, are both involved in the formation of the solution 1 complex; aspartic acids 45, 91 and 92, whose mutagenesis resulted in the complete loss of affinity between antigen and antibody, seem also to play a major role in the stability of solution 1 complex (Table 3). Other considerations also support our prediction: monoclonal antibody H107 has been raised in mouse (Luzzago et al., 1986) and is specific for human H-ferritin. It does not crossreact with mouse, rat or chicken Hferritin nor with human L-ferritin. The multiple sequence alignment of human, rat, mouse and chicken H-ferritins and human L-ferritin (Fig. 2) shows that the amino acids corresponding to the proposed regions of interaction differ in the aligned sequences, supporting the proposed solution. Together, the experimental data and the last considerations discussed strongly support the proposed solution, which can then be used as starting point for further experiments. Our data in fact suggest that a number of other residues (Table 3) are involved in the antibodyantigen interaction and new mutations can be proposed to test both the reliability of our proposed solution and our knowledge and understanding of antigen-antibody recognition rules. In particular, lysine 87 and asparagine 98 seem to be good candidates to test the model system and experiments are in progress to test their involvement in the interaction.
CONCLUSIONS The mechanism of protein-protein recognition involves a series of interactions such as shape complementarity, electrostatic interactions, hydrogen bonding etc. However, shape complementarity seems to play a major role in molecular recognition and can be used as central criterion for the proposal of solutions to the protein docking problem. The results presented here indicate that the sequence of an antibody and the structure of its antigen might be sufficient to direct the experimental work in order to precisely map the regions of interaction. The particular example used here is even more significant in that the identified epitope is conformational and, had it not been for the presence of a unique tryptophan residue in the interaction region, it would have been difficult to identify the regions to mutagenize in order to confirm the results of the epitope library. We believe that such a method can be useful to direct experimental work on antigen-antibody complexes, and we are testing a modified version of our algorithm, which should be more suitable to detect complexes between models of antibodies and homology-built models of antigens. We believe that the applications of methods such as that described here can also open the road to attractive theoretical speculations. Not much is known about the dynamics of the formation of an antibody-antigen complex, but the observation that all our proposed solutions have some contacts in common and that solutions 1 and 2 are quite similar to each other is interesting. We are currently investigating whether such a situation is commonly observed in other antigenantibody complexes.
1010
M. HELMER-CITTERICH
rt u/
GCG Package. Version 7. April 1991, 575 Science Drive, Madison. Wisconsin. U.S.A. 5371 I. Helmer-Citterich M. and Tramontano A. (1994) PUZZLE: a new method for automated protein docking based on surface shape complementarity. J. molec. Biol. 235, 1021-103 I. Hoogenboom H. R.. Griffiths A. D., Johnson K. S.. Chiswell D. J., Hudson P. and Winter G. (1991) Multi-subunit proteins on the surface of filamentous phage: methodologies for displaying antibody (Fab) heavy and light chains. Nucleic Acid.7 REFERENCES Res. 19, 41334137. Alzari P. M., Spinelli S.. Mariuzza R. A., Boulot G.. Poljiak R. Lascombe M. B.. Alzari P. M., Poljak R. J. and Nisonoff A. J.. Jarvis J. M. and Milstein C. (1990) Three-dimensional ( 1989) Three-dimensional structure of two crystal forms of 2 structure determination of an anti-?-phenyloxazolone antiFab R19.9 from a monoclonal anti-arsonate antibody. Proc,. body: the role of somatic mutation and heavy/light chain natrz. .4cud. Sci. Ci.S. A. 86, 607-612. pairing in the maturation of an immune response. Eur. rnolc~c,. Lawson D. M.. Artymiuk P. J., Yewdall S. J., Smith J. M. A.. Biol. or~9. J. 9, 3807-38 14. Livingstone J. C., Treffry A., Luzzago A.. Levi S., Cesareni Arosio P., Adelman T. G. and Drysdale J. W. (1978) On ferritin G., Thomas C. D., Shaw W. V. and Harrison P. M. ( 1991) heterogeneity: further evidence for heteropolymers. J. bin/. Solving the structure of human H-ferritin by genetically C/let??. 253,4451-4458. engineering intermolecular crystal contacts. Nature 349,541Bruccoleri R. E., Haber E. and Novotny J. (1988) Structure 544. of antibody loops reproduced by a conformational search Lesk A. M. (1986) Integrated access to sequence and structural algorithm. Nature 335, 564-568. data. In Biosciences: Perspectives and User Services in Europe. Chothia C.. Novotny J., Bruccoleri R. and Karplus M. (1985) (Edited by C. Saccone), pp. 23-28. Brussels EEC. Domain association in immunoglobulin molecules. J. m&c,. Lesk A. M. and Tramontano A. (1993) An atlas of antibody Biol. 186, 755-758. combining sites. In Structures qf’ Antigens (Edited by M. H. Chothia C. and Lesk A. M. (1987) Canonical structures for the V. Van