De novo design of transmembrane β-barrels

Abstract

Transmembrane β-barrel proteins (TMBs) are of great interest for single-molecule analytical technologies since they can spontaneously fold and insert into membranes and form stable pores, but the range of pore properties that can be achieved by repurposing natural TMBs is limited. We leverage the power of de novo computational design coupled with a “hypothesis, design and test” approach to determine TMB design principles, notably the importance of negative design to slow β-sheet assembly. We design new eight stranded TMBs, with no homology to known TMBs, that insert and fold reversibly into synthetic lipid membranes, and have NMR and X-ray crystal structures very close to the computational models. These advances should enable the custom design of pores for a wide range of applications.

One sentence summary:

De novo design reveals principles of transmembrane β-barrel folding.

Advances in de novo protein design have yielded water soluble proteins of increasing complexity (1–5), and several examples of α-helical membrane proteins (6, 7). However, the de novo design of an integral transmembrane β-barrel (TMB) has not yet been achieved. The unassisted folding of TMBs into lipid bilayers in vitro likely involves concerted membrane insertion and folding of β-hairpins (8, 9), and how this is encoded in the sequences of TMBs is not well understood because of experimental challenges in characterizing their rugged folding pathways (10, 11). To prevent misfolding and aggregation in vivo, an array of chaperones assist TMB folding and assembly in the outer membranes of prokaryotes, mitochondria and chloroplasts (12). The lipid-folding/water-aggregation trade-off places poorly understood constraints on the global sequence properties of TMBs, slowing down the development of de novo design methods. Instead, TMB engineering has proceeded by modification of naturally occurring TMBs, which has yielded nanopores for single molecule DNA sequencing (13), small-molecule sensing (14, 15) or water filtering bioinspired membranes (16).

To shed light on the sequence determinants of folding and stability of TMBs, and to enable the custom design of TMBs for specific applications, we set out to design TMBs de novo. We started by studying the constraints membrane embedding puts on both the backbone geometry and sequence of transmembrane β-barrels.

Geometric constraints on transmembrane β-barrel backbones

TMBs are formed from a single β-sheet that twists and bends to close on itself, so that all membrane-embedded backbone polar groups are hydrogen-bonded and shielded from the lipid environment. Insertion of TMBs into the lipid membrane is oriented (17), with β-strands usually connected with long loops on the translocating (trans) side of the β-barrel (extracellular in bacteria) and short β-turns on the non translocating (cis) side (Fig. 1A). The β-barrel architecture is characterized by two discrete parameters: the number of strands (n) and the shear number (S) -- the sum of residue offsets (register shifts) between the neighbour strands, starting at any strand and tracing around the β-barrel (Fig. S1A) (18). The ideal β-barrel radius r (eq. S1) and angle of the strands with the main barrel axis θ(eq. S2) are functions of S and n (Table S1) (19). S and n also define the packing arrangement of side-chains in the β-barrel. There are S continuous strips of side chain Cβ atoms perpendicular to the β-strands (Fig. S1B,C), half of the Cβ-strips point toward the lumen and the other half toward the β-barrel exterior.

Fig. 1.

Geometric principles for TMB backbone design. (A) The trans side of the β-barrel is the side that translocates through the lipid membrane during TMB folding and insertion. The cis β-turns remain on the initial protein/lipid interaction side of the membrane. (B) Comparison of side-chain packing arrangements in 8-strand β-barrels with shear numbers of 8 (left, 4-fold symmetric packing) and 10 (right, 4-fold screw resulting in a jigsaw-like packing). In panels C-E, the membrane anchoring residues are shown as salmon spheres. (C, D) Geometric model of membrane-association constraints on the β-barrel architecture. (C) Asymmetric register shifts between the β-hairpins can be accommodated by tilting the β-barrel to the transmembrane axis by an angle α = arctan(Z/C). (D) The change of level Z between two anchor residues on the cis side of the β-barrel must be match by the change of level Z’ between the two stacking anchor residues on the trans side. Because of β-strands staggering to the main β-barrel axis, an anchor residue on the cis side of a strand N stacks with an anchor residue on the trans side of strand N+3. (E) The geometric model (center) and Rosetta model (right) predict similar tilt angles (α) of the β-barrel to the membrane axis and constant hydrophobic thickness, for β-strand arrangements with matching Z and Z’ (double register shifts located on strand N in cis and on strand N+3 in trans) (top). Both models show inconsistent hydrophobic thickness for β-barrel architectures with double register shifts located on strand N in cis and on strand N+5 in trans (bottom).

