Atomic structures of TDP-43 LCD segments and insights into reversible or pathogenic aggregation.

The normally soluble TAR DNA-binding protein 43 (TDP-43) is found aggregated both in reversible stress granules and in irreversible pathogenic amyloid. In TDP-43, the low-complexity domain (LCD) is believed to be involved in both types of aggregation. To uncover the structural origins of these two modes of β-sheet-rich aggregation, we have determined ten structures of segments of the LCD of human TDP-43. Six of these segments form steric zippers characteristic of the spines of pathogenic amyloid fibrils; four others form LARKS, the labile amyloid-like interactions characteristic of protein hydrogels and proteins found in membraneless organelles, including stress granules. Supporting a hypothetical pathway from reversible to irreversible amyloid aggregation, we found that familial ALS variants of TDP-43 convert LARKS to irreversible aggregates. Our structures suggest how TDP-43 adopts both reversible and irreversible β-sheet aggregates and the role of mutation in the possible transition of reversible to irreversible pathogenic aggregation.

Introduction

The role of low complexity domains (LCDs) in proteins linked to neurodegenerative diseases is a topic of intense current investigation[1-3]. LCDs are regions of proteins that exhibit biased compositions of amino acids and frequently no secondary structure[4-6]. In some cases the LCDs are the result of nucleotide expansions that cause duplications of short sequences, as in Huntington’s Disease[7]. In other cases the LCDs are intrinsic to the protein and believed to have a functional role as in the RNA binding proteins TAR DNA binding protein 43 (TDP-43), Fused in sarcoma (FUS) and hnRNPA1[8], whose LCDs participate in formation of stress granules[1,9,10].

Stress granules (SGs) are aggregates of RNA and RNA binding proteins that form in the cytoplasm when the cell experiences physical, mechanical or chemical stress[11,12]. This process is often initiated by a halt in translation and puts the cell in a hibernating state until the stress subsides[13]. SGs are reversible, meaning that SG-proteins disaggregate when the stress is relieved, the disaggregated proteins and RNA are then free to return to their normal functions in the cell[1,2,14,15]. Phosphorylation has been shown to play a role in these SG-proteins dynamics[16,17]. Over the last decade, researchers have studied how this process occurs and what biophysical properties these aggregates exhibit. Recent studies on G3BP-1, a SG protein, have demonstrated that the SGs exist in a biphasic state with the mRNA packed at the center and the RNA-binding proteins on the periphery[18]. Also, foci composed entirely of RNA have been observed as a product of phase transition in human cells[19]. The mechanism of SGs assembly and the role of RNA are still being investigated[1,3,15,20]. Studies on FUS and hnRNPA1 have demonstrated that the LCDs are sufficient for liquid-liquid phase separation (LLPS)[1,15]. Despite progress, questions remain about the structures of the aggregates and how these may differ between SG assemblies and pathogenic states[4].

Our recent study determined atomic structures of five segments of low-complexity domains of FUS, hnRNPA1, and nup98[21]. All of these displayed kinked β-sheets with weak interactions between sheets. We gave these structures the acronym LARKS (Low-complexity Aromatic-Rich Kinked Segments), and proposed that these LARKS act as Velcro to provide adhesion between LCDs to form SGs or other membrane-less organelle assemblies. These structures contrast with steric-zippers which act as molecular glue to fasten together amyloidogenic segments in amyloid fibrils(reviewed in[22]). Our study of LARKS offered an atomic-resolution hypothesis about the molecular mechanism of SG formation.

In addition to studies aimed at determining the structure of SG assemblies, researchers have focused on the role of SG-associated proteins in neurodegenerative diseases. One of the major hallmarks of neurodegenerative diseases is the deposition of proteinaceous amyloid fibrils in cells[23]. Different from SG assembly, aggregation of proteins into the amyloid state in disease seems a virtually irreversible process[22,24]. Variants found in TDP-43, FUS and hnRNPA-1 have all been linked to Amyotrophic Lateral Sclerosis (ALS) among other diseases[25].

TAR DNA Binding Protein (TDP-43) is a 43kD protein composed of four domains: an N-terminal domain, two RNA recognition domains and a C-terminal low complexity domain (LCD)[26-28]. This protein is involved in a number of cellular functions including mRNA shuttling, transcription, miRNA processing and alternative splicing among others[29,30]. Over the past decade, research into TDP-43 has found aggregates of the protein in patients with ALS, Alzheimer’s Disease, Parkinson’s Disease, Frontotemporal Lobar Degeneration and Huntington’s Disease[31-34]. These aggregates are irreversible, ubiquitinated, hyper-phosphorylated, C-terminal cleavage products ranging in size from 25–35kD[35]. Thus it is hypothesized that the LCD of TDP-43 plays a role in both this irreversible disease aggregation and reversible SG aggregation.

Of proteins that participate in SGs and also form pathogenic amyloid fibrils, TDP-43 stands out for four reasons. First, TDP-43 has over 45 variants that are linked to ALS and other neurodegenerative diseases[25,36,37]. Second, the majority of the variants that cause disease are localized to the LCD[38,39]. Third, the majority of the aggregates found in the cytoplasm of cells of patients are cleavage products inclusive of the LCD[40,41]. Finally, TDP-43 is found to aggregate in 97% of sporadic cases and 45% of familial case of ALS[42,43]. This makes TDP-43 an intriguing candidate for understanding how its LCD participates in both reversible SG assembly and irreversible, pathogenic assembly.

Through structure determination, we can learn which segments of TDP-43 may drive aggregation and how familial variants affect this process. Here we use x-ray diffraction and micro-electron diffraction (MicroED) to determine the atomic resolution structures of ten peptide segments, seven in the reference sequence and three variants, from the LCD of TDP-43. We show that the LCD is capable of forming typical steric-zipper β-sheet structures characteristic of pathogenic aggregation. Also we show that segment 312–317 forms a kinked, labile structure—a LARKS, predicted as a protein interaction motif consistent with properties of SG assemblies[21]. We also show how this structure is perturbed by the familial mutants A315T and A315E, which provides an opportunity for understanding the interplay of formation of membrane-less organelles with pathogenic amyloid.

Results

Six segments from the LCD of TDP-43 form steric zippers

For structural studies we targeted segments throughout the LCD because there is no consensus of which region is the amyloidogenic core. The LCD of TDP-43 is an unstructured domain composed of approximately 140 amino acid residues. Researchers have demonstrated that this region plays a role in aberrant mislocalization and aggregation of TDP-43 to the cytoplasm in mice and neuroblastoma cell lines[44,45]. It was initially suggested that the entire LCD, residues 277–414, served as a prion like domain and was the minimum fragment required for aggregation[46]. However, additional studies suggested smaller regions within the LCD are sufficient for aggregation and neurodegeneration. In the first study, segment 286–331 was identified and shown to form amyloid fibrils and confer neurotoxicity on primary cortical neurons[47]. A second study identified residues 311–360, illustrating through NMR and CD studies that this segment undergoes a helix to β-sheet transition that initiates aggregation and cytoplasmic inclusion[48,49]. A third study suggested that segment 342–366 is the minimum toxic core that transitions from a random coil to a β-sheet[50]. In short, these studies illustrated that there are several regions of the LCD that can be responsible for amyloid-like aggregation.