Regenmortel). Vol. 2, pp. l-29. CRC Press, Oxford. hypervariable regions of immunoglobulins. .1. r)r&c~. Biol. Levi S.. Cesareni G.. Arosio P., Lorenzetti R., Soria M.. Sol196,901-917. lazzo M.. Albertini A. and Cortese R. ( 1987) CharacChothia C.. Lesk A. M.. Tramontano A.. Levitt M., Smith-Gill terization of human ferritin H chain synthesized in S. J., Air G.. Sheriff S.. Padlan E. A., Davies D., Tulip W. Escherichia coli. Gene 51, 269-274. R.. Colman P. M.. Spinelli S.. Alzari P. M. and Poljak R. Luzzago A.. Arosio P.. Iacobello C., Ruggeri G., Capucci L., J. (I 989) Conformations of immunoglobulin hypervariable Brocchi E., De Simone F.. Gamba D.. Gabri E., Levi S. regions. Nature 342, 877-883. and Albertini A. ( 1986) Immunochemical characterization of Clackson T.. Hoogenboom H. R., Griffiths A. D. and Winter human liver and heart ferritins with monoclonal antibodies. G. (1991) Making antibody fragments using phage display Biochim. hiophw. Actu 872, 61-71. libraries. Nature 352, 624-628. Luzzago A.. Felici F.. Tramontano A., Pessi A. and Cortese Connolly M. L. (1983) Analytical molecular surface calculation. R. (1993) Mimicking of discontinuous epitopes by phageJ. appl. Crystallogr. 16, 548-558. displayed peptides: I. Epitope mapping of human H-ferritin Dayringer H. E., Tramontano A.. Sprang S. R. and Fletterick using a phage library of constrained peptides. Gelze 128, 51R. J. (1986) Interactive program for visualization and mod57 elling of proteins, nucleic acids and small molecules. J. moles,. Martin A. C. R.. Cheetam J. and Rees A. R. (1989) Modeling Graphics 4, 82-87. antibody hypervariable loops: a combined algorithm. Proc. Felici F.. Castagnoli L., Musacchio A., Jappelli R. and Cesareni natn. Acad. Sci. U.S.A. 86, 9628-9672. Sheriff S., Silverton E. W., Padlan E. A., Cohen G. H., SmithG. (1991) Selection of antibody ligands from a large library Gill S. J., Finzel B. C. and Davies D R. (1987) Three-dimenof oligopeptides expressed on a multivalent exposition vector. sional structure of an antibody-antigen complex. Proc. natn. J. moler. Biol. 222, 301-310. Acad. Sci. U.S.A. 84, 8075-8079. Ford G.. Harrison P. M.. Rice D. W., Smith J. M. A.. Treffry Tramontano A. and Lesk A. M. (1992) Common features of the A., White J. L. and Yariv J. (1984) Ferritin: design and conformations of antigen-binding loops in immunoglobulins formation of an iron storage molecule. Phil. Truns. R. Sot,. and application to modeling loop conformations. Proteins: Lond. 304, 551-565. Structure. Function and Genetics. 13, 23 l-245. Genetics Computer Group (1991) Program Manual for the Acknowledgement.~-M. H. C and E. R. contributed equally to this work. The authors are grateful to Dr Andrew Wallace and Profs R. Cortese and G. Cesareni for useful advice. We acknowledge the expert computer support of F. Pagone and C. Moroni at IRBM and of R. de Chiara at the University of Rome. This work was supported by the E.C. BRIDGE Project.
ImmunologyVol. 32, No. 13. pp. 1001-1010,
Pergamon 0161-5890(95)00027-5
1995 Copyright ‘0 1995 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0161-5890/95 $9.50 + 0.00
MODELLING ANTIBODY-ANTIGEN INTERACTIONS: FERRITIN AS A CASE STUDY MANUELA
HELMER-CITTERICH,*§
ALESSANDRA
ERMANNA
and
LUZZAGOf
ANNA
ROVIDA,?
TRAMONTANOS
*Department of Biology, University of Rome Tor Vergata, Via della Ricerca Scientifica, 00 133 Roma, Italy; TIstituto di Tecnologie Biomediche Avanzate XNR, Via Ampere 56, Milano, Italy; and $IRBM P. Angeletti, Via Pontina Km 30.600, 00040 Pomezia, Roma, Italy (First received 20 October 1994; accepted in revised.form 16 February 1995) Abstract-In this work, we propose a model for the structure of the antigen-antibody complex formed by human H-ferritin and an antibody that specifically recognizes it. We cloned and sequenced the antibody gene, predicted the antibody three-dimensional structure, and reconstructed the Hferritin-antibody complex using an automated docking procedure previously validated on known complexes. This procedure allowed us to identify one putative complex which we carefully analysed, in order both to evaluate its likelihood, in light of a set of experimental results described in the literature, and to predict precisely which are the sites of interaction between the two molecules. Our model is compatible with the experimentally determined characteristics of the complex. Some of the residues that form the predicted antigenic site of ferritin can be found in the amino acid sequence of peptides selected from a random peptide library because of their affinity for the ferritin monoclonal antibody. Furthermore, the structural difference between the antigenic site in human H-ferritin and the corresponding region in other species permits us to rationalize the inability of the antibody to recognize human L-ferritin and rat, chicken and mouse H-ferritin. Through the analysis of our model complex, we identify a number of other residues putatively involved in the interaction. This multidisciplinary approach shows that synergy between computational and experimental methods may bring further insight into the understanding of antibody-antigen recognition rules. Key words: antigen recognition,
antibody prediction. protein docking, ferritin.
