Negative design enables cell-free expression and folding of designed transmembrane β-barrels

Significance

Membrane proteins (MPs) are essential to biology and biotechnology yet designing them from scratch remains challenging. Using synthetic transmembrane β-barrels and a cell-free expression system with lipid vesicles, we show that conventional design strategies focused solely on native state stabilization can lead to misfolding and aggregation. By incorporating “negative design” features—specific mutations that destabilize local β-strand structure—we improve β-barrel assembly in membranes without compromising stability. Remarkably, a protein language model outperformed traditional energy-based methods in predicting these beneficial mutations. Our findings highlight the critical role of folding kinetics in MP design and introduce principles for engineering folding kinetics in synthetic β-barrel nanopores.

Abstract

De novo design of membrane proteins (MPs) is a rapidly growing field with transformative potential for synthetic biology. Yet, progress has lagged behind that of soluble proteins, largely due to limited understanding of the fundamental principles governing MP folding, stability, and solubility—and the difficulty of integrating them into computational models. In cells, strict quality control mechanisms limit the expression of designed MPs with suboptimal properties, hindering iterative design-build-test cycles. Here, we use a cell-free expression system to bypass the cytotoxicity of failed or insoluble designs and investigate how sequence features influence the thermodynamically driven assembly of designed transmembrane β-barrels (TMBs) in synthetic membranes. We find that even small, idealized TMBs challenge classical protein design workflows: sequences optimized solely for thermodynamic stability misfold and aggregate, preventing membrane insertion. Aggregation in water emerges as key determinants of membrane association, instead of bulk hydrophobicity. By designing variants of a synthetic TMB, we demonstrate that suppressing aggregation-prone intermediates through local destabilization of β-strands (“negative design”) significantly improves membrane assembly. Strikingly, even substitutions typically considered highly destabilizing, such as prolines or polar threonines exposed to the bilayer core, can improve folding and assembly when strategically positioned, without significantly compromising thermodynamic stability. Based on these findings, we propose a framework for joint optimization of native stability and assembly pathways for future MP and nanopore design.

De novo protein design enables the generation of entirely new proteins with no homology to natural sequences (1, 2). Over the past decade, the field has advanced from proof-of-concept demonstrations of custom protein folds and shapes (3, 4) to the generation of fully functional proteins for diverse applications (5–9). One major approach—structure-based design—generates sequences based on 3D protein scaffolds that incorporate relevant structural or functional constraints (10). These methods typically optimize amino acid sequences for stability of the target native state, using physics-inspired energy functions, statistical potentials, or deep-learning models, while largely ignoring folding pathways and alternative energy minima (11–13). Although this simplification overlooks the complexity of protein folding, it has proven remarkably successful for designing idealized, water-soluble proteins, reinforcing the view that thermodynamics generally dominates kinetics in protein folding and design (14).

This postulate has recently been challenged by efforts to design transmembrane β-barrels (TMBs), a class of β-sheet membrane proteins (MPs) that span the lipid bilayer and form pore-like structures (15–18). TMBs are stable in membranes only when fully folded into closed β-barrel structures, in which all backbone polar groups engage in β-strand hydrogen bonding, thereby avoiding unfavorable interactions with the surrounding hydrophobic lipid chains. A central challenge for de novo design is therefore to create sequences that reliably adopt the fully connected β-barrel fold within a lipid membrane—a coupled process we refer to as “assembly.” The process requires partially folded intermediates to insert into the membrane despite transiently exposing polar groups to the hydrophobic lipid environment. As a result, and unlike soluble proteins, successful TMB design requires not only positive design (optimizing stabilizing interactions in the final folded structure) but also negative design. In globular proteins, negative design typically prevents non-native states by introducing repulsive contacts (19, 20). For TMBs, in contrast, it involves introducing a small number of features directly destabilizing the native structure, such as breaking the strict alternation of polar and hydrophobic residues along β-strands or adding disorder-promoting glycine or alanine residues. The precise mechanisms by which such subtle β-strand destabilization promotes successful β-barrel assembly remain unclear. Notably, sequences optimized solely through positive design consistently failed to express in Escherichia coli, likely due to cytotoxic behavior.

In vitro transcription-translation (IVTT) systems, such as PURExpress (21, 22), provide a powerful tool to examine the behavior of proteins that are toxic in cells (Fig. 1A). When supplemented with lipid vesicles (Fig. 1B), IVTT has been used to produce natural TMBs (23) and synthetic β-peptides (24). In this study, we use cell-free expression to dissect the role of negative design in TMB folding. We test synthetic sequences designed to adopt similar eight-stranded TMB structures but differ in biophysical properties, as well as variants carrying mutations that modulate stability, folding, and assembly (SI Appendix, Table S1). Our results suggest that negative design functions by suppressing premature folding or aggregation in aqueous solution. In contrast to conventional structure-based methods that have been broadly applied to the design of water-soluble proteins (10), we find that even minimal eight-stranded TMBs require a nuanced approach that integrates both native stability and folding pathway constraints.