Our previous studies demonstrated that adhesive cores of amyloid fibrils are composed of 6–11-residue segments that stack together as steric-zippers[22,51]. To identify which segments from the LCD of TDP-43 might be responsible for TDP-43 aggregation, we consulted ZipperDB[52], (URL: https://services.mbi.ucla.edu/zipperdb). We selected the top 10 hits scored by their adhesive energies (supplementary table 1). Additionally, we also selected segments rich in glutamine and asparagine because the C-terminus of TDP-43 is a prion-like sequence and these two residues are important for prion aggregation[53,54]. Lastly, we selected two segments based on a double mutation reported to reduce TDP-43 aggregation[48] (see methods and materials section for detailed segment selection), In total, we targeted fifteen segments in the LCD of TDP-43 (supplementary table 1).

Crystal structures revealed that six of the segments formed steric-zippers: 300GNNQGSN306, 321AMMAAA326, 328AALQSS333, 333SWGMMGMLASQ343, 370GNNSYS375, and 396GFNGGFG402 (Fig. 1A and B, Table 1). These structures demonstrated tight side chain interdigitation, like the previously reported zippers in other amyloidogenic proteins, such as β-amyloid[51] and Tau[55]. All three six structures are composed of in-register sheets[22]. Segments 321AMMAAA326 and 333SWGMMGMLASQ343 form anti-parallel sheets (Supplementary Fig. 1) while the other four segments form parallel sheets.

Figure 1

Segments from the Low Complexity Domain (LCD) of TDP-43 form steric zippers

(A) Schematic representation of full length TDP-43, the low complexity domain is enlarged and the sequences and relative positions of segments for which we determined structures are shown. (B) Six steric-zipper structures we determined in this study. For each structure, three β-sheets and two interfaces between these sheets are shown, and each sheet shows 5 of its thousands of strands. For 300GNNQGSN306, 328AALQSS333 and 370GNNSYS375, two interfaces (IF) are distinct and are labeled as IF1 and IF2; For 321AMMAAA326, 333SWGMMGMLASQ343 and 396GFNGGFG402, all interfaces are identical because of the symmetry of packing. The area buried (Ab) and shape complementary (Sc) values for each interface are shown below the structure. Water molecules found in the structure are shown as cyan droplets. All the structures are shown in top view, viewed down the fibril axis.

Table 1

Statistics of x-ray diffraction data collection and atomic refinement.

	GNNQGSN(5WKD)	AMMAAA(6CEW)	AALQSS(6CB9)	GFNGGFG(5WIQ)	GNNSYS(5WIA)	NFGAFS(5WHN)	NFGTFS(5WHP)
Data Collection
Space group	C2	P212121	P212121	P1	P212121	P212121	P21
Cell dimensions
a,b,c (Å)	50.35, 4.77, 14.75	9.52, 15.44, 44.1	4.79, 15.97, 43.43	4.81, 16.44, 22.67	4.79, 15.62,40.68	13.82, 4.85, 46.74	15.33, 4.84, 23.57
α,β,γ(°)	90, 101.73, 90	90, 90, 90	90, 90, 90	90.63, 96.37, 91.26	90, 90, 90	90, 90, 90	90.00, 96.91, 90.00
Resolution(Å)	1.80 (1.86–1.80)a	1.20 (1.23–1.20)	1.10 (1.14–1.10)	1.24 (1.29–1.25)	1.00 (1.04–1.00)	1.10 (1.14–1.10)	1.00 (1.03–1.00)
Rmerge (%)	15.8 (25.0)	18.1 (68.3)	21.1 (56.2)	16.1 (40.6)	15.5 (55.4)	17.2 (72.6)	8.6 (10.4)
I/σ(I)	6.9 (3.3)	3.2 (1.4)	5.1 (2.8)	7.1 (2.0)	9.2 (2.1)	9.7 (2.5)	15.0 (7.6)
CC1/2	97.1 (81.6)	99.5 (69.2)	97.0 (91.3)	98.5 (69.0)	99.2 (66.8)	99.1 (82.3)	99.3 (98.0)
Completeness (%)	91.6 (84.8)	96.0 (99.3)	90.7 (72.1)	94.4 (95.4)	95.2 (62.7)	94.0 (75.9)	87.6 (37.0)
Redundancy	2.3 (2.2)	3.5 (3.5)	6.9 (7.3)	4.1 (1.8)	5.4 (1.9)	8.3 (6.4)	4.9 (2.8)
Refinement
Resolution (Å)	24.32–1.80	22.05–1.20	21.22–1.10	22.53–1.24	20.34–1.00	13.25–1.10	15.22–1.00
No. reflections	345	2,078	1,288	1,796	1,836	1,300	1,639
Rwork/Rfree	18.3/19.4	19.8/21.8	19.5/24.3	15.2/16.8	15.5/18.1	13.3/16.0	8.7/9.3
No. atoms
Protein	48	79	40	94	45	50	48
Water	2	4	4	8	1	0	3
B factors (Å2)
Protein atoms	8.7	9.1	7.3	3.2	3.5	4.6	2.4
Water	12.2	20.7	20.0	7.6	15.3	-	14.8
R.m.s. deviations
Bond lengths (Å)	0.021	0.012	0.010	0.005	0.005	0.006	0.019
Bond angles (°)	1.804	1.431	1.422	0.692	1.18	0.95	1.79

Values in parentheses are for highest-resolution shell.

Segments 300GNNQGSN306, 328AALQSS333 and 370GNNSYS375 all form Class 1 steric zippers in which the β-sheets are mated face-to-face with the same edges of the sheet oriented up the fibril (Fig. 1B)[22,51]. These Class 1 structures have two distinct interfaces (labeled as IF1 and IF2 in Fig.1B). Segments 321AMMAAA326 and 333SWGMMGMLASQ343 are Class 7 steric zippers with face-to-back and up-up orientation. Segment 396GFNGGFG402 forms a Class 4 steric zipper with face-to-back β-sheets and up-down orientation (Fig. 1B)[51]. These Class 7 and Class 4 structures have one unique interface. The area buried (Ab) and shape complementary (Sc) of all interfaces are labeled in Fig. 1B[51].

Six zipper segments are important for pathogenic TDP-43 aggregation

To investigate the connection between our steric-zipper structures and the pathogenic aggregates of TDP-43, we designed mutations to disrupt the zipper structures and tested their influence on aggregation of the TDP-43 protein. The segment we chose to represent pathogenic TDP-43 aggregation contains residues D-414 because it is a pathogenic segment of TDP-43 that was identified from FTLD brains[40]. We term this segment TDP-CTF, where CTF stands for C-terminal fragment. We disrupted steric-zipper formation by substituting tryptophan at the central residue of the sheet-sheet interface; its bulk imposes steric hindrance. We designed 8 mutations, Q303W and G304W targeting interface 2 and 1 of 300GNNQGSN306, respectively; A324W targeting the interface of 321AMMAAA326; L330W and Q331W targeting interfaces 1 and 2 of 328AALQSS333, respectively; M337W targeting the interface of 333SWGMMGMLASQ343; S373W targeting interface 2 of 370GNNSYS375; and G399W targeting the interface of 396GFNGGFG402.We did not design a mutation in interface 1 of 370GNNSYS375, since the interaction is relatively weak as judged by its small area buried.

