Disulfide-compatible phage-assisted continuous evolution in the periplasmic space

CadC activates transcription upon periplasmic binding

A periplasmic protein-protein interaction selection system must convert a periplasmic binding event into a cytoplasmic transcriptional activation event. We examined transmembrane signaling proteins that physically link protein-protein binding in the periplasm with transcription in the cytoplasm. CadC is a native E. coli sensor protein and a member of the ToxR-like receptor family. Several members of the ToxR family have served as the basis for engineered sensors of periplasmic or transmembrane interactions^53,54,55. CadC consists of a periplasmic sensor domain, a transmembrane helix, and a DNA-binding cytoplasmic domain (Fig. 1b). Under stress conditions, the periplasmic sensor domains from two CadC molecules homodimerize, bringing together the cytoplasmic DNA-binding domains to generate two cooperative DNA-binding sites, which then bind two DNA motifs, Cad1 and Cad2, on the CadBA promoter (P_cadBA) to initiate gene transcription^54,56. Replacement of the sensor domain with a dimerizing protein leads to constitutive activation of P_cadBA⁵³. CadC thus converts binding in the periplasm into cytoplasmic transcriptional activation.

We reasoned that CadC could form the basis of a PACE selection for protein-protein binding in the periplasmic space (Fig. 1c). First, we optimized P_cadBA (Supplementary Fig. 1) and deleted the host genomic cadCBA operon to minimize background transcription (Supplementary Fig. 2). To validate that protein binding could trigger transcription at P_cadBA, we expressed CadC with its sensory domain replaced by the HA4 monobody, a high-affinity monobody that binds the SH2 domain of ABL1 kinase⁵⁷. We then expressed YibK, a homodimeric knottin protein, fused to the SH2 binding target of HA4. This construct was directed to the periplasm by an N-terminal signal sequence (SS) peptide derived from alkaline phosphatase A (PhoA),⁵⁸. YibK homodimerization should trigger dimerization of the CadC–HA4 fusion via binding of HA4 to the SH2 domain fused to YibK, resulting in activation of P_cadBA. Indeed, we observed that expression of SS–YibK–SH2 directed P_cadBA transcriptional activation 66-fold over expression of cytoplasmic YibK–SH2 as measured by transcription assay (Fig. 2b).

The point mutation V139R blocks YibK dimerization by disrupting hydrophobic interactions between YibK monomers and preventing a final folding transition to the native YibK structure^59,60. The K_D values for dimerization of wild-type YibK and V139R YibK are <1 nM and 360 µM, respectively⁶⁰. The introduction of V139R resulted in >8-fold loss of P_cadBA-directed LuxAB expression (Fig. 2b), establishing that protein-protein affinity determines the degree of transcriptional activation at P_cadBA.

To link binding in the periplasm to phage propagation, we drove gIII expression with P_cadBA. We challenged phage encoding SS–YibK–SH2 to propagate in overnight culture on host cells expressing CadC–HA4 and P_cadBA-driven gIII. Phage encoding wild-type YibK propagated more than three orders of magnitude more efficiently in this periplasmic PACE system than V139R YibK phage, demonstrating that pPACE links target protein binding in the periplasm to phage propagation through P_cadBA activation and production of pIII (Fig. 2c).

Periplasmic phage-assisted evolution of YibK

To validate the pPACE selection for periplasmic protein-protein binding, we challenged the system to evolve homodimeric YibK variants starting from phage encoding the monomeric V139R variant. We adapted pPACE into the format of PANCE (phage-assisted non-continuous evolution)^37,38,41,61, a non-continuous form of PACE in which host-phage populations undergo serial daily passaging in lieu of continuous flow, permitting a less stringent and more sensitive initial selection. After three PANCE passages, phage titers increased robustly (Supplementary Fig. 3b).