Fig. 1.

Cell-free expression of transmembrane β-barrels (TMBs). (A) TMBs designed with the classic positive optimization paradigm cannot be produced in E. coli inclusion bodies. (B) Overview of cell-free expression. (C) Detection using a polycysteine reactive dye shows expression of OmpT3 (Left, with characteristic heat-modifiable band) and TMB2.3 (Right). (D) Overview of sucrose density gradient fractionation. (E) OmpT3 (Top) and TMB2.3 (Bottom) expressed with 1,2-diundecanoyl-sn-glycero-3-phosphocholine (DUPC) large unilamellar vesicles (LUVs) are enriched in the lipid-bound fraction. (F) Reduced interaction of TMB2.3 with lipid vesicles composed of longer aliphatic chain 1,2-dimyristoyl-sn-glycero-3-phosphocholine (DMPC) lipids (Top) and E. coli polar lipids (Bottom, see full boiled/unboiled analysis in SI Appendix, Fig. S4B). (G) Experimental 3D structures of Outer Membrane Protein A (OmpA) (Left, PDB ID 2GE4) and TMB2.3 (Right, PDB ID 6X1K), showing transmembrane channels detected using MOLEonline (purple). The middle panel shows the constriction formed by side-chain polar contacts across the channel of OmpA. (H) Single-channel conductances recorded for OmpT3 in Fos-choline 12 (DPC) micelles (N = 71), in vitro transcription–translation (IVTT)-expressed OmpT3 (N = 122), and IVTT-expressed TMB2.3 (N = 77). Independent recordings (dots), density distributions, median, and interquartile ranges (boxplots) are shown for each sample.

Results

Expression and Assembly of TMBs in Cell-Free Systems.

In biological membranes, TMB assembly requires a dedicated Bam complex (25). By contrast, in simplified in vitro systems, TMBs can assemble spontaneously in some synthetic lipid bilayers following a process in which β-strand formation and membrane insertion likely occur in a concerted manner (26, 27). Because this assembly proceeds without cofactors and is governed primarily by thermodynamics, it has enabled extensive studies of TMB folding from chaotropic denaturants into lipid bilayers (27–29). To validate our cell-free expression system, we selected one native and one synthetic TMB based on their high assembly efficiency in vitro (15) (see SI Appendix, Fig. S1 for sequence properties comparison). The first was the natural Outer Membrane Protein A (OmpA), engineered with shortened extracellular loops to yield the more stable variant OmpT3 (SI Appendix, Fig. S2). The second was the de novo designed eight-stranded TMB2.3, which has no detectable homology to natural TMBs and whose 3D structure has been validated by solution Nuclear Magnetic Resonance (NMR) (15). Because TMB assembly efficiency depends on membrane properties such as hydrophobic thickness (27), we expressed both proteins in the presence of 1,2-diundecanoyl-sn-glycero-3-phosphocholine (DUPC, diC_11:0PC) large unilamellar vesicles [LUVs; ~100 nm diameter (Methods and Materials and SI Appendix, Fig. S3)]. DUPC, a short-chain lipid, imposes a relatively low energy barrier to MP insertion (30). Both model TMBs were successfully expressed under these conditions, with a C-terminal polycysteine tag (CCPGCC) enabling in-gel detection using the fluorogenic biarsenical probe FlAsH-EDT₂ (31) (Fig. 1C). OmpT3 ran differently on seminative sodium dodecyl-sulfate polyacrylamide gel electrophoresis (SDS-PAGE) before and after boiling, producing a characteristic “heat-modifiable” band. This behavior is widely used as evidence of native state formation for many natural TMBs, including OmpT3’s parent OmpA (32, 33). By contrast, synthetic TMBs such as TMB2.3 typically do not exhibit heat-modifiable behavior (15), making the band-shift assay uninformative in this case (Fig. 1C and SI Appendix, Fig. S4A) and necessitating additional validation experiments.

We further investigated the interaction of the in vitro expressed TMBs with lipid membranes by separating insoluble (bottom fraction), water-soluble (middle fraction), and lipid-associated proteins (top fraction) using sucrose density gradient centrifugation (Fig. 1D). As expected, most of OmpT3 was detected in the lipid fraction, where it retained its heat-modifiable behavior, indicating successful integration and folding into DUPC LUVs (Fig. 1 E, Top). TMB2.3 showed a similar strong association with the lipid fraction (Fig. 1 E, Bottom).