Our mutagenesis results indicate that all six short segments are important for TDP-43-CTF aggregation. We achieved an aggregate-free starting point by conjugating TDP-CTF wild type and its mutants with a SUMO tag (Fig.2a, top panel). Aggregation was initiated by ULP1 protease removal of the SUMO tag (Supplementary Fig.2). After 30 min of incubation on ice approximately half of the SUMO-free TDP-CTF aggregated, as judged by the pellet fraction after centrifugation (Fig.2a, middle panel). In contrast, most of the SUMO-free mutants remained soluble, except G304W. After overnight incubation, cleavage of the SUMO-tag was nearly complete and all SUMO-free TDP-CTF wild type was in the pellet (Fig.2a, bottom panel). However, aggregation was incomplete for several mutants, displaying soluble TDP-CTF mutant material in the supernatant, including Q330W, A324W, M337W, L330W, Q331W and G399W. These results indicate that interface 2 of 300GNNQGSN306, the interface of 321AMMAAA326, interfaces 1 and 2 of 328AALQSS333, the interface of 333SWGMMGMLASQ343, interface 2 of 370GNNSYS375, and the interface of 396GFNGGFG402 are important for pathogenic aggregation of TDP-43. Potentially, these segments could facilitate pathogenic aggregation of wild type TDP-43 by forming the same steric-zipper structures we report here. Three of these zippers are located within the region 318–343, previously identified as the amyloidogenic core essential for TDP-43 aggregation[48]. This correlation is significant given that it covers a relatively small region compared to the whole LCD (26 residues out of 140). Moreover, a previous discovery that the double mutation A324E-M337E reduces TDP-43 aggregation[48] is consistent with our expectation that these substitutions create steric hindrance and electrostatic repulsion, thereby disrupting steric-zipper interfaces of 321AMMAAA326 and 333SWGMMGMLASQ343. These results suggest that these segments are important for pathogenic aggregation of TDP-43, and the steric-zipper structures we report here may represent the structure of the aggregation core.

Figure 2

Validation of steric-zipper forming segments

(A) Aggregation assays of the pathological fragment TDP-CTF (208–414), both wild type (wt) and with mutations to block steric-zipper formation. TDP-CTF was conjugated with a SUMO protein tag to prevent aggregation, and ULP1 protease was added to remove the SUMO tag. TDP-CTF was incubated on ice for 30 mins after cleavage and continued to incubate at 4 °C overnight. Samples were taken after incubation and separated into supernatant (s) and pellet (p) by centrifugation, and analyzed by SDS-PAGE. Notice that TDP-CTF G304W has similar behavior to wild type TDP-CTF, whereas other mutants show reduced aggregation compared to wild type. (B) Negative stain EM and fibril diffraction pattern of TDP-LCD and steric-zipper segments. TDP-LCD was conjugated with SUMO tag and mixed with ULP1 protease to remove SUMO tag. The diffraction pattern of TDP-LCD fibrils was acquired by X-ray diffraction and the diffraction patterns of zipper forming segments were computationally simulated from their crystal structures. The diffraction pattern of TDP-LCD fibrils was overlaid with the ones of zipper-forming segments for comparison. Notice that TDP-LCD fibrils generate diffraction rings at 4.8 Å and 10 Å, indicating that they have cross-β spines. Scale bar = 200 nm.

To further validate the role of these segments in TDP-43 aggregation, we compared the fibril diffraction of the low complexity domain of TDP-43 with the simulated diffraction of the zipper structures. We carried out fibril diffraction on TDP-LCD that contains residues 274–414 and was expressed using the same SUMO-tag strategy employed previously (Supplementary Fig.2). TDP-LCD formed long extended amyloid-like fibrils around 15 nm in diameter (Fig.2B), similar to the fibrils formed by TDP-43 in ALS patients[41]. The fibril diffraction pattern shows the rings at 10 Å and 4.8 Å, the signature feature of amyloid fibrils[22]. In comparison, electron micrographs of the six zipper-forming segments reveal microcrystals (Fig. 2B). We have shown these crystals are composed of cross-β assemblies. That is, the architecture revealed by the crystal structures of the segments is consistent with the cross-β signature of the TDP-LCD. Simulated fibril diffraction patterns calculated from their crystal structures are comparable with experimental diffraction of the TDP-LCD fibril (Fig. 2B); the differences correspond to variation in sheet-to-sheet distance among segments.

Segment 312–317 of the LCD forms a kinked structure that is disrupted by the A315T and A315E familial ALS mutants

In addition to pathogenic aggregation, TDP-43 is also involved in functional aggregation, such as SG formation. In SG assembly, the LCDs of SG-proteins are reported to be the key regions responsible for this reversible aggregation[1-3]. To further explore the role of the LCD in the functional aggregation of TDP-43, we focused on segments that have the potential to contribute to SG formation. The McKnight laboratory previously discovered that glycine residues flanked by aromatic side chains are features of hydrogel formation of the LCD of the protein FUS[2]. The hydrogel they studied is a reversible aggregate formed in vitro composed of reversible amyloid fibrils that trap a large amount of water, and is potentially relevant to SG formation in vivo[56]. They showed that replacement of aromatic residues in the LCD disrupted this hydrogel formation[2], which led to the hypothesis that aromatic residues play a role in SG assembly. For this reason, our recent study focused on the segments of the LCD of SG proteins that contain an aromatic residue and two proximal small or polar residues[21]. Structural studies of these segments in FUS and hnRNPA1 revealed pairs of kinked sheets, defined as LARKS (Low complexity, Aromatic-Rich, Kinked Segments), that display weak van der Waals interactions between mating sheets. Further research suggests the connection of LARKS with protein adhesion of LCDs in reversible aggregation formation[21].

In the present study, we find 312NFGAFS317, in the LCD of TDP-43, has similar amino acid composition and also forms labile aggregates and LARKS (Fig. 3). This suggests that TDP-43 can adopt the same mechanism of SG-formation previously hypothesized for FUS and hnRNPA1. Under EM, 312NFGAFS317 forms microcrystals (Fig. 3A). These microcrystals are labile, meaning they fully dissolve when treated with ≥1% SDS and heat to 70°C for 15 minutes. Their lability was observed by diminished turbidity (Fig. 3B). This behavior is similar to hydrogels formed by LCDs of SG-proteins, which re-solubilize by treatment with SDS and heat[2], unlike pathogenic amyloid fibrils such as β-amyloid fibrils (Supplementary Fig. 3A). The structure of 312NFGAFS317 reveals a kinked backbone (Fig. 3C), similar to other LARKS[21]. The kink occurs at Phe316 and has a Ramachandran angles of a left-handed α-helix. The kinked backbone makes two phenylalanine residues at positions 2 and 5 face the same side of the sheet, whereas a canonical β-sheet places the odd numbered sidechains on one face and the even numbered sidechains on the opposite face[51]. The subsequent packing of mating strands creates two different interfaces. Here we see that all the aromatic residues pack near to each other creating a hydrophobic, aromatic interface. The other interface is held together by the hydrogen-bonded interactions of the asparagine side chains. Both interfaces are weaker than most zipper structures as indicated by the relative low surface area buried, and consistent with the observation that these assemblies are labile under treatment with SDS and heat.