YibK variants evolved mutations that restore YibK dimerization. On the YibK dimer interface, V139 forms a hydrophobic contact with A138’ of its binding partner⁶². Evolving phage did not directly revert the V139R point mutation. However, in PANCE-evolved clone 3.7, residue A138 mutated to an aspartic acid (GCC to GAT), completely restoring affinity as measured by P_cadBA transcriptional activation (Fig. 2d, f). R146, which is in close proximity to R146’, was converted to a cysteine residue in seven of eight sequenced phage (CGT to TGT; Fig. 2g; Supplementary Fig. 3d), resulting in stronger transcriptional activation of P_cadBA than wild-type YibK. Remarkably, we found that R146C results in an intermolecular disulfide bridge. The covalently bound species can be seen by SDS-PAGE in purified YibK protein, as a ~43 kDa band representing the dimeric form of the 21.6 kDa monomer (Fig. 2e, Supplementary Fig. 4b, c). In whole-cell lysates, a ~60 kDa band representing the dimer of 30 kDa YibK–SH2 can also be visualized (Supplementary Fig. 4d-e). Both dimeric species are lost upon the addition of reducing agent. Together, these results establish that the periplasmic selection platform is capable of restoring and improving the stable homodimerization of a monomeric protein by multiple mechanisms, including the evolution of novel disulfide bridges.

Periplasmic evolution of antibody–antigen affinity

Next, we applied pPACE to antibody–antigen binding. Full-length antibodies can be engineered into smaller forms such as single-chain variable fragments (scFvs), comprising only the heavy and light chain variable regions (V_H and V_L) tethered by a flexible synthetic linker^1,2. ScFvs are small (~30 kDa), can be produced in E. coli, exhibit improved tissue penetration, and can be readily conjugated to drug molecules, effector proteins and chimeric antigen receptors^1,3, making them prime candidate molecules for directed evolution approaches. Heterologous expression of scFvs in E. coli typically involves export into the periplasm using an N-terminal signal sequence^3,63.

We challenged pPACE to evolve scFv antibodies. We chose the Ω-graft antibody scFv, which targets the leucine zipper GCN4 with K_d ~500 pM^39,64. To determine whether an antibody–antigen interaction could drive CadC dimerization, we expressed CadC–HA4 and Ω-graft–SH2, with or without co-expression of a monomeric form of the leucine zipper GCN4⁶⁵ (GCN4(7P14P)) fused to SH2. In this architecture, the binding of Ω-graft to GCN4 drives dimerization of CadC–HA4 bound to Ω-graft–SH2 and a CadC–HA4 molecule bound to GCN4–SH2, creating a four-part complex (Fig. 3a). The addition of GCN4 led to a 30-fold increase in P_cadBA-driven LuxAB expression (Fig. 3b). In contrast, the substitution of a Ω-graft double point mutant L231F F232A, which impairs binding to GCN4 by >7,000-fold⁶⁴, in place of wild-type Ω-graft led to a 55-fold decrease in transcriptional activation (Fig. 5b). Collectively, these results indicate that our selection can link scFv target binding to transcriptional activation of P_cadBA.

To determine whether pPACE can distinguish between functional and nonfunctional forms of Ω-graft, we performed a competitive mock-selection experiment without mutagenesis. We seeded host cells expressing CadC–HA4 and GCN4–SH2 and encoding P_cadC-driven gene III on the AP with a mixture of selection phage containing a 1:1,000 ratio of unmutated Ω-graft–SH2 selection phage to L231F F232A mutant–SH2 selection phage and carried out PACE and PANCE. Within 12 h of PACE or following two PANCE passages, unmutated Ω-graft variants dominated both populations, enriching ≥1000-fold (Supplementary Fig. 5a), demonstrating that the selection platform can be used to selectively propagate phage encoding a target-binding antibody scFv.

Regulating scFv periplasmic export

In the small volume of the periplasm, minor changes in protein expression level have a large impact. An evolving SP might achieve increased fitness by modifying the promoter driving scFv expression to increase scFv levels and compensate for a poor K_D. We reasoned that controlling scFv export to the periplasm would be desirable to maintain selection pressure. Further, regulating the level of periplasm-targeted scFv protein could drive two simultaneous selections: for high affinity to the target and for increased solubility of the scFv, to raise the effective concentration of scFv. We therefore adapted a key aspect of a related PACE selection, soluble expression PACE or SE-PACE³⁹. SE-PACE uses a trans-splicing intein to reconstitute two fragments into a single functional protein, integrating transcription from two promoters into one output^39,66. In SE-PACE, intein-mediated splicing reconstitutes the signal sequence peptide of pIII, which must enter the periplasmic space for phage to propagate³⁹.

