Four peptides were studied, spanning the sequence of the residues 38–61 of wild-type ANXA11 and mutated versions (WT, G38R, and D40G). The peptides were purchased from PEPCEUTICALS Ltd. (Leicester, United Kingdom). The molecular weights were validated by mass spectrometry.

The coding sequence for full-length human ANXA11 in the plasmid pMCSG7-MBP-ANXA11 was kindly sent by Boris Rogelj’s laboratory (Jožef Stefan Institute, Ljubljana, Slovenia). The ANXA11 N-terminus (residues 1–191) was donated by Salvatore Adinolfi (University of Turin, Italy) in a pET His6 TEV LIC plasmid with an N-terminal thioredoxin tag. The constructs were expressed in E. coli BL21 (DE3) cells.

Transformed cultures of the proteins were grown in Luria broth (LB) supplemented with 100 μg/mL ampicillin at 37°C until the optical density at 600 nm reached 0.6 and induced with 1 mM isopropyl β-d-1-thiogalactopyranoside (IPTG) at 20°C. Cells were collected by centrifugation and resuspended in lysis buffer and lysed by sonication. The soluble proteins were recovered in the supernatant by centrifugation at 4°C, and purified by nickel affinity chromatography using 300 mM imidazole in the elution buffer. The tag of ANXA11 was cleaved by incubating the construct with tobacco etch virus protease (1:50 enzyme/protein) overnight at 4°C, while dialyzing the mixture in SEC buffer. Pure ANXA11 was obtained after a reverse Ni-NTA chromatography step, and a further size-exclusion chromatography step on an Äkta pure system (HiLoad 16/600 Superdex 200 column, GE Healthcare). Each purification step of the full-length protein was carried out at 4°C, at pH 8.5 in 20 mM Tris-HCl buffer, 150 mM KCl, and additional 5 mM EDTA and 1 mM DTT in the SEC buffer.

Pure ANXA11 N-terminus with an N-terminal Thioredoxin-(His)6 tag was obtained after a further step of size-exclusion chromatography on an ÄKTA Prime Plus system (HiLoad 26/600 Superdex 75 prep grade column, GE Healthcare). Attempts to purify the untagged ANXA11 N-terminus in 20 mM HEPES, 20 mM NaCl buffer at pH 7.0 without the Thioredoxin-(His)6 tag led to quantitative precipitation. Protein purity was assessed by SDS-PAGE (Supplementary Figure S2). Protein identity was validated by LC-MS/MS with 86% of sequence coverage. The proteins were aliquoted, flash-frozen and stored at −80°C. ¹⁵N-labelled samples were obtained by growing the cells in minimal media using ¹⁵N ammonium sulphate as the sole source of ammonium.

Far-Ultraviolet (UV) CD spectra were recorded on a JASCO-1100 spectropolarimeter equipped with a temperature control system, averaged over 10 scans and deconvoluted with the online analysis software K2D3 and BestSel. Measurements were carried out in 1 mm path-length quartz cuvettes (type S3/Q/1; Starna Scientific), applying a constant N₂ flush at 4.0 L/min.

NMR measurements were carried out on 200 µM non-labelled peptides in a 10 mM sodium phosphate buffer, at pH 6.8 and on 100 µM ¹⁵N-labelled Thioredoxin-tagged ANXA11 N-terminus sample in 20 mM HEPES, 20 mM NaCl buffer at pH 7.0. D₂O (10%) was added to the samples. NMR spectra were recorded on Bruker 800 spectrometers at 5°C and at 25°C, respectively, and processed with NMRPipe (Delaglio et al., 1995). Spectra were analyzed and assigned with NMR-Fam Sparky (Lee et al., 2015) and CARA 1.9.1.7 (Keller, 2004). TOCSY spectra were measured using the ‘dipsi2esfbgpph’ pulse sequence. NOESY spectra were recorded with a mixing time of 250 ms. Both experiments and COSY spectra (cosydfesgpphpp) were recorded with 4,096 data points in t2 and 1,024 data points in t1. For the assignment of the Thioredoxin tag a set of three spectra were recorded: ¹H-¹⁵N HSQC (hsqcfpf3gpphwg_f1180), TOCSY-HSQC (dipsihsqcf3gpwg3d) and NOESY-HSQC (noesyhsqcf3gpwg3d_cpds).

