1. Structural Biology and Molecular Biophysics
  2. Immunology and Inflammation
Download icon

Computationally-driven identification of antibody epitopes

  1. Casey K Hua
  2. Albert T Gacerez
  3. Charles L Sentman
  4. Margaret E Ackerman
  5. Yoonjoo Choi  Is a corresponding author
  6. Chris Bailey-Kellogg  Is a corresponding author
  1. Dartmouth College, United States
  2. Korea Advanced Institute of Science and Technology (KAIST), Republic of Korea
Research Article
Cite this article as: eLife 2017;6:e29023 doi: 10.7554/eLife.29023
6 figures, 8 tables and 6 additional files

Figures

Overview of computationally-driven epitope identification by EpiScope.

(A) Ab–Ag docking models are generated using computational docking methods. In the example, the green structure is the Ag human IL-18 (PDB ID: 2VXT:A), while the cartoons represent possible poses of the Ab (limited here to three for clarity). Full details including docking models and designs for this example are provided in a PyMol session (Supplementary file 1). (B) Ag variants containing a pre-defined numbers of mutations (here triple mutations, colored triangles) are generated for each docking model. (C) Variants are clustered with respect to spatial locations in the Ag, and a set of variants predicted to disrupt all of the docking models is selected. (D) Ag mutagenesis and Ag-Ab binding experiments are performed to identify which mutations result in loss of Ab recognition. (E) Examination of the disruptive variant(s) enables localization of the Ab epitope in terms of both mutated positions (pink balls) and consistent docking models, here with the model (light pink cartoon) quite similar to the actual crystal structure (dark pink cartoon).

https://doi.org/10.7554/eLife.29023.002
Figure 2 with 1 supplement
Small sets of designed Ag variants enable epitope localization for two different B7H6-targeting Abs.

(A–C). TZ47; (D-F) PB11. (A and D) Designed Ag variants, color-coded by triple mutation sets (Table 1). NKp30, a natural ligand for B7H6, is shown in grey ribbon. (B and E) Flow cytometry results from staining variant-expressing HEK cells with the relevant Ab, using NKp30-Ig as a positive control. Fluorescence is normalized to WT Ag-expressing cells. The dotted lines represent average background fluorescence measured from negative control Abs. Experiments were conducted in triplicate and error bars show the standard deviation. (C and F) Docking models (Ab cartoons of different colors) affected by the disruptive Ag variants (highlighted in red for TZ47 and green for PB11). Bar graphs depict the average (height) and standard deviation (error bars) of the MFI of 3 technical replicates, defined as the equivalent staining of a single batch of transfected cells repeated in three separate wells in the same experiment. One outlier value was excluded (PB11-staining of PB11-Ag1) where fewer than 1500 live cells were sampled and the raw MFI was two orders of magnitude higher than the other two replicates (1145.6 vs. 14.41 and 14.20).

https://doi.org/10.7554/eLife.29023.003
Figure 2—figure supplement 1
Chimeric variant (SD9) design confirmed the localization of the TZ47 epitope.

(A) Chimeric human-macaque B7H6 variant SD9 was designed based on sequence differences between the two species homologues at the region identified by EpiScope (colored green and black). (B) Histograms depicting binding of TZ47 (red), NKp30 (blue), and secondary anti-mouse fluorescent Abs (black) to chimeric human-macaque B7H6 variant (SD9) expressing cells (left) and wild-type RMA-B7H6 cells (right) are shown.

https://doi.org/10.7554/eLife.29023.004
Figure 3 with 2 supplements
A single set of Ag variants enables simultaneous localization of two different B7H6-targeting Abs.

(A) Designed Ag variants color-coded by triple-mutant design, with natural binding partner NKp30 in grey ribbon. (B) Flow cytometry results from staining variant-expressing HEK cells with the relevant Ab, using NKp30 as a positive control. Fluorescence was normalized to WT antigen-expressing cells. The dotted line represents average background fluorescence measured from negative control Abs. (C) Docking models (Ab cartoons of different colors) affected by disruptive Ag variants (highlighted in orange for TZ47 and magenta for PB11), for left: TZ47 and right: PB11. Bar graphs depict the average (height) and standard deviation (error bars) of the MFI of 3 technical replicates, defined as the equivalent staining of a single batch of transfected cells repeated in three separate wells in the same experiment. One replicate value was excluded where fewer than 1500 live cells were sampled from the well (one replicate of PB11-staining of MULTI-1) and the raw MFI was two orders of magnitude larger than the other two replicates (232.8 vs. 2.55 and 3.71).

https://doi.org/10.7554/eLife.29023.008
Figure 3—figure supplement 1
B7H6 and epitope localization for its binding antibodies.

