Previous work from the Eisenberg group^17,18 demonstrated that RNase A variants containing short amyloidogenic peptide insertions in the C-terminal hinge loop could form amyloid-like fibrils in vitro. We sought to investigate whether or not these RNase A variants could also access the fibrillar form when produced in E. coli cells. To do this, we constructed a plasmid vector that directed the inducible synthesis of RNase A (with or without an inserted amyloidogenic peptide sequence; see Methods) fused to a monomeric variant of GFP (Fig. 1a ). Cells containing these plasmids were examined by fluorescence microscopy at various times after the induction of fusion protein synthesis. Whereas most cells contained diffuse fluorescence, we observed fluorescent rod structures in a small minority (1 to 2%) of cells. Such rod-containing cells typically contained a single rod (Fig. 1b, left). These structures were observed regardless of whether or not the RNase A moiety contained the amyloidogenic peptide insertion (Fig. 1b, left and data not shown). Furthermore, we found that refrigeration dramatically stimulated formation of the rod-like structures. When we grew cells overnight after the induction of fusion protein synthesis and then incubated the cell cultures for 2 hours at 4^oC, we detected fluorescent rod structures in essentially all of the cells (Fig. 1b, right).

To define the sequence requirements for intracellular rod formation, we constructed a series of truncations from the C-terminus of RNase A (Fig. 1a). We found that as few as eight N-terminal residues of RNase A (separated from the C-terminus of GFP by a three-residue linker) sufficed to permit rod formation. When cell cultures containing this construct were induced overnight and incubated for 2 hours at 4^oC, essentially all of the cells contained fluorescent rods (Fig. 1d). In most cases, the rods extended the entire length of the cell.

We considered the possibility that the propensities of the constructed GFP fusion proteins to form fluorescent rod structures was due to the fortuitous presence of an amino acid segment with high amyloid fiber-forming propensity (termed an HP segment) either spanning the fusion junction or contained wholly within the appended sequence. We discovered an HP segment comprising the 3-residue linker and the first 3 residues of RNase A (AMAKET) that was shared by our rod-forming constructs (Fig. 1a and Fig S1)¹⁹.

Results from mutagenesis experiments were inconsistent with the hypothesis that rod formation was due to the amyloid-forming propensity of the HP segment located at the N-terminal end of the appended peptide sequence. Using a fusion protein containing RNase A residues 1–14 as our reference construct, we compared the rod-forming behaviors of the reference fusion [GFP-AMA-RNase(1–14)] and several variants bearing replacements of the linker methionine residue (residue 240 in the fusion sequence) with A, E, K, L or Q (Fig. 1c, Fig. S1). Whereas the reference construct gave rise to rods in essentially all of the cells after overnight induction of fusion protein synthesis (Fig. 1c), the variant encoding AQAKET, which has a higher predicted amyloid-forming propensity than AMAKET, gave rise to no rod-containing cells even after induced overnight cultures were refrigerated for an additional 24 hours (Fig. 1 c). Among the remaining variants, only that encoding ALAKET gave rise to rods (which formed efficiently only after refrigeration) (Fig. 1c), suggesting that a hydrophobic residue was required at the second linker position (see below). Overall, our analysis of this set of single amino acid substitution variants suggests that rod formation is not due to the amyloid-forming propensity of the appended peptide. Conclusive evidence that the rods lack amyloid architecture follows from diffraction analysis and crystal structure determination presented below.

To elucidate the intermolecular interactions that guide assembly of GFP into intracellular fluorescent rods, we crystallized a fusion protein containing RNase A residues 1–8 [GFP-AMA-RNase(1–8)] (Fig. 1d), the shortest fusion protein that efficiently formed rods. To facilitate purification of the fusion we appended a C-terminal His₆ tag [GFP-AMA-RNase(1–8)-His₆] after determining that the tag did not alter the ability of the fusion protein to form fluorescent rods (Fig. 1e).

We discovered crystals in dozens of conditions, all of which shared the same needle morphology, recapitulating the rods observed in cells (Fig. S2a). The needles were typically over 100 μm long, but seldom greater than 5 μm thick. Among the different conditions, we discovered four different crystal forms (P2₁2₁2 form1, P2₁2₁2₁ form 2, C2, and P2₁) (Fig. S2b), collected diffraction data, and solved their structures (Table 1).

