FRETraj: Structure-Guided FRET Modeling
- FRETraj is a structure- or trajectory-based FRET framework that computes dye-accessible volumes to bridge RNA 3D models with experimental FRET observations.
- It employs in silico labeling and ensemble averaging to derive mean transfer efficiencies and simulate photon-sampled FRET distributions for structure validation.
- By comparing predicted FRET observables with experiments, FRETraj enables reweighting of structural ensembles to enhance the accuracy of RNA conformational models.
FRETraj is a structure- or trajectory-based FRET framework used for in silico labeling of biomolecular conformations and for predicting FRET observables from dye-accessible spatial distributions rather than from a single fixed donor–acceptor distance. In the arXiv literature considered here, it appears most clearly as the central bridge between RNA 3D structure prediction and experimental single-molecule FRET, where it is used to compute accessible contact volumes (ACVs/mACVs) for the sCy3/sCy5 dye pair, derive mean transfer efficiencies, and simulate photon-based FRET distributions for direct comparison with experiment (Weber et al., 22 Sep 2025).
1. Definition, scope, and documentary profile
FRETraj is documented here primarily through application rather than through a standalone software paper. The most explicit description comes from its use in a workflow for FRET-guided selection of RNA 3D structures, where FRETraj performs in silico labeling and converts retained structural models into predicted FRET observables. In that workflow, FRETraj is not a generic sequence-to-structure predictor, nor is it a temporal decoder of smFRET traces. Its operational role is narrower and more specific: it attaches dyes in silico, computes ACVs or mACVs for those dyes, derives mean donor–acceptor distances and mean transfer efficiencies, and produces photon-sampled FRET distributions under experimentally motivated corrections (Weber et al., 22 Sep 2025).
This role places FRETraj in the class of forward models from structure to observable. The framework is therefore most naturally understood as a bridge between candidate conformational ensembles and experimentally measured FRET distributions. A plausible implication is that its principal epistemic function is not to generate structures de novo, but to test whether a given structural pool can support the observed FRET-state occupancies.
The surrounding literature clarifies the conceptual setting in which FRETraj operates. Nelson’s treatment of quantum decoherence in FRET provides a physical rationale for why many standard FRET observables can be interpreted with Förster-type kinetics in ordinary solution conditions, which in turn supports the common use of classical-looking rate laws and distance relations in trajectory- or structure-based workflows such as FRETraj (Nelson, 2018). By contrast, work on epsilon-near-zero metamaterials shows that geometry-only FRET models are insufficient when the photonic environment itself changes donor–acceptor coupling, indicating a clear boundary for conventional structural FRET tools (Aththanayeke et al., 24 Feb 2026).
2. Position in RNA ensemble-selection workflows
A detailed application of FRETraj is given for a ribosomal RNA tertiary-contact model construct containing a kissing loop, a highly flexible GAAA tetraloop, and a connecting poly(A)-linker. The biological target is the unbound state, in which the kissing loop remains formed while the tetraloop is not bound into its tertiary contact. Because the linker and tetraloop are conformationally heterogeneous, the study does not seek a single structure. Instead, it builds large candidate structure collections, filters them for structural plausibility, labels them in silico with FRETraj, predicts FRET observables, and then compares those predictions with experimental smFRET distributions (Weber et al., 22 Sep 2025).
The workflow has six main stages: generation of candidate RNA 3D structures from sequence and secondary structure; filtering for structurally plausible kissing-loop-containing conformers; in silico labeling with sCy3/sCy5 using FRETraj; computation of ACVs or mACVs and derivation of mean FRET efficiencies or full in silico FRET distributions; comparison of predicted distributions with experiment; and reweighting or resampling of structures according to the experimental smFRET bin probabilities. In this application, FRETraj is therefore the component that translates a structural ensemble into an experimentally comparable FRET ensemble.
The structure-generation stage used three RNA 3D modeling tools—RNAComposer, FARFAR2, and AlphaFold3—each producing 10,000 structures, together with a comparison ensemble from six independent MD simulations sampled every 100 ps. Before FRETraj-based filtering, the study asked how many structures were needed to preserve the distribution of predicted FRET observables. Using the distribution of mean transfer efficiencies derived from ACVs and a Kullback–Leibler divergence analysis with FRET bin size $0.05$, the authors concluded that 1,000 structures per tool were sufficient for later analysis.
