Structures of protein folding intermediates on the ribosome Authors: Sammy H S Chan, Julian O Streit, Tomasz Włodarski, Alkistis N Mitropoulou, Anaïs M E Cassaignau, Ivana V Bukvin, Christopher A Waudby, Lisa D Cabrita, John Christodoulou Published: 2026 ABSTRACT AbstractThe ribosome biases the conformations sampled by nascent polypeptide chains along folding pathways toward biologically active states. A hallmark of the cotranslational folding (coTF) of many proteins constitutes highly stable folding intermediates that are absent or only transiently populated off the ribosome, yet persist during translation well beyond complete emergence of the domain from the ribosome exit tunnel. Despite the importance of intermediates for folding fidelity, their structures have remained elusive and cannot be predicted by machine learning methods. Here we obtained structures of two folding intermediates of an immunoglobulin-like domain on the ribosome by developing comprehensive 19F nuclear magnetic resonance analyses using chemical shifts by rational design, paramagnetic relaxation enhancement and protein engineering, integrated with extensive molecular dynamics simulations. The resulting intermediate structures reveal native-like folded cores distinguished by nonnative termini, permitting distinct binding to a molecular chaperone and suggesting parallel folding pathways. The structures of these intermediates are conserved within the protein domain family, in contrast to their in vitro refolding mechanisms. Our detailed structural ensembles of partially folded nascent proteins on the ribosome highlight the diversity of conformations sampled during coTF, providing the ribosome with a passive means to promote efficient protein folding and maintain cellular proteostasis.Subject terms: Solution-state NMR, Ribosome Main During biosynthesis, nascent proteins emerge from the ribosome and can begin to acquire their native biologically active structures1. The ribosome itself is thought to regulate cotranslational folding (coTF) by modulating the nascent polypeptide chain’s thermodynamic stability2–8 and (un)folding rates2,6,9, through vectorial emergence10,11, steric restriction of the available conformational space12, interactions with its highly charged surface5,13–15 and with specific ribosomal proteins16–18 and the recruitment of molecular chaperones19–22. In particular, the physical chemistry of the ribosome results in the structural expansion and entropic destabilization of unfolded nascent proteins3 and enthalpic destablization of negatively charged folded states by long-range repulsive electrostatic forces23. Consequently, for negatively charged proteins (~50% proteome24), thermodynamically stable folding intermediates are the predominant conformer outside the ribosome exit tunnel2,3,23, where most tertiary folding occurs. Accordingly, CoTF contrasts with studies measuring the refolding of full-length denatured protein, where intermediates are only transiently populated3 and folding outcomes differ with generally higher propensities to misfold or aggregate; coTF is, therefore, a cellular mechanism to minimize kinetic misfolding traps25–28. Detailed descriptions of coTF pathways are absent, although structures of folding intermediates would provide mechanistic insights into how the ribosome biases nascent chain conformations away from hazardous states27,29,30 and toward conformations that are biologically active3 or accessible for cotranslational assembly31,32 or chaperoning processes19. CoTF intermediates have generally been reported to be structurally compact compared to their unfolded states33–35, with some observations of native-like structure formation2,3 within ensemble conformations15,36,37. However, unlike most mature proteins, nascent polypeptides inherently interconvert between multiple conformational states with similar thermodynamic stabilities2,38, presenting a technical challenge to their structural characterization. Solution-state nuclear magnetic resonance (NMR) spectroscopy of ribosome–nascent chain complexes (RNCs) has enabled residue-level measurements of different nascent protein conformations by using specific isotopic labeling schemes3,5,13,16,39–41, providing experimental data to guide the determination of accurate structures of unfolded3,13 and folded23 RNCs from all-atom molecular dynamics (MD) simulations. The direct observation and resolution of all conformations, including multiple partially folded coTF intermediates, has uniquely been enabled by 19F NMR2,3,16, which has the potential to extract structural information42,43. Here, we developed an array of generalizable 19F NMR analyses for structural biology and obtained direct structural contacts and intramolecular distances within two distinct coTF intermediate conformations populated by an immunoglobulin-like domain on the ribosome. These 19F NMR experiments unambiguously and quantitatively validate atomistic structural models of the coTF intermediates, obtained using biased folding simulations and extensive unbiased MD simulations, and show their conservation across immunoglobulin folds. CoTF folding intermediates identified by 19F NMR We previously mapped the folding free energy landscapes of the fifth filamin domain of the Dictyostelium discoideum gelation factor, FLN5, both off and on the ribosome, using NMR spectroscopy2,3,5,11,39. FLN5 is an immunoglobulin-like domain with two sandwiched antiparallel β-sheets (Fig. 1a). In isolation, the full-length protein is a two-state folder under denaturing conditions2 and only weakly populates an intermediate when truncated at its C terminus2,11. However, when translation-stalled on the ribosome at specific nascent chain lengths (Fig. 1b) and fully emerged from the tunnel39, FLN5 predominantly populates two stable coTF intermediates (I1 and I2; Fig. 1d,e) in slow exchange (on the NMR timescale) with the unfolded and natively folded states (Extended Data Fig. 1b,c). The intermediates are detected as broad NMR resonances (because of ribosome surface interactions2,3) with site-specific 19F-labeling using 4-trifluoromethyl-L-phenylalanine2 (tfmF; Fig. 1c) and invisible by other isotopic labeling schemes11. Measurement of 19F NMR signal integrals showed that the coTF intermediates are up to 5 kcal mol−1 more stable than the (truncated) one found off the ribosome (relative to their respective unfolded states) and persist at long nascent chain lengths during translation2 (Fig. 1d), despite complete availability of the domain for native folding to occur39. Indeed, an analysis of multiple high concentration (20 µM, ~48 mg ml−1) purified samples of FLN5+110 RNC, the final biosynthetic snapshot of the entire FLN5-6 protein, reveals high (~62%) populations of the intermediates (Fig. 1d and Supplementary Tables 1 and 2). While these data were obtained under equilibrium conditions, we used a kinetic model to assess whether the I1 and I2 states would also be expected to be formed during real-time translation. By considering the equilibrium populations together with the kinetics of translation and conservative, lower-bound estimates of folding, we anticipate large populations of I1 and I2 during real-time biosynthesis before translation termination (Extended Data Fig. 1). Structural characterization of FLN5 coTF intermediates using 19F NMR chemical shifts. a, Crystal structure of FLN5 (PDB 1QFH) with β-strands labeled. b, Schematic of FLN5 RNCs of varying L lengths. PTC, peptidyl transferase center. c, Chemical structure of tfmF. d, The 1D 19F NMR spectra of different lengths of 655tfmF-labeled FLN5 RNCs. Observed spectra are shown in gray. Native (N), unfolded (U) and intermediate state (I1, I2) resonances are indicated and were fitted to Lorentzian lineshapes (Methods). The magnification of spectra by a scaling factor is indicated. e, Folding reactions of FLN5 on and off the ribosome. f, Top, magnified view of Tyr655 and Phe675 in the FLN5 crystal structure. Bottom, 19F NMR spectrum of 655tfmF-labeled isolated wild-type and Phe675Ala mutant FLN5 (top) and FLN5+47 RNC (bottom). g, Schematic depicting ring current between two residues on adjacent β-strands and their corresponding 19F chemical shifts. h, Top, 19F NMR spectra of isolated FLN5 with 19F-label pair, overlaid with that of tfmF-labeled isolated FLN5 without accompanying aromatic or charged residue (left to right, FLN5 718tfmF 671Phe, FLN5 718tfmF 706His, 732tfmF and 726tfmF 746Phe). Bottom, 19F NMR spectra of FLN5+47 RNC with different 19F-label pairs. Arrows indicate chemical shifts of intermediate states. Observed spectra are shown in gray. The assignment of resonances is colored accordingly. Spectra were fitted to Lorentzian lineshapes (Methods). All spectra were recorded at 500 MHz and 298 K. Kinetics of coTF of FLN5. a, Kinetic models showing predicted populations of unfolded, intermediate and native states during real-time biosynthesis with a translation rate of 5 and 20 amino acids per second and experimental estimates of folding rates (see Methods). The solid lines represent the predicted values while dots show the experimentally measured equilibrium populations (this work and refs. 2,3). The schematics on the left illustrate the interconversions considered in the model, ranging from parallel pathways to sequentially formed intermediates. Simulations start at FLN5+31 (the first linker length at which the entire domain has emerged from the tunnel and begins to fold at equilibrium39) and terminate at FLN5+110 at which point the NC is released from the ribosome. b, 19F CEST measurements of FLN5+47 RNC. Given sensitivity limitations, we selected four frequencies at which to irradiate (of which one was off-resonance from all NMR peaks and shown in red, arrows indicate remaining irradiation frequencies) with the 15-Hz B1 field. Observed spectra (grey) were fitted (black), with residuals shown below, by analysing in the time domain using the Bloch-McConnell equations, as previously described2. Spectra were recorded at 298 K, 500 MHz. c, Exchange rates between FLN5+34 and FLN5+47 nascent chain states determined by CEST measurements in ref. 2 and this work respectively, using an estimated R1 of 1.1 s−1 for all RNC states. Error determined from data fits. The folding rate of isolated FLN5 (truncated variant FLN5∆6) was previously reported11. Source data Structural contacts revealed by 19F NMR chemical shifts The coTF intermediates adopt a native-like hydrophobic core, evidenced by their complete destabilization upon substitution of the natively buried residue Tyr719 (ref. 2); however, further details of their structures are sparse. To begin to extract structural information by 19F NMR spectroscopy44, we examined 19F chemical shifts and their relationship to protein structure, exploiting our recent development of solvent-exposed tfmF labels experiencing ring-current effects42, which are distance and angle dependent on nearby aromatic residues. Residue 655 on the A strand directly contacts the aromatic ring of Phe675 on the neighboring B strand; its tfmF chemical shift is consequently shielded from the random coil value by ~0.8 ppm (Fig. 1f)2 and correlated to the ring-current effect as shown by the Phe675Ala substitution42 (Fig. 1f). The tfmF655 chemical shift is, thus, a direct reporter of the interresidue 655–675 side-chain van der Waals contact42 comprising a precise, short-range (4–6 Å) distance and angular preference, as quantifiable by structural modeling42. The observed broad I2 resonance of tfmF655-labeled FLN5 RNCs shares its chemical shift with the native state and, thus, demonstrates a native interresidue 655–675 tertiary contact and complete A–B strand pairing within this intermediate structure. Conversely, the random coil tfmF655 chemical shift of I1 (−61.8 ppm) indicates the absence of the native A–B strand contact (Fig. 1f,g). Additional naturally occurring ring currents are absent in FLN5 and tfmF labels at other solvent-exposed sites on FLN5 yield chemical shifts that are poorly resolved between the folded and unfolded (random coil) states42 