The modeled loop region (Chain A 150–157; DQVGMKEN) is colored in green. The binding interface of a binding antibody (17B1.3) in PDB (4ZSO) is in red. The disruptive designs for TZ47 are in yellow and PB11 in blue. The epitopes of the Abs seem not to overlap.

https://doi.org/10.7554/eLife.29023.009
Figure 3—figure supplement 2
Docking models for TZ47 and PB11 substantially overlap.

Residues predicted to be part of the binding interface by ClusPro generated docking models are color-coded by those specific to one Ab (blue) or shared between both Abs (red).

https://doi.org/10.7554/eLife.29023.010
Figure 4 with 4 supplements
Retrospective validation demonstrates generality of efficiency and effectiveness in localizing epitopes.

(A) Over a test set of 33 diverse Ab-Ag pairs with co-crystal structures, the number of pairs in which at least one binding interface residue is included among the disruptive mutations in a set of 1–6 Ag triple-mutant variants. Ultimately, two pairs were missed when using Ag crystal structure and three pairs when using Ag homology models. (B) Violin plots of the number of Ag variants required to incorporate mutations predicted to disrupt all docking models.

https://doi.org/10.7554/eLife.29023.013
Figure 4—figure supplement 1
Ag size vs. number of Ag variants to cover docking models.

(A) Using Ag crystal structure; (B) using Ag homology model. There is a slight trend between number of surface residues on the Ag and number of variants needed to localize the epitope. The marks in the scatterplot indicate which variant sets included (open circles) or missed (solid triangles) epitope residues.

https://doi.org/10.7554/eLife.29023.014
Figure 4—figure supplement 2
Performance using size-matched sets of random triple-mutants instead of docking-based disruptive ones.

The bars show success rates for the individual targets (i.e., Ab:Ag pairs, identified by PDB ID) over 1000 runs selecting designs from 1000 different random triple mutants. Targets are sorted in terms of the size-matched number of selected designs and the number of Ag surface residues. Annotations indicate where docking-based disruptive designs failed. Violin plots depict the total number of successes over all the targets and all the runs, with dashed lines at the top indicating previously described EpiScope success rates based on docking disruption.

https://doi.org/10.7554/eLife.29023.015
Figure 4—figure supplement 3
Varying the number of mutations to include per design from 1 to 4 demonstrates that the most efficient epitope localization occurs at three mutations/design.

Violin plots depict: (A) the number of total designs needed to disrupt all docking models, (B) the number of designs overlapping true Ab epitope residues, (C) the percentage of designs overlapping true Ab epitope residues, (D) the number of docking models remaining after filtering for those overlapping disruptive designs, (E) the percentage of surface residues covered by filtered docking models, on which to focus further scanning efforts if desired, and (F) percentage of test cases for which at least one design contained a mutation overlapping the true Ab epitope.

https://doi.org/10.7554/eLife.29023.016
Figure 4—figure supplement 4
Effects of inter-mutation distances on epitope localization resolution and success rate.

(A) Average Cα distances of two and three mutation variants. The distributions are peaked at 11 ~ 15 Å. (B–D) Success rates and resolution, in terms of consistent docking models and their residue coverage, with varying distance thresholds. Each plan uses a single 1-mutation, 2-mutation, or 3-mutation variant optimized to cover docking models for a target. Bars show averages and standard deviation over the retrospective test set.

https://doi.org/10.7554/eLife.29023.017
Figure 5 with 3 supplements
A small set of Ag variants has the potential to simultaneously localize multiple Ab epitopes for a single Ag.