INTRODUCTION
The problem of mapping the interaction surface between two macromolecules that bind each other is of outstanding interest in molecular immunology. The problem can be stated as follows: given the structure (or a model of the structure) of an antigen and the structure (or a model of the structure) of an antibody raised against the antigen, can we predict the site of interaction between the two molecules? The solution to this problem, even if approximate, would help in designing experiments to precisely map the residues involved in the interface and could be instrumental both in designing peptides able to mimic the interacting surface of the antigen and in understanding where important regions of an antigen are located in its three-dimensional structure. We have previously shown that an algorithm based on surface complementarity can be used to automatically dock two proteins of known structures or one protein of known structure and a model of the other (HelmerCitterich and Tramontano, 1994). The procedure allows
§Author to whom correspondence Abbreviations: mAb, monoclonal complementarity.
should be addressed. antibody; lsc, low surface
one to obtain a limited number of relative orientations of the two proteins (between one and four) among which the correct orientation, when known from crystallographic studies, is always found. We have also shown that it is possible to build a reliable three-dimensional model of immunoglobulins taking advantage of the “canonical structure model” of hypervariable loops and of the high conservation of the framework (Chothia et al., 1989).
Here we apply our docking procedure to a complex of unknown structure, where a wealth of experimental data is available. Furthermore, the availability of the reagents should facilitate the experimental test of the predictions of our model. The chosen system is that of human H-ferritin and a monoclonal antibody, mAb H107, raised against the molecule. Ferritin is a protein involved in iron storage and is expressed as a 24-mer consisting of a mixture of two homologous subunits, H and L (Arosio et al., 1978). The two subunits share 55% sequence homology and fold into a similar structure of four tightly packed alphahelices (Ford et al., 1984; Lawson et al., 1991). The H chain of human ferritin has been overproduced in E. coli (Levi et al., 1987), and the crystallographic structure of
1001
1002
M. HELMER-CITTERICH
H-ferritin 24-mer has been obtained at high resolution (Lawson et al.. 1991). A monoclonal antibody against human ferritin (H 107) has been raised in mouse and here we report the cloning of its gene and the determination of its nucleotide sequence. This antibody reacts with homo multimers of human H-ferritin synthesized and assembled in E. coli, does not react with human L-ferritin, mouse and rat H-ferritin (Luzzago et al., 1986). The results that we obtained by docking a model of the antibody to the H-ferritin structure are in excellent agreement with experimental results coming from phage library techniques and site-directed mutagenesis and furthermore, adds a substantial amount of interesting structural information on the putative complex formed.
MATERIALS
AND METHODS
Clonirrg afzd sequefrciny qf‘HlO7 cDNA synthesis and VH and VK amplification and assembly were performed essentially as described in Clackson et al. (1991). The assembled product was reamplified with the primers VK4FOR-Not1 (Clackson et a/.. 1991) and VHlBACKSFIIS (Hoogenboom et al., 1991), for cloning into the expression vector pHEN 1. The resulting clone was sequenced with Sequenase kit (U.S. Biochemicals, Cleveland, OH), with primers: LMB3 (5’CAGGAAACAGCTATGAC-3’). MO-LINK-BACK (5’-GGGACCACGGTCACCGTCTCCTCA-3’) MOLINK-FOR (S-TGGAGACTGGGTGAGCTCAATGTC-3’) and 0L6 (5’-CCCTCATAGTTAGCGTAACG3’). HI07 single chain Fv fragment was expressed as soluble form in E. coli HB2 15 I cells (Hoogenboom et al.. 1991) and its specificity tested by dot blot (data not shown). The deduced amino acid sequence of the light and heavy chain variable regions of mAb HI 07 is reported in Fig. 1. Antiho&
modelliny
The model of the H 107 monoclonal antibody was constructed using the automatic procedure described by Lesk and Tramontano (1993). using the database of known immunoglobulin structures. We aligned the sequences of VL and VH chains of the mAb with the sequences of the corresponding domains of the known immunoglobulin structures. The alignments were performed using the program GAP from the GCG package (Genetics Computer Group, 1991). For each domain. we selected as parent structure the immunoglobulin domain with the highest sequence identity. namely HyHEL-5 (Data Bank code 2HFL, Sheriff et al., 1987) for Vi_ and R19.9 (Data Bank Code 2F19, Lascombe et al., 1989) for VH. The alignments used for model building are shown in Fig. I. We superimposed the selected models of the frameworks by the least-squares fit method of the structurally conserved regions identified by Chothia et al. ( 1985) and merged the V, domain of HvHEL-5 and the V,I. domain ,
et ml.