and, thus, do not show clear correlations with (local) protein structure, limiting their use in structural studies. Using ColabFold45, AlphaFold3 (ref. 46) and all-atom MD simulations, we rationally designed42 five additional 19F-label pairs to engineer ring currents and probe specific side-chain contacts between pairs of β-strands, encompassing the B–E, D–E and F–G strands (Fig. 1h and Extended Data Fig. 2), akin to the introduction of (bulkier) Förster resonance energy transfer (FRET) or double electron–electron resonance probes. Efforts to generate label pairs for C–F and A′–G using the ring-current strategy resulted in more modest chemical shift dispersions insufficient to resolve broad RNC resonances (Extended Data Fig. 2b). The successful FLN5 label variants, in isolation of the ribosome, remained two-state folders and showed only mild changes to their folding free energies (∆∆GN–U < 0.8 kcal mol−1 relative to nonfluorinated; Extended Data Fig. 2c) and minimal perturbations to their 1H,15N-correlated NMR resonances (Extended Data Fig. 2d), being solvent exposed. Moreover, their mildly (de)stabilizing effect is expected to be largely suppressed on the RNC because of the ribosome’s mutation buffering effect (that is, changes in folding free energies are reduced on the ribosome; Supplementary Table 2)3,29. Their 19F chemical shifts were substantially (de)shielded (>0.2 ppm) from the random coil and highly sensitive to the presence of the introduced aromatic ring (Fig. 1h), as designed. Chemical shifts engineered using ring currents can, therefore, be interpreted as a residue-pair-specific van der Waals contact42. Characterisation of solvent-exposed tfmF-labelled variants of FLN5 on and off the ribosome. a, 19F NMR spectra of FLN5 tfmF-labelled at different positions, as isolated protein, with/without urea and with/without the accompanying aromatic label residue, and as RNC. Observed spectra shown in grey were fitted and assigned to their conformational states (coloured). Total fitted spectra are shown in black, and residual spectrum after fitting are shown below. Spectra were recorded at 500 MHz and 298 K. b, 19F NMR spectra of FLN5 tfmF-labelled at different positions. Red vertical line indicates random coil chemical shift range, and therefore the limited chemical shift dispersion of the variants. Observed spectra shown in grey were fitted (blue), with the residual spectrum after fitting shown below. Spectra were recorded at 500 MHz and 298 K. c, Gibb’s free energies of folding of different tfmF-labelled FLN5, determined by lineshape analyses of spectra shown in a. Errors were determined by bootstrapping of residuals from lineshape fits (mean ± s.e.m.). Values for wild-type and 655tfmF FLN5 were previously reported2,11. d, 2D 1H,15N-SOFAST HMQC spectra of tfmF-labelled FLN5 overlaid with that of wild-type FLN5. Chemical shift perturbations (CSPs) are shown below. Spectra were recorded at 500 MHz (wild-type at 800 MHz) and 298 K. e, Simulations of FLN5 732tfmF. Left shows a structural model highlighting the 732tfmF and Glu692 sidechains. Top middle/right shows average ( ± s.e.m.) Cα-RMSF of wild-type and 732tfmF FLN5 simulated with C36m and ff15ipq parameters. Bottom middle/right shows distances between the centres of mass of tfmF CF3 and Glu692 carboxylic acid group, for 3 independent simulations using C36m and ff15ipq parameters. Source data We introduced each label into FLN5+47 RNC, the length at which the coTF intermediates are maximally populated2 (44% ± 3% and 38% ± 6% of I1 and I2, respectively, for FLN5+47 655tfmF; Fig. 1d). First, FLN5+47 RNC was labeled to probe the B–E strands (718tfmF 671Phe), yielding 19F NMR spectra with a sharp (~18-Hz linewidth) resonance at an identical chemical shift to that of the isolated protein (−62.2 ppm), corresponding to the native state as expected (Fig. 1d), and two additional broad resonances at the native and random coil (−61.8 ppm) chemical shifts attributable to intermediate states (Supplementary Note 1). The distinct chemical shifts, thus, indicate two populations of intermediates, one possessing and one lacking the native B–E strand contact. Similarly, labeling of FLN5+47 RNC at an alternative position across the B–E strands (673tfmF 716His) also produced two broad resonances, having native and random coil chemical shifts (Extended Data Fig. 2a) and suggesting that the presence or absence of native contacts persists across the interstrand network, respectively, thus completing secondary-structure formation. When labeled to probe either the D–E (718tfmF 706His) or F–G (726tfmF 746Phe) strands, FLN5 RNCs also produced broad resonances attributable to two intermediate states (Fig. 1h). In each case, the broad resonances were found at both native and random coil chemical shifts, consistent with alternative label pairs positioned along the same strands (710tfmF 716His and 728tfmF 744His; Extended Data Fig. 2a and Supplementary Note 1). In addition to the findings for the A–B and B–E strands, native interstrand contacts across the D–E and F–G strands are, therefore, only found in one intermediate. The 732tfmF (F strand) chemical shift is highly sensitive to the electrostatic charge of the native neighboring residue Glu692 (C strand) in the isolated protein42 (Fig. 1h and Extended Data Figs. 1e and 2a). In addition to the native-state resonance, the 19F NMR spectrum of FLN5+47 732tfmF showed a single broad resonance (67% ± 11% population) observed at the same chemical shift that was attributable to an intermediate state (Extended Data Fig. 2a). Only a lowly populated (11% ± 4%) peak was observed at the random coil chemical shift, having a linewidth (Supplementary Table 1) and small population (Supplementary Table 2) consistent with an unfolded state, the latter as expected from the more destabilizing nature of the 732tfmF label (relative to other 19F-label pairs; Extended Data Fig. 2c), and its identity was confirmed by its observed increase in population in response to the addition of urea and comparison to an unfolded mutant RNC (Extended Data Fig. 2a). Similar results were found for the shorter FLN5+34 RNC (Extended Data Fig. 2a and Supplementary Note 1). These data indicate that both I1 and I2 intermediates possess a native 732–692 contact. Additional structural contacts in the hydrophobic core were measured with natively buried 19F labels (Extended Data Fig. 3). Characterisation of natively buried tfmF-labelled variants of FLN5 on and off the ribosome. a, 19F NMR spectra of FLN5 tfmF-labelled at different positions, as isolated protein, with/without urea and with/without the accompanying aromatic label residue, and as RNC. Observed spectra shown in grey were fitted and assigned to their conformational states (coloured). Total fitted spectra are shown in black, and residual spectrum after fitting are shown below. Spectra were recorded at 500 MHz and 298 K. b, 2D 1H,15N-SOFAST HMQC spectra of tfmF-labelled FLN5 overlaid with that of wild-type FLN5. Chemical shift perturbations (CSPs) are shown below. Spectra were recorded at 500 MHz (wild-type at 800 MHz) and 298 K. c, Gibb’s free energies of folding of different tfmF-labelled FLN5, determined by lineshape analyses of spectra shown in a. Errors were determined by bootstrapping of residuals from lineshape fits (mean ± s.e.m.). Value for wild-type FLN5 was previously reported11. d-g, Simulations and interactions of additional FLN5 tfmF sidechains with ff15ipq parameters. Solvent-exposed sites (655tfmF and 740tfmF) were used as controls. For wild-type, 655tfmF and 740tfmF, three independent simulations of 1 μs were performed. For 691tfmF, 705tfmF, 715tfmF, 719tfmF, and 727tfmF six simulations of 1 μs (and in three simulations water coordinates were saved for analysis). d, Solvent-accessible surface area of the CF3 group (mean ± s.e.m., n = 6) for all variants. e, Population of hydrogen bonds (H-bonds) between CF3 and water (n = 3), and between CF3 and protein atoms (n = 6), for all variants (mean ± s.e.m.). f, Most common hydrogen bond contacts formed by 705tfmF, 715tfmF and 727tfmF with amide groups of residues 693, 689 and 723, respectively. g, Averaged hydrogen bond existence autocorrelation functions and estimated lifetimes (mean ± s.e.m. or upper bound, from the area under the curve). Individual lines represent simulations replicates. Source data Collectively, by using and engineering 19F chemical shifts that unambiguously probe specific intramolecular contacts, we resolved two distinct intermediate states in slow exchange with the natively folded state (Supplementary Note 1), both with native-like structure between the C–F strands and where additional native contacts across the remaining β-strand pairs and buried core are found in only one intermediate conformation. Engineering reversible crosslinks stabilizes I1 We next aimed to assign the broad resonances to either I1 or I2 to generate a dataset of chemical shifts (that is, structural contacts) specific to either coTF intermediate. Thus, we sought to selectively (de)stabilize one intermediate conformation using principles of psi-value analysis47, introducing reversible crosslinks by engineering bihistidine metal-ion-binding sites to stabilize specific β-strands. Of the seven mutants tested, one variant, 726His 746His (F–G strands), showed a progressive stabilization of up to −1.45 kcal mol−1 in its folding free energy with increasing concentrations of NiCl2 (Fig. 2a–c and Extended Data Fig. 4a–c). In the absence of NiCl2, its stability was similar to that of the wild-type protein (Fig. 2c). The metal-binding site was confirmed by 1H,15N NMR, where addition of the paramagnetic Ni(II) metal resulted in increased line broadening localized at the substitution sites (Fig. 2a and Extended Data Fig. 4b–h), in accordance with AlphaFold3 predictions (Fig. 2a). Selective stabilization of a specific coTF intermediate using an engineered ligand-binding site. a, AlphaFold3-predicted structure of FLN5 Asn726His Lys746His, highlighting metal ion coordination by the engineered histidine residues in the neighboring Asp744, and colored according to the 1H,15N-correlated NMR signal intensity in 2.5 mM NiCl2 relative to without NiCl2 (complete dataset in Extended Data Fig. 4c). b, 19F NMR spectra of 655tfmF-labeled FLN5 Asn726His Lys746His incubated in 2 M guanidine hydrochloride and varying NiCl2 concentrations. Observed and total fitted spectra are shown in gray and black, respectively. Unfolded (U), nickel-bound unfolded (Ub; Extended Data Fig. 4g) and native (N) states are in color and indicated. c, Folding free energy changes of isolated wild-type and bihistidine FLN5 in varying NiCl2 concentrations (relative to without NiCl2), determined by lineshape analysis of spectra shown in b and Extended Data Fig. 4c. Errors were determined by bootstrapping of residuals from lineshape fits. d, Schematic of equilibrium reaction between unfolded and folded β-strand pairs titrated with Ni(II). e, 19F NMR spectra of tfmF-labeled FLN5+47 RNCs with Asn726His Lys746His, with and without 3.5 mM NiCl2 (complete titration data in Extended Data Fig. 4j). Observed and total fitted spectra are shown in gray and black, respectively. The assignment of resonances is colored accordingly. f, Changes in Gibbs free energies of intermediate relative to native state, determined by lineshape analysis of spectra shown in e and Supplementary Fig. 3j (complete titration data in Extended Data Fig. 4j). Errors were determined by bootstrapping of residuals from lineshape fits. g, Schematic of equilibrium reaction among I1, I2 and native states titrated with Ni(II). Nickel binding stabilizes I1 more than I2, inferring the presence and absence of the bihistidine-binding site, respectively. I1 and I2 are also distinguished by their ring-current-sensitive chemical shifts. All spectra were recorded at 500 MHz and 298 K. Unless stated otherwise, all values in the figure represent the mean ± s.e.m. propagated from NMR lineshape fits. Source data Engineering ligand binding sites to stabilise specific conformational states. a, 