Remarkably, the structures obtained from the four crystal forms revealed essentially identical protofilament architectures, suggesting this architecture is likely to persist under many other conditions, including those within the cell. In each crystal form, the protofilament is assembled with 2₁ screw symmetry, meaning that each protomer is related to its closest neighbor protomer by a 180° rotation around the protofilament axis and translation along the axis by half the pitch length (i.e. rise per turn along the protofilament axis) (Fig 2a and c). In each crystal form, this pitch length is approximately 51.4 Å, (less than 1 Å variation) and corresponds to one of the unit cell dimensions (a=51.2 Å in P2₁2₁2₁ form 1, a=51.1 Å in P2₁2₁2₁ form 2, b=51.3 Å in C2, and b=51.8 Å in P2₁) (Fig. 3b). There is no significant structural variation in protofilament architecture among crystal forms (Fig. 2b). Indeed, the root mean square deviation is less than 0.8 Å between protofilament segments from different crystals (composed of three consecutive GFP-AMA-RNase(1–8)-His₆ protomers totaling 5667 atoms). Thus, the robust nature of this protofilament assembly is evidenced by the high degree of similarity in protofilament structures (Fig. 2b) despite clear differences in packing between protofilaments (Fig. 3b).

A closer view of the protofilament reveals that one structural feature in particular stabilizes protofilament assembly--a new, fusion-created, C-terminal helix which fastens together neighboring pairs of protomers by binding in a groove between them (Fig. 2a, c, and d). This new C-terminal helix comprises seventeen residues which span a sequence derived from three different origins: six residues from the natural GFP C-terminus (233-MDELYK-238), a three-residue linker (239-AMA-241), and eight residues from RNase (242-KETAAAKF-249) (Fig. 2d). The six residues contributed by GFP are normally disordered in crystal structures of natural-length GFP. The propensity of this segment to form a helix was evidently enhanced by its fusion with the RNase residues 1–8, which also form a helix in native RNase.

The new, fusion-created helix extends up to 40 Å away from its attachment to the barrel’s last strand, enabling the helix to reach a binding groove that is formed by the following two protomers in the protofilament. This 40 Å reach is thus directed by the relative positions of three successive protomers in the protofilament; the helix originates from protomer i, and the groove is formed between protomers i+1 and i+2. The ability of the helix to reach this groove is facilitated by the extended conformation adopted by GFP residues 228-GITHG-232, which connect the helix to strand 11, the last strand in the GFP barrel (Fig. 2d).

Protomer i+1 forms one wall of the groove (Fig. 2e, brown surface). This wall comprises the surfaces of strands 3, 10, and 11 near the waist of the barrel, and the helix docks tangentially across these. The helix-in-groove interface is snug and defined almost entirely by side-chain-to-side-chain contacts. A stripe of hydrophobic side-chains (M233, L236, Y237, M240, and T244) lines the length of the α-helix and these embed between side-chains on the surface of the barrel (T43, K45, A206, S208, V219, and L221). Loss of this hydrophobic interface upon substituting M240 with polar, charged, or small residues, such as Q, E, K, or A, would likely explain our observation that these substitutions interfere with efficient rod formation (Fig. 1c) (Fig. S3). This interface between helix i and protomer i+1 buries a total of 727 Ų on the two surfaces (Fig. S4).

Protomer i+2 forms the opposite wall of the groove. This wall comprises the end of strand 5 and residues at one cap of the barrel (Fig. 2e, blue surface). Specifically, residues L7, P89, and F114 create a small hydrophobic pocket into which docks the side chain of A245 of the helix of protomer i. This contact is flanked on both sides by salt bridges (protomer i+2 residues E90 and D117 with helix i residues K242 and K248, respectively). This interface between protomer i+2 and helix i buries a total of 426 Ų on the two surfaces (Fig. S4). Overall, the shape complementarity between the helix and groove (0.78) indicates that the fit is even tighter than typically observed among antibody-antigen complexes²⁰.

Direct contact between the sides of neighboring GFP barrels strengthens and rigidifies the contiguous helix-in-groove interaction (Fig. 2f). The interface is approximately flat. Side-chains from strands 1, 2, 5, and 6 of protomer i+1 nestle with side-chains from strands 3, 10 and 11 of protomer i+2. The interactions are primarily polar rather than non-polar. The area buried by the two surfaces of this interface is 901 Ų (Fig. S4). This patch of direct contact between barrels appears to be adventitious; it does not correspond in any way to the natural dimer interface of wild-type GFP²¹.