We split the PhoA-derived signal sequence (SS)⁵⁸ into two halves, consisting of amino acids 1–8 and 9–21 (Supplementary Fig. 6). These two halves were fused, respectively, to the N- and C-terminal portions of the Nostoc punctiforme (Npu) trans-splicing DnaE intein⁶⁷. SS amino acids 1–8 were fused to the N-terminal half of the Npu intein on a host AP1, ensuring that phages are not able to evolve the increased expression of this component. The C-terminal half of the Npu intein, fused to SS amino acids 9–20 and the evolving scFv, was encoded on the selection phage. Following translation of both fusion proteins, intein-mediated splicing reconstitute full-length SS–scFv, allowing periplasmic export (Supplementary Fig. 6).

Using the Ω-graft pPACE selection described above, we observed that expressing Ω-graft–SH2 with its SS split into two polypeptides, each fused to half of the Npu intein, led to robust phage propagation, indicating reconstitution of the SS. In contrast, when we omitted the C-terminal domain of Npu from the SS_9–20–Ω-graft–SH2 construct, phage failed to propagate (Supplementary Fig. 6c). Similarly, the expression of the SS_1–8–NpuN component is necessary for propagation (Supplementary Fig. 6d). We further found that by expressing SS_1–8–NpuN under small-molecule induction in the presence of NpuC–SS_9–20–Ω-graft (34.8 kDa), we could drive periplasmic expression of Ω-graft scFv (30.2 kDa) in a dose-dependent manner (Supplementary Fig. 6g, h). This demonstrates that Npu can regulate the reconstitution of full-length SS for periplasmic export of scFvs and P_cadBA activation.

Under this intein-regulated system, the total amount of scFv exported to the periplasm, and thus available direct CadC dimerization, is limited by the availability of the intein-SS fragment encoded on the host AP. The researcher can modify the expression level of intein–SS_1–8 fragment to limit the reconstitution of SS–scFv, and thus the amount of scFv exported to the periplasm, creating selection pressure for efficient expression of soluble scFv.

Evolution of Ω-graft and overcoming scFv homodimerization

Next, we challenged pPACE to correct the L231F F232A binding mutation in the Ω-graft antibody scFv, using both a full-length N-terminal SS sequence and the intein-SS strategy described above, in order to select for affinity alone or affinity and soluble periplasmic expression. We aimed to apply pPACE to restore binding to GCN4 by correcting mutation L231F F232A.

PACE experiments using our original selection architecture (Fig. 3a) resulted in two genotypic outcomes. First, close to half of phage reverted mutation L231F to the wild type within 96 h of pPACE. Second, scFv variants developed cysteine residues at their N-termini or within the 4X GGGS linker connecting scFv VH and VL domains (Fig. 3c). Linker cysteines in particular appeared mutually exclusive to the desired L231F reversion (Supplementary Fig. 5b). We found that at both positions, a cysteine substitution resulted in higher transcriptional activation than reversion of position 231 to Leu (Fig. 3d). The insertion of a C-terminal Cys residue has been used to manufacture stable dimeric scFvs through the formation of a covalent disulfide⁶⁸. We reasoned that an N-terminal or linker Cys residue might form a similar covalent linkage, generating stably homodimeric scFv–SH2. This homodimer circumvents the target-binding selection by binding two CadC–HA4 molecules and bring them into close proximity, without the involvement of the antigen (Fig. 3e).