19F NMR spectra of isolated FLN5 with bi-histidine mutations at different positions (across different β-strands) and with varying concentrations of NiCl2. All show no substantial stabilisation of the natively folded state with NiCl2. b, Top shows residue-resolved 1H,15N-SOFAST HMQC signal intensities of FLN5 655tfmF 726His 746His in varying NiCl2 concentrations relative to without NiCl2. Bottom shows 1H,15N-chemical shift perturbations of FLN5 655tfmF 726His 746His in 10 relative to 0 mM NiCl2. c, 19F NMR spectra of isolated FLN5 655tfmF 726His 746His with varying concentrations of NiCl2. d, Linewidths determined by lineshape analysis of spectra shown in c. Errors were determined by bootstrapping of residuals from lineshape fits. e, Residue-resolved 1H,15N-SOFAST HMQC signal intensities of FLN5 655tfmF (without bi-histidine mutations, 2.5 mM dataset mapped onto the FLN5 crystal structure, PDB 1QFH, right) and FLN5 655tfmF Tyr719Glu (the latter unfolded39) in varying NiCl2 concentrations relative to without NiCl2. f, 19F NMR spectra of isolated FLN5 (without bi-histidine mutations) in varying NiCl2 concentrations. g, 19F NMR spectra of isolated FLN5 Tyr719Glu (without bi-histidine mutations, and unfolded39) in varying NiCl2 concentrations. h, 1D 19F NMR spectra of isolated FLN5 726His 746His in varying ZnCl2 concentrations. i, 19F NMR spectra of FLN5+34 tfmF655tfmF 726His 746His RNC in varying NiCl2 concentrations, bottom shows Gibb’s free energies calculated using lineshape analyses of spectra. Errors were determined by bootstrapping of residuals from lineshape fits. j, As in i, but for FLN5+47 with different 19F-labelled and non-fluorinated variants. k, Linewidths determined by lineshape analysis of spectra shown in j. Errors were determined by bootstrapping of residuals from lineshape fits. For all 19F NMR spectra, observed spectra shown in grey were fitted and assigned to their conformational states (coloured). Total fitted spectra are shown in black, and residual spectrum after fitting are shown below. All spectra were recorded at 500 MHz and 298 K. Unless stated otherwise, all values in the figure represent mean ± s.e.m. propagated from NMR lineshape fits. Source data Having designed a conformation-specific, concentration-dependent ligand-binding site (Fig. 2d), we next introduced the bihistidine substitutions into FLN5+47 RNCs (Fig. 1h and Extended Data Fig. 4i–k). FLN5+47 tfmF655 RNCs (probing A–B strands; Fig. 1f) yielded resonances corresponding to I1, I2 and native states as expected (Fig. 2e). Addition of increasing amounts of Ni(II) to the RNC sample resulted in increasingly broader resonances, as anticipated from the metal’s paramagnetic effect (Extended Data Fig. 4k), but critically also a progressive shift in the folding equilibrium toward the I1 state (Fig. 2e,f and Extended Data Fig. 4j) only in the presence of the bihistidine substitutions (Extended Data Fig. 4j). An analysis of the populations and Gibbs free energies of each state revealed that I1 was stabilized more than I2 (each relative to N; Fig. 2f). Given the distinct chemical shifts obtained for the F–G strand 19F labels (Fig. 1h and Extended Data Fig. 4a), this indicated the presence of a folded binding site across the F–G strands in I1 that was likely absent in I2 (Fig. 2g). Therefore, I1 possesses a native F–G strand conformation but lacks A–B strand contacts (Fig. 2g), whereas I2 exhibits the opposite. The bihistidine site was introduced in FLN5+47 RNCs possessing the remaining 19F-label pairs, all resulting in only minimal perturbations in the thermodynamic stabilities of both intermediates (relative to the native state; ∆∆G(mut-wt, I1–N) = 0.3 ± 0.1 kcal mol−1 and ∆∆G(mut-wt, I2–N) = 0.0 ± 0.2 kcal mol−1) and their 19F chemical shifts at solvent-exposed (<0.02 ppm) and solvent-inaccessible (<0.15 ppm) sites. When incubated with Ni(II), all variants showed two distinct intermediate resonances (Fig. 1h and Extended Data Fig. 2a), with one intermediate (I1) consistently stabilized more than the other (I2) (Fig. 2e,f and Extended Data Fig. 4j). These data permitted us to assign the chemical shifts across all the tfmF labels to either the I1 or I2 intermediate (Fig. 1g and Extended Data Fig. 4j). Collectively, they show that I1 likely possesses a natively folded sheet 2 (CFG strands) with native contacts absent across sheet 1 (ABED strands). However, native-like strands are likely to be at least partially formed within sheet 1, as I1 is completely destabilized by the natively buried Tyr719Glu substitution in the core between the two β-sheets2. Meanwhile, on the basis of the 19F NMR contacts, I2 appears to closely resemble the intermediate found in isolated C-terminally truncated variants of FLN5 (ref. 11), comprising a natively folded A–F strand structure with a disordered C-terminal G strand; accordingly, native-like chemical shifts for the natively buried sites tfmF715 and tfmF727 also correspond to the I2 intermediate (attributed using the same bihistidine-binding approach; Extended Data Fig. 3a,j,k). Structures of folding intermediates on the ribosome Next, we aimed to convert our topological observations into detailed structures, by producing atomistic structural ensembles of the FLN5+47 RNC intermediate states. The structural similarity of contacts identified experimentally for I2 to the topology of isolated, truncated intermediate enabled us to produce a starting model of the RNC by attaching the previously determined structure of the latter11 to the ribosome (Supplementary Note 2). However, in the absence of any analogous structures to I1, we sought to simulate coTF from the unfolded to natively folded state using an all-atom (including ribosome, ions and solvent) ratchet-and-pawl MD (rMD) approach48–50 (Supplementary Note 2, Supplementary Videos 1 and 2 and Extended Data Figs. 5 and 6). The rMD simulations enabled us to successfully capture folding events, which 19F chemical exchange saturation transfer (CEST) NMR experiments show to likely occur on the timescale of >170 ms (Extended Data Fig. 1b,c), beyond the reach of conventional MD simulations, to produce a series of potential folding pathways. From the obtained coTF trajectories, we identified one on-pathway folding intermediate state (Extended Data Fig. 6), which, from our experimental dataset, appeared to possess structural features common to that of I1. To assess this further, we subsequently subjected the putative I1 and I2 models to extensive atomistic unbiased MD simulations (Extended Data Fig. 7), reducing the impact of rMD bias potentials on the final structural ensemble51. We found their folded regions to be stable on the microsecond timescale (Fig. 3 and Supplementary Videos 3 and 4) and, therefore, quantitatively compared the atomistic structural ensembles to those previously obtained for the natively folded RNC23. Structure-dependent proline isomerisation of FLN5 Pro742. a, Dihedral angles ζ and ψ used as collective variables for enhanced sampling simulations with WT-METAD. b, Sampling of ζ in three independent WT-METAD simulations of the unfolded peptide. c, Reweighted free energy landscape of Pro742 projected onto ζ and ψ, respectively, for the unfolded peptide (mean ± s.e.m., n = 3). d, Reweighted 2D free energy landscape of P742 projected onto ζ and ψ for the unfolded peptide. e, Sampling of ζ in three independent WT-METAD simulations of FLN5Δ6. f, Reweighted free energy landscape of P742 projected onto ζ for FLN5Δ6 (mean ± s.e.m., n = 3). g, Sampling of C-terminal native contacts (fraction of native contacts formed by residues I743 and D744) in three independent WT-METAD simulations of FLN5Δ6. h, Reweighted 2D free energy landscape of FLN5 Δ6 projected onto ζ and CTERM averaged across all three simulations. The energy minima corresponding to the native (N) and intermediate (I) state are annotated. i, Reweighted 2D free energy landscape of FLN5 Δ6 projected onto ζ and ψ averaged across all three simulations. j, Structural models (10 lowest free energy conformations) corresponding to FLN5Δ6 N and I states and their equilibrium populations predicted by WT-METAD including their P742 cis populations. The C-terminal residues Ile743 and Asp744 are shown in magenta. k, Sampling of ζ in three independent WT-METAD simulations of full-length, folded FLN5. l, Reweighted free energy landscape of full-length FLN5 projected onto ζ (mean ± s.e.m., n = 3). m. Reweighted 2D free energy landscape of full-length FLN5 projected onto ζ and ψ averaged across all three simulations. n. Thermodynamic cycle of the free energy changes of folding (ΔGN-U) and P742 isomerisation (ΔGcis-trans). The change in folding free energies (ΔΔGN-U,cis-trans) is equivalent to the change in isomerisation energies of the unfolded (U) and native (N) state (ΔΔGN-U,cis-trans). Using the value of the isomerisation free energies for U and N computed in panels c and l, we obtain a value of 3.70 ± 0.38 kcal mol−1. This is quantitatively consistent with the experimental value (ref. 11), 4.00 ± 0.10 kcal mol−1, for ΔΔGN-U for the P742A mutant (with A742 in trans). All-atom folding simulations of FLN5 using rMD. a, FLN5 crystal structure (PDB 1QFH) annotated with the main β-strands. b-c, Transition path free energy landscape of isolated FLN5 (b) and FLN5+47 RNC (c) as a function of the fraction of native contacts (Q), all-atom RMSD with respect to the native structure, and native contacts in sheet 1 (ABED) and 2 (CFG). The 1D energy plot shows the mean and error, estimated as the difference between the first and second half of the dataset (n = 2). 321 and 226 folding trajectories obtained from rMD simulations were used for the analysis of isolated FLN5 and FLN5+47, respectively. d, Average fraction of native contacts calculated for structures attributed to the folding intermediate, I1, energy minimum. Error bars were estimated as the difference between the first and second half of the dataset (n = 2). e-f, Transition path free energy landscape of isolated FLN5 (e) and FLN5+47 RNC (f) projected onto the native contacts of the main β-strand pairs. g, Five structural models of I1 on the ribosome, selected for unbiased MD simulations (see Methods). h, Aligned structural models of I1 using residues 685–750 (strands C to G) showing the folded core of I1 and disordered N-terminus (transparent, strands A to B). i, Table summarising the tested force constants (κ) and observed folding success rates for isolated FLN5 (total number of trial trajectories given in parentheses). j, Average fraction of strand pair contacts (strand pair Q) as a function of global folding progress (Q) for different force constants. Region of Q corresponding I1 (identified with κ = 100 kJ / mol, b) is shaded. k, Transition path free energy landscape of isolated FLN5 for two different force constants as a function of Q and the all-atom RMSD with respect to the native structure. Source data Structures, dynamics, and interactions of I1 and I2 coTF intermediates. a, Root mean square fluctuations (RMSF, measured at Cα atoms) across the sequence of FLN5 (mean ± s.e.m. from n = 15 independent simulations for I1/I2 and n = 8 independent simulations for N) and comparing all conformational states. Secondary structure regions (β-strands) of native FLN5 are annotated. b, (left) Structures of I1 from the lowest free energy structural cluster aligned using the common core of strands C to G where the B-strand is also associated with the core. (right) Energy landscape of I1 correlating native contacts between strands A-B (AB) and B-E (BE), showing two parallel folding routes towards the N state. The relative populations of these pathways were estimated from the rMD transition path ensemble (errors represent the difference between the first and second half of the dataset). c, Difference contact probability of I1 and I2 compared to N (N-I) highlighting loss of native (blue) and gain of non-native (red) contacts. The black contours represent the native contacts calculated from the crystal structure. Contacts were defined as heavy atom distances of less than 0.5 