Principles for computational design of binding antibodies

A database of natural antibody conformations and sequences.

A set of 924 antibodies from the structural antibody database (SAbDab) (34) was aligned to the Fv of antibody 4m5.3 [Protein Data Bank (PDB) ID code 1x9q]. The dihedral angles (ϕ, ψ, ω) from each aligned segment were extracted, modeled on PDB ID code 1x9q, and relaxed using the methods described by Lapidoth et al. (14). In the current study, only κ antibodies were built, although the methods were recently extended to λ-chains as well (35).

During the course of algorithm development, we changed the conformational segments’ boundaries. Specifically, in design cycles 1–4, each of the six CDRs constituted an independent conformation segment, and the framework was based on the coordinates in PDB ID code 1x9q. The segment boundaries in cycles 1–4 were [using the amino acid position numbers of Wu and Kabat (36)] as follows: L1, 23–34; L2, 45–54; L3, 87–100; H1, 27–35; H2, 45–57; and H3, 91–104. Boundaries in cycle 5 were as follows: vL, 1–88; L3, 89–105; vH, 7–92; and H3, 93–112.

In cycles 4–5, backbone conformations were clustered by rmsd with a cutoff of 1.0 Å, and for each conformation class, a sequence alignment and PSSM were computed, essentially as described by Lapidoth et al. (14).

Sequence constraints during design.

Similar to previous binder and enzyme design studies (6, 17), in design cycles 1–3, we disallowed design at positions in which the natural segments had Gly, Pro, or Cys. In cycle 2, we further disallowed Gly, Pro, Cys, or His to be introduced during design. In cycle 3, we added specific sequence constraints that mainly targeted positions in the antibody framework, where we noticed persistent design flaws, such as underpacking or buried but unsatisfied polar groups. For instance, the following restrictions were implemented (numbering in accordance with PDB ID code 1x9q): Phe/Tyr (position 90), Leu (position 91), Val (position 174), Phe (positions 178 and 180), and Tyr (position 233). In this design cycle, we additionally restricted core positions to apolar residue identities. For each position, we calculated the number of neighboring Cα atoms (<8 Å), excluding sequence neighbors (10 positions N- or C-terminally). Any position for which there were more than five such neighbors was considered core, and design was restricted in these positions to apolar residues.

With the implementation of conformation-specific sequence constraints using PSSMs in design cycle 4, all of the above sequence constraints were lifted. Instead, identities with PSSM cutoffs of −2 and 0 were used in the CDRs and framework, respectively.

A set of conformationally representative designed antibodies.

Using the conformation databases above, a set of representative antibody structures was designed (2,000–8,000 structures, depending on the design cycle). For each segment, a single representative conformation was arbitrarily selected from the conformation databases, and all representative conformations were recombined with one another to assemble backbone conformations of the complete Fv. The sequence of each antibody was then optimized using RosettaDesign.

Docking and fixed-backbone design.

We used as antigens human insulin and Mycobacterium tuberculosis ACP2 (PDB ID codes 2a3g and 2cgq, respectively). The structure of PDB ID code 2a3g was truncated on the termini of both chains after initial design runs strongly favored interactions with the terminal regions, and interaction with those would experimentally be hindered by the N-terminal biotinylation of the ligand. The structures were docked against the set of representative antibody backbones using the feature-matching algorithm PatchDock (37). Briefly, the parameters were set to prediction of antibody–ligand interaction, and poses that did not bind the antigen through CDRs H3 and L3 were eliminated.

For each representative antibody, the first 2,000 PatchDock entries were used as seeds for Rosetta simulations, in which all-atom docking, sequence design, backbone minimization, and combinatorial sidechain packing were carried out. At the end of this trajectory, the designed complexes were automatically filtered according to the following criteria: shape complementarity (Sc) > 0.5 (38), binding energy less than −10 Rosetta energy units (R.e.u.), and binding surface area > 1,100 Ų.

Stability and binding optimization through modular backbone design.

Designed antibodies that passed the previous filters were optimized through modular backbone design, essentially as described by Lapidoth et al. (14). Briefly, for each starting design, 150 moves were conducted, in each of which a backbone segment was chosen at random and replaced with another segment from the conformation databases. Following segment replacement, the sequence of the segment and every residue within 8 Å was designed, and iterations of backbone and side-chain minimization, side-chain packing, and rigid-body minimization at the binding interface were conducted.