Figure 3

The 312NFGAFS317 segments forms a kinked β-sheet that is strengthened by familial variants A315T and A315E

(A) Negative stain EM images of aggregates of 312NFGAFS317 and its variants after 4 days of shaking at 37 °C. Scale bar = 200 nm. (B) Denaturing assays of aggregates of 312NFGAFS317 and its variants. The aggregates of each segment were treated with the indicated percentage of SDS, and heat to 70 °C for 15 mins. The absorbance at wavelength 340 (OD340) of the solution after treatment was measured to represent turbidity. The readings from same segment were shown as percentage of the mean values of the segment treated with 0% SDS and without heat, and were shown as mean ± SD (n=3 independent experiments) with individual data overlaid. (C) Structures of 312NFGAFS317 and its variants. For each structure, six sheets are shown and two interfaces (IF1 and IF2) are labeled between three parallel packed sheets. For 312NFGEFS317, the two interfaces between the parallel packed sheets are identical because of its face-to-back packing, so the interfaces are not labeled. The area buried (Ab) and shape complementary (Sc) values for each interface are shown below the structure. Water molecules found in the structure are shown in cyan droplets. All the structures are shown viewed down the fibril axis. Notice that the larger interfaces in both 312NFGEFS317 and 312NFGpTFS317 are increased over those in 312NFGAFS317 and 312NFGTFS317, explaining in part their greater stability. See Supplementary Fig. 3 for additional hydrogen bond stabilization.

We observe that 312NFGAFS317 forms a LARKS structure and its microcrystalline aggregates are labile, thus providing another example of LARKS with a potential role in SG formation, and an opportunity to investigate the influence of pathogenic variants of the putative key segment of functional aggregation. Within the NFGAFS segment, there are two ALS-linked familial variants: A315T and A315E[47,57]. Both of these variants have been shown to increase redistribution of TDP-43 to the cytoplasm and increase neuronal death as compared to the reference sequence[58-61]. Additionally, previous studies have shown that disease-causing mutations in SG-proteins are associated with accumulation of persistent SGs[62-64]. However the molecular mechanism of transformation of SGs to irreversible aggregates has remained elusive. Here we studied the structure and biochemical behavior of 312NFGTFS[31], 312NFGEFS317, and 312NFGpTFS317 where pT stands for phosphorylated threonine. We studied 312NFGpTFS317 because TDP-43 is found to be hyperphosphorylated in ALS patients[35], and the A315T mutation may create a new phosphorylation site in vivo.

All three of these mutant segments formed microcrystals under EM (Fig. 3A). 312NFGEFS317 and 312NFGpTFS317 aggregates were irreversible under the same SDS and heat denaturing assay as 312NFGAFS317. 312NFGTFS317 aggregates, however, were still labile (Fig. 3B). These results indicated that the stabilities of the mutant segments A315E and A315pT were significantly enhanced as compared to that of 312NFGAFS317.

To better understand the biochemical findings, we determined the structures of all three mutant segments. All three structures are composed of kinked, in-register, parallel sheets; however the kink positions differ. In 312NFGAFS317 the kink occurs at Phe316. In the case of 312NFGTFS317 and 312NFGEFS317, the kink occurs at Gly314; and the 312NFGpTFS317 structure is kinked at two residues, Phe313 and Gly314. Despite the different positions of the kinks, all of the kinks result in the two phenylalanine residues facing the same side of the sheet (Fig. 3C).

Further analysis of the structures illustrates that the packing of sheets in both 312NFGTFS317 and 312NFGpTFS317 are face-to-face, as in 312NFGAFS[31]. Both have two distinct interfaces of mating strands, one dry interface formed by hydrophobic residues (mainly phenylalanine residues, labeled as IF1 in Fig. 3C) and one wet interface formed by mostly hydrophilic residues (labeled as IF2 in Fig. 3C). The packing of 312NFGEFS317 is face-to-back so its two interfaces are identical, with hydrophobic interactions in the middle and hydrophilic interactions on both sides. For the 312NFGpTFS317 structure, the phosphate is clearly observed in the Fo-Fc omit map (Supplementary Fig. 3B). To our knowledge, this is the first time a post-translational modification has been reported for an amyloid mimetic structure. The enhanced stability of the A315E and A315pT of 312NFGAFS317 assemblies can be explained by the increased interface strengths, indicated by area buried for the dry interface and hydrogen network of the wet interface. The area buried for the dry interface increased significantly from 63.0 Å2/chain in 312NFGAFS317 to 137.0 Å2/chain in 312NFGEFS317 and 175.6 Å2/chain in 312NFGpTFS317 (Fig. 3C). In the case of 312NFGpTFS317, increased buried area results from the four phenylalanine residues of two strands perfectly interdigitating with each other. Additionally, on the wet interface, the side chains of Glu315 and pThr315 create a hydrogen bonding network, which further strengthens the assembly (Supplementary Fig. 3C). This stacking of negatively charged Glu315 and pThr315 is facilitated by the acidic pH of crystallization conditions, in which the residual TFA from peptide synthesis significantly decreases the pH, and results in partially protonated glutamate and P-threonine side chains (see discussion). For 312NFGTFS317, the area buried for dry interface and hydrogen network for the wet interface is similar that of 312NFGAFS317, which is consistent with the observation that 312NFGTFS317 has similar behavior in our denaturing assay as 312NFGAFS317.

In summary, we find that the 312NFGAFS317 segment of TDP-43 forms a kinked β-sheet structure—a LARKS—which is labile. We find that two familial variants within this segment, A315T and A315E, together with phosphorylation, strengthen the reference sequence assembly and making it irreversible by creating a more stable structure with stronger interaction between each the sheets. These results further support our hypothesis that LARKS are the potential molecular mechanism of functional, reversible aggregate formation, by showing that LARKS exist in another SG-forming protein, TDP-43. Moreover, these results indicate that the pathogenic mutations and or post-translational modifications can strengthen LARKS and drive SG-proteins towards pathogenic aggregation.