nm. Two regions in the box highlight non-native contacts (C-G strand contacts in I1 and A’-F strand contacts in I2). d, Probability (mean ± s.e.m.) of hydrogen bonds and sidechain contacts (0.35 nm heavy atom cut-off) between non-native structural elements (highlighted in c) compared to the native state. e, Representative structures of I1 showing non-native C-G strand contacts. f, Representative structures of I2 showing non-native A’-F strand contacts. g, Secondary structure propensity of coTF intermediates (lines) compared to the native state shown with bars (mean ± s.e.m.). h, Ribosome interaction probability (mean ± s.e.m.) shown along the sequence of FLN5, with native β-strands annotated. Negatively and positively charged residues are denoted with red and blue circles, respectively. i, FLN5 (residues 646–750) interactions mapped onto the ribosome surface for all conformational states. j, Distance distributions and mean ± s.e.m. of residue sidechains corresponding to 19F-label pairs calculated for I1, I2 and native state structural ensembles. k, Light bars show rotational correlation time of FLN5 calculated for all conformational states (mean ± s.e.m.), defined using the natively folded regions for I1 and I2. Dark bars show rotational correlation time calculated using experimental linewidths and taken from ref. 3. l, Ring current geometric factor (1–3cos2θ)/r3) (de)shielding contribution to the 19F chemical shift of probes reporting on strand pair contacts calculated from MD simulations of the native and I1 state using isolated proteins (mean ± s.e.m.). m, Tyr-OH hydrogen bonding probability calculated for all conformational state ensembles (mean ± s.e.m.). n, Population of hydrogen bonds (H-bonds) between CF3 and protein atoms calculated using isolated protein MD simulations (mean ± s.e.m.). Source data Atomistic structures of FLN5+47 coTF intermediates on the ribosome. a, Structural ensemble of FLN5+47 I1 on the ribosome. Inset, representative structure. b, Lowest-energy structural cluster of I1 (Supplementary Table 4) aligned against the common C–G strand core. c, Structural ensemble of FLN5+47 I2 on the ribosome. Inset, representative structure. d, Lowest-energy structural cluster of I2 (Supplementary Table 4) aligned against the common A–F strand core. e, Average fraction of native contacts (Q) of FLN5 and across different β-strand pairs and β-sheets for I1 and I2 (mean ± s.e.m. from n = 15 independent simulations of 1 µs). f, Radius of gyration probability distribution and mean ± s.e.m. of I1, I2 and native states. g, Difference in SASA (mean ± s.e.m.) of I1 and I2 (n = 15) relative to the native state (n = 8). h, Distance distributions and mean ± s.e.m. (n = 15) of residue side chains corresponding to 19F-label pairs calculated for structural ensembles in I1, I2 and native states. Source data The structures of the putative I1 state (Fig. 3a,b and Extended Data Fig. 7a–g) possess a native-like C–G strand folded core (Fig. 3e,f); the remaining N-terminal A, A′ and B strands are largely detached from the structure and form either a completely disordered tail (Supplementary Video 1) or a β-hairpin (Fig. 3e, Extended Data Fig. 7b and Supplementary Video 2). The loss of the A–B strands from the folded structure results in transient nonnative contacts, predominantly between the β-sheets at the exposed E–G and A′–F strand interfaces (~20–30%; Extended Data Fig. 7c–e). However, we also find a subpopulation (~24%) where the B strand folds onto the C–G core, with the A–A′ strands remaining disordered (Extended Data Fig. 7b); this conformational heterogeneity suggests multiple accessible pathways from I1 to the native state, as also reflected in our folding simulations (Extended Data Fig. 6), although these two I1–N routes cannot be resolved by our experiments. Overall, detachment of the N-terminal strands causes an expansion of the structural ensemble, with a larger radius of gyration and increased solvent-accessible surface area (SASA) compared to the native state (Fig. 3g). Structures of the putative I2 state (Fig. 3c,d and Extended Data Fig. 7a–g) adopt a more ordered and compact native-like conformation, with a smaller increase in SASA (Fig. 3f,g). Native contacts are found across the A–F strands (Fig. 3e), with a nonnative trans-Pro742 and disordered G strand (residues 743–747) detached from the rest of the domain permitting some nonnative interactions between A′ and F strands (Extended Data Fig. 7c–f). We note that energetic calculations show that native G strand folding is dependent on the cis–trans isomerization state of Pro742 (Supplementary Note 2 and Extended Data Fig. 5). To assess whether the obtained structures faithfully represent our experimental data (Figs. 1 and 2), we examined their intramolecular contacts and compared them to the obtained 19F chemical shifts, reporting on precise, short-range interresidue distances42. We calculated distance distributions between the pairs of residues labeled for NMR analysis (Methods, Fig. 3h and Extended Data Fig. 7j) and found that all could account for the observed chemical shifts of I2; those with a native chemical shift showed short (~0.5 nm), native-like interresidue distances, consistent with the measured van der Waal’s contact, whereas residues labeling the F–G strands were distanced far (>2 nm) from each other, as reflected by their random coil chemical shifts (Fig. 1f,h and Extended Data Fig. 2a). Analyses for I1 also showed good agreement between interresidue distances and their ring-current-induced chemical shifts (Fig. 3d and Extended Data Fig. 7j). The exceptions were the two D–E strand labels, which could instead be resolved using more precise, explicit modeling of the 19F-label pair; the D–E strand pairs form native backbone contacts but lack side-chain interactions in line with the random coil chemical shifts found in this region (Supplementary Note 3 and Extended Data Fig. 7l). Likewise, such simulations entirely validate residue contacts in the core of the domain (Supplementary Note 3 and Extended Data Fig. 7m,n). In summary, our residue contact analyses show that both I1 and I2 structures quantitatively account for all of our 19F chemical shift data and, thus, that the rMD folding simulations captured an intermediate conformation corresponding to I1. Interactions and orientations of intermediates with the ribosome The intermediate-state structures show transient interactions with the ribosome surface, predominantly with ribosomal proteins uL24, uL29 and RNA H59 lining the vestibule of the exit tunnel and more favorably than the native state (Extended Data Fig. 7h,i). The dynamics of these interactions are consistent with line broadening of their 19F NMR resonances2,3 and the interaction sites and orientations were found to be in accord with both chemical crosslinking and ensemble-reweighting52 analyses of cryo-electron microscopy (cryo-EM) maps of FLN5+47 RNC53 (Extended Data Fig. 8 and Supplementary Note 4). Validation of inter-molecular interactions and orientations of structural ensembles. a-c, Cryo-EM fitting and reweighting of structural ensembles. a, Probability distributions of the correlation coefficients obtained by fitting single conformations from all-atom MD trajectories of the native, I1, and I2 states to one of four cryo-EM maps, corresponding to the AP1, AP2, AP5 and P2 classes of FLN5+47 (Methods). b, Bayesian reweighting of the collective I1, I2 and N trajectories using the cryoEM maps of FLN5+47, obtained by cryoENsemble52. The reweighted structural ensembles are shown, coloured according to their conformational state, and their opacity set to their populations. Total posterior populations of each state are plotted in the bar chart. c, Cross-correlations obtained for cryoENsemble reweighting shown in b. We note that high cross-correlations of individual states do not necessarily correlate with their populations in the combined reweighting. States which agree only moderately with the cryo-EM map individually can still be in good agreement with the maps when averaged together. d-e, Cross-linking experiments of FLN5+47 RNC on uL24 Asn53Cys ribosomes. d, Top shows probability distributions and averages (mean ± s.e.m., vertical lines) for all distance pairs (uL24 Asn53 with cysteine residue on FLN5) and conformational states calculated from structural ensembles, middle shows averaged distances (mean ± s.e.m. from n = 15 independent simulations for I1/I2 and n = 8 independent simulations for N), and bottom shows probability of contacts using a cut-off of 2 nm. e, Anti-hexahistidine detected western blot of FLN5+47 cysteine variant (pseudo wild-type Cys747Val, Methods) cross-linked to uL24 Asn53Cys using BM(PEG)2 (Methods). Ribosome-bound nascent chain (covalently linked to its tRNA, labelled tRNA-NC) and same but covalently linked to uL24 are indicated. Source data 19F paramagnetic relaxation enhancement (PRE) NMR measurements validate intermediate structures We sought to complement the 19F chemical shift data by measuring long-range (10–30 Å) distances using 19F PRE NMR. Having identified detachment of the A–B strands as a key structural feature in I1 (Fig. 3a,b), we covalently linked paramagnetic spin labels on the adjacent B strand (Thr673Cys) and F–G loop (Asn740Cys) and measured their effect on the tfmF655 (A strand) resonance (Methods and Fig. 4a). As expected, the isolated, folded protein showed line broadening in the presence of either paramagnetic label but only marginal changes when the protein was fully unfolded (Fig. 4b and Extended Data Fig. 9a). Validation of coTF intermediate structures by 19F PRE NMR. a, Location of tfmF and paramagnetic spin labels shown on the FLN5 crystal structure (PDB 1QFH). b, 19F NMR spectra of isolated 655tfmF-labeled FLN5 (pseudo-wild-type Cys747Val (Methods) and unfolded with destabilizing substitutions (A3A3)39, spin-labeled at positions 673Cys (left) and 740Cys (right). Observed spectra are shown in gray. c, 19F NMR spectra of 655tfmF-labeled FLN5+47 RNC (pseudo-wild-type Cys747Val), spin-labeled at positions 673Cys (left) and 740Cys (right). Observed and total fitted spectra are shown in gray and black, respectively. The assignment of resonances is colored accordingly. d, Distances between the 19F and spin labels at 673Cys (left) and 740Cys (right); bars show the experimental values determined by lineshape analysis (mean ± s.e.m.) of spectra shown in c (Methods), while dots show r6-averaged values (mean ± s.e.m.) from the intermediate structures. Errors of experimental values were determined by bootstrapping of residuals from lineshape fits. All spectra were recorded at 500 MHz and 298 K. Source data Validation of structural ensembles using 19F PRE measurements. a, PRE rates measured for isolated variants of FLN5 (pseudo wild-type Cys747Val, Methods), spectra shown in Fig. 4. Errors were determined by bootstrapping of residuals from lineshape fits (mean ± s.em.). b, 19F NMR spectra of FLN5+47 tfmF655 RNC variants under paramagnetic and diamagnetic conditions. Observed spectra shown at top, and also in grey were fitted and assigned to their conformational states (coloured). Total fitted spectra are shown in black, and residual spectrum after fitting are shown below. Spectra were recorded at 500 MHz and 298 K. c, Linewidths (mean ± s.d.) measured by lineshape analyses of diamagnetic spectra shown in b, used to calculate restraints for lineshape analyses of native state paramagnetic resonances and to determine rotational correlation times for conversion of PRE rates to distances (see Methods and Supplementary Note 5). d, PRE rates calculated from lineshape analyses of spectra shown in b. Errors were determined by bootstrapping of residuals from lineshape fits (mean ± s.e.m.). The native-state resonance was also broadened in paramagnetic spin-labeled FLN5+47 RNC. However, lineshape analysis of the resonances (Methods) revealed differential effects between the I1 and I2 linewidths and the two spin label sites (Fig. 4c and Extended Data Fig. 9b–d). We used the measured linewidths to calculate