of R19.9 in the resulting relative position. This is justified by the observation that, if one superimposes the conserved regions of the VL domains of any pair of immunoglobulins of known structure, the R.M.S. distance between the conserved regions of their V,, domains is always lower than 1 A (data not shown). We next analysed the canonical structures for each hypervariable loop. The Ll, L2, Hl and H2 loop main chain conformations were left unchanged with respect to their framework templates since they have the same canonical structure. The canonical structure of the HI07 L3 loop is different from that of the parent HyHEL-5 structure. This loop in H 107 is six residues long and is characterized by the presence of a proline in position 95 (in the cis conformation in all known immunoglobulins with this canonical structure) and by a group in position 90 able to hydrogen bond to main chain atoms of the loop. NQlO (Alzdri et al., 1990) is the structure with the highest sequence identity and the same canonical structure which was then selected as a template (Fig. 1). No canonical structure has been identified for the H3 hypervariable loop so far. This loop is the most variable in length, sequence and conformation. It is 15 residues long in the parent R19.9 and 11 in H107. Long H3 loops differ most at their tip (A. T., unpublished observation) so we reasoned that modelling this loop using R19.9 as a template and deleting the central four residues would produce a reasonable structure at least in the initial and final region of the loop, while reducing the unavoidable errors introduced by splicing into a framework a loop from a different structure. Furthermore. database loop searching techniques have been shown not to be reliable in the case of antibody loops (Tramontano and Lesk, 1992) and combined database searching and conformational energy calculations (Bruccoleri et al., 1988; Martin et al., 1989) are very CPU intensive and not yet very reliable for loops longer than eight residues. In the case of L3. where the canonical structure selected for the framework is different from that selected for the loop. we used the weighted least-squares fit of the main chain atoms of the four residues preceding and following the loop described by Tramontano and Lesk (1992) and grafted the loop into the model. The conformations of the side chains were modelled as follows: where the parent structure and the model had the same residue, we retained the conformation of the parent structures. If the side chain differed. we took the side chain conformation (if possible) from another immunoglobulin having the same residue in the corresponding position. When more than one structure was a candidate for this step. that more similar in sequence to H107 was selected. In all other cases, we used the most common conformer for the amino acid side chain. The model was subjected to 100 cycles of energy refinement (with the default parameters of Discover. Biosym Technologies. San Diego, CA) to obtain a reasonable stereochemistry. Complex predictiofz The PUZZLE automated Citterich and Tramontano,
docking procedure (Helmer1994) is based only on geo-
Antibody-antigen
1003
interaction
A) * * ** 1 DIELTQSPAIMSASLGEEITLTCSASSSVSYIHWYQQKSDTSPKLLIYST II 1lIIIIIIIII II.:I:IIIIIIIl.I:.IIIIII:IIlI :Il.I 1 DIVLTQSPAIMSASPGEKVTMTCSA~~QQKSGTSPKRWIY~ l
H107 hyHEL- 5
H107
* * * 51 SSLASGVPSRFSGSGSGTFYSLTISSVEAEDAADYYCHQWSFYPWTFGGG I.IIlIII
HyHEL - 5 nql0
IIIIIIIII
I
IIIIIII:I.IIII:III:II:
IIIII
51 SKLASGVPVRFSGSGSGTSYSLTISSMETEDAAEYYCQQFP,TFGGG
H107
101
HyHEL-5
100
TKLEIK Il1llI TKLEIK
B) H107 R19.9
* ** * 1 DVQLQESGTELVKPGASVKLSCKASGYTFTSYWMHWVKQRPGQGLEWIGE : IIII:II.III:: I.lII:IIIIIIIIIIII ::llllllllllllll 1 QVQLQQSGAELVRAGSSVKMSCKAS GYTFTSXGVNWVKQRPGQGLEWIGY
H107
* 51 IYPSNGATNYNEKFKSKATLTVDKSSNTASMQLSSLTSEDSAVFFCASSG
R19.9
51 INaYL&‘NEKFKGKTTLTVDRSSSTAYMQLRSLTSEDAAVYFCAw
I
I:.l
IIIIII:I.IIIII:II.II
IIl.IIIIII.II:III.I
H107
101
AYY.. ..GNPFTYWGQGTTVTVSS.
R19.9
101 &LAVYYA&WGQGTTLTVSS
IIIIII:llll
Fig. 1. Alignment of the V, (A) and VH (B) chain sequences of mAb H107 with the corresponding sequences of HyHEL-5 and RL9.9, respectively. The sequence ofthe L3 loop of NQIO is also reported. Asterisks indicate positions mainly responsible for the canonical structures of the hypervariable loops (underlined).
metric criteria applied on the solvent accessible surface of the two proteins (Connolly, 1983). Each one of the two proteins is projected onto a cylinder, cut in slices orthogonal to the cylinder axis and opened onto a grid.