To prevent circumvention of the target-binding selection, we modified the selection architecture by fusing the GCN4(7P14P) antigen directly to CadC in place of HA4, to eliminate the possibility of scFv homodimerization resulting in selection survival (Fig. 3f). We created obligate homodimeric scFvs by removing the now-redundant SH2 domain fusion and either pre-installing an N-terminal cysteine in the Ω-graft scFv (Fig. 4a), or, as a more general strategy, by fusing a homodimerizing GCN4 leucine zipper domain C-terminal to the scFv (Fig. 3f; Fig. 5a, Supplementary Fig. 7). This strategy ensures that efficient dimerization does not depend on the properties of the scFv being evolved, since different scFvs homodimerize at different rates⁶⁹. In this second-generation selection architecture, a dimeric scFv antibody must bind two CadC-fused antigens to activate P_cadBA. Transcriptional assays suggested that mutation L231F accounts for the loss of binding, and that F232A alone has little effect (Supplementary Fig. 8e). We therefore considered reversion of F232A to be unnecessary in desired selection outcomes.

Using this second-generation architecture, phage encoding Ω-graft showed three orders of magnitude higher levels of propagation in overnight enrichment assays than phage encoding Ω-graft L231F (Fig. 4b, c). Incorporation of a nonsense mutation at position 100 (W100*) also led to strong de-enrichment of phage (Fig. 4b).

We challenged pPACE using the second-generation architecture to correct a stop codon at W100 in addition to the L231F binding defect mutation. Within 96 h of pPACE, phage reverted mutations correcting both deleterious mutations in population 1 (Fig. 4e, Supplementary Fig. 9a). In population 2, we used the split-intein signal sequence strategy described above to regulate periplasmic scFv expression in host cells (Fig. 4a, f). Due to the decreased fitness of intein-SS phage compared to phage with full-length SS (Fig. 4b), we did not challenge population 2 to correct a stop codon. Mutation F231L was present in ~50% of this population by 96 h and dominated the population by 156 h (Supplementary Fig. 9b). Phage in different populations accessed leucine codons at position 231 via two distinct point mutations, converting TTC to TTA or CTC (Fig. 4e, f). Importantly, we observed no new cysteines arising during evolution (Fig. 4g; Supplementary Fig. 9a). These results suggest that the second-generation pPACE selection prevents phage from passing the selection by evolving stable scFv–SH2 homodimers alone, and requires a tight scFv-antigen interaction in order to activate P_cadBA.

In population 2, which used the intein-SS strategy, two point mutations, F231L and L224S, enriched as separate solutions present at a similar frequency at 96 h. Mutations F231L and L224S were observed in the same variant by 112 h (Supplementary Fig. 9b). Mutation L224V was previously reported to enhance cytoplasmic solubility of the Ω-graft scFv³⁹. We compared the soluble expression of Ω-graft scFv and variant L224S with an N-terminal PhoA SS from the promoter P_T7Lac in BL21*DE3 cells by western blotting. We found that L224S increased soluble expression of Ω-graft by roughly 8-fold (Supplementary Fig. 9c, d).

Together, these findings demonstrate that pPACE can restore affinity of an antibody to an antigen, and that regulating periplasmic export of the evolving species using a split-intein signal sequence can support the evolution of improved soluble expression and improved binding. These results also show that the second-generation pPACE system avoids outcomes that circumvent the selection by homodimerizing the evolving protein, rather than by binding the target.

Periplasmic PACE of novel trastuzumab scFv variants

We used the second-generation pPACE selection to evolve an scFv form of the antibody trastuzumab to bind a new target antigen. Trastuzumab targets the oncogenic receptor Her2 and is a successful first-line treatment for Her2⁺ breast cancers. Most trastuzumab-responsive tumors develop resistance to the drug within one year⁷⁰. Second-line treatments can overcome resistance using multi-specific engineered antibodies, which combine variable domains of two or more mAbs to simultaneously target several epitopes, such as Her3, EGFR and VEGF kinase receptors^71,72,73. The ability of pPACE to rapidly evolve affinity to novel epitopes could further broaden the targeting capacity of engineered multi-specific antibodies.

The Her2 mimetic peptide H98 was identified in a peptide library screen for trastuzumab binding, and bears structural similarity but no sequence homology to Her2⁷⁴. We sought to apply pPACE to evolve an scFv form of trastuzumab with a higher affinity for the H98 peptide. We evolved trastuzumab scFv in the second-generation pPACE selection using either full-length SS or the split-intein SS strategy, resulting in mutually exclusive outcomes within 96 h of evolution. The H98 antigen was presented as a CadC–H98 fusion driven by a weak constitutive promoter on the AP. Trastuzumab was expressed as an scFv–GCN4 fusion to ensure dimerization, as we found that the use of a larger domain such as YibK to direct dimerization resulted in poor phage propagation (Fig. 5b), possibly due to excessive crowding of the periplasmic space.