We focused on the simplest and smallest β-barrel architecture of 8 β-strands. We first considered a shear number of 8 (n=S). In this configuration, the total register shift is distributed equally among the four β-hairpins (2-residue offset between each β-hairpin) and the four Cβ-strips pointing toward the lumen of the barrel are arranged in 4-fold symmetric rungs with the Cα-Cβ vectors (which indicate the direction of the sidechains) pointing at each other (Fig. 1B, left). This symmetric arrangement combined with a small β-barrel radius does not allow tight jigsaw-puzzle like packing - the side-chains clash with each other rather than interdigitating. To enable better packing, we broke the symmetry in the core by increasing the register shift between two β-hairpins from 2 to 4 residues resulting in a shear number of 10. In this case, there are five intertwined Cβ-strips which spiral around the barrel axis, and the Cα-Cβ vectors point between rather than at each other so that the side-chains can pack in a more interdigitated fourfold screw-like pattern (Fig. 1B, right).

The uneven distribution of register shifts between β-hairpins complicates interactions with the lipid membrane, which can be approximated as two planes that must be parallel to ensure constant membrane thickness. In natural TMBs the cis (periplasmic) β-turns are close to the periplasmic lipid/water boundary (Fig. S2A–D). While the β-turn residues closely match the sequence preferences observed in water-soluble β-barrels (mostly polar residues), the lipid-exposed residues flanking these β-turns are predominantly hydrophobic (Fig. S2H–K) (20) and define the cis boundary of the transmembrane region (“membrane anchor residues”, Fig. S2A–D). The geometric challenge is that differences in the register shifts between β-hairpins result in a screw-like arrangement of the four anchor residues with a translation Z along the main β-barrel axis (eq. S3–S5), hence if the β-barrel axis is along the membrane normal the anchor residues cannot all be in the same plane. The vertical offset of the anchor residues can be made more compatible with the planarity requirement by tilting the β-barrel in the membrane by an angle α = arctan (Z/C) where the denominator is the length of the arc between anchor residues 1 to 4 projected onto the plane perpendicular to the main axis (eq. S6) (Fig. 1C, Fig. S3B). In the case of a β-barrel with symmetry (n=8, S=8), the vertical offset between each anchor residue, approximated from the geometric model, is close to zero (supplementary text) and no tilt is required. When S is increased to 10 by increasing the register shift between one pair of hairpins to 4 residues, the barrel must be tilted by approximately 6.7° to the transmembrane axis (Fig. 1E, top) to bring the anchor residues into the same plane. To test the validity of this simple geometric model, we explicitly assembled TMB backbones using Rosetta (21) and predicted their placement in the membrane (Methods (22)), which yielded an average tilt angle (8.1°, Fig. 1E, top) close to that of the geometric model. Placing the four-residue register shift after each of the four cis hairpins resulted in tilts with similar amplitude but different directions relative to the membrane axis (Supplementary text, Fig. S4B–G); we focused on the placement in which the 4-residue register shift is in the middle of the β-sheet.

We next investigated the structural consequences of the fact that the planes representing the cis and trans membrane boundaries must be roughly parallel to each other to keep the hydrophobic thickness constant. To achieve this, the offset Z between any two neighbour anchor residues on the cis face must be matched by a similar offset Z’ between the anchor residues above it on the trans face. For barrel topologies of (n=8, S=10) spanning a membrane of 24 Å, an anchor residue on the cis side of strand N stacks along the main β-barrel axis with the anchor residue on the trans side of strand N+3 because of the staggered β-strands (Fig. 1D, Supplementary text). Hence, to maintain constant thickness, the register shift between strands N and N+1 on the cis side must be equal to the register shift between strands N+3 and N+4 on the trans side. To confirm this prediction of our geometric model, we set the cis side register shift between strands N and N+1 to four residues, and ran Rosetta design simulations and transmembrane plane predictions on backbones with a matching 4-residue register shift on the trans side between (i) strands N+3 and N+4 and (ii) strands N+5 and N+6. We averaged planes representing the membrane boundary in cis and trans and found, consistent with the model, parallel planes and constant hydrophobic thickness for the N+3 case (i), but a 3 Å change of thickness in the N+5 case ((ii), Fig. 1E, bottom).