The combined surface area buried by protomers i, i+1, and i+2 appears large enough to maintain protofilament architecture in solution, free of the crystal lattice. Coordination of neighboring protomers by the C-terminal helix buries a total of 1153 Ų. Direct contact between protomers buries an additional 901 Ų of surface area. Either interface by itself would likely be insufficient to support protofilament assembly. But, together, these total to 2054 Ų, which exceeds the 1712 Ų threshold value that discriminates between biological and artificial dimers²². The division of the protomer-protomer interface over three surfaces implies cooperativity in protofilament assembly.

As a further test of whether or not the intermolecular interface observed in the crystals is relevant to rod formation in the cell, we mutated a solvent exposed residue on the GFP barrel that was predicted to contribute significantly to the interface formed by the binding of the C-terminal alpha helix of the neighboring protomer. Specifically, we introduced a charged residue (E) in place of V219, which should disfavor the docking of M233, L236, and M240 in the C-terminal alpha helix (Fig. S3). We then induced the synthesis of GFP-AMA-RNase(1–8)-His₆ with or without the V219E substitution in the GFP moiety. Whereas the unmutated fusion protein formed rod structures in essentially all of the cells after overnight induction followed by incubation for 2 hours at 4^oC, the mutant fusion protein failed to form rods even after overnight cultures were refrigerated for several days, and instead gave rise to diffuse fluorescence (Fig. 1f and Fig. S5). These findings provide strong support for the hypothesis that the intermolecular interface observed in the four crystal forms is also the basis for fluorescent rod formation in bacterial cells.

We performed in cellulo diffraction experiments to evaluate the pattern of protofilament bundling in fluorescent rods produced in E. coli. X-ray diffraction patterns from E. coli producing GFP-AMA-RNase A(1–8)-His₆ revealed six reflections with Bragg spacings between 55 Å and 25 Å (Fig. 3c). Remarkably, these spacings observed from intracellular filaments coincide with those of the crystal structure of purified GFP-AMA-RNase(1–8)-His₆ in space group C2 (Fig. 3e). Moreover, the relative intensities of the six observed reflections correlate between the observed and calculated patterns, as well as the geometric disposition with respect to the meridional and equatorial directions (Fig. 3f). Structurally, we find that the Bragg planes for these six reflections coincide well with the spacings between protomers in the crystal (Fig. S6a-g). The sum of these six Fourier terms produces a low-resolution image of the protofilament in close agreement with the crystal structure (Fig. S6h). These correlations, like matching fingerprints, indicate that the protofilaments bundle together in the fluorescent rods in the same pattern as observed in the C2 crystal (Supplementary note 1).

Poor correlation is observed between the in cellulo diffraction and the remaining three crystal forms (P2₁, P2₁2₁2₁ form 1 and P2₁2₁2₁ form 2; Fig. 3d). This poor correlation is not an indication of dissimilarity among the protofilament structures. Recall that all four crystal forms are composed of the same protofilament structure (Fig. 2b and 3a). Rather, the differences among the simulated diffraction patterns arise from alternative packing interactions between structurally identical protofilaments (Fig. 3b). Apparently, the cellular milieu favors the same packing of protofilaments as the C2 crystal form.

The symmetry of the protofilaments themselves likely makes them prone to crystal-like bundling, because it corresponds to one of the most prevalent symmetry elements in crystallography—a two-fold screw axis. Fibers of hemoglobin E6V are also rigidified by crystal-like bundling²³, and also feature a two-fold screw protofilament symmetry²⁴. Conversely, many functional protofilaments, such as actin and tubulin, have resisted crystallization because their symmetry is imprecise or does not coincide with common crystallographic symmetry elements.

GFP fusions were produced from pBR322-derived plasmids under the control of the arabinose-inducible promoter pBAD. Specifically, the constructs depicted in Fig. 1 (top to bottom) were produced from plasmids pLM163, pLM168, pLM180, pLM179, pLM177, pLM178, pLM176, pLM187, pLM192, and pLM198. Plasmid pLM136 encodes the full-length RNaseA fusion with the amyloidogenic peptide GGVVIA (derived from Aβ) inserted between RNaseA residues 113 and 114¹⁸. Constructs with only eight RNaseA residues included an additional aspartate (D) residue with or without a C-terminal hexa-histidine tag. Overnight cultures of E. coli BW27785^49,50 transformed with the appropriate GFP-fusion construct were diluted to an OD600 of 0.02 in 50 mL LB supplemented with carbenicillin (100 μg/mL), grown for 30 min at 30 °C, and induced with L-arabinose (0.2% wt/vol). For experiments in which early stationary phase cultures were examined, cells were harvested after 4 h induction at 30 ºC. Otherwise, cultures were induced overnight (~20 h) at 30 ºC. Where indicated, both early stationary phase and overnight cultures were refrigerated at 4 ºC for 2 h or ≥24 h in order to enhance formation of GFP rods.