(A) Heat map of competitive binding data (Sela-Culang et al., 2014) for 12 antibodies directed against the vaccinia virus D8 protein, with the extent of cross-blocking ranging from 0.0 (white, no effect) to 1.0 (black, complete blocking). Colors in all panels refer to the four Ab groups identified by this competition assay (I: purple, II: blue, III: yellow, and IV: red). (B) Heat map of the overlap between ClusPro-generated docks for each pair of Abs, ranging from 60% (white) to 100% (black). (C) Heat map of the average Hausdorff distance between Ag variants designed for each Ab, ranging from 0 (identical mutation sites, black) to 12 (white). (D–F) Ag variants designed to disrupt one Ab from each group (I: JE11, II: CC7.1, III: EE11, IV: LA5) are represented as triangles. Four designs were sufficient to cover all docking models, and the designs overlapped all of the epitope groups. True epitopes are color coded by group on the surface of the antigen; epitopes in group II and III overlapped, and are colored in green. Design residues overlapping the true epitopes are indicated with circles. (E and F) Zoomed views of epitope faces.

https://doi.org/10.7554/eLife.29023.026
Figure 5—figure supplement 1
ClusPro generates docking models that cover nearly all surface residues for the set of 12 VACV anti-D8 envelope targeting Abs.

Bars show fraction of surface residues in contact with a docking model; 78% on average (dashed line).

https://doi.org/10.7554/eLife.29023.027
Figure 5—figure supplement 2
Identification of vaccinia D8 epitopes against LA5.

The LA5 Ab is in green and the D8 Ag is in magenta (PDB id 4ETQ). The modeled loop is in red. Five designs were generated by EpiScope and two of them are in the binding interface region.

https://doi.org/10.7554/eLife.29023.028
Figure 5—figure supplement 3
Heatmaps of sequence identity between the selected 7 VACV anti-D8 Envelope targeting Abs.

Different panels are restricted to different CDRs, with ‘All’ for average over all CDRs. Colored square outlines represent the 4 groups of antibodies (epitope bins) identified by competitive binding assays (Sela-Culang et al., 2014).

https://doi.org/10.7554/eLife.29023.029
Figure 6 with 2 supplements
Success of EpiScope and the quality of docking models.

In general, docking using Ag crystal structures is better than using Ag homology models according to the fnat value; it is above ‘medium’ for crystal structures but only ‘acceptable’ for model structures. Poor docking models are necessary, but not sufficient, for the failure of the EpiScope approach: EpiScope still identifies epitopes for some poorly docked models, but all failed cases have low fnat values.

https://doi.org/10.7554/eLife.29023.035
Figure 6—figure supplement 1
Examples of success and failure depending on qualities of Ab structure (A and B) and Ag structure (C and D).

Crystal structures are colored in blue and homology models in yellow. (A) In the case of 1FE8, though Ab modeling was highly successful (TM-score: 0.96 and backbone RMSD of CDR-H3: 0.63A), the docking models were very poor (fnat: 0.1) and epitope localization was not successful. The crystal structure of CDR-H3 is highlighted in red with stick representation. (B) Though modeling of Ab structure 2XQB was poor (TM-score: 0.92 and backbone RMSD of CDR-H3: 7.21A), moderate quality docking models were generated (fnat: 0.32) and the identification of epitopes was successful. (C) The Ag model of 4LVH was extremely accurate (TM-score: 0.9; template seq. ID: 41%) but poor modeling of the loop in the binding interface region (Ab structure is in ribbon) likely contributed in failure of epitope localization. (D) The receptor (4JZJ) was poorly modeled (TM-score: 0.33; template seq. ID: 32%) due to highly flexible loops, but the Ab binding region was modeled well and epitope identification succeeded.

https://doi.org/10.7554/eLife.29023.036
Figure 6—figure supplement 2
Two examples of EpiScope failure cases.

All ClusPro-generated docks (yellow) are shown with the crystal structures of Ab (cyan)-Ag complexes. (A) Failure due to poor docking quality (1H0D). (B and C) Failure due to insufficient mutational choice information (1OAZ). The Ag has a large long flexible loop involved in Ab binding (red in stick, panel B). Since the loop has no mutational information in closely related protein sequences (panel C), mutations that could disrupt binding are not considered in the design process.

https://doi.org/10.7554/eLife.29023.037

Tables

Table 1
Summary of mutations in EpiScope Ag designs for each Ab.

Designs that disrupted binding for each Ab are highlighted.

https://doi.org/10.7554/eLife.29023.007
DesignMutations
TZ47-Ag1F47Y, N49Q, W98E
TZ47-Ag2F184D, I188Q, V225T
TZ47-Ag3T71K, K74E, V76H
TZ47-Ag4M154E, N157G, S217H
PB-Ag1M30V, Q132V, Q136L
PB-Ag2F51H, Y52D, R99G
PB-Ag3A88T, F89T, G111R
PB-Ag4T176K, V194I, R231E
PB-Ag5N216K, S217A, Q219V
Table 2
Summary of mutations in Multi-Ab specific EpiScope Ag designs.