We used this constant hydrophobic thickness constraint to guide the distribution of the register shifts around the β-barrel. The cis hairpins were closed with short β-turns associated with a β-bulge which match the local twist of the β-strands (these are abundant in water-soluble (5) and transmembrane β-barrels (Fig. S2A)). On the trans side, the strands were connected with canonical β-turn sequences with strong β-hairpin nucleating properties (3:5 type I β-turns + G1 bulge with canonical SDG sequence (21–23)) in place of the long loops found in native TMBs; such turns were previously used to design water-soluble β-barrels (Fig. S2E,G) (5). To relieve strain from high β-sheet curvature, we placed glycine kinks (5) - glycine residues in a fully extended conformation within β-strands - into the blueprint such that a) every Cβ-strip pointing to the core of the barrel contains a glycine and b) no Cβ-strip contains more than 4 non-glycine residues in a row (¼ of the average barrel circumference). The glycine kinks in the designed β-barrel blueprint stack along four vertical lines (Fig. 2A,B) such that the resulting Rosetta models have four regions of strong β-sheet bending that delimit a lumen with a distinctive square-shaped cross-section not observed in naturally occurring β-barrels (Fig. S2F).

Fig. 2.

Sequence features defining de novo TMB fold and shape. (A and B) 2D schematic representation of the connectivity (hydrogen bonds as dashed lines) between β-strands in the TMB designs. Side-chains are shown as grey spheres and glycine residues as yellow dots. Aromatic girdle motifs are shown in red; tyrosines of the mortise/tenon motifs in blue; and prolines as black pentagons. Glycine kinks were arranged to bend the β-sheet into four corners (vertical arrows). (C) Hydrogen bond geometries between pairs of residues involving a glycine kink. Left: Examples from crystal structures of water-soluble (PDB ID: 6CZH) and transmembrane β-barrels (PDB ID: 1BXW). Glycine are yellow and water molecules are red dots. Right: distributions of the C-O-H-N and O-H-N angle values describing the hydrogen bond geometry in crystal structures.

Sequence design and initial experimental characterization

To delimit the upper and lower membrane boundaries, four tyrosine residues were placed two positions upstream of the anchor residues on the cis side, and alternating tyrosine and tyrosine/tryptophan motifs were placed at the trans boundary (Fig. 2A, supplementary text; such “aromatic girdles” are observed in native TMBs (24)). To design the remainder of the sequence, we first experimented with the approach we took for helical transmembrane proteins (6), using standard Rosetta design methods to design core residues (which results in largely hydrophobic interiors, as in helical transmembrane proteins) and re-surfacing the outside with hydrophobic residues. However, this resulted in sequences that had strong amyloid propensity (Fig. S5). To reduce the amyloid propensity and the hydrophobicity (native TMBs are usually less hydrophobic than α-helical membrane proteins (25)), we experimented with requiring all residues in the interior of the barrel (excluding the glycine kink positions) to be polar and the surface residues to be hydrophobic, resulting in the hydrophobic/polar sequence pattern characteristic of the β-sheet secondary structure but inside-out compared to water-soluble β-barrels. To help define the register between β-strands we placed tyrosine residues adopting (+60,90) rotamer angles to closely interact with the groove formed by a neighbour glycine kink (26) (the “mortise/tenon” motif (27, 28)). We placed two such motifs in the regions where defining the register shift seemed likely to be particularly important: Y69 on strand 5 where the 4-residue register shifts in cis and trans produce a larger vertical offset in the β-sheet, and Y11 on strand 1, where the β-sheet closes on itself but lacks a register-defining β-turn between the first and last strands (Fig. 2B, Fig. S6). Finally, we designed full β-barrel sequences using Rosetta combinatorial sequence design and the ref2015 energy function (29) with increased weight on the electrostatics term to favor sidechain-sidechain hydrogen bonds in the core of the β-barrels. As expected, the secondary structure of the resulting designs was accurately recapitulated by secondary structure prediction programs (Fig. 3A).

Fig. 3.