distances between the 19F and paramagnetic spin labels (Supplementary Note 5). Native-like distances between A–B strands and A-loop(FG) were evaluated for I2 (Fig. 4d), consistent with its determined structural ensemble (Fig. 3c,d). In contrast, larger distances were calculated for I1—in particular, a substantial increase (to >20 Å) for the 740Cys–MTSL spin label (measuring between A–B strands and A-loop(FG)) compared to the native state (Fig. 4d). These data agree with the structures (Fig. 3a,b), where the A strand remains detached from the C terminus across all I1 conformations, while the formation of a β-hairpin in subpopulations of I1 rationalizes the shorter ensemble-averaged distance between A and B strands. Thus, collectively, the NMR chemical shifts (Figs. 1 and 2), crosslinking (Extended Data Fig. 8d), cryo-EM (Extended Data Fig. 8a–c) and PRE data (Fig. 4 and Extended Data Fig. 9) experimentally confirm that the MD-derived structural models of I1 and I2 (and their native-like character) are accurate. Folding modulated by a molecular chaperone We next examined possible interactions with the ribosome-associated molecular chaperone trigger factor (TF). In contrast to the isolated (destabilized) FLN5 variant (Fig. 5a)11, we observed substantial alterations in the 19F NMR spectra of FLN5 RNCs on addition of TF (Fig. 5a). Resonances of both I1 and the native states showed increases in their integrals and linewidths in the presence of equimolar TF, while no peaks corresponding to I2 or the unfolded states could be quantified (Fig. 5b,c). These measurements indicate selective binding of TF to I1 and the native state, increasing their populations, particularly at the shorter FLN5+34 length. This modulation by TF is most consistent with FLN5 folding through parallel (I1 or I2) rather than sequential (I1 to I2 to N) pathways (Fig. 5d), although this cannot be shown unambiguously with these experiments (Supplementary Note 6). Our data add to emerging evidence of the dependence on nascent chain length and selectivity of TF for partially folded, rather than unfolded, nascent proteins20,54. Effect of TF on coTF of FLN5. a, 19F NMR spectra of isolated FLN5∆6, FLN5+34 and FLN5+47 in the absence and presence of equimolar concentration of TF (all species were at 20 μM). Observed and total fitted spectra are shown in gray and black, respectively. The assignment of resonances is colored accordingly. All spectra were recorded at 298 K and 500 MHz. b, Changes in free energy differences between the intermediate and native states upon addition of TF, determined by analysis of spectra shown in a. The asterisk indicates the lower bound, assuming <5% population detection limit2. c, Linewidths of I1 and native states determined by lineshape analysis of spectra shown in b. d, Schematic depicting I1 and I2 along parallel pathways from unfolded to folded states and change in flux by TF. Unless stated otherwise, all values in the figure represent the mean ± s.e.m. propagated from NMR lineshape fits. Source data Conservation of intermediate structures across filamin domains Lastly, we sought to examine whether the intermediates of FLN5 are conserved across the coTF pathways of domains with a similar fold. Therefore, we studied the folding of three additional immunoglobulin-like domains (of a multidomain protein) on the ribosome (Fig. 6a): FLN4, the domain preceding FLN5 (47% identity relative to FLN5), FLNa21, the 21st domain of human filamin A (the human homolog of gelation factor; 31% identity), and I27, the 27th immunoglobulin-like domain of titin (24% identity), whose smaller size (89 versus 105 residues compared to FLN5) is accompanied by a subtly different topology. For each protein, we rationally designed and experimentally validated 19F-label pairs with ring currents in the natively folded state42, permitting us to assess interstrand contacts across each immunoglobulin-like fold. 19F NMR of RNCs of immunoglobulin domains shows that coTF intermediates of filamins are conserved. a, Structures of FLN4 (AlphaFold3), FLNa21 (PDB 2W0P) and titin I27 (PDB 1TIT) and alignment with FLN5 structure (PDB 1QFH). b–d, 19F NMR spectra of 555tfmF 575His-labeled FLN4+L, 2244tfmF 2262Trp-labeled FLNa21+L and 14tfmF 87His-labeled I27+L RNCs, with linkers of L amino acids deriving from the subsequent domain. Top, spectra showing isolated protein spectra, with random coil value (−61.8 ppm) indicated. Observed and total fitted spectra are shown in gray and black, respectively. The assignment of resonances is colored accordingly. The magnification of spectra by a scaling factor is indicated. e, Summary of 19F NMR data of other labeling sites, shown as a schematic of coTF intermediates and pathways of the filamins (FLN4, FLN5 and FLNa21) and I27 (all NMR data shown in Extended Data Fig. 10a–c). All spectra were recorded at 500 MHz and 298 K. To identify nascent chain conformations populated on the ribosome, we produced a library of RNCs varying in length L of the subsequent domain (domain + L RNC). FLN4 RNCs 19F-labeled at the A–B strands (555tfmF 575Phe) showed a sharp resonance with a random coil chemical shift at the shortest RNC length (FLN4 + 21), which reduced in population with increasing RNC length and was, thus, attributable to the unfolded state; concurrently, a sharp resonance with a native chemical shift progressively increased in signal intensity and was ascribed to the natively folded state (Fig. 6b and Extended Data Fig. 10a). Importantly, two broad resonances, maximally populated at intermediate RNC lengths (FLN4 + 34 and FLN4 + 47), were also found at each of the same chemical shifts (Fig. 6b and Extended Data Fig. 10a). Thus, identically to FLN5 (Fig. 1f), FLN4 populates two coTF intermediates on the ribosome, distinguished by the presence of A–B strand contacts. Indeed, analysis using other 19F-label pairs (B–E and F–G strands) indicated that the same interresidue contacts are found in FLN4 as in FLN5 coTF intermediates (Fig. 6e and Extended Data Fig. 10a). We also probed the A′–G strand using a 19F-label pair in FLN4 (565tfmF 646His; Extended Data Fig. 10a) and found the absence of contacts for either intermediate state; these data are consistent with the determined structures of FLN5 I1 and I2, having detached A′ and G strands respectively (Fig. 3a–d). Similar 19F NMR analyses of FLNa21 RNCs with analogous labeling pairs (Fig. 6c,e and Extended Data Fig. 10b) also concluded the population of two coTF intermediates with the same structural features as those of FLN5. Collectively, all the 19F chemical shifts measured for our filamin domain coTF intermediates show that these states contain substructures of the native topology. Given the identical contacts observed by 19F NMR analyses, the determined structures of FLN5 coTF intermediates appear to be remarkably conserved across the folding of filamin domains. Characterisation of coTF intermediates of other immunoglobulin-like domains. a, 19F NMR spectra of isolated FLN4 and FLN4 RNC variants. b, 19F NMR spectra of isolated FLNa21 and FLNa21 RNC variants. The Tyr2297Glu mutation, analogous to Tyr719Glu in FLN5, was used to destabilise the domain. c, 19F NMR spectra of isolated titin I27 and I27 RNC variants. The Trp34Glu mutation was used to destabilise the domain3. Observed spectra shown in grey were fitted and assigned to their conformational states (coloured). Total fitted spectra are shown in black, and residual spectrum after fitting are shown below. Spectra were recorded at 500 MHz and 298 K. d-f, Structure, dynamics and aromatic ring currents of the titin I27 folding intermediate (titin I27 ΔA-strand) observed in MD simulations. d. Representative structures of the native and intermediate state of titin I27 ΔA. Native was defined as having Cα-Cα distances of less than 10 Å for both Leu8-Glu22 and Leu8-Ala81, and intermediate was defined as having at least one of these distances above 10 Å. The sidechains of 14tfmF and 87His are shown in stick representation. e, From top to bottom, plot shows centre-of-mass distance between the 14tfmF CF3 and 87His aromatic ring coloured by the angle, θ, formed between the CF3-ring interaction vector and aromatic ring normal vector. Angles below and above 54.7° are considered perpendicular and in-plane CF3-ring interactions. Second plot shows Cα-RMSD with respect to the starting structure across all three simulations. Third plot shows distances between the N-terminal Cα (residue Leu8) and Cα atoms of residues Glu22 and Ala81 calculated in three simulations. Bottom left shows average (mean ± s.e.m.) Cα -RMSF of full-length and ΔA titin I27 14tfmF 87His. The full-length simulations were analysed from previous work42. The location of β-strands is annotated along the sequence. Bottom right shows geometric factor (1-3cos2θ)/r3, r is the distance from panel in top left) quantifying the ring current effect experienced by 14tfmF due to 87His for the native and intermediate state. The distribution and average (mean ± s.e.m. from block averaging) are shown. f, Identical analyses as in e for titin I27 ΔA-strand 59tfmF (with wild-type His20). We previously identified two coTF intermediates of titin I27 on the ribosome3 and show here that their populations alter with RNC length, as expected (Fig. 6a and Extended Data Fig. 10c). However, in contrast to the filamin domains, 19F NMR spectra of I27 RNCs labeled across the A′–G strands (14tfmF 87His) reveal an intermediate-state resonance (I1) at the random coil chemical shift but, in contrast to the filamin domains, also at a notably more shielded, nonnative, chemical shift compared to that of the native state (I2). A weakly populated intermediate having the same chemical shift was also found in a destabilized variant of I27 (Phe73Ala) in isolation3 (Extended Data Fig. 10c), suggesting that its conformation was similar to the ribosome-bound I2 intermediate, whose structural features have previously been examined55,56. Indeed, all-atom MD simulations of the isolated I27 intermediate reconcile the observed chemical shifts and contacts of the I2 intermediate (Extended Data Fig. 10d–f and Supplementary Note 7). Overall, our experimental data suggest that the I27 coTF intermediates I1 and I2 are defined by detached A–B and A strands, respectively, from a common native-like β-sandwich core comprising β-strands C–G (Fig. 6e and Extended Data Fig. 10c). Thus, while the filamin I1 state is apparently conserved, I2 is subtly distinct to those adopted by filamin domains. Discussion We developed a range of quantitative and direct 19F NMR analyses and used these together with MD simulations to determine atomistic structural ensembles of coTF intermediates on the ribosome. A particular strength of 19F NMR, crucial to obtaining direct, precise structural contacts, is its ability to resolve interconverting states2. We extracted short-range and long-range distances from 19F NMR chemical shifts (Fig. 1 and Extended Data Figs. 2 and 3) and PRE measurements (Fig. 4 and Extended Data Figs. 8 and 9), respectively, and obtained systematic structural analyses specific to each intermediate state using a psi-value-based analysis (Fig. 2 and Extended Data Fig. 4). Extensive MD simulations of the putative models of the partially folded intermediate-state conformations (Fig. 3) show striking and entire consistency with all experimental NMR, cryo-EM and biochemical datasets (Extended Data Fig. 9). Overall, our study advances the utility of 19F NMR in structural studies and coTF and establishes a generalizable approach to validate atomistic structures of large macromolecular complexes. Our structural ensembles of I1 and I2 show a native-like, reduced β-sandwich core with strands C–G and A–F natively folded, respectively. Thus, while I1 and I2 are distinct in lacking their N-terminal (A–B strands) and C-terminal (G strand) structure, they share a common core comprising C–F strands. This latter feature is shared with the other immunoglobulin-like domains studied here. Given that the filamin intermediates have distinct subsets of native contacts at