To distinguish native membrane assembly from mere adsorption to the membrane surface, we tested TMB2.3 interactions with bilayers of varying thickness, which present different kinetic barriers to insertion. True membrane integration is expected to occur less efficiently in thicker bilayers (30). When TMB2.3 was expressed in the presence of the longer hydrocarbon chain lipid DMPC (1,2-dimyristoyl-sn-glycero-3-phosphocholine, diC_14:0PC), membrane association was reduced, as indicated by a more even distribution across the lipid-associated and water-soluble fractions (Fig. 1 F, Top). When expressed with E. coli native lipids [where TMB assembly requires the Bam complex (34–36)], TMB2.3 aggregated, producing a partition profile similar to that observed in the absence of lipid vesicles (Fig. 1F and SI Appendix, Fig. S4 B and C). A comparable dependence of lipid composition was observed for OmpT3 (SI Appendix, Fig. S4 B–D). Additional evidence for integration in DUPC bilayer was obtained via protease challenge (SI Appendix, Fig. S5). The transmembrane domains of both OmpT3 and TMB2.3—but not their water-accessible polycysteine tags—were resistant to proteolysis by the broad-spectrum proteinase K (SI Appendix, Fig. S5 A and B). Taken together, our results support transmembrane β-barrel assembly for both proteins in the cell-free system.

We next sought to confirm that the IVTT-expressed TMBs adopted the same conformation as those generated by detergent refolding, a validated preparative method. To do this, we compared the nanopore activity of OmpT3 produced under both conditions using single-channel electrophysiology. In this set-up, insertion of a pore into an electrically insulating lipid membrane (by LUV fusion or directly from a detergent-solubilized sample) produces a discrete current jump whose amplitude reflects the pore structural state. OmpT3 reconstituted in detergent micelles exhibited a symmetric base current of ~±8 pA (Fig. 1H). The same unitary conductance was observed for OmpT3 produced by cell-free expression, although distributions were broader and multimodal, consistent with multiple pores inserting simultaneously through a single LUV fusion event (Fig. 1H and SI Appendix, Fig. S6A). These results indicate that both TMB assembly methods yield equivalent functional states. TMB2.3 expressed in LUVs also produced robust, voltage-dependent conductance steps (Fig. 1H and SI Appendix, Fig. S6B), confirming its assembly into stable nanopores. Its higher unitary conductance (~±16 pA) relative to OmpT3 likely reflects the absence of side-chain gating seen in OmpA-derived pores (37, 38), as TMB2.3 experimental and predicted structures feature a continuous transmembrane channel [Fig. 1G, channels detected with MOLEonline (39)]. Together, these findings demonstrate that both natural and synthetic TMBs can assemble into functional nanopores directly from cell-free expression, with conductance properties consistent with their expected architectures.

Strong Stabilization of the TMB Folded State Promotes Aggregation.

Next, the cell-free system was used to express two TMB sequences designed with the classic native-state optimization approach: TMB0.1 and TMB1.1 (15). Both designs were predicted by AlphaFold2 (40) to form well-defined β-barrel structures with high confidence, whereas TMB2.3 predictions yielded only partial, low-confidence β-sheets (Fig. 2A). Despite this strong sequence-structure compatibility, neither TMB0.1 and TMB1.1 had previously been expressed successfully in E. coli as inclusion bodies, which is typically considered as the most basic success criterion for de novo TMB design.

Fig. 2.

Aggregation of TMB0.1 and TMB1.1, designed with conventional native-state optimization. (A) AlphaFold2-predicted structures of TMB0.1, TMB1.1, and TMB2.3, colored by local pLDDT scores. (B–C) TMB0.1 (B) and TMB1.1 (C) are enriched in the aggregated fraction after cell-free expression with DUPC LUVs and sucrose-gradient separation (boiled lanes are shown, full gel images in SI Appendix, Fig. S7B). (D) Hydropathy [GRAVY scale (41)] and predicted β-strand propensity [percentage of total sequence length, predicted by RaptorX (42)] for TMB0.1 and TMB1.1 compared with TMB2.3 and OmpT3. “NMR” indicates the experimentally determined β-strand content of TMB2.3 based on NMR chemical shifts (15).