Cells were imaged after overnight induction of fusion protein synthesis either without any subsequent incubation or following incubation for the time specified at 4^oC. Cells were harvested at 3,000 × g, resuspended in PBS, and spotted onto agarose pads (1% wt/vol in PBS; Seakem LE Agarose, Lonza) for visualization with a UplanFLN 100× phase contrast objective on an Olympus BX61 microscope as described elsewhere⁵¹. Images were captured with a CoolSnapHQ camera (Photometrics) and the Metamorph software package, version 6.1 (Universal Imaging). Exposures were typically 50–100 ms. Images were cropped and adjusted using Metamorph or ImageJ⁵².

Cultures (50 mL) were grown as described above, and the cell densities were recorded (OD600). The cultures were centrifuged 10 min at 3,000 × g at 4 ºC, and the pellets were resuspended in BugBuster Protein Extraction Reagent (EMD Millipore) to normalize for cell density (800μL of lysis buffer for an OD600 of 0.8.). rLysozme (EMD Millipore) was added to 3,000 units/mL and Omnicleave endonuclease (Epicentre) was added to 200 units/mL; the mixture was incubated, rocking, at room temperature (RT) for 30 min. For SDS-PAGE and Western blot analysis, samples were diluted 1:25 in standard SDS-PAGE loading buffer and equal volumes were loaded. Blots prepared in duplicate were probed with either an anti-His mouse monoclonal antibody (Genscript A00612), to detect His₆-tagged GFP-fusion protein, or with an antibody specific for the RpoA (α) subunit of E. coli RNA polymerase (Neoclone, clone 4RA2), which served as a loading control. Blots were detected using an ECLplus Western Blotting Detection System (GE Heathcare).

E. coli strain BW27785 was transformed with pLM192 (pBAD-gfp-AMA-KETAAAKFD-his₆) to produce the GFP fusion, GFP-AMA-KETAAAKFD-His₆, for crystallization. We term this fusion protein GFP-RNase(1–8). A culture of LB supplemented with100 μg/mL carbenicillin was inoculated with a single transformed colony and grown at 30 ºC overnight. This overnight culture was back diluted to an OD₆₀₀ of 0.03 in 1 L of fresh LB medium supplemented with 100 μg/mL carbenicillin and grown at 30 ºC 220 rpm until the culture reached mid-log phase (OD₆₀₀ 0.5). L-arabinose was added to a final concentration of 0.2%, and the culture was grown at 30 ºC at 220 rpm for an additional 5 h. Cells were harvested from a 500 mL aliquot of this culture by centrifugation and stored at −80 ºC until used. On the day of purification, the cell pellet was thawed on ice, washed once in cold, sterile PBS, and lysed in 8.5 mL BugBuster supplemented with Santa Cruz Complete^™ Protease Inhibitor Cocktail Tablet (1 tablet per 50 mL lysis buffer), 5 μL rLysozyme, and 5 μL Omnicleave. The lysate was centrifuged for 20 min at 12,600 rpm to remove cell debris, and the supernatant fraction was incubated with 3 mL pre-washed Ni-NTA resin (Qiagen) at 4 ºC for 1 h with rocking. After this incubation step, the resin/supernatant mixture was added to a gravity column and drained. The column was washed three times with a buffer containing 50 mM Hepes pH 7.9, 300 mM NaCl, 5 mM BME, and increasing concentrations of imidazole (0, 20 mM, and 60 mM, respectively). Bound protein was eluted in 8 mL elution buffer (50 mM Hepes pH 7.9, 300 mM NaCl, 5 mM BME, 200 mM imidazole), and 1 mL aliquots were frozen on dry ice. Purification protocol was adapted from Ormö et al.⁴⁸

The protein was concentrated to approximately 40 mg/mL in a centrifugal concentrator. The buffer was exchanged three times with 20 mM HEPES pH 7.5 and 10 mM NaCl. Crystals were grown by the hanging-drop vapor diffusion method. The drops were set up using a Mosquito robot, dispensing drops composed of 200 nL of protein and 200 nL reservoir. The reservoir volume was 100 μL. Crystals typically grew within a few hours. Crystals in space group P2₁ and P2₁2₁2₁ were cryoprotected by the addition of glycerol to a concentration of 35%.