The grid is equivalent to a matrix, in which each protein slice is described as a series of values in a row, Each protei! slice (2 A thick) is described as a polygon with sides 2 A in length. All vertices of the polygon coincide with a protein surface point. Each element of a matrix row is calculated as the difference between adjacent radial coordinates of the polygon vertices. Each position in the matrix row describes a fixed area of the protein surface with a number which is related to the protein surface shape in that slice. The search for protein surface complementarity becomes a search for submatrices which describe regions of surface complementarity. A complete search in the space of all possible protein/protein orientations can be made by rotating one of the two proteins, projecting its surface onto the un-
rotated cylinder and applying the comparison algorithm to the new matrix obtained. Firstly, a coarse sampling of the rotation space at 10” steps is performed; regions around the obtained solutions are subsequently scanned at 10 steps. Lacking any a priori knowledge about how many monomers are involved in the 24-met/antibody interaction, one should in principle, use the unique symmetry element of ferritin, that is the heptamer. However, this choice would substantially increase the CPU time required by the procedure. Consequently, we decided to try and use the ferritin monomer as a docking probe and look for relatively small regions of surface complementarity with the target antibody. Since, as we show later, this allowed us to detect contacts between the antibody and more than one monomer, we concluded that this faster approach is indeed adequate. We allowed the listing of areas of surface complementarity lower than the threshold used in our
M. HELMER-CITTERICH
1004
Table 1. List of solutions
test runs. The procedure then became able to construct more than just a few solutions, even with a relatively low area of surface complementarity (> - 100 AZ/protein). We will call these solutions low surface complementarity (kc) solutions. After the coarse run, the solutions were clustered in groups whenever they differed less than 5 A translation and 20” rotation and carefully analysed at the graphics. Solutions not discarded at this stage (see Results) were subjected to the fine run of the programme and finally subjected to energy refinement (1000 iterations of steepest descent algorithm with the default parameters of Discover). Structural
Rank Cl
RESULTS In order to address the problem of precisely identifying the surface of interaction between an antibody and its antigen we selected a well studied system as a case study, namely that of human H-ferritin and of the monoclonal antibody H107. We cloned and sequenced the anti-ferritin H 107 monoclonal antibody gene. modelled its three-dimensional structure (Tramontano and Lesk, 1992) and used the automatic docking method described above (HelmerCitterich and Trdmontano, 1994) to predict the antibody-antigen complex structure. The search for complementarity between the ferritin H monomer and the monoclonal antibody H 107 ended up with 20 Isc solutions, which were carefully analysed at the graphics in order to discard the matches proposed if they were not compatible with the overall 24-mer structure or with an antibody-antigen interaction involving the antigen binding site of the antibody. The results are reported in Table 1. Only eight solutions out of 20 correctly positioned the antibody outside the 24-mer cavity and were compatible with the overall 24-mer structure. These solutions were compared, clustered in three groups (Table 1) and ranked by PUZZLE according to a scoring system based on the dimension of the matching submatrices. The member of the cluster with the largest buried area then became representative of its cluster and underwent the fine run analysis, which resulted in three solutions listed in Table 2 where the value of buried surface area upon complex formation is also shown. Solutions 1 and 2 are relatively similar to each other
Group ‘
Notes ”
I
I
2 3 4 5 6
A
7
A B
II D I
I
8 9 IO II 12 13 I4 15 I6 17 18 19 20
analysis qf’the putative compleses
The proposed complexes were analysed by measuring their buried surface area upon complex formation, the number of contacts and the number of hydrogen bonds between the antibody and the antigen. Buried surface area upon complex formation was measured using the program Insight (Dayringer et al., 1986) and the values reported always refer to the H-ferritin 24mer buried surface area upon complex formation. All other analyses were performed using the program PINQ (Lesk, 1986).
et (11.
I III D 1 B A II C B D D A
“The solutions are ranked here on the basis of the approximate complementary surface evaluation performed automatically by the program. *Discarded solutions are indicated by a letter in this column: A. antibody positioned inside the 24-mer cavity; B, antibody partially positioned inside the 24-mer cavity; C. part of the antibody colliding with a monomer other than the one used as probe protein; D. contacts not involving the hypervariable loops of the antibody. ‘Solutions are subsequently clustered in groups of similarity (indicated by roman numerals).
(3.7 A and 14.4’ of translation and rotation, respectively), while solution 3 substantially differs from them (Table 2). Interestingly, all three of them involve the interaction with another ferritin monomer (no experimental data are available on this point). It is worth noting that the surface of contact of each of these solutions with the second monomer is below the threshold used in the docking procedure, and for this reason they do not appear in the final list of solutions. The three solutions have been analysed for antibody-antigen contacts and potential hydrogen bonding (Table 3). Solution 1 shows the more extended buried surface area upon formation of the dimer-antibody complex (554 Table 2. List of accepted
Solution Solution Solution
1 2 3
solutions
Group
T (A) il
R ( ‘) ”
A (A’) *
II 1 III
3.7 22.0
14.4 34.0
554 283 224
“For each solution, T and R are the translations and rotations calculated with respect to solution 1. ‘A is the H-ferritin 24-mer buried surface area upon complex formation.
Antibody-antigen
interaction
1005
006
M. HELMER-CITTERICH
er rrl
Fig. 2. (a) Ribbon representation of the H-ferritin 24-mer complexed with the V,_ and Vn variable portion of the mAb H107. The ferritin dimer directly involved in binding according to our solution 1 is depicted in yellow (monomer 1) and purple (monomer 2). In blue and white, the remaining portion of H-ferritin and in cyan, the antibody moiety. (b) Solution 1 is shown here in more detail with the same colours as in (a). The side chains of the amino acids directly involved in binding are shown in ball and stick representation.
Antibody-antigen
interaction
1007
Table 3. List of close contacts and hydrogen bonds between the H-ferritin 24-mer and the antibody for each accepted solution Solution number Solution 1
CDR
FW
Ll L2 L3
H2
Monomer I”
Monomer II”
Total number of H-bonds
Y40, R43, D45, D91, D92 KS7 Y40, C90, D91, D92, W93, E94. N98 N109
R79, 180, F81
15
T5, R9. E17,
S66, G67, T68
Solution 2
R79 D89, C90, N98, C102, H105 D84, KS7
H3 Ll
F81, L82, Q83
F81, L82,
17
Q83,DW 486 Y29,486, K87, P88, D89, C90. D91 El16 K87
L3
H2
Solution 3
Ll
Y29, Q86, K87, D89
L3 Dl
Q83, D84
F81, 483,
6
D84, Q86 483, D84
D89, C90, D91
“Two residues are considered in close contact when the distance between any pair of their atoms is less than their van der Waals radii plus 0.6 A; amino acids involved in putative hydrogen bonding are in bold.