Discussion

Our previous studies suggested connections of cross-β structures to both amyloid aggregation through steric-zippers forming irreversible pathogenic fibrils[22] and LARKS forming reversible hydrogels and possibly membrane-less organelles[21]. In this study, we determined structures of segments from the LCD of TDP-43 and found that that this domain has the capacity to form both steric-zippers and LARKS. This is consistent with the dual nature of the TDP-43 LCD which can drive aggregation of the protein into either functional SG aggregates or into pathogenic amyloid fibrils as in ALS or FTLD. In Figure 4 we present a speculative model that incorporates our findings and explains how SG assembly might lead to pathogenic assembly. In physiological conditions, TDP-43 can execute its normal functions without irreversible aggregation, even though steric-zipper (red bars) and LARKS (kinked blue bars) segments exist in their sequences (Fig. 4, top panel). In pathogenic conditions, the steric-zipper forming segments interact with each other to form irreversible assemblies, driving the protein into pathogenic amyloid fibrils (Fig. 4, left pathway). In stress conditions, the LARKS forming segments interact with each other to form reversible assemblies and drive the protein to functional aggregates such as SGs. Since functional assemblies are reversible[1,2,14], the aggregation of proteins can be reversed when stress is relieved (Fig. 4, right pathway). Numerous factors are believed to influence the fate of the proteins, including intra/inter-cellular signals, chaperones, RNA binding, protein folding, microenvironments such as pH, protein concentration, etc. Here, our study suggests that the mutations and or post-translational modifications can also change the pathway of assembly by directly changing the character of assembly-driving segments (Fig. 4, middle pathway). As exemplified by A315E and A315pT, the mutation and phosphorylation strengthen the 312NFGAFS317 assembly, converting a labile LARKS to irreversible aggregates. Our hypothesis is that under conditions that favor SGs, LARKS segments interact with each other, and should one or more of them convert to an irreversible structure, the protein may form a pathological, irreversible assembly.

Figure 4

Speculative model showing the alternative pathways of formation of stress granules with pathogenic amyloid

TDP-43 forms several irreversible steric zippers (represented by red bars) and several reversible LARKS (represented by blue kinked bars). When performing its normal functions, the protein is not aggregated (top panel). Under pathogenic conditions (left pathway), the steric zipper segments interact with each other, acting as molecular glue to drive the protein into irreversible aggregation, and forming amyloid fibrils; under stress conditions (right pathway), the LARKS interact with each other, and acting as molecular Velcro, providing adhesion to form a hydrogel as in stress granules; these interactions are reversible. The crystal structures of 312NFGAFS317, 312NFGEFS317 and 312NFGpTFS317 show that A315E and A315T with phosphorylated Thr can strengthen the LARKS to irreversible aggregates. Either mutation or post-translational modifications can act as the agent of conversion (middle pathway). Since the involvement of 312NFGAFS317 in stress granules formation is yet to be validated, the middle pathway is shown by gray dashed lines, emphasizing that this pathway is speculative.

The mechanism represented in Figure 4 may apply for other regions of TDP-43 and other SG-forming proteins. The 312NFGAFS317 segment is not the only region proposed to be important for reversible aggregation of TDP-43; residues 320–340 have been reported to be involved in granule formation[3,65]. Applying a threading algorithm for predicting LARKS from amino acid sequence, we predicted additional LARKS in the LCD of TDP-43[21]. We compared the LARKS predictions with the familial variants reported on TDP-43 LCD, and found that 28 out of 42 reported familial variants are in the regions of predicted LARKS (supplementary table 2). This observation supports our hypothesis that the mechanism of conversion of A315E and A315pT on 312NFGAFS317 may be a general mechanism.

Notice that both 312NFGEFS317 and 312NFGpTFS317 crystals were grown at acidic pH, which facilitated the stacking of glutamate and phosphorylated threonine residues. A recent study also shows that decrease in pH drives phase separation in yeast and in in vitro systems[66]. This observation suggests the importance of pH on SG formation and the role of pH in transition of SG to pathogenic aggregation.

In this study, we also reported six steric-zipper structures from LCD of TDP-43, and showed the connection between these structures with pathogenic aggregates of TDP-43 by mutagenesis experiments. The observation that all six segments are important for TDP-43 aggregation, suggests a segmental polymorphism, meaning the aggregation of TDP-43 can be formed via multiple cores. Although recent structural studies have shown that the amyloid fibril core usually contains tens of residues (e.g. cryo-EM structures of Tau[67]), our study on short peptides facilitates the research of amyloidogenic proteins by identifying the high-resolution structures of core aggregation-prone segments, and more importantly, by providing structural insights for inhibitor designs. By designing peptide inhibitors that bind to the ends of steric-zipper structures and cap their growth, we have successfully developed the inhibitors for aggregation of several amyloidogenic proteins, including p53[68] and Tau[55]. The zipper structures we reported here will provide the opportunity for structure-based design of TDP-43 inhibitors.

In summary, the present study provides atomic-resolution structures of irreversible and reversible aggregates of segments of TDP-43, and the possible roles of familial mutations and or post-translational modifications in the interconversion of these two distinct aggregation pathways.

Online Methods

Selection and Generation of TDP-43 Peptide Targets

We chose our peptides of interest by a combination of computational prediction and literature reports (supplementary table 1). For computational prediction, we selected the segments that contain high rank predictions from ZipperDB (ranked by the predicted Rosetta energy). From literature reports, we have selected the segments containing the features important for LCD aggregation. First, the LCD of TDP-43 is believed to be a prion-like domain, and glutamine and asparagine residues are found to be important for prion aggregation[53,54], so we selected three segments from LCD of TDP-43 that have the highest Gln and Asn density. These three segments are GNNQGSN (4 out of 7), ASQQNQ (4 out of 6) and GNNQNQ (5 out of 6); the latter two were also identified by ZipperDB. Second, aromatic residues are reported to be important in LCD aggregation[2], so we selected two segments from the LCD of TDP-43 that have multiple aromatic residues close to each other; these are NFGAFS (two residues between two phenylalanine residues) and GFNGGFG (three residues between two phenylalanine residues). Third, the A324E-M337E double mutation is reported to reduce TDP-43 aggregation[48], so AMMAAA and SWGMMGMLASQ were selected to study the molecular mechanism of this double mutation. All peptides were purchased from Genscript at a purity of 95% or higher. Notice that from fifteen peptides we selected, we determined the structures for ten of them. The other segments we selected (supplementary table 1), failed to generate structures because of poor crystal formation or poor quality of crystal diffraction.

Peptide Aggregates Formation

All peptide segments were dissolved in phosphate buffered saline (PBS) at pH 7.5 to a final concentration of 20mM. Samples were then incubated at 37°C in a Torrey Pine Scientific shaker at level 9 for 80 hours. Samples were examined by TEM following fibril formation.

Construct design

TDP-CTF (208–414) was designed to mimic the pathological fragment of TDP-43 found in FTLD brain, and the low complexity domain of TDP-43 (TDP-LCD, 274–414) was also designed. Both TDP-43 fragments was conjugated to the carboxyl-terminus of small ubiquitin-like modifier (SUMO) proteins to increase its solubility and prevent its aggregation during expression and purification. The SUMO protein was also conjugated with (His)6-tag on its amino-terminus for easier purification. The SUMO-TDP constructs were inserted into a pET28a vector with following expressed sequences:

SUMO-TDP-CTF: MGSSHHHHHHGSGLVPRGSASMSDSEVNQEAKPEVKPEVKPETHINLKVSDGSSEIFFKIKKTTPLRRLMEAFAKRQGKEMDSLRFLYDGIRIQADQTPEDLDMEDNDIIEAHREQIGGREFFSQYGDVMDVFIPKPFRAFAFVTFADDQIAQSLCGEDLIIKGISVHISNAEPKHNSNRQLERSGRFGGNPGGFGNQGGFGNSRGGGAGLGNNQGSNMGGGMNFGAFSINPAMMAAAQAALQSSWGMMGMLASQQNQSGPSGNNQNQGNMQREPNQAFGSGNNSYSGSNSGAAIGWGSASNAGSGSGFNGGFGSSMDSKSSGWGM

