Principles for computational design of binding antibodies

Abstract

Natural proteins must both fold into a stable conformation and exert their molecular function. To date, computational design has successfully produced stable and atomically accurate proteins by using so-called "ideal" folds rich in regular secondary structures and almost devoid of loops and destabilizing elements, such as cavities. Molecular function, such as binding and catalysis, however, often demands nonideal features, including large and irregular loops and buried polar interaction networks, which have remained challenging for fold design. Through five design/experiment cycles, we learned principles for designing stable and functional antibody variable fragments (Fvs). Specifically, we (i) used sequence-design constraints derived from antibody multiple-sequence alignments, and (ii) during backbone design, maintained stabilizing interactions observed in natural antibodies between the framework and loops of complementarity-determining regions (CDRs) 1 and 2. Designed Fvs bound their ligands with midnanomolar affinities and were as stable as natural antibodies, despite having >30 mutations from mammalian antibody germlines. Furthermore, crystallographic analysis demonstrated atomic accuracy throughout the framework and in four of six CDRs in one design and atomic accuracy in the entire Fv in another. The principles we learned are general, and can be implemented to design other nonideal folds, generating stable, specific, and precise antibodies and enzymes.

Keywords: AbDesign; Rosetta; V(D)J recombination; expressibility; stability.

Fig. 1.

Improved antibody expressibility through five design/experiment cycles. The 114 insulin-targeting designs were formatted as scFvs and their yeast surface expression levels were evaluated in five successive cycles of algorithm development (expression levels are normalized to those of the high-expression antibody 4m5.3 tested under identical conditions). Molecular representations show flaws observed in early design cycles: cavities (gray) in the protein core (1ins01) (Left), a buried but unpaired arginine (1ins10) (Center), and failure to maintain a buried hydrogen-bonding network between segments distant in sequence (3ins17) (Right). Starting in design cycle 4, conformation-dependent sequence constraints were used to guide Rosetta design choices. In this cycle, the entire Fv (framework and CDRs) was subjected to Rosetta design. Additionally, in cycle 5, the Fv backbone was segmented in two parts in each chain: one comprising the framework and CDRs 1 and 2 and another comprising CDR 3. Side chains in gray show identities typical of natural antibodies in the relevant positions. The backbone is rainbow-colored from the amino terminus (blue) to carboxy terminus (red). HC, heavy chain; LC, light chain.

Fig. 2.

Backbone segmentation of the antibody Fv. (Left) Conventional antibody design and engineering studies segmented the Fv into seven parts (a framework and six CDRs) and generated antibodies by combining segments from various antibodies. (Right) Segmentation used by AbDesign in design cycle 5, by contrast, uses four parts: two comprising CDRs 1 and 2 and the framework and two comprising CDR 3. The latter segmentation maintains the structural interactions between CDRs 1 and 2 and the framework, resulting in improved core packing. The conserved disulfides are shown as sticks in structural representations and as yellow dots on the primary-sequence representation. vH, heavy-chain variable domain; vL, light-chain variable domain.

Fig. S1.

Core-packing improvements over the course of algorithm development and testing. Packstat provides a measure of core packing density; higher values indicate fewer core cavities and more realistic structures. Designs from cycles 4 and 5 showed higher core packing than those from cycle 1.

Fig. S2.

Structural features of the design models (antibody in lime, ligand in teal/gray). (A) Long loops in the light chain and heavy chain form a deep cleft, engulfing the ligand and creating a large binding surface (2,050 Ų). The designed interaction between antibody and ligand uses both hydrogen bonds and aromatic stacking. (B) CDRs are stabilized by buried polar interactions similar to those observed in natural antibodies, and ligand binding is mediated by aromatic stacking and charge complementarity. (C) CDRs of 5ins16 are stabilized by a tightly knit network of hydrogen bonds between polar atoms of backbone as well as the side chain atoms within and between loops, as they are observed in natural antibodies.

Fig. 3.