The standard iterative protocol was used with the ARIAweb 07d7d10a (2021-06-08) service (Brünger et al., 1998; Allain et al., 2020) on the Pasteur@Galaxy cluster (doi 10.7490/f1000research.1114334.1) using default settings. ARIAweb implements the latest release of ARIA version 2.3 (Rieping et al., 2007) in combination with CNS version 1.21 (Brunger, 2007) modified with dedicated ARIA routines. 143 NOE based distance restraints were used (75 intra-residue and 68 sequential). The number of structures calculated was twenty for iterations 0–8 of which the seven best - based on the total energy - were used in the proceeding iteration. After nine iterations were completed, the 10 lowest energy conformers were refined in a shell of water molecules.

Further Molecular dynamics (MD) simulations were performed using the NAMD 2.13 package (Phillips et al., 2020) with the CHARMM36m force field. Input files were generated with CHARMM-GUI (Jo et al., 2008; Lee et al., 2016). The structures were solvated with the TIP3P water model in a rectangular box such that the minimum distance to the edge of the box was 10 Å under periodic boundary conditions. An appropriate number of Na⁺ counterions were added to neutralize the protein charge. The replicas were used for three separate production runs: ii) one imposing all NOE restraints, ii) one imposing only NH-NH NOE restraints, and iii) one with no restraints. For all active restraints the lower and upper walls were defined below and above 2.5 and 5.5 Å, with constants of 2 and 10 kcal mol⁻¹ Å⁻² respectively. Each replica was subject to 1 ns of equilibration at 278.1 K and normal pressure. The production runs (100 ns) were performed under the same conditions.

The structure evaluation and analysis procedures were performed with NMRBox (Maciejewski et al., 2017). Ten thousand structures were generated.

Structure predictions were carried out either running AlphaFold 2.2.0 in NMRBox (https://nmrbox.nmrhub.org/) or running the prediction on the Baker’s lab Robetta server (https://robetta.bakerlab.org/results.php?id=550799) using the full-length protein. Per-residue estimates of the confidence of the models for the AlphaFold models are given by per-residue predicted local distance difference test (pLDDT) scores of the final model. This score is a scale 0–100 and represents the confidence of the predicted structure compared to the “true” (ground truth) structure. Likewise, confidence in RosETTAFold models is given by the global distance test (GDT) (100.0 good, 0.0 bad). The structures were visualized using Pymol (Schrödinger, L. & DeLano, W. (2020), retrieved from http://www.pymol.org/pymol).

Structure predictions of the peptides based on chemical shift information was achieved by CS-Rosetta (https://csrosetta.chemistry.ucsc.edu/) (Shen et al., 2008). CS-Rosetta produces reliable structural ensembles from NMR observables (chemical shifts, J-couplings, NOEs, residual dipolar couplings, etc.). To do this, it performs a selection of protein backbone fragments from high-resolution structures from the PDB, which are used in conjunction with Rosetta’s high-resolution energy function to model the structures of proteins up to 35 kDa.

The catGRANULE software (http://service.tartaglialab.com/new_submission/catGRANULE) was used to predict the protein tendency to phase separate (Bolognesi et al., 2016). catRAPID signature was used to predict the propensity of TDP-43 to interact with RNA and identify RNA-binding domains (Livi et al., 2016).