Designs in cycles 1 and 2 showed poor packing between Fv light and heavy chains. Starting with design cycle 3, we therefore added rigid-body minimization between these two chains (in conjunction with rigid-body minimization of the Fv relative to the antigen).

In design cycles 3–5, at the end of each design move comprising backbone exchange, sequence design, packing, and minimization, the stability and binding energies were computed and evaluated through the “fuzzy”-logic design framework with an objective function O = B(Ab,L) ∧ S(Ab), where B and S are computed energies of binding and stability, respectively, Ab is the antibody, and L is the ligand (15). The new pose was accepted if it passed the Metropolis criterion with a gradually decreasing temperature, thereby effecting simulated-annealing Monte Carlo optimization (14).

At the end of the optimization trajectory, the structures were filtered by structure and energy criteria. Insulin-binding designs were filtered by Sc > 0.6 (38), binding energy less than −15 R.e.u., and binding surface area > 1,300 Ų. ACP-binding designs were filtered by stricter criteria reflecting the larger size of ACP: Sc > 0.6 (38), binding energy less than −18 R.e.u., and binding surface area > 1,400 Ų.

Manual selection and design improvement.

Designs were ranked according to Sc for the antigen (38) and visualized using PyMOL. Designs with unpaired charges or cavities at the binding interface or within the antibody core were eliminated, and the remaining designs were manually optimized using FoldIt (39). Briefly, in rounds 1–3, mutations were mainly introduced to optimize interactions between the framework and the CDRs, and structures of natural antibodies of similar conformation were often reviewed in making these design decisions. With the introduction of PSSM constraints in design cycle 4, such manually introduced mutations were no longer necessary. Additional mutations were introduced to improve charge complementarity with the antigen through long-range electrostatics.

High-performance computing using Google Exacycle.

Design cycles 1–3 produced few models that passed the structure and energy filters described above. With the introduction of conformation-dependent PSSM constraints in cycle 4 and the improved backbone segmentation of cycle 5, the productivity of the design algorithm improved dramatically. To address the low productivity of the algorithm in initial design cycles, we deployed the algorithm on the Google Exacycle infrastructure. Briefly, during algorithm development, the modified Rosetta source code was regularly committed to the Rosetta development community’s common code repository, and the conformation and sequence databases, as well as the PatchDock output files, were computed on the local computer cluster at the Weizmann Institute of Science. The resulting intermediate data files, as well as RosettaScript (40) protocols detailing the design trajectory, were electronically transferred to Google, where the updated Rosetta source code was compiled and deployed on the Exacycle platform.

Exacycle is a highly parallel infrastructure using hundreds of thousands of idle cores at Google. The high likelihood of preemption of the low-priority design jobs required runtimes to be kept to <40 min. Because the strict filtering following design cycles 1–3 yielded very few good candidates but executed relatively quickly, this computation was a perfect fit for the Exacycle platform. We were able to cast an extremely wide net, rapidly exploring a huge number of possible candidates using very low-cost computing resources. The vast majority of these candidates were quickly rejected, but the remaining ones (in the thousands) were of high quality and were further refined and optimized, now requiring only moderate compute resources.

Designs computed on Exacycle were saved in compact Rosetta-formatted binary silent files and electronically transferred to the Weizmann Institute for further analysis.

Rosetta energy function.

Throughout all modeling calculations, the Rosetta score12 energy function was used. This energy function is dominated by van der Waals packing, hydrogen bonding, and an implicit solvation model (41). RosettaScripts and commandline instructions are listed in Dataset S2.

Cloning and yeast cell surface display.

Genes encoding each of the designs were optimized for Escherichia coli expression using DNAWorks (42), and ordered from Gen9Bio as double-stranded DNA oligos. Each DNA sample was amplified by PCR to include at least 30 bases overlapping with the vector. The pETCON plasmid (6) was linearized using NdeI/BamHI (New England Biolabs), and 100 ng of linearized plasmid and 50 μL of PCR product of each gene were transformed into yeast EBY100 cells by lithium-acetate transformation, essentially as described by Chao et al. (19). Plasmid was then extracted using the Zymoprep Yeast Plasmid Miniprep I kit (Zymo Research) and transformed into E. cloni cells. Single clones were selected for sequence verification, and the verified plasmids were then transformed into yeast again for yeast surface display.