Using the cell-free system, we successfully expressed the first design, TMB0.1, allowing us to directly assess its behavior and biophysical properties. Density gradient ultracentrifugation revealed that TMB0.1 strongly aggregated across all lipid conditions tested (Fig. 2B for DUPC, SI Appendix, Fig. S7 for E. coli lipids). TMB0.1 was designed with a highly charged lumen to promote pore hydration and was overall relatively hydrophilic [GRAVY hydrophobicity index: −0.541 (41)]. To test whether excessive hydrophilicity prevented lipid integration (by comparison, OmpT3 and TMB2.3 have GRAVY scores of −0.259 and −0.256, respectively), we expressed a second design, TMB1.1. This construct featured a polar but less charged lumen, resulting in a more hydrophobic sequence (GRAVY index: 0.08). Although TMB1.1 was also successfully expressed in vitro, it also failed to associate with membranes and instead aggregated (Fig. 2C). Negative staining Electron Microscopy revealed that both designs formed dense aggregates that colocalized with ribosome clusters from the cell-free reaction (SI Appendix, Fig. S8). These results indicate that strong optimization of sequence features stabilizing the native state can, paradoxically, promote cross-β aggregation rather than productive membrane assembly. Earlier work linked efficient in vitro folding of native TMBs to high overall sequence hydrophilicity (43), but our findings suggest that membrane association is not dictated by bulk hydrophobicity alone. Despite having distinct hydrophobicity (Fig. 2D), both TMB0.1 and TMB1.1 preferentially aggregated instead of interacting with lipids. We propose that this high aggregation propensity arises from relatively high β-strand content and structural order, which exceed those of TMB2.3 and OmpT3 (Fig. 2D and SI Appendix, Fig. S1).

Notably, even natural TMBs vary in their tendency to form higher-order oligomers (29, 44, 45) or amyloid-like fibers (46), processes that compete with proper membrane insertion and can reduce native β-barrel assembly efficiency over time (43). TMB0.1 and TMB1.1, however, lie at the extreme end of this membrane partitioning–aggregation spectrum: Although their sequences strongly encode β-barrel-like 3D structures, they show little or no detectable interaction with membranes. Together, these observations highlight kinetic competition between aqueous aggregation and membrane assembly as a key constraint in TMB design.

Disorder-Promoting Mutations Improve the Assembly of a Synthetic TMB.

Natural TMBs have evolved to achieve exceptional thermostability and assembly efficiency (30, 47). By contrast, synthetic TMBs often display a wide range of behaviors. Understanding how stability and assembly efficiency are jointly encoded within a sequence is therefore a critical step toward designing stable, functional MPs and nanopores that can be produced at high yield. We identified one 8-stranded design, TMB2.17 (15), that offered the opportunity to probe this tradeoff: It was highly stable but assembled inefficiently in LUVs. Its β-barrel structure had been previously validated by high-resolution X-ray crystallography, and it was the most stable design in the batch, with a folding free energy of 56 kJ mol⁻¹ in DUPC LUVs, compared to 38 kJ mol⁻¹ for TMB2.3 (15). Despite this high native state stability, TMB2.17 displayed incomplete folding in vitro, partitioning between lipid-associated, water-soluble, and aggregated fractions after cell-free expression (Fig. 3C and SI Appendix, Fig. S9). On SDS-PAGE, it consistently produced two bands: one at the expected molecular weight and another faster-migrating band (SI Appendix, Fig. S10), likely reflecting partial folding or altered detergent binding, a phenomenon among MPs (48).

Fig. 3.

ESM3-predicted mutations improving TMB2.17 assembly in LUVs. (A) Predicted β-aggregation hotspot in the first β-strand of TMB2.17. (B) Predicted effects of single-point substitutions along β-strand 1 on folding free energy (Rosetta ΔΔG, Left) and evolutionary fitness (ESM3 log-likelihood ratio (LLR), Right). [*indicates the wild-type (WT) residue at each position]. (C) Representative SDS–PAGE gels showing partitioning of protein between lipid-associated and water-soluble fractions after IVTT and sucrose-gradient separation. (D) Quantification of membrane association, calculated as Protein_lipid/(Protein_lipid + Protein_soluble). Membrane association increases from 58% for WT TMB2.17 to 71 to 75% for variants V12T and V10P and remains at ~55% for V12P (3 to 8 biological replicates (dots); error bars show SD; *P < 0.05; **P < 0.01).