SAXS experiments were performed at the BM29 beamline at the ESRF in Grenoble, France (Tully et al., 2023). The wavelength of the beamline was 0.99 Å (12.5 KeV), with the distance between sample and detector (PILATUS3 2M) set to 2,812 mm, giving the scattering vector q 0.007–0.55 Å⁻¹. This vector is defined as q = 4π sin(θ)/λ, where 2θ is the scattering angle and λ is the wavelength of the incident beam. Ten successive frames of the scattering from the samples were recorded in batch mode with an exposure time of 2 s for each frame due to 75 mA beam intensity. The scattering from the corresponding buffer was measured before and after each sample for the same exposure time, and subtracted from the sample scattering. Measurements were performed at 20°C, and the forward scattering, I ₀, was converted to an absolute scale by water calibration. The data were automatically reduced using FreeSAS (Kieffer et al., 2022) and further processed and analyzed using the Scatter IV (Tully et al., 2021) and ATSAS program packages (Manalastas-Cantos et al., 2021). I ₀, D _max and R _g were determined from P(r), although the Guinier approach was also used for comparison. The molecular weight of the species in solution was determined from I ₀.

Solutions of full-length ANXA11 were measured at protein concentrations of 3.8, 3.0, 1.9 mg/mL (70.0, 55.3, 35.0 µM) in the absence of calcium and at 3.6, 2.6 and 1.8 mg/mL (66.4, 47.9, 33.2 µM) in the presence of calcium (500 µM). Interparticle interactions were seen at higher concentrations. The curves were extrapolated to 0 and scale-merged using the PRIMUS software. Ab initio models were constructed using the programs DAMMIN (Svergun, 1999), DAMMIF (Franke and Svergun, 2009) and GASBOR (Svergun et al., 2001). DAMMIN and its reimplementation DAMMIF represent the protein molecule by compact beads connected to each other. In GASBOR, proteins are represented as an ensemble of dummy residues instead of dummy atoms. The crystal structure of rat ANXA11 (6tu2) as a monomer was fitted manually into the GASBOR bead-model to observe the volume occupied by the N- and C-terminus. Molecular ensemble models were generated by the Ensemble Optimization Method (EOM) software (Bernado et al., 2007; Tria et al., 2015). This package generates an ensemble of protein conformations and a theoretical average scattering intensity curve based on the ensemble. Finally, it fits the theoretical curve onto the experimental SAXS data. High-resolution structures of individual protein domains can be used as rigid bodies, while intrinsically disordered protein segments are modelled with completely random configurations. The RANCH, FFMAKER and GAJOE programs, all available at https://www.embl-hamburg.de/biosaxs/, allow respectively generation of an ensemble of models, computation of the scattering intensities based on PDB structures, and the selection of the ensemble of conformations whose computed scattering curve best-fits the experimental SAXS curve. We separated the sequence of ANXA11 to an N-terminal (chain B) disordered region and a C-terminal globular domain (chain A). The C-terminal part was fixed as rigid body using the PDB file 6tu2 as a monomer. The last residue of the N terminus was kept in steric proximity of the first residue of the C-terminus by defining a 5–7 Å distance constraint between them. The input scattering curves were extrapolated to 0 concentration in the absence and presence of calcium. The raw data were deposited to the SASBDB database with accession codes SASDTV5 (apo) and SASDTW5 (holo).

We produced the recombinant N-terminal domain of ANXA11 by E. coli expression of a fusion protein with an N-terminal thioredoxin tag. Since attempts to cleave the tag led to quantitative precipitation of the protein, we decided to keep the tag and analyze the fusion protein by NMR. The spectrum of this ¹⁵N-labelled protein showed a good dispersion but with a large number of resonances overlapping in the center of the spectrum (Figure 1). Comparison of the spectrum with the spectral assignment of thioredoxin retrieved from the BMRB data base (27636) allowed us to establish that the vast majority of the resonances with good chemical shift spreading correspond to residues in the thioredoxin tag, indicating that the residues of the ANXA11 N-terminus mainly contribute to the spectrum by the overlapping resonances. Absence of appreciable shifts of the thioredoxin peaks from their positions in the isolated protein indicated lack of significant interactions between the two proteins. This observation tells us that the ANXA11 N-terminus is mostly unstructured, as expected from the sequence composition.