Negative design is critical for de novo TMB folding. (A) Successful design of TMBs requires reducing β-sheet propensity of the transmembrane β-strands. X axis: β-sheet propensity of the transmembrane region (calculated with RaptroX (57)). Y axis: hydrophobicity of the core (GRAVY hydropathy index (58)). Grey spheres, non-expressing TMB designs (repeated twice); black circles, expressing designs that do not fold; red, TMB designs that pass biochemical folding screening (repeated in two different detergents) -- labels indicate folded species was validated by HSQC; Green, naturally occurring TMBs with 8 strands. Circle size - aggregation propensity of the sequence predicted with TANGO (59). (B) Experimental workflow. The number of unique designs (excluding loop doublons) satisfying each criteria is shown in brackets. (C and D) Proper folding of tOmpA requires negative design against strong β-turn nucleating sequences on the trans side. Left: Rosetta energy landscapes of designs with canonical low energy (C) or sub-optimal (D) sequences substituted in a 3:5 type I β-turn with a G1 β-bulge. Conformational perturbations were generated using Kinematic Loop Closure (60); the inset shows the backbone conformations of the twenty-five lowest-energy models. Center: After refolding in 2X CMC DDM detergent, OmpTrans3 elutes on SEC similarly to tOmpA (arrow, 14.62 ml for OmpTrans3 and 14.53 ml for tOmpA) and runs as a heat modifiable species on SDS-PAGE characteristic of folded tOmpA, while the OmpAAG peak elutes earlier (13.96 ml) and does not show a band shift (band shift assay repeated three times; SEC repeated two times). Right: The far-UV CD spectrum of OmpTrans3, but not OmpAAG, is similar to that of tOmpA (repeated two times).

Folding of TMBs is chaperone-mediated and catalyzed in vivo by the β-barrel assembly machinery (BAM) complex in Gram-negative bacteria, the sorting and assembly machinery (SAM) complex in mitochondria, and the OEP80 foldase/insertase in the outer chloroplast membrane (30). Because it was unclear whether our TMB designs would be able to interact with the chaperone machinery to fold in the outer membrane of E. coli, we expressed the designed sequences in the cytoplasm, anticipating that they would form inclusion bodies that could then be solubilized in urea/guanidinium chloride (both natural and engineered TMBs have been produced in this way (31)). We obtained E. coli codon optimized synthetic genes for 9 designs (set TMB0, Fig. S7), but no protein of the correct molecular weight was produced upon the induction of protein expression (Data S2). Reasoning that the designed sequences may have had too much positive charge, which can impair translation (32), in a second round of 16 designs we reduced the number of charged residues in the core of the protein (set TMB1, Fig. S7). Again none expressed in E coli (Data S2).

Because of the failures at the expression stage, experimental feedback to improve the design methodology could not be obtained. To gain insight, we instead compared our designs to sequences of natural 8-strand TMBs. We noted two differences: first, the natural TMBs often have long and disordered trans loops rather than short β-turns (20) and second, the secondary structure propensity of their transmembrane β-strands was lower than the designs we had tried to express (Fig. 3A). We hypothesized that the high β-turn and/or β-strand propensities of our designed sequences could result in rapid formation of off-target β-sheet structures when expressed in the cytoplasm, which could be cleared rapidly or hinder growth of expressing cells.

Role of trans β-turns

We first explored the role of the trans loops in TMB folding/expression by redesigning the native TMB of the protein OmpA (tOmpA), replacing its trans loops with the canonical SDG β-turn sequence used in our designs (Fig. S8A,B). The re-looped tOmpA construct (OmpSDG) expressed at high levels in E. coli (where it was found in inclusion bodies), but it could not be correctly refolded (Fig. 3C, Fig. S10C,D). To understand this observation, we carried out Rosetta energy landscape calculations on short β-turns at the trans membrane boundary of natural TMB structures, and observed that their sequences have relatively low propensity to form β-turn structures (supplementary text) compared to the β-turns of soluble β-barrels and the SDG β-turns of OmpSDG. We constructed and tested four variants of tOmpA (OmpTrans1–4) that each contain two such 3:5 type I β-turns with suboptimal sequences (these designs are shorter than the shortest variant of tOmpA previously reported, which has trans connections of 5 to 18 residues (33)). The proteins were again expressed at high levels in inclusion bodies (Table S2), but this time all four of these sequences showed a heat-modifiable band (analyzed by cold SDS-PAGE) when folded into DDM detergent micelles and Large Unilammelar Vesicles (LUVs), characteristic of properly folded tOmpA (Fig. 3D, Fig. S10C,D). The best expressed design OmpTrans3 was characterized in more detail; following refolding in detergent micelles it had a similar retention time to native tOmpA on a Size Exclusion Chromatography (SEC) column (Fig. 3D, Fig. S12), a similar native mass spectrometry (nMS) profile (Fig. S13), well-dispersed resonance peaks by ¹H-¹⁵N-HSQC NMR in Fos-choline-12 (DPC) detergent (Fig. S10B) and a similar CD spectrum to tOmpA in both DDM micelles (Fig. 3D) and in 1,2-diundecanoyl-sn-glycero-3-phosphocholine (DUPC, diC11:0PC) LUVs with distinctive peaks at ~220 nm and 231 nm (Fig. S10A) (34). These data suggest that either long loops or short suboptimal β-hairpin loops on the trans side are necessary to slow down nucleation of the trans β-hairpins, and allow proper folding of TMBs in vitro. However, simply replacing the trans loops on four of the TMB0 designs with the extracellular loops of tOmpA or scrambled versions of these loops did not significantly increase protein expression (Fig. S9), suggesting there was a further problem with the properties of the transmembrane β-strands in the original designs.