A2), 40 amino acid contacts have been detected between the antibody and the ferritin dimer (27 with monomer I and 13 with monomer II). Fifteen hydrogen bonds can be formed (10 with monomer I and five with monomer II). Five CDRs out of six are involved in the binding and each of them always forms two or more hydrogen bonds. Also, three residues of the framework contact amino acids of the antigen and possibly form hydrogen bonds with them. Visual inspection of the complex shows a good shape complementarity between the protein surfaces and no major steric hindrance is detected. Solution 2 has a considerably less extended buried surface area (283 A’); nevertheless, 23 contacts between amino acids of the complex are detected (12 with monomer I and 11 with monomer II). Seventeen hydrogen bonds can be formed to stabilize complex formation. Only three out of six hypervariable loops of the antibody are involved in antigen binding and this is in contrast with what is generally seen in other known antibodyantigen complexes (Chothia and Lesk, 1987). Solution 3 shows a less extended buried surface area upon formation of the putative complex (224 A’) and the analysis of all the parameters which can be measured suggests that it is not a good candidate as a stable complex. Eighteen amino acid contacts can be detected (11 with monomer I and seven with monomer II); the formation of only six hydrogen bonds can be proposed and only two CDRs out of six are involved in the for-
mation of the putative complex. Shape complementarity is very good, even if much less extended than what is observed in known antibody-antigen complexes. Taken together, the results of our analysis show that the solution to be preferred on the basis of the extension of surface complementarity, number of contacts, number of hydrogen bonds and number of antigen hypervariable loops involved in the interaction is indeed the first solution proposed by the automatic procedure. The structure of the proposed complex is shown in Fig. 2. According to the structure of the resulting complex, we propose that the complex epitope recognized by the H107 monoclonal antibody comprises residue 29, residues 40 through 46, 87 through 95,98 through 103, 105 through 109 in one monomer, and residues 5 through 9, 17 through 21 and 77 through 82 in the second monomer. By screening a nonapeptide random library inserted in the major coat protein of the filamentous phage fl (Felici et al., 1991), Luzzago et al. (1993) affinity-purified 11 independent clones with the H107 monoclonal antibody (Luzzago et al., 1986) and the deduced amino acid sequences were then arranged into two groups on the basis of their similarity. The first group was formed by three sequences, and characterized by the consensus Y,D,N,-X-X-X-X-W, the second cornprized eight sequences and had the consensus GSA2SI-S8-X-F,Y, (subscripts indicate the number of occurrences of the amino acid at that position, and dashes separate positions in the
M. HELMER-CITTERICH
1008
et ul. ++
1 Frih_Human
0
Frih_Mouse Frih_Rat Frih_Chick Fril_Human
TTASTSQVRQ NYHQDSEAAI NRQINLELYA SYVYLSMSYY FDRDDVALKN .. ..P ........ ..A .................... ..C. .......... .. ..P ............................... ..C. .......... .ATPP ........ ..C .................................. .. ..S..I .. ..ST.V ...V .SLV..Y.Q. ..T...LGF. ...... ..EG
Frih_Human Frih_Mouse Frih_Rat Frih_Chick Fril_Human
51 FAKYFLHQSH EEREHAEKLM KLQNQRGGRI FLQDIKKPDC ....................................... . ....................................... . ....................................... . VSHF.RELAE .K..GY.R .L .M..... ..A LF......A E
Frih_Human Frih_Mouse Frih_Rat Frih_Chick Fril_Human
101 ECALHLEKNV NQSLLELHKL ATDKNDPHLC DFIETHYLNE QVKAIKELGD ...... ..S. ............................... ..S ...... ..S ...... R ..... ..S. ............................... E ................... ..E ............. ..D. .... ..Q ... KA.MA...K L ..A..D..A. GSART ..... ..L...F.D. E..L..K.M..
Frih_Human Frih_Mouse Frih_Rat Frih_Chick Fril_Human
151 HVTNLRKMGA PESGLAEYLF DKHTLGDSDN ES ........................ ..HG .E S. ............ ..M ......... ..HG .E S. .......... .KY.M ......... ..E ..S .. .L...H.L .G ..A..G .... ERL..KHD ....
M
. .
DDWESGLNAM .......... .......... ..... ..T .. .E.GKTPD ..
Fig.3.Multiple alignment ofhuman.mouse.rat, chicken H-ferritin and human L-ferritin. Swissprot data bank codes identify the sequences. Only positions human H-ferritin are shown. Amino acids in bold indicate positions our modelled complex.
alignment). H-ferritin was able to compete the interaction of all the 11 selected phage clones with mAb H 107, indicating that binding occurred at the level of the antigenbinding site. The attempt of mapping the conformational epitope of ferritin was facilitated by the observation that a set of the selected clones contained a tryptophan and that a unique tryptophan was present in the linear sequence of H-ferritin. To definitely prove this point, the authors subsequently proceeded to the site-directed mutagenesis of residues surrounding Trp 93 in the threedimensional structure of H-ferritin. They purified and probed two groups of H-ferritin mutants with mAb H107. The first group involved mutations surrounding Trp 93 in the primary structure of H-ferritin (D91V. D92N, E94G. S95K), the second included residues neighbouring the tryptophan in the three-dimensional structure (F411, Y40F, F41 S, Y40G, Y4OC, D45A) (see Figs 3 and 4). Mutants in the first region abolished binding; significant reduction of the binding was observed for the mutants of the second group. These results strongly support the hypothesis that these two regions are part of the discontinuous epitope identified through phage library experiments and are compatible with our results (Table 4). Phenylalanine 41 was not directly involved in antigen-antibody contacts in our model. However, its
different from the sequence in contact with the antibody
The of in
mutation caused a significant reduction of binding. Since residues surrounding this position (Table 3) are in contact with the antibody, we propose that the F411 mutation can affect the conformation of the adjacent interface residues.
