Bacterial strains and growth conditions

All strains and plasmids used in this work are listed in Supplementary Table 2. F. johnsoniae was routinely grown aerobically in Casitone yeast extract (CYE) medium46 at 30 °C with shaking. For some physiological studies, the cells were cultured in motility medium (MM)47 or PY2 medium48 as indicated below.

Genetic constructs

Plasmids were constructed by Gibson cloning49 or Q5 site-directed mutagenesis (New England Biolabs) using the primers and target DNA in Supplementary Table 3. Suicide and expression plasmids were introduced into the appropriate F. johnsoniae background strain by biparental mating, using E. coli S17-1 (ref. 50) as donor strain as previously described46. Chromosomal modifications were introduced using the suicide vector pYT313 harbouring the counter-selectable sacB gene as previously described16. All plasmid constructs and chromosomal modifications were confirmed by sequencing.

Purification of SprA and SprE complexes

To purify complexes containing Twin-strep-tagged SprA, the relevant strain was cultured for 22 h in 12 l of CYE medium using 500 ml culture volume in 2 l flasks. Cells were collected by centrifugation at 12,000 g for 30 min and stored at −20 °C until further use. All purification steps were carried out at 4 °C. Cell pellets were resuspended in buffer W (100 mM Tris-HCl, pH 8.0, 150 mM NaCl, 1 mM EDTA) containing 30 μg ml−1 DNase I, 400 μg ml−1 lysozyme and 1 mM phenylmethylsulfonyl fluoride (PMSF) at a ratio of 5 ml of buffer to 1 g of cell pellet. Cells were incubated on ice for 30 min with constant stirring before being lysed by three passages through a TS series 1.1 kW cell disruptor (Constant Systems) at 30,000 PSI. Unbroken cells were removed by centrifugation at 20,000 g for 35 min. The supernatant was recovered and total membranes were collected by centrifugation at 230,000 g for 75 min. Membranes were resuspended in buffer W to a protein concentration of 6.5 mg ml−1 and solubilized by incubation with 1% (w/v) lauryl maltose neopentyl glycol (LMNG, Anatrace) for 2 h. Insoluble material was removed by centrifugation at 230,000 g for 75 min. Endogenous biotin-containing proteins were masked by addition of 1 ml BioLock (IBA Lifesciences) solution per 100 ml of supernatant and incubation for 20 min with constant stirring. The solution was then circulated through a Strep-TactinXT Superflow column (IBA Lifesciences) overnight. The column was washed with 10 column volumes (CV) of buffer W containing 0.01% LMNG (buffer WD) and bound proteins were eluted with 6 CV Strep-TactinXT BXT buffer (IBA Lifesciences) containing 0.01% LMNG. The eluate was concentrated to 500 μl using a 100-kDa molecular weight cut-off (MWCO) Amicon ultra-15 centrifugal filter unit (Merck) and then injected onto a Superose 6 Increase 10/300 GL column (Cytiva) previously equilibrated in buffer WD. Peak fractions were collected and concentrated using a 100-kDa-MWCO Vivaspin 500 column (Sartorius).

Complexes containing Twin-strep-tagged SprE were purified similarly, but the final size exclusion chromatography step was omitted.

Purification of recombinant proteins

mCherry-CTDRemA and mCherry-CTDFjoh_2389 were purified from E. coli BL21 Star (DE3) cells containing pFL125 or pFL156, respectively. The cells were grown in 2.5 l LB medium at 37 °C to an optical density at 60 nm (OD600) of 0.5 and protein expression induced by addition of 40 μM isopropyl-β-d-thiogalactoside (IPTG). The cells were then cultured for an additional 5 h at 37 °C. Cells were collected by centrifugation at 12,000 g for 30 min and stored at −20 °C until further use. All purification steps were carried out at 4 °C. Cell pellets were resuspended in buffer A (10 mM Na2HPO4, 1.8 mM KH2PO4, 300 mM NaCl, 2.7 mM KCl, 20 mM imidazole, pH 7.4) containing 30 μg ml−1 DNase I, 400 μg ml−1 lysozyme and 1 mM PMSF at a ratio of 5 ml of buffer to 1 g of cell pellet. Cells were incubated on ice for 30 min with constant stirring before being lysed by two passages through a French pressure cell at 12,000 PSI. Cell debris was removed by centrifugation at 24,000 g for 30 min. The supernatant was then clarified using a 0.22 µm syringe filter unit (Millipore) and circulated through a 5 ml HisTrap HP column (Cytiva) for 2 h. The column was washed with 9 CV of buffer A and bound proteins were eluted with a 10–500 mM linear gradient of imidazole over 20 CV of buffer A. Peak fractions were collected and concentrated to 500 μl using a 10-kDa-MWCO Amicon ultra-15 centrifugal filter unit (Merck), then injected onto a Superdex 75 10/300 GL column (Cytiva) previously equilibrated in PBS pH 7.4 (10 mM Na2HPO4, 1.8 mM KH2PO4, 150 mM NaCl, 2.7 mM KCl). Purified fusion proteins were concentrated using a 10-kDa-MWCO Amicon ultra-15 centrifugal filter unit.