Designs that disrupted binding for each Ab are highlighted.

https://doi.org/10.7554/eLife.29023.012
DesignMutations
MULTI-1N57D, D84N, W98E (PB11)
MULTI-2F66Y, T71K, F72D
MULTI-3V78L, F89T, G111R
MULTI-4M154E, N157E, N216K (TZ47)
MULTI-5A172H, R231E, A233E
MULTI-6T176K, R231E, H236S
Table 3
Retrospective test cases.

Columns indicate the PDB ID of each Ab-Ag pair; the number of residues for various subsets of the Ag; the number and success of EpiScope designs based on crystal and model Ag structures; a measure of the quality of the closest native-like docking model among ClusPro generated models (fnat[Lensink et al., 2007]); the quality of the homology models built for Ab and Ags (TM-score [Zhang and Skolnick, 2004]); and the number of docking decoys generated by ClusPro.

https://doi.org/10.7554/eLife.29023.021
PDB codeNumber of residuesCrystal structureModel structureFnatTM-scoreNumber of docking decoys
WholeSurfaceEpitopesNumber of designsOverlap with epitopesNumber of designsOverlap with epitopesCrystalModelAntibodyAntigenCrystalModel
1FE8196124274Y3N0.10.040.960.843024
1FNS196120125Y2Y0.390.090.960.862620
1H0D12396143N2Y00.050.980.793030
1LK3160102263Y3Y0.730.440.970.742329
1OAZ123101142N2N0.10.10.970.773029
1OB19974133Y4Y0.620.610.970.853029
1RJL9582133Y2Y0.290.30.960.893027
1V7M163113206Y3Y0.450.190.960.782724
1YJD14086142Y3Y0.420.130.980.772530
2ARJ12390173Y3Y0.630.260.970.752430
2VXQ9671213Y3Y0.320.210.890.893030
2VXT157116193Y3Y0.830.130.950.931319
2XQB11487182Y3Y0.160.250.920.891721
3D9A12993193Y2Y0.090.630.930.932229
3HI1290246204N3N0.280.020.970.833029
3L5X1138386Y3Y0.180.240.980.833030
3MXW169108222Y2Y0.520.470.960.922023
3QWO5748103Y4Y0.320.530.960.863030
3RKD146105185Y3Y0.550.480.970.493030
4DN47650122Y2Y0.490.280.920.882027
4DW2222175204Y3Y0.10.350.920.853030
4ETQ226186224Y3Y0.450.570.960.943030
4G3Y157114123Y4Y0.350.120.940.863030
4G6J158109133Y3Y0.570.150.970.853030
4I3S190163234Y4Y0.050.090.810.823030
4JZJ252210184Y6Y0.310.250.950.333030
4KI518310871Y1Y0.390.290.950.932420
4L5F1117992Y3Y0.520.10.970.83030
4LVH223184135Y6N0.120.050.930.93029
4M6215510562Y2N0.120.110.920.812424
4NP4272230253Y5Y0.090.070.960.673030
4RGO226187173N6Y0.160.210.970.963030
5D96235198224Y4Y0.230.150.950.963030
Average162.88122.5216.483.303.180.330.240.950.8227.1227.67
STD57.5751.535.541.191.210.210.180.030.134.533.55
Table 4
Ab modeling quality.

Antibody structures were generally highly accurately predicted both overall (average TM-score: 0.95) and for CDRs (all-backbone-atom, including N, C, Cα and O, RMSDs reported). Overall, non-CDR-H3 loops were very well predicted based on the canonical rules, and even for CDR-H3 loops the average RMSDs was <2 Å.