We hypothesized that a local region in TMB2.17 retained a strong β-strand signature despite our prior negative design efforts, potentially interfering with β-barrel assembly. Aggregation propensity analysis using Tango (49) revealed a hotspot in the first β-strand, driven by clustered valines (V8, V10 to 12), with an aggregation score ~threefold higher than the corresponding region in TMB2.3 (Fig. 3A and SI Appendix, Fig. S1). In silico mutational scanning with Rosetta cartesian_ddg protocol (50) failed to identify substitutions that reduced the aggregation score while preserving the predicted stability (Fig. 3B and SI Appendix, Fig. S11). We therefore turned to the ESM3 protein language model (PLM) (51), which predicts evolutionary fitness rather than thermodynamic stability. Unexpectedly, several substitutions at V10 and V12 that Rosetta predicted to be highly destabilizing were scored as favorable by ESM3. These included replacements with β-strand disrupting residues (proline, glycine, or alanine) as well as polar residues (threonine or serine) introduced at the otherwise hydrophobic barrel surface (Fig. 3B). These substitutions reduce aggregation propensity and increase predicted disorder within β-strand1 [IUPred3 (52), SI Appendix, Table S1], though at the expense of Rosetta-predicted native state stability.

Because TMB assembly in vitro has been described as a concerted insertion and folding process (27, 28) involving a partially disordered, membrane-embedded, transition-state ensemble (29, 53), we reasoned that premature ordering of β-strand1 (either through early folding or aggregation) may interfere with this ensemble. To test this hypothesis, we constructed five variants (V12T, V10G, V12G, V10P, V12P) and expressed them in vitro with DUPC LUVs. We analyzed their partitioning on sucrose density gradient using SDS-PAGE and quantified membrane association as the intensity of the bands in the lipid fraction divided by the sum of intensities in the lipid and soluble fractions. Variants V10P and V12T showed significantly enhanced membrane association, whereas V10G, V12G, and V12P behaved similarly to WT TMB2.17 (Fig. 3 C and D and SI Appendix, Fig. S12). A control substitution, V10L, slightly reduced membrane association. Several β-strand-disrupting variants also decreased the intensity of the faster-migrating SDS-PAGE band, consistent with destabilization of a partially folded intermediate involving β-strand1 (SI Appendix, Figs. S13 and S14).

Together, these results show that residual β-aggregation hotspots can persist even in designs that incorporate negative design principles, hindering proper membrane assembly. Such hotspots were likely overlooked in earlier folding studies of TMB2.17 performed in 2 M urea, a condition that may destabilize the relevant intermediate. Mutations that disrupt these regions improve TMB assembly and were more accurately predicted with ESM3 (evolutionary fitness) than Rosetta (energy-based stability), underscoring the importance of kinetic factors in TMB assembly that traditional stability models do not capture.

TMB2.17 Variants Preserve Native Structure and High Stability.

Next, we evaluated how assembly-enhancing mutations affected the structural integrity and stability of purified proteins refolded in detergent. The five TMB2.17 variants were expressed in E. coli as inclusion bodies and subsequently refolded in Fos-choline 12 (DPC) at twice the critical micelle concentration (CMC; Methods and Materials). Size exclusion chromatography (SEC) showed that all variants eluted as major peaks with retention volumes similar to the WT TMB2.17 (SI Appendix, Fig. S15). The most notable shift was observed for V12P, potentially indicating structural changes or differences in detergent interaction. To assess secondary structure content, we performed far-UV circular dichroism (CD) spectroscopy. All five variants displayed CD spectra characteristic of all-β proteins and were nearly superimposable with that of WT TMB2.17. Deconvolution using ChiraKit (54) revealed a slight increase in stable β-sheet content for variants V12T (42%) and V10P (41%) relative to TMB2.17 (39%) (Fig. 4A and SI Appendix, Fig. S16 A and B). In contrast, variants V10G and V12G showed reduced β-sheet content (33% and 34%, respectively; SI Appendix, Fig. S15 C and D), consistent with the introduction of disorder-promoting glycine. Variant V12P exhibited a small but distinct spectral shift at 230 nm (SI Appendix, Fig. S16E), suggesting a conformational change affecting the local environment of lipid-facing tryptophan residues. To examine tertiary structure, we used NMR spectroscopy. ¹H-¹⁵N HSQC spectra of four variants—V10G, V12G, V12T, and V10P—showed well-dispersed chemical shifts consistent with a folded β-barrel and resembling the spectrum of WT TMB2.17 (Fig. 4B and SI Appendix, Fig. S17 B–E). In contrast, clustered chemical shifts were observed for V12P. We further validated these results with ¹H NMR that allows rapid, direct comparison across samples and minimizes experimental variability. TMBs exhibit characteristic ¹H NMR signals between 8.5 to 10.5 ppm, which were clearly present in the spectra of V10G, V12G, V12T, and V10P (SI Appendix, Fig. S17A). By comparison, the spectrum of V12P lacked these characteristic signals, consistent with disrupted β-barrel formation (SI Appendix, Fig. S17A).