Fig. 4. Ribbon representation of the H-ferritin monomer. Only the side chains of residues F41 and W93 are shown to illustrate their proximity in the three-dimensional structure of the protein.
Antibody-antigen
Table 4. Comparison
interaction
1009
of experimental and theoretical results on the ferritin-H107
complex
Mutant ‘I
Source
Binding h
Interface contacts ‘
D91V, D92N E94G, S95K F411 Y40F. F41S Y40G Y4OC
Mut Mut Mut Mut Mut Mut Mut Mouse
No No No No No No No No
Yes Yes No Yes Yes Yes Yes Yes
Rat
No
Yes
Chicken
No
Yes
D45A
T5P, Y38C, C90R, N109S, A144S, D177H, S178G, N180E, E181S T5P, Y38C, C90R, ElOlR, N109S, A144S, L165M, D177H, S178G. N180E, E181S T2A, A3T. S4P, T5P, C90R, N98T, N139D, E147Q, El62K, S163Y, L165M. D177E, Nl80S
Both the sequence variations in site-specific and natural isoforms are included. Bold case indicates residues predicted to be in contact in our modelled complex. *Adapted from Luzzago et al. (1986) and Levi et al. (1987). ‘Interface residues have their surface within 2.5 A from the surface of at least one Ab atom.
DISCUSSION The experimental results reported in the literature (Luzzago et al., 1993) and summarized above are con-
sistent with our theoretical prediction of the antibodyantigen complex structure (solution l), in that they identify the very same amino acids that are involved in the formation of our predicted complex. Tyrosine 40 and Trp 93, identified as discontinuous epitopes in the phage library experiment, are both involved in the formation of the solution 1 complex; aspartic acids 45, 91 and 92, whose mutagenesis resulted in the complete loss of affinity between antigen and antibody, seem also to play a major role in the stability of solution 1 complex (Table 3). Other considerations also support our prediction: monoclonal antibody H107 has been raised in mouse (Luzzago et al., 1986) and is specific for human H-ferritin. It does not crossreact with mouse, rat or chicken Hferritin nor with human L-ferritin. The multiple sequence alignment of human, rat, mouse and chicken H-ferritins and human L-ferritin (Fig. 2) shows that the amino acids corresponding to the proposed regions of interaction differ in the aligned sequences, supporting the proposed solution. Together, the experimental data and the last considerations discussed strongly support the proposed solution, which can then be used as starting point for further experiments. Our data in fact suggest that a number of other residues (Table 3) are involved in the antibodyantigen interaction and new mutations can be proposed to test both the reliability of our proposed solution and our knowledge and understanding of antigen-antibody recognition rules. In particular, lysine 87 and asparagine 98 seem to be good candidates to test the model system and experiments are in progress to test their involvement in the interaction.
CONCLUSIONS The mechanism of protein-protein recognition involves a series of interactions such as shape complementarity, electrostatic interactions, hydrogen bonding etc. However, shape complementarity seems to play a major role in molecular recognition and can be used as central criterion for the proposal of solutions to the protein docking problem. The results presented here indicate that the sequence of an antibody and the structure of its antigen might be sufficient to direct the experimental work in order to precisely map the regions of interaction. The particular example used here is even more significant in that the identified epitope is conformational and, had it not been for the presence of a unique tryptophan residue in the interaction region, it would have been difficult to identify the regions to mutagenize in order to confirm the results of the epitope library. We believe that such a method can be useful to direct experimental work on antigen-antibody complexes, and we are testing a modified version of our algorithm, which should be more suitable to detect complexes between models of antibodies and homology-built models of antigens. We believe that the applications of methods such as that described here can also open the road to attractive theoretical speculations. Not much is known about the dynamics of the formation of an antibody-antigen complex, but the observation that all our proposed solutions have some contacts in common and that solutions 1 and 2 are quite similar to each other is interesting. We are currently investigating whether such a situation is commonly observed in other antigenantibody complexes.