Reducing β-sheet propensity

We sought to further use negative design (35, 36) to disfavor off-target states and slow down folding, this time through reduction of the high secondary structure propensity of the β-strands in our designs. We increased the hydrophobicity of the side-chains in the β-barrel lumen, thereby disrupting the strict alternation of polar and hydrophobic residues along the β-strands. To do so, we experimented with the design of networks of hydrogen bonds surrounded with scattered hydrophobic patches. We extended the mortise/tenon motifs to include a hydrogen bond between the tyrosine and a negatively charged Asp or Glu residue and to seed the design of the hydrogen bonds networks. Possible positions for the Asp or Glu were exhaustively searched using the Rosetta HBNet protocol (37) to design β-barrel backbones with pre-installed YGD/E motifs at one or both of the locations identified above. We used Rosetta combinatorial sequence optimization to design the remainder of the positions facing the core of the barrel, allowing all 18 amino acids other than Cys and Pro.

To further lower the β-sheet propensity, we experimented with incorporation of glycine residues (which destabilize β-strands (38)) on the hydrophobic outer-surface of the β-barrels. To guide placement of the glycines, we compared crystal structures of natural TMBs with those of water-soluble β-barrels, which rarely have water-exposed surface glycine (supplementary text). We found that the extended backbone conformation of core glycine kinks in the water-soluble β-barrels result in non-canonical out-of-plane backbone hydrogen bond geometries characteristic of a left-hand twist (O--H--N angle ~ 130°; C--O--H--N dihedral ~ −100°, Fig. 2C, top, Fig. S14A, center), while the surface residues preceding the glycine kink have more pronounced right-hand twist (C--O--H--N dihedral > 0°, Fig. S14A, right) than canonical in-plane backbone hydrogen bonds (O--H--N angle ~ 155°; C--O--H--N dihedral ~ 0° Fig. S14A, left) (39). The backbone carbonyls of glycine kinks involved in extremely out-of-plane hydrogen bonds (resulting in strongly bent β-strands) in water-soluble β-barrels is often exposed to solvent to interact with a water molecule or other hydrogen bond donor in crystal structures (Fig. S15E,F). Such exposed carbonyls are likely disfavored on the lipid-buried surface of TMBs because there are no water molecules in the bilayer to stabilize them, and TMBs indeed have a smaller population of glycine kinks and pre-glycine hydrogen bonds deviating from in-plane geometry (Fig. 2C, bottom, Fig. S14B, center). We hypothesized that glycines in positions preceding glycine kinks could allow more canonical in-plane hydrogen bonds and hence reduce unfavorable surface exposure of the carbonyls to the apolar lipid environment, and confirmed this with explicit Rosetta design calculations (supplementary text). In the subsequent sequence design calculations, we identified strongly bent glycine kinks in the β-barrel blueprint and placed glycines in surface-exposed positions directly preceding them.

Negative design of loops and strands enables de novo TMB design

We carried out three iterations of core and surface design according to the above principles using the suboptimal loops from OmpTrans3 on the trans side, allowing the backbone to relax based on the current sequence by gradient based energy minimization at each step (glycine placement in particular allows local backbone rearrangement). The design calculations converged on 20 distinct core hydrogen bond network architectures with overall amino acid composition similar to that of natural 8-strand TMBs (Fig. S7D). Codon optimized synthetic genes were obtained for several representatives of each network architecture for a total of 90 designs (set TMB2). In sharp contrast with the lack of expression in our previous unsuccessful design rounds, 66 of these designs were well expressed in inclusion bodies as intended. To test the influence of the trans loops on expression, we expressed variants of 20 of these designs incorporating the extracellular loops of tOmpA. The same designs expressed/did not express with the short sub-optimal β-turns or the long tOmpA loops (Data S2), indicating that the transmembrane β-strands - rather than the β-strand connections - carry the main sequence determinants of cytoplasmic expression (the turn sequence does matter for subsequent assembly and membrane insertion, as exemplified by the failure of OmpSDG to fold properly).