Mutation analysis of designed binding modes. (Top) Design models were visually inspected, and mutations were introduced manually to improve packing and/or solvation (red spheres). For assessing the binding mode, three to four single-point mutations (cyan spheres) were introduced in designed antibodies 2acp12 and 5ins16 (5acp14 was not subjected to mutagenesis because of low affinity). (Bottom) Relative binding signal of variants containing single-point mutations in ligand-binding sites (gray bars) or outside the ligand-binding sites (black bars). Additional manual mutations were introduced on the ligand ACP (cyan spheres) in the intended binding surface (gray bars) or outside the binding surface (black bars) and tested for binding of the affinity-matured antibodies. Finally, mutations were introduced in the designed antibodies also through random mutagenesis and in vitro selection of improved binders (blue spheres). Mutations at the binding surface reduced affinity, except for Val71Glu on ACP, which enhanced binding of the affinity-matured 2acp12_ev. Mutations away from the interface, including those isolated from in vitro selection, increased binding affinity. Binding signals were tested in yeast surface display with the antibodies expressed as scFv and the ligand at 50 nM concentration for 2acp12_ev and at 1 μM concentration for 2acp12, 5acp14_ev, and 5ins16. SEs from the mean for two independent experiments are indicated.

Fig. S3.

Ligand binding, affinity maturation, and thermal resistance of the designed binders. (A) Antibody affinity was first determined by yeast display titrations (● or ■), and then improved through random mutagenesis and in vitro selection using FACS, yielding affinity improvement from an apparent K_d = 900–50 nM and from an apparent K_d = 300 to 30 nM in 2acp12 and 5ins16, respectively (○ or □). (B) Thermal denaturation of the designs expressed as Fabs resulted in apparent T_ms of 71 °C, 57 °C, and 79 °C for 2acp12, 5acp14, and 5ins16, respectively. Designs 2acp12 and 5ins16 show high background, but only one sharp transition over the measured range of temperatures (50). (C) Surface plasmon resonance traces. The antibodies were expressed as Fab- and amine-coupled to CM5 chips. The affinities determined from the kinetic fits were 100 nM for 2acp12, 50 nM for 5acp14, and 50 nM for 5ins16.

Fig. 4.

Comparison of design models and experimental structures of 5ins16_ev and 5ins14. In both designed antibodies, the light chain (LC), the backbone conformation of the framework, and the LC–heavy chain (HC) heterodimer interfaces are atomically accurate. (A) 5ins16_ev: Backbone and side-chain packing deviations occur in H1 and H3, but other regions, including buried hydrogen-bonding networks (dashed lines) involving L1, are atomically accurate. (B) 5ins14: Core packing of hydrophobic residues on the framework specifying the L1 conformation, as well as a buried polar network specifying H1, are atomically accurate. Model and experimental structures are colored in lime and purple, respectively. Antibodies were expressed and crystalized as Fabs, and only the Fvs are shown.

Fig. S4.

Electron density maps and computational docking of the designed antibodies. (A) The 2F_o − F_c electron density maps (violet mesh) of 5ins16_ev and 5ins14 allow unambiguous assignment of the backbone conformation and many of the bulky side chains in the antibody core and the CDRs, indicating low flexibility. Where differences occur between the model (lime) and experimental structure (violet), they are likely to be the result of alternative packing. (B) CDRs H1 and H3, which differ in conformation relative to the design model, form packing interactions with crystallographic-symmetry neighbors (shown as surface). (C) Insulin was docked against the experimental structure of 5ins16_ev using Rosetta starting from insulin’s designed orientation relative to 5ins16 (light cyan), yielding a low-energy predicted conformation (dark cyan). The docking orientation is mostly conserved (0.4 Å), with a few rotamer changes at the interface.

Fig. S5.

Threshold for expression level calculations. Expression levels are given as percentage of cells exhibiting higher fluorescence levels than the negative control. Given are fluorescence densities of unstained cells (negative) and stained cells expressing the reference antibody 4m5.3.