1010
M. HELMER-CITTERICH
rt u/
GCG Package. Version 7. April 1991, 575 Science Drive, Madison. Wisconsin. U.S.A. 5371 I. Helmer-Citterich M. and Tramontano A. (1994) PUZZLE: a new method for automated protein docking based on surface shape complementarity. J. molec. Biol. 235, 1021-103 I. Hoogenboom H. R.. Griffiths A. D., Johnson K. S.. Chiswell D. J., Hudson P. and Winter G. (1991) Multi-subunit proteins on the surface of filamentous phage: methodologies for displaying antibody (Fab) heavy and light chains. Nucleic Acid.7 REFERENCES Res. 19, 41334137. Alzari P. M., Spinelli S.. Mariuzza R. A., Boulot G.. Poljiak R. Lascombe M. B.. Alzari P. M., Poljak R. J. and Nisonoff A. J.. Jarvis J. M. and Milstein C. (1990) Three-dimensional ( 1989) Three-dimensional structure of two crystal forms of 2 structure determination of an anti-?-phenyloxazolone antiFab R19.9 from a monoclonal anti-arsonate antibody. Proc,. body: the role of somatic mutation and heavy/light chain natrz. .4cud. Sci. Ci.S. A. 86, 607-612. pairing in the maturation of an immune response. Eur. rnolc~c,. Lawson D. M.. Artymiuk P. J., Yewdall S. J., Smith J. M. A.. Biol. or~9. J. 9, 3807-38 14. Livingstone J. C., Treffry A., Luzzago A.. Levi S., Cesareni Arosio P., Adelman T. G. and Drysdale J. W. (1978) On ferritin G., Thomas C. D., Shaw W. V. and Harrison P. M. ( 1991) heterogeneity: further evidence for heteropolymers. J. bin/. Solving the structure of human H-ferritin by genetically C/let??. 253,4451-4458. engineering intermolecular crystal contacts. Nature 349,541Bruccoleri R. E., Haber E. and Novotny J. (1988) Structure 544. of antibody loops reproduced by a conformational search Lesk A. M. (1986) Integrated access to sequence and structural algorithm. Nature 335, 564-568. data. In Biosciences: Perspectives and User Services in Europe. Chothia C.. Novotny J., Bruccoleri R. and Karplus M. (1985) (Edited by C. Saccone), pp. 23-28. Brussels EEC. Domain association in immunoglobulin molecules. J. m&c,. Lesk A. M. and Tramontano A. (1993) An atlas of antibody Biol. 186, 755-758. combining sites. In Structures qf’ Antigens (Edited by M. H. Chothia C. and Lesk A. M. (1987) Canonical structures for the V. Van Regenmortel). Vol. 2, pp. l-29. CRC Press, Oxford. hypervariable regions of immunoglobulins. .1. r)r&c~. Biol. Levi S.. Cesareni G.. Arosio P., Lorenzetti R., Soria M.. Sol196,901-917. lazzo M.. Albertini A. and Cortese R. ( 1987) CharacChothia C.. Lesk A. M.. Tramontano A.. Levitt M., Smith-Gill terization of human ferritin H chain synthesized in S. J., Air G.. Sheriff S.. Padlan E. A., Davies D., Tulip W. Escherichia coli. Gene 51, 269-274. R.. Colman P. M.. Spinelli S.. Alzari P. M. and Poljak R. Luzzago A.. Arosio P.. Iacobello C., Ruggeri G., Capucci L., J. (I 989) Conformations of immunoglobulin hypervariable Brocchi E., De Simone F.. Gamba D.. Gabri E., Levi S. regions. Nature 342, 877-883. and Albertini A. ( 1986) Immunochemical characterization of Clackson T.. Hoogenboom H. R., Griffiths A. D. and Winter human liver and heart ferritins with monoclonal antibodies. G. (1991) Making antibody fragments using phage display Biochim. hiophw. Actu 872, 61-71. libraries. Nature 352, 624-628. Luzzago A.. Felici F.. Tramontano A., Pessi A. and Cortese Connolly M. L. (1983) Analytical molecular surface calculation. R. (1993) Mimicking of discontinuous epitopes by phageJ. appl. Crystallogr. 16, 548-558. displayed peptides: I. Epitope mapping of human H-ferritin Dayringer H. E., Tramontano A.. Sprang S. R. and Fletterick using a phage library of constrained peptides. Gelze 128, 51R. J. (1986) Interactive program for visualization and mod57 elling of proteins, nucleic acids and small molecules. J. moles,. Martin A. C. R.. Cheetam J. and Rees A. R. (1989) Modeling Graphics 4, 82-87. antibody hypervariable loops: a combined algorithm. Proc. Felici F.. Castagnoli L., Musacchio A., Jappelli R. and Cesareni natn. Acad. Sci. U.S.A. 86, 9628-9672. Sheriff S., Silverton E. W., Padlan E. A., Cohen G. H., SmithG. (1991) Selection of antibody ligands from a large library Gill S. J., Finzel B. C. and Davies D R. (1987) Three-dimenof oligopeptides expressed on a multivalent exposition vector. sional structure of an antibody-antigen complex. Proc. natn. J. moler. Biol. 222, 301-310. Acad. Sci. U.S.A. 84, 8075-8079. Ford G.. Harrison P. M.. Rice D. W., Smith J. M. A.. Treffry Tramontano A. and Lesk A. M. (1992) Common features of the A., White J. L. and Yariv J. (1984) Ferritin: design and conformations of antigen-binding loops in immunoglobulins formation of an iron storage molecule. Phil. Truns. R. Sot,. and application to modeling loop conformations. Proteins: Lond. 304, 551-565. Structure. Function and Genetics. 13, 23 l-245. Genetics Computer Group (1991) Program Manual for the Acknowledgement.~-M. H. C and E. R. contributed equally to this work. The authors are grateful to Dr Andrew Wallace and Profs R. Cortese and G. Cesareni for useful advice. We acknowledge the expert computer support of F. Pagone and C. Moroni at IRBM and of R. de Chiara at the University of Rome. This work was supported by the E.C. BRIDGE Project.