PorD was purified similarly using plasmid pFL175, except that protein expression was induced by addition of 250 μM IPTG and the cells were then cultured for an additional 5 h at 20 °C. HisTrap HP peak elution fractions were concentrated to 5 ml using a 10-kDa-MWCO Amicon ultra-15 centrifugal filter unit, then injected onto a HiLoad 16/60 Superdex 75 PG column (Cytiva) previously equilibrated in PBS pH 7.4.

Reconstitution of substrate–SprA complexes

SprA complexes were purified from a Δplug background (or from a ΔporV background for the control experiment in Extended Data Fig. 8b) using the procedure described above and then transferred into buffer B (PBS pH 7.4 containing 0.005% LMNG) by dialysis. Of the SprA complexes, 300–400 pmol were mixed with the selected mCherry–CTD fusion at a 1:2–1:5 molar ratio in a final volume of 400 μl of buffer B and incubated overnight at 4 °C with constant agitation. The samples were then injected onto a Superose 6 Increase 10/300 GL size exclusion column (Cytiva) previously equilibrated in buffer B and the fluorescence (587 nm excitation, 610 nm emission) of the eluted material monitored using a Prominence fluorescence detector (RF-20AXS, Shimadzu). Samples for cryo-EM were crosslinked directly before the size exclusion step by addition of 3.5 μl of a 25% (w/v) glutaraldehyde solution. The samples were then incubated for 60 min on ice before the reaction was quenched by addition of 100 μl of buffer W. Peak fractions from the size exclusion column containing SprA–mCherry–CTD complexes were concentrated using a 100-kDa-MWCO Vivaspin 500 column.

Peptide mass fingerprinting

Samples were excised from Coomassie-stained gels. For whole-sample proteomic analysis, SDS–PAGE was carried out only until the sample had fully entered the gel and the protein smear at the top of the gel was excised. Samples were subjected to in-gel trypsin digestion and electrospray mass spectrometry at the BSRC mass spectrometry and proteomics facility (University of St Andrews, United Kingdom).

Immunoblotting and far western blotting

The following commercial antisera were used: anti-StrepTag (34850, Qiagen), anti-GroEL (G6532, Merck), anti-mCherry (Ab167453, Abcam), anti-HaloTag (G921A, Promega), anti-mouse IgG peroxidase conjugate (A4416, Merck) and anti-rabbit IgG peroxidase conjugate (31462, Pierce). PorD antibodies were raised in rabbits against the purified recombinant PorD protein.

For whole-cell immunoblots, F. johnsoniae strains were cultured in MM to OD600 = 0.4 and analysed by SDS–PAGE and immunoblotting as previously described23. The following modifications to the general protocol were made for whole-cell immunoblots of SprA. The cells were suspended in NuPAGE LDS sample buffer (ThermoFisher), incubated at 100 °C for 10 min and separated on NuPAGE 3 to 8% tris-acetate gradient gels (ThermoFisher). Proteins were transferred onto polyvinylidene difluoride (PVDF) membranes by wet transfer using a TE22 tank (Hoefer) run at 20 V overnight at 4 °C in 25 mM Tris, pH 8.3, 192 mM glycine containing 10% methanol (v/v) and 0.05% Tween 20 (v/v).