their domain termini, we suggest that they are likely populated along two parallel, rather than sequential, folding pathways. Although this cannot be shown unambiguously here with our equilibrium experimental data, this interpretation is supported by several other lines of indirect evidence (Supplementary Note 6)57–63. Moreover, G strand folding is coupled to the isomerization state of Pro742 (Extended Data Fig. 5)11, trans in I2 but cis in I1 and the native states (Fig. 3). The selective destabilization of I1 by the Pro742Ala substitution2 is also in line with an interpretation of parallel folding pathways and identifies this cis-proline as a conformational switch11. Notably, both the natively cis-Pro742 (ref. 11) and the coTF intermediate structures appear to be conserved across filamin domains (Fig. 6). In contrast, the I2 intermediate of I27 is structurally distinct, suggestive of a sequential coTF pathway (Supplementary Note 8). The coTF intermediates appear to share structural similarities to those found transiently off the ribosome by FLN4 (I1-like, on-pathway and obligate for folding56,64), N-terminal fragments of FLN5 resembling I2 (also on-pathway2,11), C-terminal fragments of I27 (I2-like55,56) and those of refolding transition states of immunoglobulin-like domains65. On the basis of our experimental data, this observation suggests that isolated intermediate conformations are explored on the ribosome but additional conformational states not observed in isolation also become thermodynamically more accessible (Supplementary Note 9). This thermodynamic stabilization of intermediates on the ribosome persists for long durations of translation2, even after translation of the subsequent domain (Fig. 1d and Extended Data Fig. 10a). The prolonged lifetimes of intermediates, in contrast to highly transient populations off the ribosome, may provide an opportunity for chaperones19–21 (Fig. 5), cofactors64, enzymes and quality control machinery66 to survey, engage and modify67 specific nascent chain conformations or for recruitment of other nascent proteins during cotranslational assembly31,32. Intriguingly, filamin function is highly regulated by serine phosphorylation at consensus sequences in humans68. Strand detachment and a trans-proline in the I2 intermediate exposes a natively buried phosphorylation site, potentially implicating the formation of stable coTF intermediates to cotranslational modifications69,70. Future detailed characterizations of nascent chain structure will permit mechanistic insights into folding pathways of other protein folds, misfolding29,30,71,72 and other cotranslational events. Methods Sample preparation DNA constructs of FLN5, FLN4, FLNa21 and titin I27 were previously described2,3,39,42. Mutations and amber stop codons were introduced using standard site-directed mutagenesis procedures and constructs were verified by DNA sequencing. For PRE NMR experiments, the additional substitution Cys747Val was introduced to yield a cysteine-free background for site-specific spin labeling, as previously described3. Isolated proteins were expressed as His-tagged proteins (FLNa21 as GST-fused protein42), isotopically labeled in Escherichia coli BL21 DE3-Gold cells (using amber suppression for incorporation of tfmF) and purified as previously described2,39,42. RNC constructs comprised a His-tag for purification and an arrest-enhanced variant of the SecM stalling sequence, FSTPVWIWWWPRIRGPP, as previously described2. RNCs were expressed, isotopically labeled and purified as previously described2. For intermolecular crosslinking experiments, RNCs were expressed in CRISPR-modified E. coli BL21 strains with cysteine substitutions in uL24, as previously described3. MTSL spin labels for PRE NMR were introduced into purified samples, as previously described3. Sample purity and integrity were assessed using SDS–PAGE and western blot analyses with an anti-His5 horseradish-peroxide-conjugated antibody (Invitrogen, 1:5,000 dilution), as previously described73. Unlabeled TF was purified as previously described74. NMR spectroscopy NMR samples were prepared in Tico buffer (10 mM HEPES, 30 mM NH4Cl, 12 mM MgCl2 and 1 mM EDTA), supplemented with 10% (v/v) D2O and 0.001% (w/v) DSS as a reference. NMR data were recorded using a 500-MHz (19F and 1H,15N NMR) or 800-MHz (1H,15N NMR) Bruker Avance III spectrometer, both equipped with a TCI cryoprobe and Topspin 3.5pl2, at 298 K, unless stated otherwise. Data were processed using NMRPipe (version 11.7)75, CCPN (version 2.4), MATLAB (R2017a, Mathworks) and Julia (version 1.5)76. To obtain quantitative insights into the structure (short-range contacts) and populations of intermediates on the ribosome, one-dimensional (1D) 19F pulse-acquire experiments were recorded using a 350-ms acquisition time and 1.5-s recycle delay. Multiple short experiments were recorded in succession to monitor changes in RNC sample integrity and data indicating nascent chain release or sample degradation through changes in linewidths, signal intensities or chemical shifts were not used to produce the final summed spectrum, as previously described2,3. Processed spectra were baseline-corrected and peaks were fitted to Lorentzian lineshapes, with errors of the linewidths and integrals (populations) estimated using bootstrapping (200 iterations, calculating the s.e.m.), as previously described2. The time-domain 19F NMR spectra were multiplied with an exponential window function with a line broadening factor of 5 (isolated protein) or 10 Hz (RNC), unless stated otherwise, before Fourier transformation. Integrals were used to calculate folding free energies using ∆G = −RTln(Keq). Two-dimensional (2D) 1H,15N band-selective optimized flip angle short transient definition (SOFAST) heteronuclear multiple-quantum correlation (HMQC) experiments77 were recorded using an acquisition time of 50 ms in the direct dimension and interscan delays of 100 ms. Cosine-squared window functions were used to process the spectra. Long-range distance measurements were obtained using 19F PRE experiments, where PRE rates were measured using 1D pulse-acquire experiments with samples in the paramagnetic and diamagnetic state, the latter by addition of 2.5 mM (RNC) or 100× molar excess (isolated protein) sodium ascorbate. For PRE measurements of RNCs, spin-labeled samples were prepared and split into two halves; one half was immediately recorded by NMR at 298 K to produce the paramagnetic spectrum, while the other half was incubated at 277 K. Once the paramagnetic spectrum was acquired (typically 16–24 h), the latter half sample was reduced in sodium ascorbate and the diamagnetic spectrum acquired at 298 K (typically also 16–24 h). To improve the signal-to-noise ratio, 5–6 separate RNC samples were prepared and the datasets summed together (using only data corresponding to intact samples) to produce the final spectra. The spectra were fitted to Lorentzian lineshapes, as previously described2, which provide accurate measurements of transverse relaxation rates (R2) and identical to those obtained from direct 19F R2 measurements3. Line broadening resulting from the PRE effect resulted in highly overlapped I2 and native-state resonances, which were deconvolved by restraining the lineshape fit of the native state using the expected PRE rate bounds (Supplementary Note 5). 19F CEST measurements were performed and analyzed as previously described2. These experiments aimed to quantify the exchange rates between different conformational states. Experiments were recorded with an acquisition time of 200 ms and a recycle delay of 30 ms, with a weak B1 field of 15 Hz applied for a saturation time of 800 ms at saturation frequencies of −41, −61.8, −62.2 and −63.0 ppm. AlphaFold structure predictions AlphaFold3 (https://alphafoldserver.com)46 was used to predict structures of FLN5 bihistidine mutants in complex with a metal ion, using Co(II) to replace Ni(II) that is otherwise unavailable. AlphaFold3 was also used to predict the structure of flexible regions of FLN4 that are absent in its crystal structure (PDB 1WLH). Default settings were applied. Crosslinking experiments To probe RNC interactions site-specifically, crosslinks were introduced into FLN5+47 RNC using a similar approach as described for PRE spin labeling3. Samples were reduced overnight at 277 K in Tico supplemented with 2 mM TCEP. TCEP was then removed by buffer exchange into labeling buffer (50 mM HEPES, 12 mM MgCl2, 20 mM NH4Cl and 1 mM EDTA, pH 8.0). The RNC sample was then incubated with variable concentrations of the crosslinker 1,8-bismaleimido-diethyleneglycol (BM(PEG)2) at 298 K and the reaction was quenched with DTT. Crosslinking was assessed by a mass increase of ~16 kDa (that is, the mass of uL24) by western blot. Simulating coTF during real-time synthesis To assess whether folding is expected to occur during real-time translation (compared to equilibrium measurements performed by 19F NMR), we considered a kinetic model. We simulated coTF during real-time synthesis for FLN5 using a system of ordinary differential equations (ODEs) representing a four-state model with unfolded, I1, I2 and native states. Folding was modeled as a function of nascent chain length during vectorial synthesis at fixed elongation rates of 5 and 20 aa per second (approximate eukaryotic and prokaryotic translation rates, respectively). Simulations were initiated at FLN5+31, corresponding to full emergence of the FLN5 domain from the ribosome39 and 100% unfolded state. Calculations were then performed until the maximum chain length the polypeptide achieves on the ribosome before release of full-length gelation factor (FLN5+110)39. We set a maximum folding rate, kf, of 0.8 s−1, corresponding to the apparent folding rate previously measured for isolated FLN5 experimentally11. As the ribosome has been shown to reduce folding rates at short NC linker lengths, we used a decrease in folding rate by two orders of magnitude (0.008 s−1) on the basis of experimental measurements9. We then linearly interpolated the folding rate between these two values as a function of chain length and synthesis time. Prior experimental studies have indicated that folding rates, in fact, accelerate more rapidly than expected by linear interpolation34; therefore, this represents a lower bound (that is, slowest folding rate feasible given experimental information). We assumed all forward rates to be equal to the length-dependent value of kf. This also represents a conservative choice of a lower bound for individual rates as apparent, measured folding rates are generally faster than conversion rates to and from intermediates. Experimentally determined equilibrium populations for unfolded, I1, I2 and native states at FLN5+31, FLN5+34, FLN5+47, FLN5+67 and FLN5+110 (refs. 2,3 and this study) were used to calculate length-dependent free energy differences (ΔG) between states and interpolated linearly between these discrete chain length measurements. Reverse (unfolding) rates (ku) were obtained from the NC length-specific values of kf and ΔG using ku = kf·exp(ΔG/RT). The ODE system was numerically integrated over time using solve_ivp (SciPy, version 1.3.1) with default settings and 1,000 evenly spaced time steps. Simulated folding populations were plotted as a function of chain length and compared to experimental equilibrium measurements. The following system of ODEs was used to model NC populations during real-time biosynthesis, where U and N denote the unfolded and native states: To model different scenarios of intermediates being populated along parallel pathways or sequentially along a common pathway, we set the rate constants of ‘forbidden’ transitions to 0 s−1. Proline free energy calculations To quantitatively assess the regulatory role of P742 isomerization on the stability and folding of FLN5, we used well-tempered metadynamics78 (WT-METAD) simulations and calculated the free energy landscapes capturing the isomerization state of FLN5 proline P742. All-atom simulations were run using the DES-Amber force field and TIP4P-D