Yeast display experiments were conducted essentially as described by Chao et al. (19). Yeast cells were taken from the transformation plate and grown in selective medium (SDCAA: 20 g of dextrose, 6.7 g of Difco yeast nitrogen base, 5 g of bacto casamino acids, 5.4 g of Na₂HPO₄, 8.56 g of NaH₂PO₄ * H₂O, and 1 L of H₂O, sterilized by filtration) overnight at 30 °C. From the overnight culture, another 5 mL of culture in SDCAA was inoculated and grown at 30 °C to OD₆₀₀ = 0.6. The cells were then harvested and resuspended in 10 mL of induction medium (SG/RCAA: 20 g of galactose, 20 g of raffinose, 6.7 g of Difco yeast nitrogen base, 5 g of bacto casamino acids, 5.4 g of Na₂HPO₄, 8.56 g of NaH₂PO₄ * H₂O, and 1 L of H₂O, sterilized by filtration) and incubated at 20 °C for 20 h. One hundred microliters of culture was then used for yeast cell surface display experiments: Cells were harvested, washed twice in ice-cold PBS-F (1 g of BSA, 100 mL of Dulbecco’s 10× PBS, and 1 L of H₂O, sterilized by filtration), and subjected to primary antibody [mouse monoclonal IgG₁ anti–c-Myc (9E10) sc-40; Santa Cruz Biotechnology] for expression monitoring and biotinylated ligand at the desired concentration in ice-cold PBS-F for 20 min at 4 °C. The cells were then washed twice in ice-cold PBS-F and underwent a second staining with a fluorescently labeled secondary antibody [AlexaFluor488 goat–anti-mouse IgG₁ (Life Technologies) for scFv labeling, Streptavidin-APC (SouthernBiotech) for ligand labeling] for 20 min at 4 °C. After another wash in ice-cold PBS-F, the cell fluorescence was measured on an Accuri C6 flow cytometer, and expression levels were calculated as shown in Fig. S5.

Error-prone gene diversification and selection.

Designs that showed initial binding underwent genetic diversification by error-prone PCR using a GeneMorph II Random Mutagenesis Kit (Agilent) and subsequent library generation (43). The libraries contained 10⁷–10⁸ clones, with an average of two to three mutations per clone.

The libraries were then sorted three times on an Aria II fluorescence-activated cell sorter (Becton Dickinson), and 10⁵ cells were collected each time. After the third sort, single clones were isolated and evaluated separately.

Ligand expression and purification.

A gene encoding ACP2 optimized for E. coli expression was ordered from GenScript in pET29 and cotransfected with the biotin-ligase BirA in pET28 for in vivo biotinylation in BL21 cells. An overnight culture was grown in LB with 50 ng/mL kanamycin, 30 ng/mL chloramphenicol at 37 °C, and a 1-L culture with 50 ng/mL kanamycin and 0.5% glucose (to prevent premature expression of the proteins) inoculated from it. The culture was incubated at 37 °C until OD₆₀₀ = 0.6, when protein expression was induced by addition of 1 mL of 1 M isopropyl β-d-1-thiogalactopyranoside (IPTG). For in vivo biotinylation, at the same time, d-biotin was added to a final concentration of 50 nM from stock (5 mM in 10 mM bicine, pH 8.3, sterile-filtered). The culture was further incubated at 37 °C overnight and harvested in the morning.

Cells were resuspended in 20 mL of HBS [20 mM Hepes, 150 mM NaCl (pH 7.4)] lysed by sonication; the lysate cleared by centrifugation (13,000 × g, 20 min, 4 °C) and loaded onto nickel-nitrilotriacetic acid (Ni-NTA) resin (Ni-NTA His-Bind Resin; Merck). After two washing steps with HBS/100 mM imidazole, the protein was eluted with HBS/500 mM imidazole and dialyzed against HBS. After concentration, the protein was aliquoted and stored at −80 °C.

For experiments with insulin binders, commercially available biotinylated insulin was used (Immunological and Biochemical Testsystems).

IgG and Fab production.