For PorD far western blots, samples containing 0.25 μg of SprE–Twin-strep complexes were prepared for SDS–PAGE without heating. Following protein transfer to PVDF membranes, the membranes were blocked with 20 mM Tris-HCl, pH 7.6, 100 mM NaCl, 0.5 mM EDTA, 10% glycerol (v/v), 0.1% Tween 20 (v/v) and 2% milk powder (w/v), followed by overnight incubation with 30 μg purified PorD in the same buffer. Membranes were then probed with anti-PorD antibodies, followed by secondary anti-rabbit IgG antibodies.

Phenotypic analysis methods

Measurement of gliding motility on agar and analysis of the secreted proteome were performed as previously described23.

A protease accessibility assay was used to assess the cell surface exposure of HaloTag-SprB448 in strain AK_067. This strain contains a ΔporV mutation to prevent the export of endogenous secreted proteases which digest HaloTag-SprB448. SprB has a Type B CTD, hence does not require PorV for export. Cells were cultured in MM to OD600 = 0.4 and resuspended in phosphate buffered saline (PBS) containing 10 mM MgCl2. Samples were supplemented as appropriate with 200 μg ml−1 proteinase K (ThermoFisher) and 1% (v/v) Triton X-100 (Merck) and incubated for 15 min at 30 °C. Reactions were stopped by the addition of 5 mM phenylmethylsulfonyl fluoride (ITW Reagents), followed by incubation at 100 °C for 5 min, the addition of SDS–PAGE sample buffer and further incubation at 100 °C for 5 min before analysis by immunoblotting.

Measurement of mCherry–CTD secretion

For the analysis of secretion by immunoblotting, cells containing an mCherry–CTD fusion-expressing plasmid were cultured either in MM to mid-log phase (OD600 = 0.4) or in CYE medium to stationary phase (OD600 = 4.5). The cultures were separated into cell and supernatant fractions by centrifugation at 12,000 g for 2 min. The supernatant fraction was filtered using an Ultrafree-MC centrifugal filter unit, followed by centrifugation at 210,000 g for 75 min at 24 °C. The samples were then analysed by immunoblotting using mCherry antibodies.

Live-cell fluorescence microscopy

Single-molecule imaging was realized by under-labelling the cells with fluorophore. Strains were cultured in MM to OD600 = 0.4, at which point 2.5 nM Janelia Fluor 646 HaloTag ligand (Promega) was added and the cells cultured for another 20 min. Cells were washed five times with PY2 medium, supplemented with 0.5 ng ml−1 SynaptoGreen (Biotium) to label the OM, and 2 μl spotted onto PY2 agar pads. Where indicated, 25 μg ml−1 chloramphenicol was added to the cells 30 min before imaging.

All imaging data were acquired using HiLo (glancing TIRF) illumination on a Nanoimager (Oxford Nanoimaging) equipped with a 640 nm 1W DPSS laser. Optical magnification was provided by a ×100 oil-immersion objective (Olympus, numerical aperture (NA) 1.4) and images were acquired using an ORCA-Flash4.0 V3 CMOS camera (Hamamatsu). All fluorescence images were collected at 15% laser power.

Raw data were analysed using the Fiji plugin ThunderSTORM51 to determine single-molecule localizations. Cell outlines were determined using custom Python codes and single-molecule trajectories within cells computed using the Trackpy Python package (https://zenodo.org/record/3492186#.Y3ZWpH2ZNPY). Finally, apparent diffusion coefficients were determined using custom Matlab codes.

Cryo-EM grid preparation and data acquisition

Before grid preparation, SprA complexes purified from various genetic backgrounds were concentrated to an absorbance at 280 nm (A280) ranging from 0.5 to 3.1 and in vitro-reconstituted SprA–model substrate complexes were concentrated to an A280 of 1.0 (mCherry–CTDRemA complex) and 2.3 (mCherry–CTDFspA complex). Four microlitres of each sample was applied onto glow-discharged (30 s, 15 mA) 300 mesh Quantifoil Au R1.2/1.3 holey carbon coated grids, adsorbed for 10 s, blotted for 2 s at 100% humidity at 4–8 °C and plunge frozen in liquid ethane using a Vitrobot Mark IV (ThermoFisher).