water model79,80. The calculations were prepared and performed with GROMACS (version 2020)81 and PLUMED (version 2.6)82,83. Three different models were used to investigate the effect of local structure on P742 isomerization energetics: an unfolded peptide with P742 (underlined) in the middle of the sequence corresponding to FLN5 residues N736-I748 (Ac-NPIKNMPIDVKCI-NMe), a C-terminally truncated FLN5 construct, FLN5 Δ6, populating a folded and disordered C terminus at equilibrium11, and full-length, folded FLN5 (PDB 1QFH, residues K646–G750)84. Simulations of the unfolded peptide were initiated from a linear chain built in PyMOL 2.3 (Schrödinger). All systems were initially centered in a dodecahedral simulation box at least 1.2 nm away from the box edge, solvated and neutralized with MgCl2. Energy minimization of the solvated systems was performed using the steepest descent algorithm using a tolerance of <1,000 kJ mol−1 nm−1. Equilibration was then performed for 1 ns in the NVT and 1 ns in the NPT ensemble, where the temperature (298 K) was maintained using velocity rescaling85 with a coupling constant of 0.1 ps and the pressure (1 bar) was regulated with the Berendsen barostat86 a compressibility of 4.5 × 10−5 bar−1 and coupling constant of 2 ps. During equilibration all protein heavy atoms were position-restrained using a force constant of 1,000 kJ mol−1 nm−2. Position restraints were then removed and pressure coupling switched to the Parrinello–Rahman method87 with the same compressibility and coupling constant. A time step of 2 fs was used together with the leapfrog integrator and the LINCS algorithm88 to constrain all bonds involving hydrogen atoms. Production simulations with WT-METAD were run in triplicates from different initial velocities for 1 μs each. Van der Waals interactions were computed up to a cutoff distance of 0.9 nm and short-range electrostatics were computed up to 1 nm. Long-range electrostatics were calculated using the particle mesh Ewald method89 with cubic interpolation and a grid spacing of 0.125 nm. As collective variables for WT-METAD, we used two dihedral angles, ζ (Cαi−1–Oi−1–CδI–Cαi) and ψ (Ni–Cαi–Ci–Ni+1), where i corresponds to P742. ζ describes the isomerization state (~0 for cis and ~±π for trans in radians) and ψ describes the the C-terminal amide orientation. Previous work has shown this combination to be efficient for simulating proline isomerization with WT-METAD90. Gaussians were deposited every 1 ps with a width of 0.07 rad for each dihedral angle. The Gaussian height and bias factor were set to 0.8 kJ mol−1 and 10, respectively. We used the method described by Tiwary et al. to reweight all snapshots to correct for the WT-METAD bias potential (saved every 5 ps)91, discarding the first 200 ns (20%) of each trajectory. All ensemble properties including cis/trans populations were then calculated using these weights, with the mean and s.e.m. calculated from the three independent WT-METAD simulations initiated with different starting velocities. For full-length FLN5 and FLN5 Δ6, we used a collective variable to quantify the fraction of native contacts (Q) formed by the two downstream residues I743 and D744 to correlate C-terminal structure with P742 isomerization. SMOG2 (ref. 92) was used to obtain a list of native heavy atom contacts and Q was calculated using the following switching function93: The distance between two heavy atoms is defined as ri,j, whereas r0i,j corresponds to the reference distance. β and λ were set to 50 nm−1 and 1.8, respectively. WT-METAD simulations of full-length FLN5 were restrained to have a C-terminal Q > 0.8 using a lower wall restraint in PLUMED (force constant = 10,000 kJ mol−1) to maintain the native structure of FLN5 during the simulations. All-atom simulations of protein folding With the approach described here, we aimed to sample starting structures and models of coTF intermediates. We used rMD simulations48,51,94,95 to sample folding transition paths starting from unfolded conformations. For isolated FLN5, we used SMOG2 (ref. 92) to generate a simple structure-based model without any contacts (that is, no nonbonded attractive potentials) and unfold FLN5. A total of 50 random unfolded conformations were then picked for rMD simulations. Using the DES-Amber force field and TIP4P-D water model79,80,96, these structures were then energy-minimized and equilibrated using an equivalent protocol as detailed above at 350 K. For the FLN5+47 RNC, we performed thermal unfolding simulations starting from 50 different conformations of folded FLN5+47 on the ribosome obtained in previous work23. The RNC was simulated with using the DES-Amber force field and TIP4P-D water model79,80,96, as we showed that this force field results in good agreement of folded FLN5 simulations with NMR data23. For thermal unfolding and subsequent relaxation runs, we used hydrogen mass repartitioning97 and a 4-fs timestep with LINCS. GROMACS 2020 and PLUMED 2.6 were used for these simulations. Temperature and pressure coupling settings were identical to the production runs used for the proline free energy calculations described above. All ribosome heavy atoms (and C-terminal nascent chain residue) were position-restrained at all times using a force constant of 1,000 kJ mol−1 nm−1. Unfolding was performed at 400 K and 1 bar for 25 ns in the presence of lower wall restraints on the radius gyration calculated from residues 646–750 Cα atoms at 3 nm with a force constant of 10,000 kJ mol−1. The total fraction of native contacts, as defined in the previous section, was restrained to be below 0.02 using a force constant of 10,000 kJ mol−1. To ensure all starting structures had a trans peptide bond for P742 (P742 is in the cis conformation in the native structure), we additionally restrained the absolute value of the ζ dihedral angle to be above 2.62 rad (~150°). These unfolded structures were then relaxed at 350 K and 1 bar for 1 ns in the absence of wall restraints (but with ribosome position restraints). Enhanced sampling is achieved with the rMD method (adiabatic-bias MD in PLUMED). We used GROMACS 2020 and PLUMED 2.6 for rMD simulations. In this approach, simulations proceed in the absence of a biasing potential when spontaneous progression toward a predefined target state (that is, the folded state of FLN5) occurs. However, a soft biasing potential is applied when the system moves in the opposite direction48,94: The force constant, κ, controls the magnitude of the bias potential, CV(t) is the instantaneous value of a defined collective variable at time t and CVtarget is its value in the target (folded) state. Therefore, rMD effectively promotes transitions from an initial (unfolded) state to the final (native state) using a reaction coordinate. In our rMD simulations, we set κ and CVtarget to 100 kJ mol−1 and 0, respectively. As a collective variable, we used the difference contact map between a given conformation and the native state48,94, defined by heavy atom native contacts, j, within a 0.6-nm distance cutoff obtained the shadow algorithm implemented in SMOG2 (ref. 92): Here, rj is the native contact distance for atom pair j and r0 was set to 0.5 nm. For FLN5, this approach includes 1,009 heavy atom contacts calculated using the energy-minimized crystal structure of FLN5. For all 50 starting structures, we ran 36 independent rMD simulations starting from randomly assigned initial velocities for a time of 5 ns, totaling 1,800 trajectories and 9 µs of biased sampling. Simulations were run in the NPT ensemble at 350 K, using the same temperature and pressure coupling parameters as described above. The integration time step was 2 fs. For the RNC, position restraints on the heavy atoms of the ribosome and C-terminal NC residue at the peptidyl transferase center were kept at 1,000 kJ mol−1 nm−1. We then calculated the root-mean-square deviation (r.m.s.d.) over Cα atoms of residues 646–750 for all trajectories with respect to the energy-minimized crystal structure of FLN5. Trajectories reaching an r.m.s.d. of less than 0.2 nm were defined as successful folding transition paths, resulting in 321 and 226 successful folding transitions for isolated FLN5 and FLN5+47 RNC, respectively. These trajectories were analyzed further and we calculated the total fraction of native contacts and native contacts between specific structural elements of the protein to project the transition path landscape onto these variables to investigate the folding pathway and intermediate states. Unbiased all-atom simulations of coTF intermediates Models of coTF intermediates were subjected to unbiased MD simulations to assess the stability of their structures and their dynamic ensembles. To select starting structures of the I1 coTF intermediate, we considered all potential structures belonging to the local free energy minimum observed in rMD simulations between a fraction of a global Q (0.55 < Q < 0.68; Extended Data Fig. 6b–f). As this intermediate free energy minimum corresponds to structures with a folded G strand (Extended Data Fig. 6f) and our proline free energy calculations show that P742 is in the cis configuration in this case (Extended Data Fig. 5), we only selected structures with P742 in cis. To ensure the selection of physically reasonable structures, we sorted the structures by ascending distance r.m.s.d. (relative to the energy-minimized native-state crystal structure) of all heavy atoms belonging to residues F691–G750. These residues correspond to the natively folded region of this intermediate state predicted by rMD (Extended Data Fig. 6c). The top 100 structures were then clustered using the GROMOS algorithm98 implemented in gmx cluster, using all-atom r.m.s.d. of the entire nascent chain without alignment and a cutoff of 0.9 nm to select structures with different orientations on the ribosome. The top five cluster centroids (Extended Data Fig. 6g) were then used for unbiased MD simulations. As rMD simulations only sampled I1 (Extended Data Fig. 6) and because of the structural similarity between the I2 coTF intermediate and an intermediate previously characterized for isolated FLN5 Δ6 (refs. 3,11,41), we randomly selected five structures from the lowest free energy cluster of the NMR-derived structural ensemble of isolated FLN5 Δ6 (ref. 11). We then built the C-terminal residues of FLN5 in an extended conformation and joined these models with the NC linker for FLN5+47 used in previous work23. Unbiased MD simulations were then initiated from these structural models. For all input structures of I1 and I2, we placed these complexes in the center of a dodecahedron box and added solvent atoms and MgCl2 to neutralize the charge of the system. The box size was determined by gmx editconf using minimum distances of 3 and 2 nm from the box edge for I1 and I2, respectively. This resulted in system sizes of 1.2–1.7 million atoms. After energy minimization, we first gently heated the simulation boxes from 0 to 298 K over a 500 ps using a time step of 0.5 fs and LINCS constraints for all bonds involving hydrogen atoms. The systems were then equilibrated at 298 K in the NVT ensemble (velocity rescaling thermostat with a coupling constant of 0.1 ps) for 500 ps using a 2-fs time step, followed by a 1-ns equilibration step in the NPT ensemble at 1 bar (Berendsen barostat with a time constant of 2 ps). During the equilibration protocol, all RNA and protein heavy atoms were position-restrained (1,000 kJ mol−1 nm−2). These restraints were then removed for all NC atoms (except the C-terminal residue at the peptidyl transferase center) for a further 1-ns equilibration simulation (NPT, Parrinello–Rahman barostat, 2-ps coupling constant) and production simulations. We performed three unbiased simulations starting from different initial velocities for each starting structure lasting 1 μs, resulting in a total of 15 μs for each coTF intermediate state. Root-mean-square fluctuation (r.m.s.f.) and SASA analyses were performed with gmx rmsf and sasa81, respectively. Alignment before r.m.s.f. calculation was performed using Cα