SUMO-TDP-LCD: MGSSHHHHHHGSGLVPRGSASMSDSEVNQEAKPEVKPEVKPETHINLKVSDGSSEIFFKIKKTTPLRRLMEAFAKRQGKEMDSLRFLYDGIRIQADQTPEDLDMEDNDIIEAHREQIGGGRFGGNPGGFGNQGGFGNSRGGGAGLGNNQGSNMGGGMNFGAFSINPAMMAAAQAALQSSWGMMGMLASQQNQSGPSGNNQNQGNMQREPNQAFGSGNNSYSGSNSGAAIGWGSASNAGSGSGFNGGFGSSMDSKSSGWGM

Protein purification and validation

Both SUMO-TDP-CTF and SUMO-TDP-LCD protein was expressed in Escherichia coli BL21 (DE3) strain. Bacterial cells were cultured in LB media supplemented with 50 µg/ml kanamycin at 37 °C to an OD600 of 0.6–0.8. Protein expression was induced with 1 mM isopropyl β-D-1-thiogalactopyranoside (IPTG) at 25 °C for 3 hours. Cells were harvested and re-suspended in 20 mM Tris-HCl, pH 8.0, 500 mM NaCl, 20 mM Imidazole, 10% (v/v) glycerol, supplemented with 1% (v/v) Halt Protease Inhibitor single-use cocktail (Thermo Scientific). The re-suspension solution was sonicated (3s on/3s off cycle, 10 mins) and the supernatant solution was collected after centrifugation (24,000 × g for 20 mins). The supernatant solution was mixed with homemade NucA nuclease (5000 U per liter of cell culture) and filtered before load to HisTrap HP column (GE health care). The HisTrap column was equalized with 20 mM Tris-HCl, pH 8.0, 500 mM NaCl, 20 mM Imidazole before loading of supernatant solution, and washed with 20 mM Tris-HCl, pH 8.0, 500 mM NaCl, 200 mM Imidazole after loading. The protein expected was eluted by 20 mM Tris-HCl, pH 8.0, 500 mM NaCl, 500 mM Imidazole and concentrated by Amicon Ultra-15 centrifugal filters (Millipore) to at least 2 mg/ml (10 fold higher than the working concentration of aggregation assays). SUMO-TDP-LCD was concentrated to 10 mg/ml (10 fold higher than its fibril formation concentration).

To validate the cleavage of SUMO-tag and existence of TDP fragment, purified SUMO-TDP-CTF and SUMO-TDP-LCD was diluted into 0.2 mg/ml with 20 mM Tris-HCl, pH 8.0, 150 mM NaCl, 10 µM DTT, mixed with 100:1 (weight basis) homemade ULP1 protease, and incubated on ice. Samples were taken before and 1 hour after adding ULP1 protease, and separated by NuPAGE 4–12% Bir-Tris gel (Invitrogen). The pellet of SUMO-TDP-LCD was also collected and sampled from fibril formation assay and loaded to the same gel. The SDS-PAGE was stained by coomassie blue, or probed with TDP-43 polyclonal antibody (Invitrogen, catalog number PA5-27221, 1:1000 dilution) and Anti-Rabbit IgG (whole molecule)-Peroxidase antibody (Sigma, catalog number A0545, 1:5000 dilution) after transferred to a nitrocellulose membrane (Invitrogen). The membrane was visualized with SuperSignal West Pico Chemiluminescent Substrate kit (Thermo Scientific) and BIOMAX Light Film (Carestream).

Fibril formation and diffraction of TDP-LCD

SUMO-TDP-LCD was diluted into 1 mg/ml with the same dilution buffer used in aggregation assays of TDP-CTF, and mixed with 100:1 (weight basis) homemade ULP1 protease. The fibril was formed by shaking the solution at 37 °C for 3 days and sitting it on bench for another 2 days. The fibril was collected by centrifugation at 9,000 × g for 5 mins and washed by water for 2 times. The fibril was sampled for TEM analysis and SDS-PAGE, and then concentrated to 2% of the original volume by centrifugation and removal of supernatant. The concentrated solution was loaded between two siliconized glass capillaries with their heads facing each other. The solution was then dried out and the fibril remained on one of the capillary heads in solid form. The capillary with fibril was mounted on an in-house X-ray machine and the fibril was shot by X-ray for 3 mins. The diffraction pattern was detected and recorded by a CCD detector.

Fibril diffraction simulation of zipper forming segments

The crystal structures of zipper forming segments were used for generation of fibril diffraction simulation. To simulate the fibril diffraction patterns, the space group of each crystal structure was expanded into P1, and structure factors were calculated using the CCP4 program sfall[69]. The diffraction patterns were cylindrically averaged 180 degrees around the fibril axis using custom written software.

Negative Stain Transmission Electron Microscopy (TEM)

Samples for TEM were prepared by applying 5 µL of sample on hydrophilic 400 mesh carbon-coated formvar support films mounted on copper grids (Ted Pella, Inc.). The fibrils were allowed to adhere for 4 min and washed twice with water. The samples were then stained for 2 min with 2% uranyl acetate, leaving a few nanoliters of stain on the grid. Grids were allowed to dry for 1 minute before storage. Each grid was inspected on a T12 (FEI) microscope.

Aggregation assays of TDP-CTF

SUMO-TDP-CTF wild type and its mutants were diluted into 0.2 mg/ml with 20 mM Tris-HCl, pH 8.0, 150 mM NaCl, 10 µM DTT and filtered by 0.1 µm Ultrafree-MC-VV centrifugal filters (Millipore). The filtered solution was mixed with 100:1 (weight basis) homemade ULP1 protease and incubated on ice to remove the SUMO-tag from TDP-CTF. The samples were taken before and 30 mins after adding the protease, and the 30-minute-samples were divided into supernatant and pellet fraction by centrifugation at 18,000 × g for 3 mins at 4 °C. The pellet fraction was re-suspended with the same buffer used in dilution and with the same volume of the original samples. The samples were mixed with 3:1 (v/v) of NuPAGE LDS sample buffer (Invitrogen) and heated to 100 °C for 10 mins. After the 30-minute-samples were taken, the cleavage solutions were moved to refrigerator (4 °C) and incubated overnight. The overnight samples were then taken and prepared same as 30-minute-samples. All samples were loaded to and separated by NuPAGE 4–12% Bis-Tris gels (Invitrogen) and stained by coomassie blue. To make better comparisons, the samples of wild type TDP-CTF were loaded to both gels when the samples taken from the same time point were split into two gels because of loading well limitation.