Electron microscopy was performed on a Titan Krios G3 (ThermoFisher) operating at 300 kV and equipped with a BioQuantum imaging filter (Gatan) and 20 e−V slit width. Data for endogenous SprA complexes were collected in counted super-resolution mode on a K3 detector (Gatan), real pixel size of 0.832 Å per pixel, using a total dose of 58.0–62.4 e− A−2 over 40 fractions. Data for reconstituted SprA–model substrate complexes were collected in counting mode on a K2 detector (Gatan), real pixel size of 0.822 Å per pixel, using a total dose of 51.2–52 e− A−2 over 20 fractions.

Cryo-EM data processing

For the endogenous SprA complexes, patched (15 ×10) motion correction, contrast transfer function (CTF) parameter estimation, particle picking and initial two-dimensional (2D) classification were all performed in SIMPLE (3.0)52. For the reconstituted SprA–model substrate complexes, motion correction and dose weighting were performed using MotionCor-2 implemented in Relion (v.3.0)53, CTF parameters were estimated using CTFFIND4 (ref. 54), and particle picking and initial 2D classification was performed in SIMPLE (2.0)55. All downstream processing was carried out in Relion v.3.0. Gold-standard Fourier shell correlations using the 0.143 criterion and local-resolution estimations were calculated within Relion.

For the mCherry–CTDRemA reconstituted SprA complex (Extended Data Fig. 9a), 1,245,976 particles were extracted from 10,619 movies. After one round of reference-free 2D classification, 330,442 particles were classified in 3D (3 classes) against a 60 Å lowpass-filtered map of the substrate-free PorV complex, EMD-0133. The major class containing 179,790 particles was then subjected to masked 3D auto-refinement, yielding a 3.6 Å map. CTF refinement (per-particle defocus and beamtilt estimation) followed by another round of 3D auto-refinement improved map resolution to 3.3 Å. Bayesian polishing plus an additional CTF refinement yielded further improvements in map quality, to 3.1 Å. Improvement in CTD density was observed after alignment-free 3D classification performed against 2 classes. Particles (37,140) belonging to the class with strong CTD density underwent masked 3D auto-refinement, generating a 3.2 Å volume.

For the mCherry–CTDFspA reconstituted SprA complex (Extended Data Fig. 9b), 1,004,533 particles were extracted from 4,950 movies. Following 2D cleanup in SIMPLE, 192,766 particles were classified in 3D (3 classes) against a 60 Å lowpass-filtered map of the PorV complex, EMD-0133. The major class containing 104,716 particles was then subjected to masked 3D auto-refinement, yielding a 3.7 Å map. Bayesian polishing plus CTF refinement (per-particle defocus estimation and beamtilt fitting) yielded a 3.2 Å volume. Alignment-free 3D classification (2 classes) followed by 3D auto-refinement generated a 3.2 Å volume with strong CTD density from 42,981 particles.

For the Extended Translocon (ΔgldL peak I) (Extended Data Fig. 2b–d), 9,287,798 particles were extracted from 28,642 movies and then subjected to 2D cleanup in SIMPLE and a round of reference-free 2D classification in RELION, yielding 3,395,677 pruned particles. A 224,276-particle subset was 3D classified (3 classes) against a 40 Å lowpass-filtered map of the SprA–mCherry–CTDRemA complex volume described above. Particles (93,016) from the prominent class that displayed clear density protruding from the barrel of SprA were selected and 3D auto-refined to generate a 3.2 Å volume. This volume was used as reference for 3D classification (6 classes) against the entire 2D-cleaned dataset of 3,395,677 particles. Particles (1,105,952) from the strongest, most populated class were subjected to 3D auto-refinement using the corresponding 40 Å lowpass filtered reference volume, generating a 3.3 Å map that demonstrated strong, extramembraneous density on the periplasmic rim of SprA. Focused classification without alignment was performed (4 classes) using a mask encompassing this extramembraneous density, generating two similar classes with clear secondary structure elements. Particles from both classes were combined (546,490 particles) and 3D auto-refined against the 3.3 Å map described above (lowpass-filtered to 8 Å), yielding a 2.7 Å map. CTF refinement incorporating per-particle defocus, beamtilt, trefoil and fourth-order aberrations estimations followed by 3D auto-refinement with local searches generated a 2.55 Å map. Bayesian particle polishing and subsequent 3D auto-refinement improved map resolution to 2.45 Å. Local refinements were then performed using a mask encompassing SprA–PorV–PPI–CTD–substrate or Skp–SprE, generating 2.4 Å and 2.7 Å maps, respectively. Particle subtraction and recentring was performed for the Skp–SprE volume, followed by global 3D auto-refinement to yield a 2.8 Å map. A composite map of the Extended Translocon was then generated by Phenix (phenix.combine_focused_maps) using global B-factor sharpened SprA–PorV–PPI–CTD–substrate and Skp–SprE maps as input.