atoms of residues in the folded region of the intermediate state, F691–G750 (C strand to C terminus) and K646–L733 (N terminus to F strand) for I1 and I2, respectively. The fraction of native contacts (Q) was calculated using PLUMED as described above. The radius of gyration was computed using all Cα atoms of the FLN5 domain (residues K646–G750) with PLUMED to compare the compactness of the coTF intermediates to the native state. Contact maps, ribosome interactions, distances and secondary-structure content were calculated using the MDTraj99 and MDAnalysis100 Python libraries. The effective rotational correlation times of the coTF intermediates were calculated as previously described23. The rotation of the domain was defined using backbone heavy atoms belonging to the same regions as defined for the r.m.s.f. analysis of I1 and I2. Structural analyses were compared to the native-state ensemble of FLN5 previously obtained23 with the DES-Amber force field. All ensemble properties were averaged over the entire 15-μs ensemble and errors are reported as the s.e.m. obtained from the 15 independent 1-μs trajectories. The structural ensembles were clustered using for the purpose of structural visualization using gmx cluster81 and the GROMOS algorithm98. Alignment and r.m.s.d. calculations for clustering were performed using the same residue ranges as for the r.m.s.f. analysis and all backbone heavy atoms. Similarity cutoffs of 0.2 and 0.1 nm were chosen for I1 and I2, respectively, because of these values resulting in similar cluster populations (Supplementary Table 4). Structures from the most populated cluster were then used for visualization in the main text figure (Fig. 3a–d). Simulations of fluorinated protein variants in isolation Additional MD simulations of fluorinated protein variants were used to help reconcile experimental observations of 19F NMR chemical shifts and the origins of (de)shielding effects in different conformational states. We used the CHARMM36m (ref. 101) and ff15ipq (refs. 102,103) force fields to simulate the following isolated FLN5 variants labeled with tfmF: 655tfmF, 691tfmF, 705tfmF, 715tfmF, 719tfmF, 727tfmF, 732tfmF and 740tfmF. We used an identical simulation protocol as described previously42. Production simulations were run for 1 μs, saving coordinates for analysis every 0.1 ns. Titin I27 variants lacking the A strand (residues 1–7 deleted) were simulated with the ff15ipq (ref. 102) force field for 10 μs in triplicate to sufficiently sample the folding intermediate. GROMACS tools gmx sasa, rmsd and rmsf were used for general analyses of the CF3 group (SASA) and protein structure and dynamics (rmsd and rmsf). Hydrogen bonds between fluorine (acceptor) and donor atoms were analyzed with gmx hbond. To estimate hydrogen-bond lifetimes, we calculated the area under the averaged autocorrelation function from all hydrogen-bond existence functions104,105. Distances were calculated with in-house Python scripts using MDAnalysis100 and CF3–aromatic interactions were analyzed with an open-source Python script (https://github.com/julian-streit/RingCurrents19F). Simulations isolated FLN5 fragments mimicking coTF intermediates I1 and I2 were performed using the ff15ipq force field102 and minimal constructs comprising the folded core of each intermediate. We used residues Ac-A683–G750-NMe and S645–D744 for I1 and I2, respectively. Three different starting structures were taken from the same rMD-generated models as used for the unbiased RNC simulations (I1) and FLN5 Δ6 structural ensemble11 (I2). Independent simulations were performed using the same protocol as for full-length fluorinated proteins from these starting structures with random initial velocities to obtain three replicates where the I1/I2 topology was maintained during the 1-μs simulations. This was assessed by having at least one 19F NMR residue pair with a native-like distance per β-strand pair (within 0.1 nm of the average distance calculated for the FLN5+47 native state23. The native β-strand pairs measured by 19F NMR for I1 were CF/FG and AB/BE/DE/CF for I2. 19F PRE distance calculations 19F NMR PRE rates Γ2 were converted to distances, r, using the Solomon–Bloembergen equation106, where ωF and τc are the Larmor frequency of fluorine-19 and rotational correlation time, respectively: K (1.092 × 10−44 m6 s−2) is defined as a product of physical constants including the vacuum permeability (μ0), gyromagnetic ratio of fluorine-19 (γF) electron g factor, Bohr magneton (μB) and the spin number of fluorine-19 (S). Values of τc were determined in previous studies40 and those for RNCs were determined from the linewidths using an empirical relationship recently determined3. Therefore, to extrapolate the expected PRE of native FLN5 on the ribosome, we used the measured PRE distance in isolation and RNC-specific τc: Cryo-EM fitting To orthogonally assess whether our structural models are compatible with existing cryo-EM data, we fitted the ribosome model used in the MD simulations to each cryo-EM map representing the four FLN5+47 dataset classes (AP1, AP2, AP5 and P5)53. The obtained rotation–translation matrix was applied in Python using the MDAnalysis library to fit the NC structure from each frame of the MD trajectory (native, I2 and I1 states) into the NC density. Next, we used scripts from cryoENsemble52 to generate synthetic cryo-EM maps for each MD frame at 8-Å resolution. The correlation coefficient between each synthetic map and the corresponding experimental cryo-EM map was then calculated following the definition of map correlation by ChimeraX107. As an alternative approach, we also used the cryo-EM maps to reweight the combined trajectories corresponding to native, I2 and I1 states using the cryoENsemble iterative strategy52. Before reweighting, we clustered each trajectory using the GROMOS algorithm in gmx cluster with a 1-nm cutoff. The obtained clustered trajectory, comprising 2,294 conformations (1,091 from I1, 706 from I2 and 497 from native state), was used in four cryoENsemble reweighting runs in each of the cryo-EM maps of FLN5+47, following the previously described protocol52. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Online content Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41594-026-01814-7. Supplementary information Supplementary Notes 1–9 and Tables 1–5. Reporting Summary Peer Review file Exemplar rMD trajectory of FLN5+47 RNC showing folding sequence through I1 conformation with disordered A–B strands. Exemplar rMD trajectory of FLN5+47 RNC showing folding sequence through I1 conformation with a detached A–B β-hairpin. Exemplar unbiased MD trajectory (1 µs) of the FLN5+47 RNC I1 state with a detached A–B β-hairpin. Exemplar unbiased MD trajectory (1 µs) of the FLN5+47 RNC I2 state with a detached G strand. Source data Statistical source data. Acknowledgements This work was supported by a Wellcome Trust Investigator Award (206409/Z/17/Z). J.O.S. acknowledges support from a Biotechnology and Biological Sciences Research Council London Interdisciplinary Biosciences Doctoral Program studentship. We acknowledge use of the University College London (UCL) Biomolecular NMR Center and thank A. M. Figueiredo for technical assistance. This project made use of time on high-performance computing (HPC) resources on Archer2 (ARCHER2 UK National Supercomputing service, https://www.archer2.ac.uk) granted through the UK High-End Computing Consortium for Biomolecular Simulation, HECBioSim (https://www.hecbiosim.ac.uk), supported by Engineering and Physical Sciences Research Council (EPSRC; grant no. EP/R029407/1 and EP/X035603/1). We also acknowledge the EuroHPC Joint Undertaking for awarding this project access to the EuroHPC supercomputer LUMI, hosted by CSC (Finland) and the LUMI consortium through a EuroHPC Regular Access call and the Baskerville Tier 2 HPC service (https://www.baskerville.ac.uk/). Baskerville was funded by the EPSRC and UK Research Institute through the World Class Labs scheme (EP/T022221/1) and the Digital Research Infrastructure program (EP/W032244/1) and is operated by Advanced Research Computing at the University of Birmingham. We additionally acknowledge the use of the UCL Myriad, Kathleen and Young High-Performance Computing Facility and associated support services in the completion of this work. Extended data Author contributions S.H.S.C., J.O.S. and J.C. designed and conceptualized the project. S.H.S.C. and J.O.S. produced the samples and performed the NMR experiments. S.H.S.C. analyzed the NMR experiments. J.O.S. performed and analyzed the MD simulations. T.W. and A.M. performed the cryo-EM analyses. A.M.E.C. and I.V.B. contributed to the initial 740Cys PRE experiments with S.H.S.C. and J.O.S. S.H.S.C. and C.A.W. analyzed the CEST NMR data. S.H.S.C., J.O.S., L.D.C. and J.C. sourced the computational resources and acquired the funding. S.H.S.C. and J.C. supervised the project. S.H.S.C., J.O.S. and J.C. administered the project. S.H.S.C. wrote the original draft, S.H.S.C., J.O.S. and J.C. edited the manuscript and all authors reviewed the manuscript. Peer review Peer review information Nature Structural & Molecular Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available. Primary Handling Editor: Sara Osman, in collaboration with the Nature Structural & Molecular Biology team. Data availability The structural ensembles of FLN5+47 I1 and I2 are available from Zenodo (https://doi.org/10.5281/zenodo.16601045)108. The NMR data are available from Zenodo (https://doi.org/10.5281/zenodo.19210765)109. This study made use of public datasets deposited to the Protein Data Bank under accession codes 1QFH (FLN5), 1TIT (titin I27) and 2W0P (FLNa21). Source data are provided with this paper. Code availability Pulse sequences used for the acquisition of NMR data and Matlab and Julia codes used to fit the 19F NMR data are available from Zenodo (https://doi.org/10.5281/zenodo.15169089)110. Python scripts used to predict aromatic ring interactions for fluorinated protein variants are available from GitHub (https://github.com/julian-streit/RingCurrents19F) and Zenodo (https://doi.org/10.5281/zenodo.15173797)111. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Sammy H. S. Chan, Julian O. Streit. Change history 10/12/2026 In the version of the article initially published, the second affiliation of Tomasz Włodarski (Institute of Biochemistry and Biophysics, Polish Academy of Sciences, Warsaw, Poland) was missing and has now been added in the HTML and PDF versions of the article. Extended data is available for this paper at https://doi.org/10.1038/s41594-026-01814-7. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41594-026-01814-7. References Associated Data Supplementary Materials Supplementary Notes 1–9 and Tables 1–5. Reporting Summary Peer Review file Exemplar rMD trajectory of FLN5+47 RNC showing folding sequence through I1 conformation with disordered A–B strands. Exemplar rMD trajectory of FLN5+47 RNC showing folding sequence through I1 conformation with a detached A–B β-hairpin. Exemplar unbiased MD trajectory (1 µs) of the FLN5+47 RNC I1 state with a detached A–B β-hairpin. Exemplar unbiased MD trajectory (1 µs) of the FLN5+47 RNC I2 state with a detached G strand. Statistical source data. Data Availability Statement The structural ensembles of FLN5+47 I1 and I2 are available from Zenodo (https://doi.org/10.5281/zenodo.16601045)108. The NMR data are available from Zenodo (https://doi.org/10.5281/zenodo.19210765)109. This study made use of public datasets deposited to the Protein Data Bank under accession codes 1QFH (FLN5), 1TIT (titin I27) and 2W0P (FLNa21). Source data are provided with this paper. Pulse sequences used for the acquisition of NMR data and Matlab and Julia codes used to fit the 19F NMR data are available from Zenodo (https://doi.org/10.5281/zenodo.15169089)110. Python scripts used to predict aromatic ring interactions for fluorinated protein variants are available from GitHub (https://github.com/julian-streit/RingCurrents19F) and Zenodo (https://doi.org/10.5281/zenodo.15173797)111.