Denaturing assays of 312NFGAFS317 and its variants

Twenty-percent (m/v) SDS solution (in water) was diluted to 15%, 10% and 5% (m/v) by water, and 20% (v/v) of SDS-water stocks or water were mixed with 80% (v/v) of PBS to make SDS-PBS stocks with 4%, 3%, 2%, 1% and 0% (m/v) of SDS. SDS-PBS stocks with varies SDS content were mixed with 20 mM peptide-PBS solution in 1:1 volume ratio to get final peptide solutions with 2%, 1.5%, 1%, 0.5% and 0% (m/v) SDS. All solutions were divided into 3 PCR tubes (5 µl per tube), and the solutions contain 2%, 1.5%, 1% and 0.5% (m/v) SDS were heated to 70 °C for 15 mins by a thermocycler. The OD340 of solutions in each tube were measured by P330 NanoPhotometer (Implen), and the readings from same peptide were normalized to the mean values of the solutions contain no SDS and without heating.

Aβ fibril denaturing assay

Aβ purification and preparation please see our previous study[70]. Aβ1–42 was diluted into 50 µM with PBS buffer, and incubated at 37 °C overnight to form Aβ fibrils. The fibril solution was mixed with 10% (v/v) of 20% (m/v) SDS stock to make 2% (m/v) final SDS concentration, and heated to 70 °C for 15 mins. The samples were taken before and after adding SDS and heating, and analyzed by TEM.

Crystallization

All crystals, except 312NFGEFS317, 312NFGpTFS317 and 333SWGMMGMLASQ343 were grown by hanging drop vapor diffusion. 300GNNQGSN306 was dissolved in water to a final concentration of 10 mg/ml. The reservoir solution was optimized around the initial hit of 0.2M sodium sulfate, 0.1M Bis tris propane pH 7.5, and 20% (w/v) PEG 3350. Hanging drops contained a 2:1 solution of protein: reservoir.

321AMMAAA326 was dissolved in water to a final concentration of 10mg/ml, supplemented with 29.4 mM of LiOH. Crystals were grown over a reservoir solution containing 1.3 M Ammonium Phosphate and 0.1 M sodium acetate pH 6.27. Hanging drops contained a 1:2 solution of protein: reservoir.

328AALQSS333 was dissolved in water to a final concentration of 10mg/ml. Crystals were grown over a reservoir solution containing 0.01 M nickel chloride, 0.1 M Tris pH 8.5 and 20% (w/v) PDG 2000. Hanging drops contained a 2:1 solution of protein: reservoir.

370GNNSYS375 was dissolved in water to a final concentration of 10mg/mL. Crystals were grown over a reservoir solution containing 0.2M sodium nitrate, 0.1M Bis tris propane pH 8.5, and 20% (w/v) PEG 3350. Hanging drops contained a 2:1 solution of protein: reservoir.

396GFNGGFG402 was dissolved in water to a final concentration of 12 mg/mL. Crystals formed with the reservoir solution: 800 mM Sodium phosphate monobasic/ 1200 mM Potassium phosphate dibasic, 100 mM Sodium acetate/ Acetic acid pH 4.5. Hanging drops contained a 2:1 solution of protein: reservoir.

312NFGAFS317 was dissolved in water to a final concentration of 11mg/mL. Crystals formed with a reservoir solution of 0.1M phosphate/citrate pH 4.2, 40% w/v ethanol and 5% w/v PEG 1,000. The drops contained a 2:1 ratio of protein solution: reservoir.

312NFGTFS317 was dissolved in water to a final concentration of 11mg/mL. Crystals formed with a reservoir solution of 0.2M ammonium acetate, 0.1M sodium acetate pH 4.6, and 30% w/v PEG 4,000. The drops contained a 2:1 ratio of protein solution: reservoir.

312NFGEFS317, 312NFGpTFS317 and 333SWGMMGMLASQ343 were dissolved in PBS (pH 7.5) to a final concentration of 20mM. The peptide solution was shaken at 37°C in a Torrey Pines shaker at level 9 for 80 hours. To better understand the stacking of glutamate and phosphorylated threonine in 312NFGEFS317 and 312NFGpTFS317 structures, the pH of both peptide solutions was tested by MColorpHast pH-indicator strips (Millipore) and the pH of both solutions was around 4. For 312NFGpTFS317, initial crystals growing from shaking diffracted poorly, and in order to get better diffracting crystals, an additional round of crystal seeding was performed. Fresh made 20 mM 312NFGpTFS317 PBS solution was filtered and mixed with 2% (v/v) of pre-formed 312NFGpTFS317 crystals, and incubate at 37°C without shaking for 120 hours. Diffraction data was collected from seeded crystals, and the initial crystals were used in denaturing assay and EM observation showing in Fig.3.

X-ray Diffraction Data Collection and Processing

All x-ray diffraction data were collected at beamline 24-ID-E of the Advanced Photon Source, Argonne National Laboratory, Argonne, IL, USA, at a wavelength of 0.971 Å and temperature of 100 K. For NFGAFS, data were collected at 5° oscillations and 130 mm detector distance. Detector ADSC Q315 was used. Indexing and integration were done using XDS in P212121 and scaled with XSCALE to 1.1 Å resolution[71]. The data were imported using the CCP4 suite[69]. Molecular replacement was done using Phaser[72]. Initial models of different idealized polyalanine β-strands did not yield a structure. The final model used was a kinked strand with the sequence NFGAAS, which was based on the truncated structure of NNFGAIL[73]. The model was manually adjusted over 3 cycles to the correct sequence and rotamer orientation in COOT[74]. The atomic model was refined by Phenix[75] with a final Rwork/Rfree of 13.3/16.0% and 100% of Ramachandran angles favored.

For NFGTFS, data were collected at 5° oscillations and 125 mm detector distance. Detector ADSC Q315 was used. Indexing and integration was done using XDS in P21 and scaled with XSCALE to 1.0Å resolution[71]. The data were imported using the CCP4 suite[69]. The structure was solved by direct methods using the program SHELX[76]. The model was manually adjusted over 3 cycles to the correct sequence and rotamer orientation in COOT[74]. The atomic model was refined by Refmac[77] with a final Rwork/Rfree of 8.7/9.3% and 100% of Ramachandran angles favored.

For GFNGGFG, data were collected at 5° oscillations and 150 mm detector distance. Detector ADSC Q315 was used. Indexing and integration was done using Denzo in P1 and scaled with SCALEPACK to 1.25Å resolution[78]. The data were imported using the CCP4 suite[69]. Molecular replacement was done using Phaser[72]. Initial models of different idealized polyalanine β-strands did not yield a structure. The final model used was two kinked strands with the sequence GFAGG, which was based on the truncated structure of SYSGYS[21]. The model was manually refined using Phenix[75] over 4 cycles to the correct 7 residue sequence and rotamer orientation in COOT[74]. The atomic model has a final Rwork/Rfree of 15.2/16.8% and 100% of Ramachandran angles favored.

For AMMAAA, data were collected at 5° oscillations and 130 mm detector distance. Detector EIGER 16M was used. Indexing and integration was done using XDS in P212121 and scaled with XSCALE to 1.2Å resolution[71]. The data were imported using the CCP4 suite[69]. Molecular replacement was done using Phaser[72] with an idealized polyalanine β-strand (AAAAAA). The model was manually adjusted over 14 cycles to the correct 6 residue sequence and rotamer orientation in COOT[74]. The atomic model was refined by Refmac[77] with a final Rwork/Rfree of 19.8/21.8% and 100% of Ramachandran angles favored.