Phages were allowed a 24-h period of evolutionary drift when pIII was provided freely with elevated mutagenesis³⁰, to generate a large and diverse phage library. Phages were then subjected to high-stringency pPACE at increasing flow rates until titers plateaued (Fig. 5c, Supplementary Fig. 10). In populations 1 and 2, phage encoded the full-length SS. Both populations converged on a single point mutation, H91Y (variant 1.1, Fig. 5d). In population 3, periplasmic export was restricted through the split-intein strategy described above, leading to enrichment of a single variant (3.2) with mutations A34D Y49S. Periplasmic PACE experiments carried out at further increased stringencies did not result in the enrichment of any additional point mutations in the scFv (Supplementary Fig. 11).

Computational modeling indicates that trastuzumab interacts with H98 through heavy chain residues V33, R50, and Y105, and light chain residues T94 and N30⁷⁴. In the trastuzumab crystal structure, residue T94 is proximal to residue H91 (H91Y in variant 1.1), and residue N30 is proximal to residue A34 (A34D in 3.2) (Supplementary Fig. 10e). Light chain residue Y49 is adjacent to residue A34 in a β-sheet, and mutation Y49S (variant 3.2) may help to accommodate the substitution of alanine for a bulky, charged aspartic acid at position 34 (PDB ID: 1N8Z⁷⁵).

Trastuzumab and evolved variants show a similar, characteristic change in mobility consistent with the reduction of disulfides during SDS-PAGE under reducing conditions, suggesting that intra-chain disulfides are retained in evolved variants (Supplementary Fig. 12a). To examine the role of intra-chain disulfides in the stability and binding of trastuzumab variants, we abrogated the possibility of disulfide formation by replacing the disulfide-forming cysteine residues with serines. In the absence of disulfide bonds, both trastuzumab and evolved variants failed to induce transcription from P_cadBA (Supplementary Fig. 12b). These results further suggest that trastuzumab binding is likely dependent on intra-chain disulfides, in agreement with the findings of Wörn and Plückthun that expression of trastuzumab scFv without disulfide bonds results in insoluble protein⁷⁶, and that these disulfides are preserved through pPACE. To ensure that the accumulation of insoluble cytoplasmic scFv does not impair host cell fitness, we carried out a growth timecourse, and found that scFv expression, with or without split-intein SS, had little to no effect on host cell growth (Supplementary Fig. 13).

Evolved variant 1.1 showed ~2.5-fold improved binding to H98 as measured by ELISA and little change in soluble expression (Fig. 5f-h, Table 1, Supplementary Fig. 10c, d). Evolved variant 3.2 was selected using the split-intein SS selection and showed a ~2-fold increase in affinity (Fig. 5f-h, Table 1, Supplementary Fig. 10c, d). To support ELISA data, we also carried out microscale thermophoresis (MST) (Supplementary Fig. 14, Supplementary Table 5). We note that in MST experiments, the upper bound of the binding curve was not accessible due to solubility limits of the H98 peptide in aqueous buffer (Supplementary Fig. 14), which may affect the accuracy of K_d determination. However, EC₅₀ values determined by MST and ELISA reflect similar fold improvements for evolved variants over the starting scFv.

Variant 3.2 showed substantial increases in soluble periplasmic expression (~5-fold as measured by western blotting and 2.5-fold as measured by Coomassie staining; see Supplementary Fig. 15), suggesting that restricting the level of scFv export to the periplasm selected for enhanced solubility. Evolved variants showed unchanged binding to Her2 in ELISA compared to that of trastuzumab scFv (Supplementary Fig. 10b). Evolved variants showed relatively unchanged thermal stability. Unevolved trastuzumab scFv had a melting temperature of 68.5 °C. We observed T_M increase of +4.0 ° for variant 1.1 and a T_M decrease of –5 °C for variant 3.2 (Supplementary Fig. 15f, Table 1).