https://doi.org/10.7554/eLife.29023.022
TargetSpeciesCDR-L1L2L3H1H2CDR-H3TM-score
RMSDSequenceLength
1FE8MOUSE0.420.220.741.010.510.63AGNYYGMDY90.96
1FNSMOUSE0.540.180.930.270.602.10VRDPADYGNYDYALDY160.96
1H0DMOUSE1.430.570.420.441.110.66TRLGDYGYAYTMDY140.98
1LK3RAT0.410.430.520.571.511.00TRGVPGNNWFPY120.97
1OAZMOUSE1.150.440.881.300.561.25ARMWYYGTYYFDY130.97
1OB1MOUSE0.580.310.630.420.631.97ARNYYRFDGGMDF130.97
1RJLMOUSE1.430.574.960.691.001.16ARMRYGDYYAMDN130.96
1V7MMOUSE0.700.260.830.651.100.59SGWSFLY70.96
1YJDMOUSE0.880.511.340.621.191.76TRSHYGLDWNFDV130.98
2ARJRAT0.710.671.120.460.700.65TPLIGSWYFDF110.97
2VXQHUMAN0.350.740.960.901.271.05ARLDGYTLDI100.89
2VXTMOUSE0.470.371.140.450.530.43ARGLRF60.95
2XQBHUMAN1.610.430.981.190.897.21ARDPAAWPLQQSLAWFDP180.92
3D9AMOUSE0.400.611.180.991.880.51ANWDGDY70.93
3HI1HUMAN0.800.860.830.610.441.25ARGPVPAVFYGDYRLDP170.97
3L5XHUMAN0.560.610.911.050.901.73ARMGSDYDVWFDY130.98
3MXWHUMAN0.580.710.711.090.820.96ARDWERGDFFDY120.96
3QWOHUMANIZED0.480.281.090.870.501.13ARDMIFNFYFDV120.96
3RKDMOUSE0.620.420.521.060.651.45ARIKSVITTGDYALDY160.97
4DN4HUMAN2.110.371.581.642.402.36ARYDGIYGELDF120.92
4DW2MOUSE1.200.434.120.851.143.18ERGELTYAMDY110.92
4ETQMOUSE1.070.291.590.350.910.94TRSNYRYDYFDV120.96
4G3YCHIMERIC0.680.710.570.900.981.22SRNYYGSTYDY110.94
4G6JHUMAN0.720.440.900.410.351.14ARDLRTGPFDY110.97
4I3SHUMAN1.340.460.644.331.083.49ARQKFYTGGQGWYFDL160.81
4JZJHUMAN0.590.540.980.841.042.96ARSHLLRASWFAY130.95
4KI5MOUSE0.740.520.782.140.441.49AREDDGLAS90.95
4L5FMOUSE0.760.421.030.490.941.83TKRINWALDY100.97
4LVHMOUSE1.630.712.821.462.701.91ARHGSPGYTLYAWDY150.93
4M62HUMAN2.080.791.402.552.788.26AREGTTGSGWLGKPIGAFAY200.92
4NP4HUMAN2.210.872.840.880.551.53ARRRNWGNAFDI120.96
4RGOMOUSE0.531.010.750.710.312.20VRDLYGDYVGRYAY140.97
5D96MOUSE0.740.570.530.620.893.43ASDSMDPGSFAY120.95
Average0.920.521.250.991.011.920.95
STD0.530.201.010.780.621.710.03
Table 5
The quality of Ag models and their template structures.

Failed cases are highlighted in red.

https://doi.org/10.7554/eLife.29023.024
TargetTemplateTemplate chainSeq. ID.TM-score
1FE83PPYA28.090.84
1FNS4IGIA24.730.86
1H0D3MWQA33.880.79
1LK34DOHA27.940.74
1OAZ2PUKC48.040.77
1OB11N1IA49.440.85
1RJL2FKJC62.110.89
1V7M1CN4C23.740.78
1YJD1AH1A30.700.77
2ARJ4XMNF26.260.75
2VXQ1N10A41.300.89
2VXT4XFSA94.230.93
2XQB2PSMA69.910.89
3D9A2EQLA49.220.93
3HI12BF1A33.940.83
3L5X3BPOA99.050.83
3MXW2IBGB70.000.92
3QWO1EDKA50.940.86
3RKD3RKCA88.190.49
4DN43FPUB41.670.88
4DW22ODQA25.940.85
4ETQ2ZNCA30.560.94
4G3Y1TNRA36.430.86
4G6J3NJ5A35.370.85
4I3S2B4CA61.960.82
4JZJ4RS1A31.970.33
4KI54QDRA44.970.93
4L5F2HG0A45.920.80
4LVH5BNYA40.890.90
4M624GQXA23.940.81
4NP42GJ6A35.860.67
4RGO5FKAC34.230.96
5D963G6OA80.770.96
Average46.130.82
STD21.060.13
Table 6
Success rates with epitopes defined according to IEDB or according to contacts in the binding interface.