For AALQSS, data were collected at 5° oscillations and 130 mm detector distance. Detector EIGER 16M was used. Indexing and integration was done using XDS in P212121 and scaled with XSCALE to 1.1 Å resolution[71]. The data were imported using the CCP4 suite[69]. The structure was solved by direct methods using the program SHELX[76]. The model was manually adjusted over 13 cycles to the correct 6 residue sequence and rotamer orientation in COOT[74]. The atomic model was refined by Refmac[77] with a final Rwork/Rfree of 19.5/24.3% and 100% of Ramachandran angles favored.

For GNNQGSN, data were collected at 5° oscillations and 125 mm detector distance. Detector ADSC Q315 was used. Indexing and integration was done using XDS[71] in C2 and scaled with SCALEPACK[79] to 1.8Å resolution. The data were imported using the CCP4 suite[69]. Molecular replacement was done using Phaser[72] with an idealized β-strand of the sequence GNNAGA. The model was manually adjusted over 38 cycles to the correct 7 residue sequence and rotamer orientation in COOT[74]. The atomic model was refined by Refmac[77] with a final Rwork/Rfree of 18.3/19.4% and 80% of Ramachandran angles favored as well as 20% allowed.

For GNNSYS, data were collected at 3° oscillations and 125mm detector distance. Detector ADSC Q315 was used. Indexing and integration was done using Denzo in P212121 and scaled with SCALEPACK to 1.00Å resolution[78]. The data were imported using the CCP4 suite[69]. Molecular replacement was done using Phaser with an idealized 5 residue polyalanine β-strand[72]. The model was manually refined using Phenix over 2 cycles to the correct 6 residue sequence and rotamer orientation in COOT[74,75]. The atomic model has a final Rwork/Rfree of 15.5/18.1% and 100% of Ramachandran angles favored.

MicroED Data Collection and Processing

The MicroED data were collected using the same protocol as Rodriguez et al. with minor differences detailed below[80-82]. Data was collected at HHMI Janelia Research Campus using the Tecnai F20 TEM operating at 200 kV recorded using a bottom mount TVIPS F416 CMOS camera.

For NFGEFS, Individual image frames were exposed for 3s per image. During each exposure, crystals were continuously rotated within the beam at a rate of 0.2° per second. Diffraction data were collected and merged from 4 individual crystals. Indexing and integration were done using XDS[71] in P21212 and scaled with SCALEPACK[79] to 1.0Å resolution. The data was imported using the CCP4 suite[69]. The structure was solved by direct methods using the program SHELX[76]. The model was manually adjusted over 4 cycles to the correct sequence and rotamer orientation in COOT[74]. The atomic model was refined by Refmac[77] with a final Rwork/Rfree of 22.0/27.0% and 75% of Ramachandran angles favored as well as 25% allowed.

For NFGpTFS, Individual image frames were exposed for 3s per image. During each exposure, crystals were continuously rotated within the beam at a rate of 0.295° per second. Diffraction data was collected on a single crystal. Indexing and integration were done using XDS in P212121 and scaled with XSCALE to 0.75 Å resolution[71]. The data was imported using the CCP4 suite[69]. The structure was solved by direct methods using the program SHELX[76]. The model was manually adjusted over 16 cycles to the correct sequence and rotamer orientation in COOT[74]. The atomic model was refined by Refmac[77] with a final Rwork/Rfree of 23.2/25.1% and 100% of Ramachandran angles favored.

For SWGMMGMLASQ, individual images were exposed for 2s. During each exposure, the crystals were continuously rotated within the beam at a rate of 0.315° per second. Diffraction data were collected on seven crystals. Indexing and integration were performed XDS in space group P1 and scaled with XSCALE to 1.5 Å resolution[71]. The data were imported using the CCP4 suite[69]. The structure was solved by molecular replacement using the program phaser[72]and a search model consisting of an 9-alanine-long ideal beta-strand. The model was manually adjusted in COOT[74]. The atomic model was refined by Refmac[77] and Buster[83] with a final Rwork/Rfree of 28.0/31.3% and 94.5% of Ramachandran angles favored as well as 5.5% allowed. These R-factors are higher than expected for a data set of this quality. We think the discrepancy is due to a disorder in the crystal. The P1 unit cell dimensions of SWGMMGLASQ allow pseudo-merohedral twinning (pseudo space group C2). Also, the native Patterson map suggests that the crystal contains a translocation disorder (peak at 0.379, −0.264, −0.471 with height of 36% of the origin peak). Attempts to correct for these disorders did not improve the R-factors significantly.

Structure Analysis

All images of crystallographic structures were generated in Pymol. Area buried and shape complementarity values were calculated based on published methods[84,85].

Data Availability

The density maps and atomic coordinates of all models reported here have been deposited to the Protein Data Bank (PDB). The MicroED model has also been reported to the EM Data Bank. Accession codes are provided here: GNNQGSN (5WKD); AMMAAA (6CEW); AALQSS (6CB9); SWGMMGMLASQ (6CFH, EMD-7467); GFNGGFG (5WIQ); GNNSYS (5WIA); NFGAFS (5WHN); NFGTFS (5WHP); NFGEFS (5WKB, EMD-8857); NFGpTFS (6CF4, EMD-7466). All other data are available upon reasonable requests to the authors.

Accession Codes Availability

The fibril diffraction simulation code is available upon request.

Table 2

Statistics of MicroED data collection and atomic refinement.

	SWGMMGMLASQ(EMD-7467, 6CFH)	NFGEFS(EMD-8857, 5WKB)	NFGpTFS(EMD-7466, 6CF4)
Data Collection
Space group	P1	P21212	P212121
Cell dimensions
a, b, c (A°)	8.56, 9.60, 39.97	42.77, 17.42, 4.90	23.65, 4.72, 30.06
α, β, γ (°)	97.2, 92.9, 105.9	90, 90, 90	90, 90, 90
Resolution (Å)	1.50 (1.55–1.50)a	1.00 (1.03–1.00)	0.75 (0.77–0.75)
Rmerge (%)	20.8 (85.5)b	28.3 (99.3)c	17.2 (66.1)
I/σ(I)	3.3 (0.7)	4.6 (1.1)	3.9 (1.3)
CC1/2	98.7 (72.0)	99.2 (36.0)	98.9(59.9)
Completeness (%)	93.5 (89.4)	88.7 (78.0)	86.6(87.2)
Redundancy	4.2 (2.4)	9.5 (4.8)	3.8 (4.0)
Refinement
Resolution (Å)	13.17–1.50	21.39–1.00	7.65–0.75
No. reflections	1,819	1,801	4,178
Rwork/Rfree	28.0/31.3	22.0/27.0	23.2/25.1
No. atoms
Protein	162	90	93
Water	0	2	1
B factors
Protein atoms	18.1	5.3	19.6
Water	-	25.3	3.1
R.m.s. deviations
Bond lengths (Å)	0.007	0.016	0.021
Bond angles (°)	0.91	1.09	2.04

Highest resolution shell shown in parenthesis.

Data are from 7 crystals.

Data are from 4 crystals.