Diffraction data from GFP-RNase(1–8) crystals in space group P2₁2₁2₁ (form 1), C2, and P2₁ were collected at the Advanced Photon Source beamline 24-ID-C using a Dectris Pilatus 6M-F pixel detector. The data set from crystal form 2 of space group P2₁2₁2₁ was collected at the beamline 24-ID-E using an ADSC Quantum 315 CCD detector. Each data set was collected from a single crystal at a temperature of 100 K. A cryoprotectant mixture of 35% glycerol and 65% reservoir was used for crystal form P2₁. The remaining crystal forms required no additional cryoprotectant. The following X-ray wavelengths were used for crystal forms P2₁2₁2₁ (form 1), P2₁2₁2₁ (form 2), C2, and P2₁, respectively: 0.9795 Å, 0.9791 Å, 0.9795 Å, and 1.4760 Å. Data was processed using the XDS package⁵³ for all crystals. In addition, autoPROC⁵⁴ was used for processing data sets in P2₁2₁2₁ (form 1), and C2; aimless⁵⁵ was used for scaling in P2₁2₁2₁ (form 1). Data collection statistics are reported in supplemental Table S1.

The first crystal form we obtained of GFP-RNase(1–8), P2₁2₁2₁ form 1, was isomorphous with a previously determined GFP structure, PDB ID code 4P1Q⁵⁶. We used these coordinates as a starting model in the refinement of the GFP-RNase(1–8) Buster⁵⁷. Difference density maps revealed the position of the N-terminal fusion, and these atoms were built using the graphics program Coot⁵⁸. The remaining structures were determined by molecular replacement using the program Phaser⁵⁹. The search model for molecular replacement in space group C2 was the refined structure from space group P2₁2₁2₁ form 1. The search model for molecular replacement in space groups P2₁ and P2₁2₁2₁ form 2 was the refined structure from space group C2. All models were built and refined using the same software as described above for P2₁2₁2₁ form 1. Structures were illustrated using the program Pymol⁶⁰.

Crystal diffraction intensities were cylindrically averaged around the filament axis to simulate the radial disorder of a filament using a custom-written Fortran code.

To identify how many globular, monomeric proteins are poised to form protofilaments, we began by screening 130,446 PDB entry headers to identify monomers. We defined “monomers” as entries that contained only a single protein chain, were determined by crystallography, had resolution better than 3.0 Å, and had no biological assembly operators other than the identity matrix. We also excluded redundant entries defined as having the same MOLECULE record and unit cell lengths identical to ± 3Å and unit cell angles identical to ±5°. We found 12,933 entries that passed these criteria for monomers.

To determine how many of these monomers were poised to form screw-related protofilaments, we performed the following steps. First, we applied crystallographic operations to bring the molecule as close as possible to the origin, generating a reference molecule. Then, we applied crystallographic screw operations to the reference molecules, allowing for up to 3 unit cell translations in any combination along positive and negative a, b, and c. Screw operators were accepted as potential protofilament generators only if the reference molecule and the operated molecule were close enough to touch. We defined touching distance as any distance less than twice the distance of the furthest atom of the reference molecule from its center of mass. These criteria produced 73,968 screw-related pairs. Surface areas buried by these interfaces were calculated using the CCP4 program, areaimol⁵⁵. Of these, 26,375 screw-related pairs meet the criterion of a quasi-stable interface as defined by equaling or surpassing the 901 Ų of buried surface area (450 Ų per surface) observed between barrel domains of our GFP fusion protein protofilament. That is, 9119 PDB entries, or 74% of the monomer entries. Including a constraint that one of the termini be located within 12 Å of the intermolecular interface (a threshold defined by our GFP fusion protein) leaves 53% of monomeric PDB entries meeting these combined criteria. That is, over half the monomer entries in the PDB (6816 entries) are poised for filament formation.

We used a random number generator and probability weighted amino acid frequencies in vertebrates to calculate one million peptide sequences. We scored the sequences for helix propensity using Chou & Fasman rules⁶¹. Using two different seeds for the random number generator, we found 127,952 and 128,171 sequences passed the criteria for helix. That is approximately 13%.