Fig. 4.

Biophysical characterization of TMB2.17 and its variants. (A) Far-UV CD spectra of variants V12T (Left), V10P (Right) (gray) overlaid to the spectrum of TMB2.17 (purple). (B) HSQC NMR spectra of TMB2.17 and V12T are characteristic of TMB proteins. (C) Thermal denaturation experiments in 3 M urea show cooperative two-state unfolding transitions. V12T and V10P mutations retain high thermostability.

Thermal stability in DPC micelles was assessed by monitoring intrinsic tryptophan fluorescence during thermal denaturation. All proteins remained stable up to 95 °C and required 3 M urea for detectable unfolding, following apparent two-state transitions (Fig. 4C). Under these conditions, WT TMB2.17 had a melting temperature (Tm) of ~92.5 °C. The variants showed only modest reduction in thermal stability, with ΔTm values of 3.2 to 4.5 °C. We further evaluated stability across urea concentrations ranging from 2.5 to 7 M for TMB2.17, V12T, V12G, and V10P. All proteins unfolded between 6 and 7 M urea (SI Appendix, Fig. S18).

In summary, four ESM3-identified mutations predicted to be destabilizing by traditional stability models preserved the high thermodynamic stability and structural integrity of the TMB2.17 scaffold. Remarkably, V12T and V10P also enhanced assembly in LUVs in the cell-free system. By contrast, the V12P mutation disrupted β-barrel formation, as evidenced by subtle changes in SEC and CD profiles and a strong loss of NMR signal. These findings highlight the difficulty of distinguishing properly folded TMBs from near-native, molten states using standard biophysical methods alone.

Conclusion

Our results show that designing TMBs requires more than optimizing thermodynamic stability. Positive design alone can redirect the pathways toward misfolding or aggregation, whereas incorporating negative design elements can delay premature structural ordering and promote productive membrane insertion (Fig. 5). Even so, residual β-aggregation hotspots can persist in designs that otherwise correctly assemble into β-barrel structures and exhibit high thermodynamic stability. This is illustrated by TMB2.17, in which a strong β-strand signature in the first strand created a bottleneck for membrane integration under the cell-free conditions used in this study. Selectively destabilizing this region with disorder-promoting mutations improved assembly in LUVs while only modestly reducing overall thermodynamic stability, suggesting that the membrane environment can buffer these destabilizing substitutions and revealing gaps in our understanding of protein–lipid interactions.

Fig. 5.

Proposed model illustrating the interplay between native state stability and folding kinetics in designed TMBs. (A) Classic native-state energy optimization can drive premature β-strand formation and promotes aggregation in water. (B) Incorporating negative-design features locally destabilizes secondary-structure elements, delaying misfolding in aqueous solution and enabling productive membrane insertion and assembly.

We also find that the fitness-modeling ESM3 PLM outperforms traditional thermodynamic stability approaches for identifying such negative-design mutations. It is important to emphasize that many PLM-identified mutations in this study correspond to classical β-strand–disrupting residues (e.g., proline, glycine), which modern PLMs tend to score with high confidence (either strongly favorably or strongly unfavorably). Improving this approach, particularly for detecting more subtle or context-dependent substitutions, will likely require PLMs fine-tuned on mutation–fitness datasets specific to MPs, which are currently limited.

Finally, we note that our experiments were conducted entirely in vitro. Thermodynamically driven assembly of TMBs in synthetic membranes differs fundamentally from their biogenesis in vivo, which depends on coordinated interactions with chaperones and foldases such as Bam. Whether synthetic TMBs (and the negative-design mutations identified here) are compatible with different in vivo assembly pathways remains unknown. Because PLMs distill evolutionary constraints from natural sequences, future work may explore how PLM-identified mutations influence the trafficking and membrane-insertion profiles of designed TMBs in cellular environments. Interestingly, a recent study in yeast reported deficient mitochondrial import and outer membrane insertion of TMB2.17 relative to TMB2.3, similarly to their relative assembly efficiencies observed in vitro in this study (55).

Together, these design principles underscore the importance of computational strategies that integrate both thermodynamic and kinetic considerations. Future approaches may benefit from combining physics-inspired models with PLMs capable of evaluating sequence fitness, particularly as fine-tuned or next-generation PLMs become supported by larger mutation–fitness datasets. Such advances will be especially important for designing larger or more complex β-barrels, which remain challenging for traditional methods due to computational limitations. Extending design capabilities to these systems could enable the creation of synthetic nanopores suitable for sequencing and molecular sensing. More broadly, the principles identified here may inform the design of other MPs, dynamic soluble folds, or intrinsically disordered proteins.