Characterization of folding, stability and structure

To test the ability of the designs to stably fold to TMB structures in vitro, we followed procedures used to fold tOmpA and other natural TMBs (Fig. 3B) (40, 41). Briefly, the inclusion bodies were dissolved in 8 M urea and rapidly diluted into DDM, DPC or n-octyl-β-D-glucopyranoside (OG) detergents at Twice the Critical Micelle Concentration (2X CMC) (Data S4). Out of the sixty-six expressing designs, sixty-two formed soluble species in such conditions. We purified the protein/detergent complexes by SEC and characterized the fifty designs which had a SEC retention volume expected for a monomeric TMB (similar to the 8-stranded tOmpA monomer and OmpTrans3) and a far-UV CD spectrum characteristic of a β-sheet protein. The band-shift assay used to monitor native TMB folding (42) was uninformative for the identification of folded de novo designed TMBs. Instead, we found a good agreement between the resistance of a design to protease digestion and thermostability up to 95 °C of the characteristic β-sheet far-UV CD spectrum. In total, twenty-three designs satisfied the biochemical screening criteria, suggesting that they fold into a TMB structure. Eleven such designs were randomly selected for analysis by ¹H-¹⁵N HSQC NMR in DPC detergent micelles, and seven had well dispersed chemical shifts profiles, characteristic of a folded protein in this detergent (Fig. S16 and Fig. S17 - validated designs, Fig. S18 - designs that failed biochemical tests, Fig. S19 - designs that passed the biochemical tests but appear misfolded by NMR).

We selected two de novo designs, TMB2.17 (highest BLAST E-value to the non-redundant protein database: 0.10) and TMB2.3 (BLAST E-value: 0.035) and the OmpTrans3 construct for detailed biophysical characterization in a lipid bilayer to determine whether the proteins exhibit the expected properties for a membrane spanning β-barrel (using tOmpA as a control). After refolding into 100 nm DUPC LUVs, all four proteins had far-UV CD spectra characteristic of a β-sheet protein both in 0.24 M and 2 M urea, and distinct from the spectra of the fully unfolded proteins in 8 M urea and from the proteins refolded in the absence of lipid (Fig. S10A, Fig. S20). We next determined the stability of the folded proteins by monitoring their ability to fold into/unfold out of LUVs at increasing urea concentrations, monitored by the change of fluorescence intensity between water-exposed and lipid embedded surface tryptophans (43). The designed TMB proteins are more thermodynamically stable (midpoint urea concentration for folding (Cm^F) 5.7 M and 7.2 M for TMB2.3 and TMB2.17, respectively, Fig. 4A) than tOmpA (Cm^F = 4.7 M), while OmpTrans3 is the most stable protein as it appears folded even in 9 M urea (Fig. S21), in agreement with the far-UV CD data. It has been previously shown that the folding/unfolding transitions of many natural TMBs exhibit hysteresis due to the high kinetic barrier to unfolding and extraction from the membrane environment (11, 44, 45). Under the conditions tested here, this behavior was observed for tOmpA but not for the designs TMB2.3 and TMB2.17 which showed superimposable and reversible unfolding/folding transitions, suggesting reduced kinetic stability relative to tOmpA. These observations likely explain the lack of a band-shift in SDS-PAGE: the de novo designs unfold during electrophoresis due to lower kinetic barriers to unfolding (46) (Fig. S22). The equilibrium unfolding curves for TMB2.3 and TMB2.17 fitted well to a two-state transition (Fig. S23) with unfolding free energies (ΔG⁰_UF) of 38 kJ.mol⁻¹ and 56 kJ.mol⁻¹. These ΔG⁰_UF values fall within the range of natural TMBs (ΔG⁰_UF 10–140 kJ.mol⁻¹ (43, 47–49)).

Fig. 4.

Folding of de novo designed TMB2.3 and TMB2.17 compared to tOmpA in synthetic lipid membranes. (A) Urea dependence of folding and unfolding in DUPC LUVs. The fluorescence intensity at 335 nm was plotted against urea concentration to determine the midpoint urea concentration for folding (C_m^F) (open circles, dashed line) and unfolding (C-_m^UF) (filled circles, solid line). Kinetics of folding into (B) DUPC and (C) DMPC LUVs at a Lipid to Protein Ratio (LPR) of 3200:1 (mol/mol) in 50 mM glycine-NaOH pH 9.5, 2 M urea at 25 °C monitored by tryptophan fluorescence at 335 nm over 30 minutes (red line). Data were fitted with a single exponential function to determine folding rate constants (black dashed line). Three technical replicates each.