Adherent HEK293T cells were cotransfected with the heavy and light IgG chains or with the heavy and light Fabs (heavy chain fused to C-terminal His tag) using linear polyethylenimine as a transfection reagent (12.5 µg/12.5 µg/50 µg, respectively, per 15-cm plate). Seventy-two hours posttransfection, the medium containing the secreted protein was collected (∼1.5 L).

Fab purification from IgG (5ins14, 2acp12).

The filtered medium was concentrated to 100 mL using a diafiltration device (QuixStand Benchtop System; GE Healthcare). The medium composition was exchanged to buffer containing 50 mM Tris (pH 8) and 100 mM NaCl using the same device. The clarified supernatant was subsequently loaded onto a 5-mL HisTrap_Protein A_HP column (GE Healthcare). The IgG was eluted from the column with buffer containing 20 mM citric acid (pH 3). Fractions (1.5 mL) were collected into wells containing 200 μL of neutralizing buffer (1 M Tris, pH 8.8). Fractions containing IgG were dialyzed against sample buffer [20 mM sodium phosphate, 10 mM EDTA (pH 7)] and lyophilized. Immobilized papain (Thermo Scientific) was activated according to the manufacturer’s instructions with digestion buffer (sample buffer containing 20 mM cysteine at pH 7). The lyophilized IgG (10 mg) was dissolved in 1 mL of digestion buffer and incubated with 1 mL of activated papain beads for 5 h at 37 °C. Dilution with 10 mM Tris (pH 7.5) stopped the reaction and prepared the cleavage products for subsequent purification on a protein A column (equilibrated with 10 mM Tris, pH 7.5). Fab fragments were collected from the flow-through.

Secreted Fab purification (5ins16).

Filtered medium (1.5 L) was loaded onto a HiTrap_chelating_HP_5 mL column (GE Healthcare) equilibrated with 50 mM Tris (pH 8) and 100 mM NaCl. The column was washed with the same buffer containing 100 mM imidazole, and the Fab fragments were eluted in one step with 500 mM imidazole and desalted on a HiPrep_26/10_desalting column (GE Healthcare) equilibrated with 50 mM Tris (pH 8) and 100 mM NaCl. The pure Fab was concentrated to 20 mg/mL for crystallization.

Periplasmatic expression and purification (5acp14).

The design was transformed into RH2.2 plasmid for expression as Fabs, where the heavy chain was N-terminally His-tagged and the light chain was expressed as a separate protein. Both chains contain a secretion sequence for direction to the periplasmatic space, where they fold and dimerize. Restriction-free cloning was done using Kapa HiFi Hotstart Readymix (Kapa Biosystems) according to the manufacturer’s protocol.

Cells were induced with 1 mM IPTG at OD₆₀₀ = 0.6, transferred to 16 °C, and harvested after 20 h. The cells were then resuspended in sucrose buffer [50 mM Tris, 1 mM EDTA, 20% wt/vol sucrose (pH 7.4); 100-mL/L culture], shaken for 20 min at 4 °C, and harvested, and the periplasm was lysed in low-osmolar solution (500 nM MgCl₂; 50-mL/L culture) for 30 min at 4 °C. The supernatant was harvested by centrifugation (15,000 × g, 15 min) and filtered, and 20× HBS [400 mM Hepes, 3 M NaCl (pH 7.2)] was added to reach a 1× HBS concentration. The osmotic shock fluid was then loaded onto Ni-NTA resin and washed twice with five bed volumes of 15 mM imidazole in HBS and once with five bed volumes of 30 mM imidazole in HBS, and then eluted with three bed volumes of 500 mM imidazole in HBS. Imidazole was removed from the eluate by dialysis against HBS (1:500).

The sample was then concentrated (Amicon Ultra-15 Centrifugal Filter; Merck) and purified by gel filtration in HBS over a HiLoad 16/600 Superdex 75-pg column.

Apparent T_m measurements.