We then analyzed the structure of the region 38–59 using synthetic peptides: we used a peptide spanning the sequence of wild-type ANXA11 (hereafter referred to as WT) and three peptides in which the clinically important mutations G38R and D40G were introduced. These mutants are hereafter indicated as G38R and D40G peptides. We screened different pH, temperatures, and buffer conditions by far-UV CD to understand how they could affect the WT peptide. When the peptide was dissolved in 10 mM phosphate buffer, it gave a CD spectrum with a negative minimum in ellipticity at 200 nm (Figure 2). This behavior is typical of an unfolded conformation. However, the spectrum also had a weak negative band around 220 nm which was compatible with a residual helical structure in equilibrium with a random coil conformation in the CD time scale (native helix) (Dyson et al., 1988a). Changes in temperature and buffer/pH did not significantly affect the amount of secondary structure of the peptide. The K2D3 webserver (Louis-Jeune et al., 2012) estimated 3%–6% α-helical and 15%–17% beta strand content. When 1%–30% (v/v) trifluoroethanol (TFE) was added, the minimum at 223 nm became deeper, corresponding to an appreciable increase in the propensity to a helical secondary structure. This alcohol is known to stabilize helical structure in peptides and is often used to enhance their helical propensities (Vincenzi et al., 2019). The G38R and D40G mutant peptides did not show appreciable differences to the WT. These results support the hypothesis of a helical element in this region of the ANXA11 N-terminus.

CD spectroscopy is an excellent technique to screen conditions, but NMR is a much more powerful means as it works in a different average time scale and can provide sequence-specific information on the structure of peptides. We thus studied the structural behavior of the peptides by NMR. Virtually complete assignment of the NMR spectrum of the non-labelled WT peptide in aqueous buffer was obtained using standard 2D techniques (Wuthrich, 1986; Redfield, 1993), except for the highly mobile and solvent exchangeable first and second N-terminal residues. Numerous Nuclear Overhauser Effects (NOEs) were observed that are typical of an α-helical conformation. This is unusual for a peptide of this relatively small size in aqueous solutions, even more at neutral pH and without the addition of co-solvents. In particular, an almost uninterrupted network of sequential HN-HN effects was observed along the whole sequence (region L39-N69) (Figure 3). This behavior indicated a strong tendency of the peptide to fold in a helical conformation throughout the sequence and confirmed an overall behavior typical of a nascent helix (Dyson et al., 1988b). The 10 lowest energy models of the WT peptide obtained by CS-Rosetta, a structure prediction program that uses chemical shift information, contained flexible termini (residues 37–38, 66–69), two distinct alpha-helices (residues 39–47, 52–65) and a short turn connecting the helices. This prediction does not however reflect the uninterrupted NH-NH sequential NOE cross-peak pattern. Accordingly, no long-range NOEs characteristic for a hairpin-like structure was identified. A plot of the secondary chemical shifts of the Hα of the peptides as defined by Pastore and Saudek (1990), a simple but effective method to detect secondary structural tendencies, suggested an uninterrupted helical structure for the isolated peptide. Conversely, extensive restrained and unrestrained molecular dynamics simulations provided trajectories with only transient formation of local helical regions, in agreement with the transient nature of a native helix (Supplementary Figure S3). Taken together, these results indicate that the peptides have a uniform tendency along the sequence to fold as a helix, but this secondary structure is not stably formed in water in the absence of stabilizing tertiary contacts.

When we analyzed the mutants, only minor chemical shift differences were observed at and around the residues affected by the mutations. Accordingly, the overall NOE patterns remained unchanged, indicating that the mutations do not affect the peptide structure and thus suggesting that D40 has a functional role. This is in agreement with the observation that the D40G mutation abolishes calcyclin binding (Smith et al., 2017).

To gain more information on the full-length protein, we first consulted AlphaFold (Jumper et al., 2021; Varadi et al., 2022) and RoseTTAFold predictions (Yang et al., 2020). Both predictions detect the presence of two distinct domains. The reliability of the C-terminus is high, which reflects high pLDDT values, whereas the non-conserved N-terminal domain has low confidence (Figure 4A). This is reasonable since, while the multiple alignment of the C-terminus comprises sequences and structures from all members of the vast annexin family, the N-terminus is specific to ANXA11 and the domain is intrinsically disordered as shown above by NMR.