For ΔgldL peak II (Extended Data Fig. 4), 2,075,818 particles were extracted across a total of 6,547 movies and subjected to 2D classification within SIMPLE then RELION, yielding 630,907 pruned particles. These particles were then 3D classified against a 60 Å lowpass-filtered PorV complex volume (EMD-0133). Particles corresponding to CTD-bound PorV–SprA (across 2 classes) were combined and subjected to 3D auto-refinement, generating a 2.7 Å volume. Focused 3D classification without alignment (6 classes) using a soft mask encompassing the CTD was then performed, which resulted in one class (containing 57,573 particles) demonstrating strong β-sandwich density after 3D auto-refinement. Particles belonging to this class were Bayesian polished, 2D classified and CTF refined (fitting per-particle defocus, beamtilt, trefoil), generating a 2.8 Å volume following 3D auto-refinement.

For ΔgldLΔsprE peak I′ (Extended Data Fig. 3a,b), 8,098,702 particles extracted from 21,159 movies were 2D classified in SIMPLE and then RELION to yield 4,215,403 curated particles. These particles were classified in 3D (6 classes) against a 60 Å lowpass-filtered volume of the PorV complex (EMD-0133). Particles belonging to two prominent, similar classes were combined (2,952,793 total particles) and 3D auto-refined to yield a 2.3 Å map. Bayesian particle polishing followed by an additional round of 2D classification and then multiple rounds of CTF refinement (per-particle defocus, beamtilt, trefoil and fourth-order aberrations) yielded a 2.0 Å volume after 3D auto-refinement from 2,798,799 particles. Alignment-free 3D classification using a mask encompassing the CTD (6 classes) generated strong CTD density for three classes, two of which were similar enough to be combined. 3D auto-refinement of the two remaining isolated classes (422,344 and 820,880 particles each) yielded 2.2 Å maps demonstrating clear conformational differences in the β4–β5 CTD loop.

For ΔgldM peak I (Extended Data Fig. 10a), 218,615 particles were recovered following two rounds of 2D classification and then 3D classified (3 classes) against a 60 Å lowpass-filtered Extended Translocon volume. Two classes, corresponding to either an Extended Translocon complex (50,045 particles) or a RemZ-bound PorV–SprA complex (128,033 particles), were recovered and independently refined, yielding 4.3 Å and 3.8 Å volumes, respectively.

For ΔsprE peak I′ (Extended Data Fig. 10b), 355,991 particles were recovered following two rounds of 2D classification and then 3D classified (4 classes) against a 60 Å lowpass-filtered PorV complex reference volume. Particles (135,611) belonging to the strongest class were 3D auto-refined against their corresponding volume, yielding a 3.4 Å volume for the RemZ-bound PorV complex.

For ΔgldLsprA(Δnterm) peak I′ (Extended Data Fig. 10c), 1,667,681 particles were recovered following two rounds of 2D classification and then 3D classified (6 classes) against a 60 Å lowpass-filtered PorV complex reference volume. Particles (639,717) belonging to the strongest class were 3D auto-refined against their corresponding volume, yielding a 2.4 Å volume for the RemZ-bound PorV complex.

For ΔgldL ΔskpA peak I′ (Extended Data Fig. 10d), 1,419,226 particles were recovered following two rounds of 2D classification and then 3D classified (3 classes) against a 60 Å lowpass-filtered PorV complex reference volume. Particles (654,831) belonging to the strongest class were 3D auto-refined against their corresponding volume, yielding a 2.6 Å volume for the RemZ-bound PorV complex.

For ΔgldLΔsprE peak II (Extended Data Fig. 10e), 794,873 particles were recovered following two rounds of 2D classification, and then 3D classified (6 classes) against a 60 Å lowpass-filtered PorV complex reference volume. Particles (291,323) belonging to the strongest class were 3D auto-refined against their corresponding volume, yielding a 2.7 Å volume for the substrate-bound PorV complex. Alignment-free 3D classification using a mask encompassing the CTD yielded either (1) empty PorV complexes (41.0% of total particles), (2) undefined CTD–PorV complexes (39.4% of total particles) or (3) strong CTD–PorV complexes (19.6% of total particles). Particles belonging to the strong CTD-bound class were polished, CTF refined and then 3D auto-refined against the consensus 2.7 Å PorV-complex volume, yielding a 2.8 Å volume for which CTD density could be assigned to the substrate NucA.