The designed TMBs fold more than an order of magnitude more rapidly than tOmpA ((50), folding rate constant of 3 × 10⁻³ s⁻¹ for tOmpA); too rapid to allow accurate measurement of the folding rate constant (Fig. 4B). Tryptophan fluorescence emission spectra of the end point of the folding reactions confirmed that the TMBs were indeed fully folded (Fig. S24). To confirm that the designs integrate into the lipid bilayer rather than folding on the lipid surface or in the absence of lipid, proteins dissolved in 8 M urea were diluted into 2 M urea without lipid or into LUVs composed of 1,2-dimyristoyl-sn-glycero-3-phosphocholine (DMPC, diC_14:0PC). Consistent with previous results showing that the folding rates of natural TMBs are inversely correlated with lipid chain length (9, 51), the designed TMBs fold more slowly into lipids of longer acyl chain length (Fig. 4C), and do not fold in the absence of lipid (Fig. S25B), confirming that they indeed integrate into the lipid bilayer.

To characterize the structure of the designed TMBs in solution, we solved the structure of TMB2.3 folded into DPC detergent micelles using NMR spectroscopy (Table S3). Resonance peaks for 107 of the 117 non-proline residues of TMB2.3 were fully assigned; 6 more were partially assigned (Fig. 5A, Fig. S26A). Four out of six non-assigned residues are located in the trans β-turn regions - the remaining two are the N- and C-terminal residues of the protein. The secondary structure TMB2.3 calculated using TALOS-N (52) consists of eight β-strands that closely match the β-strand boundaries in the designed model (Fig. S26C). Nine out of eleven glycine residues pointing toward the core of the β-barrel (glycine kink residues) have the designed torsional irregularities based on the positive Cα chemical shifts [41] (Fig. S27A,B) and the more extended predicted backbone conformations (ɸ and Ѱ closer to 180°; Data S5). To validate the residue connectivity between the β-strands, we collected a total of 81 unique Nuclear Overhauser Effects (NOEs) between amide protons; these suggest 72 inter-strand backbone hydrogen bonds that are in agreement with the β-strand connectivity of the design and the shear number of 10 across the β-barrel (Fig. 5C, Fig. S26D). The NMR structure ensemble generated based on the chemical shifts and NOE information agrees closely with the design model (average backbone RMSD of 2.2 Å, Fig. 5B). We observed low-intensity additional resonance peaks for a subset of residues, indicating the presence of a (minor) secondary conformation. The secondary signals strong enough for analysis were consistent with the secondary structure assignment and NOEs of the main conformation, indicating that the secondary conformation does not involve modification of the β-barrel architecture. The residues with double peaks cluster in the cis region of strands 1, 2 and 8 (Fig. S26B); these could result from close proximity to the flexible N-terminus or transient dimeric interactions identified by native mass spectrometry in detergent micelles (Fig. S28, Fig. S29).

Fig. 5.

NMR structure of TMB2.3 in DPC detergent micelles. (A) Assigned ¹⁵N-¹H TROSY spectrum of TMB2.3. (B) NMR constraints mapped on the TMB2.3 design model. Residues predicted to have β-sheet secondary structure are colored in blue. Collected inter-residues NOEs are shown as red sticks. (C) TMB2.3 Rosetta design model (magenta) aligned to the 20 lowest energy models generated with NMR constraints (grey).

To determine structure at the atomic level, we crystallized TMB2.17 and solved the structure at 2.05 Å resolution (Table S4). All but two residues located in one trans β-turn were resolved in the electron density map. The crystal structure of TMB2.17 closely matches the design model (1.1 Å backbone RMSD over all residues, Fig. 6A), and the β-barrel has a wide lumen delimited by glycines in an extended conformation, which form kinks in the β-strands as designed (Fig. 6B,C). The two YGD/E interactions (Y69, Y11, G27, G89, D39, E103) belonging to the extended mortise/tenon motifs are present in the crystal structure and the second shell of interactions, involving K71, E53 and Q29, is also properly recapitulated with additional interactions to water molecules (Fig. 6D); these extended side-chain hydrogen bond networks fill the lumen of the β-barrel. Overall, the buried amino acid side chain conformations and interactions in the design model are in very good agreement with the crystal structure (Fig. 6E; compare pink and gray). We compared TMB2.17 to the transmembrane region of tOmpA, the only natural TMB sequence with a known structure returned by a BLAST search for sequences similar to TMB2.17 in the non-redundant sequence database (the E-value of 1.6 is in the range expected from random matches, alignment shown in Fig. S30). The shape of the β-barrel lumen is quite different in the two proteins (Fig. 6B), as are the amino acid identities and packing arrangements of the core sidechains (Fig. 6F).