The predictions by the two servers are different but have one feature in common (Figure 4B): all models predict an overall disordered structure for the N-terminus with a helix around residues 38–59. In the Alphafold structures, the helix tends to be interrupted around residues 47–49, whereas in the RosETTAFold structures the helix is uninterrupted. In the AlphaFold structures, the N-terminus consistently wraps around the C-terminal domain creating a more globular, though expanded, structure with the possibility of making contacts also between the region 38–59 and the first two annexin repeats. In the RosETTAFold structures, the N-terminus is completely separated from the C-terminus and does not form interactions with it.

SAXS is a low-resolution structural technique which can provide information about overall shape and domain orientation of proteins and is well suited to investigating flexible or intrinsically disordered proteins (Kachala et al., 2015; Lenton et al., 2023). We characterized full-length ANXA11 both in the presence (holo) and in the absence (apo) of a 5 M excess of CaCl₂. The excluded volumes of the hydrated particles (V_p) for the holo and apo ANXA11 were reasonably consistent with the values expected for a monomeric species and with the masses estimated from the primary sequence (Table 1). The higher than expected, estimated values of the molecular weight, ∼74–62 kDa, as compared to the theoretical one (54.4 kDa) are likely explainable by the flexibility of the system and to a minor degree of aggregation that appeared to be more accentuated in the calcium-free samples, as it is evident from some concentration dependence of the SAXS observables in the experiments in the absence of calcium.

The log I(q) versus q curves showed a high degree of similarity at low q, with the holo form having a slightly larger R_g. A small deviation was seen at high q (Figure 5A and Supplementary Figure S4) that is consistent with small changes in buffer matching. The main peak of the normalized Kratky plots has an approximately gaussian shape that indicates a globular, elongated domain, whereas the tail suggests an unfolded region (Figure 5B), as explained in more detail below. Pair-distribution curves for the holo and apo forms showed a similar bell-shape with elongated tails with an increased D_max 154.5 ± 0.5 Å for the apo form as compared with 150 ± 0.5 Å with the holo form. The computed distance distribution functions P(r) displayed a single peak with a tail, a pattern that is typical of proteins with elongated shapes (Figure 5C).

Shape reconstruction of full-length ANXA11 was performed by ab initio modelling using two complementary programs, DAMMIN (Svergun, 1999) and GASBOR (Svergun et al., 2001). The crystal structure of the C-terminus (6tu2) fitted well into one end of the more detailed GASBOR model leaving a region of extra density for one possible conformation in the ensemble of flexible N-terminus intrinsically disordered region to fill (Figure 5D).

The flexibility of the N-terminus was further investigated using two different approaches. For a qualitative approach, we used the normalized Kratky plot (Figure 5B) (Durand et al., 2010), that allows direct comparison of objects with different shapes and sizes. In this plot, folded compact globular proteins provide a bell-shaped curve at low angles with a maximum at q*R_g of 1.75 (Durand et al., 2010). We observed instead Kratky plots with maxima at q*R_g of 2.6 and 2.8 for the apo and holo ANXA11. These deviations from the standard behaviour are consistent with an appreciable level of flexibility. Both plots showed a broadening of the bell-shaped curve and a shift of the maxima to larger q*R_g values, as expected for extended and flexible molecules. The plots were also characterized by upward trends at higher q*R_g values (i.e., higher scattering angles), which is also indicative of flexibility. To obtain a more quantitative approach, the ensemble optimization method (EOM) (Bernado et al., 2007) was used to analyze the flexibility and size distribution of possible multiple configurations and to obtain optimized ensembles with a fit to the experimental scattering data (χ² ∼1.68 and 1.95) (Figure 6). The values of the ensemble average R_g from the EOM analysis were 39.7 and 39.8 Å in the absence and in the presence of calcium. Likewise, no significant variation was observed between the ensemble average D_max values, 148.4 and 149.2 Å. The values showed reasonable agreement with the experimental data. The degree of flexibility of apo and holo ANXA11 was estimated by comparing the corresponding Shannon entropy R_flex values of the ensemble distributions to that of a random pool and is a reference for flexibility. The comparison between apo and holo ANXA11 revealed almost identical ensemble R_flex values (77% versus 74%, respectively) and confirmed random motions of the N-termini. Looking in greater detail to the R_g distributions of these ensembles, we see the ensembles were shifted to the left of the distribution of randomly generated models of the initial pool (Figure 6). This indicates that the ensemble of conformations is more often at a lower R_g and thus more globular rather than completely elongated, suggesting that the flexible regions may be predominately around the core than fully extended (Supplementary Figure S5). This conclusion leads us to consider the solutions from the Alphafold server more dominant within the conformational ensemble compatible with the SAXS measurements.