Materials and Methods

TMB Expression in Inclusion Bodies and Refolding in Detergent.

TMBs were produced in E. coli as inclusion bodies by overnight cultures incubation in Studier autoinduction medium (56) (37 °C, 150 rpm). The cells were harvested, disrupted by sonication, and inclusion bodies were isolated following a previously described approach (16). The final pellet was dissolved in 6 M guanidine chloride and the TMBs were refolded by fast dilution into a buffer containing 0.1 % Fos-cholin12 (DPC) as previously described (15).

Liposome Preparation.

LUVs of DUPC, DMPC, or E. coli Polar Lipid Extract were prepared by dried film suspension, three freeze/thaw cycles in liquid N₂ and extrusion through 100 nm membranes. Extruded liposomes were analyzed using a DynaPro NanoStar DLS instrument to determine their size.

Cell-Free Protein Synthesis.

The PurExpress^® in vitro protein synthesis kit (New England Biolabs) was used according to the manufacturer’s instructions and for a total reaction volume of 30 µL. The reaction mix was supplemented with liposomes for a final concentration of 8 mg/mL and with RNAse inhibitors. The IVTT reaction was initiated by adding 200 ng of plasmid DNA (200 ng/µL) and allowed to proceed for 3 h at 37 °C.

Sucrose Density Gradient Centrifugation.

Sucrose solutions were prepared as described (22). The gradient was prepared by depositing 50 µL of 60% sucrose in a 175 µL Ultra-Clear Tube (5 × 20 mm, Beckman Coulter), overlaid with 100 µL of 30% sucrose. Following IVTT, 30 µL of the PurExpress in vitro synthesis reaction were mixed with 20 µL of assay buffer and gently layered on top of the gradient. The tubes containing the gradients were centrifuged 3 h at 250,000 g to separate the lipid vesicles (floating at the interface between 30% sucrose and 0% sucrose layers) and the unreacted components of the IVTT kit at the bottom of the gradient. After centrifugation, 75 µL, 75 µL, and 50 µL were collected from the top to the bottom of the gradient to separate lipid-bound, water-soluble, and aggregated fractions, which were analyzed by SDS-PAGE.

SDS-PAGE and FLAsH-EDT₂ Labeling.

Samples were boiled (10 min, 95 °C) and cooled down to room temperature before adding 25 mM TCEP and 0.1 µM FLAsH-EDT₂ dye. After 30 min incubation at room temperature, the samples were analyzed on 12% Tris-Tricine SDS-PAGE gels. Visualization was performed on an Odyssey M scanner (LICORbio) with the following settings: excitation: 685 nm; emission: 721 to 740 nm to visualize the ladder. Excitation: 488 nm; emission: 519 to 543 nm for the labeled proteins. Brightness and contrast adjustments, when applied, were performed uniformly across the entire image. Images were cropped only to focus on the analyzed bands. In some scans, vertical line artifacts might originate from dust, water, or SDS crystals contaminating the fluorescence scanner calibration surface.

Pulsed Protease Challenge.

All protease challenges were performed directly on the cell-free reaction. Trypsin was used at 0.1 mg/mL. For the positive control, 2 M of urea was added to the sample prior to trypsin digestion. The reactions were incubated at 37 °C for 3 h and stopped by boiling the sample for 10 min at 95 °C with Laemmli sample buffer. Proteinase K was used at 0.3 mg/mL. The reaction was incubated at room temperature for 30 min and stopped by incubating the sample at 70 °C for 10 min.

Electrophysiology.

Single-channel conductance measurements were performed using the Orbit mini instrument from Nanion Technologies (150 μm MECA 4 chips). Planar bilayers of Di-Phytanoyl-Phosphatidyl-Choline were painted to achieve capacitance between 10 to 20 pF. Refolded proteins were diluted to a final concentration of ~0.25 mg/mL, and 0.2 µL of the proteins were added to the cis chamber of the chip. For cell-free expressed samples, 1 µL of the reaction was added to the cis chamber without further sample processing. Insertion pulses (100 to 200 mV) of 100 ms were used to help protein insertion. Spontaneous insertions of the protein were recorded at a temperature of 25 °C using the Element Data Reader 4 software, setting a sampling rate of 10 kHz.

Site-Directed Mutagenesis.

The primers used for site-directed mutagenesis are listed in SI Appendix, Table S1. Mutagenic PCRs were done using PrimeStar GxL polymerase according to the manufacturer’s instructions and using 20 ng of template DNA, followed by DpnI treatment and religated using Gibson assembly.