Fig. 6.

Crystal structure of TMB2.17 is nearly identical to the design model. The crystal structure (pink) was determined in DPC detergent, superimposed on the design model (grey) and compared to the crystal structure of the naturally occurring tOmpA (blue, PDB ID: 1BXW). (A) Full backbone superposition. (B) Comparison of the transverse β-barrel cross-section geometries. (C) Superposition of the β-strands around a mortise-tenon motif, showing the extended backbone conformation of the glycine kink (G27) and the rotamer of the tyrosine involved in the aromatic rescue interaction (Y11) which are nearly identical in crystal structure and design model. (D) Superposition of the side-chains involved in the core network of polar interactions around the two mortise-tenon motifs. The black lines indicate the locations of the four transverse slices for which core packing is shown in panel E for the design model and crystal structure; the two are very similar. Corresponding slices in tOmpA (F) are quite different in both shape and amino acid composition and placement. Cα atoms are shown as spheres and glycine kink residues are colored in yellow; the positions of the tyrosines in the mortise/tenon folding motifs are labeled.

Conclusions

Both the initial failures and the ultimate success of our hypothesize, design and test approach to de novo TMB design inform our understanding of the sequence determinants of TMB folding and structure. The sequential approach previously used to build helical transmembrane proteins (6) - the design of proteins with hydrophobic cores compatible with folding of water-soluble proteins and subsequent hydrophobic residue re-surfacing to convert them to membrane proteins - yielded sequences strongly predicted to form amyloid. Designs with more polar cores which had high β-sheet propensity because of the perfect alternation of hydrophobic and polar residues systematically failed to express in E. coli. Iterative improvement of the design protocol ultimately enabled the generation of a set of sequences with more than 8 % of sequences encoding proteins able to adopt a β-barrel fold (based on ¹H-¹⁵N-HSQC NMR). The NMR structure and high-resolution crystal structure of two of these designs are very close to the design model. The key to this success was introducing glycine kinks, β-bulges and register-defining sidechain interactions - also critical for the folding of water-soluble β-barrels (5, 53) and hence important for defining β-barrel architecture irrespective of the solvent environment - and balancing the hydrophobicity and β-sheet propensities of the sequences.

Our results suggest that, to enable TMB expression and folding, the β-hairpins of outer membrane β-barrels need to be sufficiently unstable in water that they do not form off-target β-sheet-containing species, and become populated at high levels only in the context of the fully folded state in the hydrophobic environment of the membrane. Slowing down the folding and assembly of trans hairpins could also allow more time for passage of the mostly hydrophilic amino acids in these β-strand connections across the lipid membrane, which likely has a large activation barrier. The overall β-sheet propensity and hydrophobicity of our successful designs are in the range of those of naturally-occurring TMBs sequences, suggesting that the natural TMBs might be under a similar negative selection pressure against formation of non-native β-sheet structures in an aqueous environment (54). Further work is required to determine whether the design principles applied here enable TMB folding into biological membranes (whose properties present a formidable kinetic barrier to folding (55)). In Gram-negative bacteria, the BAM complex is responsible for accelerating the assembly of natural TMB substrates into the outer membrane by lowering the kinetic barrier to folding (55). Our design strategy incorporates neither signals for BAM complex association, such as the conserved β-signal (56), nor evolution-conserved functional motifs and hence represents a “blank slate” for probing the tradeoffs between TMB folding, stability and function, as well as the evolutionary constraints on OMP trafficking and biogenesis.

Larger TMBs share similar sequence properties with the 8-strand TMBs considered in this study (supplemental text), suggesting that the general design principles and methods we have described here should be applicable to the design of larger pore-containing β-barrels, after the generalization of the β-barrel architecture definition rules (glycine kinks, β-bulges, etc ) to different combinations of n and S. The extent to which essentially all of the key design features are recapitulated with atomic level accuracy in the crystal structure of TMB2.17 suggests considerable control over TMB structure, which should enable custom design of transmembrane pores with geometric and chemical properties tailored for specific applications.