Altogether, these results tell us that there is little difference between the structures of the apo and holo forms of ANXA11, suggesting that calcium regulation does not involve major conformational changes between the calcium free and calcium loaded forms. This is in agreement with what is observed in other members of the annexin family with a much shorter N-terminus (Burger et al., 1996; Sopkova et al., 2002; Rosengarth and Luecke, 2003; Hong et al., 2020). It also tells us that the N-terminus does not appreciably participate to the calcium regulation of ANXA11 which seems to involve mainly the C-terminus.

Finally, we analyzed the question of which region(s) of ANXA11 is/are involved in soluble (liquid-to-liquid) phase separation and in RNA-binding. We assessed the potential of ANXA11 to form protein granules using the catGRANULE approach (Bolognesi et al., 2016). catGRANULE is a machine learning software trained on granule-forming proteins, utilizing features such as RNA binding, structural disorder, and amino acid composition to predict phase separation propensity. It specifically considers factors such as structural disorder, nucleic acid binding affinity, and amino acid motifs like arginine-glycine and phenylalanine-glycine as indicative of a protein’s tendency to coalesce into granules (Bolognesi et al., 2016).

The catGRANULE analysis indicated a strong propensity for granule formation, predominantly localized within the first ∼200 N-terminal residues, with limited contribution from the remainder of the protein (Figures 7A, B). This finding aligns with recent research proposing that the N-terminus is both necessary and sufficient for driving concentration-dependent ANXA11 phase transitions from dispersion to condensation (Nixon-Abell et al., 2023). Notably, the region with the highest phase separation propensity contains a proline-rich domain (MSGTFGGANMPNLYPGAPGAGYPPVPPGGF). This is noteworthy as proline-rich domains have been implicated in phase separation (Zhang et al., 2020) and RNA binding (Wang et al., 2006) as observed in the case of Tau.

To complement these findings, we conducted an analysis with the catRAPID signature program, which predicts RNA-binding regions within a protein sequence. catRAPID signature leverages physicochemical properties, secondary structure characteristics, and hydrophobicity profiles (Livi et al., 2016). The analysis identified three regions in the C-terminal core domain with the potential for RNA binding, while suggesting minimal RNA-binding capability at the N-terminus (Figure 7C). The absolute maximum is around residue 290 that is in the second annexin repeat. The other two lower maxima are in repeats 3 and 4. Notably, the last maximum encompassing residues around position 460 contains a sequence homologous to the consensus motif (F/Y)XXX (F/Y)XKSL, known to interact with nucleic acids/RNA in ANXA2 and ANXA6 (Bandorowicz-Pikula et al., 2003; Aukrust et al., 2007). Two smaller maxima were observed, one of which is in the N-terminus. They are however very close to the threshold and do not support a strong tendency to stoichiometric interactions. The one in the N-terminus is detected only in the context of the full-length protein.

These results point towards a primary role of ANXA11 N-terminus in granule formation, primarily triggered by structural disorder rather than RNA binding.