Success is indicated as ‘T’ and failure as ‘F’. In test cases colored blue, EpiScope failed to find IEDB epitopes but did find binding interface residues.

https://doi.org/10.7554/eLife.29023.025
TargetCrystal structureModel structureTargetCrystal structureModel structure
IEDBInterfaceIEDBInterfaceIEDBInterfaceIEDBInterface
1FE8TTFF3QWOTTTT
1FNSTTTT3RKDTTTT
1H0DFFTT4DN4TTTT
1LK3TTTT4DW2TTTT
1OAZFFFF4ETQTTTT
1OB1TTTT4G3YTTTT
1RJLTTTT4G6JTTTT
1V7MTTTT4I3STTTT
1YJDTTTT4JZJTTTT
2ARJTTTT4KI5TTTT
2VXQTTTT4L5FTTTT
2VXTTTTT4LVHTTFT
2XQBTTTT4M62TTFT
3D9ATTTT4NP4TTTT
3HI1FFFF4RGOFTTT
3L5XTTTT5D96TTTT
3MXWTTTTTotal29 (88%)30 (91%)28 (85%)
Table 7
Comparison of residues predicted by PEASE for TZ47 and PB11 to mutations included in disruptive EpiScope designs.

Residue score cut-off 0.43 was used for PEASE.

https://doi.org/10.7554/eLife.29023.033
PatchPredicted patch residue positionsPatch scoreDisruptive EpiScope design mutation positions
TZ47-Patch 1158,159,160,161,1620.41154, 157, 217 (TZ47-Ag4)
TZ47-Patch2158,160,161,162,1630.4154, 157, 216 (MULTI-4)
TZ47-Patch31,29,30,31,320.4
TZ47-Patch41,2,30,31,1060.39
PB-Patch11,2,30,31,1060.4751, 52, 99 (PB-Ag2)
PB-Patch246,47,48,49,500.4157, 84, 98 (MULTI-1)
PB-Patch3158,160,161,162,1630.4
PB-Patch4195,196,197,198,2030.38
PB-Patch5123,124,125,126,1390.38
Table 8
Comparison of predictive components of PEASE and EpiScope on retrospective test set of 33 non-redundant Ab-Ag pairs.

The number of designs needed/considered indicates the number of designs generated by EpiScope to cover all ClusPro docking models. An equivalent number of the top ranked PEASE patch predictions are considered for each Ab. Coloring highlights the cases in which Episcope (green) or PEASE (red) succeeded where the other method failed. Grey coloring indicates cases in which both methods failed.

https://doi.org/10.7554/eLife.29023.034
TargetCrystal structure of agModeled structure of ag
# of Designs Needed/Considered# of EpiScopeDesigns Overlapping True Epitope# of PEASE patches Overlapping True Epitope# of Designs Needed/Considered# of EpiScopeDesigns Overlapping True Epitope# of PEASE patches Overlapping True Epitope
1FE8424300
1FNS525212
1H0D300210
1LK3310310
1OAZ202202
1OB1310424
1RJL313212
1 V7M610312
1YJD210310
2ARJ313313
2VXQ313313
2VXT313313
2XQB212313
3D9A310210
3HI1400300
3L5X616323
3MXW212212
3QWO323434
3RKD523312
4DN4210210
4DW2412311
4ETQ411311
4G3Y310423
4G6J312313
4I3S423412
4JZJ410640
4KI5110110
4 L5F210310
4LVH510602
4 M62210200
4 NP4310530
4RGO302613
5D96410410

Additional files

Source code 1

TINKER minimization key file.

https://doi.org/10.7554/eLife.29023.038
Source code 2

OSPREY configuration files.

https://doi.org/10.7554/eLife.29023.039
Supplementary file 1

PyMol session file for an example of 2VXT as shown in Figure 1.

https://doi.org/10.7554/eLife.29023.040
Supplementary file 2

PyMol session file for B7H6 binding disruptive designs against TZ47 and PB11.

https://doi.org/10.7554/eLife.29023.041
Supplementary file 3

Full sequences of TZ47, PB11, and all B7H6 variants.

https://doi.org/10.7554/eLife.29023.042
Transparent reporting form
https://doi.org/10.7554/eLife.29023.043

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)