For ΔgldL sprA(Δnterm) peak II (Extended Data Fig. 10f), 1,017,948 particles were recovered following two rounds of 2D classification, and then 3D classified (6 classes) against a 60 Å lowpass-filtered PorV complex reference volume. Particles (382,353) belonging to the strongest class were 3D auto-refined against their corresponding volume, yielding a 2.9 Å volume for the substrate-bound PorV complex. Alignment-free 3D classification using a mask encompassing the CTD yielded either (1) empty PorV complexes (50.9% of total particles), (2) undefined CTD–PorV complexes (33.9% of total particles) or (3) strong CTD–PorV complexes (15.2% of total particles). Particles belonging to the strong CTD-bound class were polished, CTF refined and then 3D auto-refined against the consensus 2.9 Å PorV-complex volume, yielding a 2.8 Å volume for which CTD density could be assigned to the substrate NucA.

For ΔgldL ΔskpA peak II (Extended Data Fig. 10g), 1,223,999 particles were recovered following two rounds of 2D classification, and then 3D classified (3 classes) against a 60 Å lowpass-filtered PorV complex reference volume. Particles (744,533) belonging to the two strongest classes were combined and 3D auto-refined against one of their corresponding volumes, yielding a 2.5 Å volume for the substrate-bound PorV complex. Alignment-free 3D classification using a mask encompassing the CTD yielded a minor class of undefined ‘junk’ particles (12.3% of total particles) and either empty PorV complexes (42.4% of total particles) or CTD–PorV complexes (45.3% of total particles). Particles belonging to the strong CTD-bound class were 3D auto-refined against the consensus 2.5 Å map, yielding a 2.6 Å volume for which clear CTD density was apparent but could not be explicitly assigned to a specific substrate.

For ΔsprE peak II (Extended Data Fig. 10h), 1,010,782 particles were recovered following two rounds of 2D classification, and then 3D classified (6 classes) against a 60 Å lowpass-filtered PorV complex reference volume. Particles (616,051) belonging to the strongest class were 3D auto-refined against their corresponding volume, yielding a 3.1 Å volume of substrate-bound PorV complex. Alignment-free 3D classification using a mask encompassing the CTD yielded a minor class of undefined ‘junk’ particles (40.4% of total particles) and either empty PorV complexes (47.4% of total particles) or CTD–PorV complexes (12.2% of total particles). Particles belonging to the strong CTD-bound class were 3D auto-refined against the consensus 3.1 Å map, yielding a 3.5 Å volume for which CTD density was apparent but could not be explicitly assigned to a specific substrate.

Model building, structure refinement and figure preparation

Atomic models of PorV complex + RemZ (with CTD in two alternate conformations), PorV complex + NucA substrate, PorV complex + mCherry–CTDRemA and PorV complex + mCherry–CTDFspA were generated by rigid-body fitting our previously deposited PorV complex model (PDB 6H3I) and building the additional components de novo into their corresponding sharpened, local-resolution filtered maps generated within Relion. The maps used for PorV complex + NucA substrate and PorV complex + RemZ conformer modelling were derived from the ΔgldL peak II and ΔgldLΔsprE peak I′ datasets, respectively. An atomic model of the Extended Translocon was generated by rigid-body fitting the RemZ-bound PorV complex and building the remaining components (SprE, SkpA, SprANterm) de novo into the sharpened composite map of the ΔgldL peak I dataset. Because density for the Extended Translocon RemZ CTD could not be unambiguously assigned to either of the conformations, both CTD conformations were modelled with equal occupancy. Iterative model building and real-space refinement using secondary structure, rotamer and Ramachandran restraints was performed in Coot (v.0.9)56 and Phenix57, respectively. Validation was performed in Molprobity58 within Phenix. Cryo-EM data collection, image processing and structure refinement statistics are listed in Supplementary Table 1. Figures were prepared using UCSF ChimeraX (v.1.4)59.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.