All experiments were performed on a ViiA 7 real-time PCR machine (Applied Biosystems). Fabs were diluted to 0.125 mg/mL in a 1:5,000 dilution of SYPRO Orange in HBS, and 20 μL of the solution was measured. The temperature was ramped from 25 °C to 100 °C at 0.05 K⋅s⁻¹. The maximal unquenching of the fluorophore occurred at 71 °C for 2acp12, at 57 °C for 5acp14, and at 79 °C for 5ins16 (Fig. S3B). The slow decrease of fluorescence after the peak in each curve (at high temperatures) is typical of protein aggregation, excluding the dye from formerly exposed hydrophobic core regions. For 5ins16, a five-bin, unweighted, sliding-average smooth was applied.

Surface plasmon resonance.

Surface plasmon resonance experiments on the ACP binders were carried out on a Biacore T200 instrument (GE Healthcare) at 25 °C with HBS-N EP+ [10 mM Hepes, 150 mM NaCl, 3 mM EDTA, 0.005% vol/vol surfactant P20 (pH 7.4)]. For binding analysis, 1,000–2,800 response units (RU) of Fab were captured on a CM5 sensor chip. Samples of different protein concentrations were injected over the surface at a flow rate of 30 μL⋅min⁻¹ for 240 s, and the chip was washed with buffer for 1,200 s. If necessary, surface regeneration was performed with one 30-s injection of 1.5 mM NaOH at a flow rate of 30 μL⋅min⁻¹. One flow cell contained no ligand and was used as a reference.

The acquired data were analyzed using the device’s software, and kinetic fits were performed using the two-state reaction model to avoid overfitting of the late (asymptotic) dissociation phase.

Experiments for the insulin binder were carried out on a ProteOn XPR36 Protein Interaction Array System (BioRad) at 25 °C with HBS+ [20 mM Hepes, 100 mM NaCl, 0.005% Tween (pH 7.5)]. For binding analysis, 8,000 RU of Fab was captured on a GLC Sensor Chip (BioRad). Samples of different concentration were injected over the surface at a flow rate of 30 μL⋅min⁻¹ for 200 s and dissociation-monitored while the chip was flushed with buffer for 2,000 s. One flow cell contained no ligand and was used as a reference.

The data were analyzed using the software’s kinetic fit for heterogeneous ligands to account for the undirected immobilization of the Fab on the chip surface.

Structure determination and refinement of the 5ins16 and 5ins14 structures.

Crystals of 5ins16 and 5ins14 were obtained using the sitting-drop vapor-diffusion method with a Mosquito robot (TTP LabTech). The crystals of 5ins16 were grown from 12% PEG 3,500, 0.1 M (NH₄)₂SO₄, and 0.05 M Tris (pH 8.5). The crystals formed in the orthorhombic space group P2₁2₁2₁, with two monomers per asymmetric unit. A complete dataset to 3.0 Å resolution was collected at 100 K on a single crystal on in-house RIGAKU RU-H3R X-ray.

Crystals of 5ins14 were grown from 15% PEG 8,000, 0.1 M NH₄CH₃CO₂, 0.005 M Mg(CH₃CO₂)₂, and 0.05 M cacodylate (pH 6.5). The crystals formed in the trigonal space group P3₁2₁, with one monomer per asymmetric unit. A complete dataset to 2.1 Å resolution was collected at the European Synchrotron Radiation Facility beamline, ID23_1.

Diffraction images of the 5ins16 and 5ins14 crystals were indexed and integrated using the Mosflm program (44), and the integrated reflections were scaled using the SCALA program (45). Structure factor amplitudes were calculated using TRUNCATE (46) from the CCP4 program suite. Both structures were solved by molecular replacement with the program PHASER (47). The model used to solve the 5ins16 structure was the light chain of the anti-human respiratory syncytial virus and the heavy chain of the human CB2 Fab (PDB ID codes 3qq9 and 2hff, respectively). The model used to solve the 5ins14 structure was the light chain of the anti-human respiratory syncytial virus and the heavy chain of the human phosphoserine (pSAb) scaffold (PDB ID codes 3qq9 and 4jfz, respectively).

All steps of atomic refinement of both structures were carried out with the CCP4/REFMAC5 program (48). The models were built into 2mF_obs − DF_calc and mF_obs − DF_calc maps by using the Coot program (49). Details of the refinement statistics of the 5ins16 and the 5ins14 structures are described in Table S2. The coordinates of 5ins16 and the 5ins14 were deposited in the RCSB PDB with ID codes 5NB5 and 5BNI, respectively.