Negative Stain Electron Microscopy.

For negative stain data, 2 µL of the sample was applied on a glow discharged continuous carbon EM grid (FCF400-CU, EMS). After three short wash and blot steps with MilliQ water, the grids were immersed in 20 µL drops of 2% uranyl acetate for 10, 2, and 60 s with short blotting in between. Grids were visualized with a JEOL JEM-1400Plus electron microscope operating at 120 kV and equipped with a LaB6 cathode. Images were recorded on a CMOS TemCam-F416 camera (TVIPS, Germany) at a nominal magnification of 2,000 (pixel size 5.9 nm) and 60,000 (defocus of approximately 2.5 μm and a corresponding pixel size of 0.19 nm).

Isotopic Labeling for NMR Experiments.

TMBs were produced as inclusion bodies in M9 minimal media supplemented with 1 g/L of ¹⁵N-NH₄Cl isotope. Protein expression was induced with 1 mM IPTG at OD600 0.8 to 1.0. TMBs were reconstituted in DPC micelles as described above, concentrated to ∼200 µL and dialyzed 20 mM sodium phosphate buffer, 0.1% DPC, pH 6.8. All samples were centrifuged for 2 min at 8,000 rpm before the measurement. Unlabeled samples for ¹H NMR were produced in Studier autoinduction medium, as described above, and used at concentrations between 50 to 100 μM.

NMR Spectroscopy.

Protein samples were prepared in 3 mm NMR tubes and contained 20 mM sodium phosphate, 0.1% DPC (pH 6.8) and 6% D2O for the lock. All NMR spectra were acquired at 298 K on a Bruker Avance III HD 800 MHz spectrometer equipped with a TCI cryoprobe for enhanced sensitivity. One-dimensional ¹H spectra were recorded with 16 ppm spectral width, 4.7 ppm pivot, and 256 scans; water suppression was achieved by gradient-based excitation sculpting. Two-dimensional (¹H, ¹⁵N) BEST-HSQC spectra were recorded with 800 and 256 points, 16 ppm and 35 ppm spectral widths, and 4.7 ppm and 117 ppm pivots in the direct and indirect dimensions, respectively, and 256 scans. All experiments were acquired and processed in Burker TopSpin 3.7.

Thermostability Assay.

TMBs refolded in DPC micelles were prepared at a concentration of 1 mg/mL and supplemented with urea at concentrations ranging from 2.5 M to 7 M. 9 µL of sample were analyzed for tryptophan fluorescence using an Uncle instrument (Unchained Labs), ramping up the temperature 25 °C to 95 °C at a rate of 0.5 °C/min (number of acquisitions: 5; acquisition time: 5 s). A two-state sigmoidal transition was fitted to the recorded change of tryptophan fluorescence ratio (F_350nm/F_330nm) to extract the T_m.

Circular Dichroism.

CD spectra were recorded for TMBs refolded in DPC micelles at a concentration of 0.15 mg/mL, at a temperature of 25 °C and with the following settings: wavelength range = 190 to 260 nm; wavelength steps = 0.5 nm; Acquisition period = 0.5 s; number of repeats = 5. The spectra were analyzed with ChiraKit (54).

Structural Modeling and Bioinformatics.

Alphafold2 (31) was used through the ColabFold implementation to assess TMB structure with the following settings: recycles=3;single_sequence=True; template_PDB=False and AMBER_relaxation=True. The Rosetta cartesian_ddg (50) protocol was used to evaluate the impact of point mutations on protein stability. The input structure (TMB2.17, PDB ID: 6X9Z) was prepared using Rosetta FastRelax in cartesian mode to ensure a physically realistic starting conformation. Three independent cartesian_ddg trajectories were calculated, averaged for each tested substitution and reported in a heat map. Aggregation-prone regions in protein sequences were identified using TANGO (49). Predictions were performed using the following conditions: pH 7.0, temperature 298 K, and ionic strength 0.02 M. A standalone version of RaptorX-Property was implemented for secondary structure prediction (42). All predictions were executed with no-profile mode, which does not rely on multiple sequence alignments. ESM3 was used to predict mutational pseudolikelihoods using the “esm3-sm-open-v1” model with openly available weights retrieved from HuggingFace (https://huggingface.co/). The transformer model was initialized in PyTorch using the ESM Python modules. The sequence was also tokenized using the “EsmSequenceTokenizer” module, then in silico site saturation mutagenesis was performed from the beginning to the end of the sequence with likelihoods converted to LLRs.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The experimental and simulated datasets, the scripts, and the Colab notebooks used for data analysis are deposited on GitHub (57) and the raw data archived in Zenodo (58).