Structural validation preceded FRET comparison. The kissing loop was checked against the cryo-EM reference structure PDB ID 3JCT with Barnaba-based Watson–Crick base-pair analysis, and candidate structures were retained only if they preserved the relevant kissing-loop base pairs and satisfied . Each filtered model was then subjected to a short 1 ns MD simulation and re-annotated for base pairing and eRMSD. This ordering is important: FRET consistency was explicitly not treated as sufficient by itself.
3. ACVs, mACVs, and the structure-to-FRET forward model
Within this workflow, FRETraj computes the accessible contact volume of the sCy3/sCy5 dye pair for each retained structure. The study states that ACVs were calculated for every structure from all three prediction tools and for every 100 ps along the MD trajectories, and that associated ACV files were saved as .pkl files. Figure-level descriptions further refer to prediction of the mACVs of sCy3 and sCy5. The immediate consequence is that FRET prediction is not based on a single fixed donor–acceptor atom distance from the RNA backbone, but on dye-accessible spatial distributions attached to the structure through modeled linkers (Weber et al., 22 Sep 2025).
From these ACVs or mACVs, the workflow derives structure-specific mean transfer efficiencies and corresponding mean donor–acceptor distances. Across an ensemble of structures, these become distributions of predicted FRET observables. FRETraj is then used again at the photon level: photon emission events are simulated with shot-noise broadening, direct excitation, gamma correction, and the experimental burst size distribution. This extends the forward model from idealized structure-based efficiencies to observables that are directly comparable to corrected smFRET histograms.
The same study makes an important dynamical assumption. Predicted structures, whether obtained from MD or from RNA 3D prediction tools, are treated as if they were rapidly sampled during the photon detection time. Bursts are therefore determined by averaging rather than by segmenting them into trajectory splits of individual MD runs or subsamples of predicted structures. The method thus corresponds to an ensemble-averaged fast-exchange regime. The authors later identify this assumption as one reason why simulated weighted distributions remained narrower than experiment: the real RNA may undergo slower microsecond-to-millisecond dynamics.
This forward-modeling logic distinguishes FRETraj from approaches that attach one deterministic distance to each structural state. It also distinguishes it from methods that infer kinetic pathways directly from traces. A plausible implication is that FRETraj is best viewed as an observable generator for candidate structural ensembles rather than as a state-space inference engine.
4. Experiment-guided weighting and quantitative performance
The RNA study used FRETraj outputs in two modes. In the unweighted mode, each filtered structure contributed equally, with bin population
In the weighted mode, structures were resampled so that the occupancies of ACV-derived bins matched the experimental smFRET probabilities as closely as possible: If a bin contained more structures than needed, a subset was sampled; if it contained too few, the available structures were selected repeatedly and randomly permuted until the required occupancy was achieved (Weber et al., 22 Sep 2025).
The main quantitative outcome was that unweighted structure collections did not reproduce the experimental smFRET distribution, whereas FRET-guided weighting moved all methods much closer to the experimental mean. The study reports the following mean FRET efficiencies:
| Ensemble | Without photon sampling | Photon sampling, unweighted | Photon sampling, weighted |
|---|---|---|---|
| RNAComposer | |||
| FARFAR2 | $0.05$0 | $0.05$1 | $0.05$2 |
| AlphaFold3 | $0.05$3 | $0.05$4 | $0.05$5 |
| MD simulation | $0.05$6 | $0.05$7 | $0.05$8 |
| smFRET experiment | — | — | $0.05$9 |
These numbers establish two distinct uses of FRETraj. First, it is a validation layer: it reveals when a structural generator does not naturally reproduce the observed FRET distribution. Second, it is a selection or reweighting layer: it can identify compatible subensembles even when the initial unweighted pool is inconsistent with experiment.
The same analysis also exposed differences among upstream structure generators. RNAComposer and AlphaFold3 mostly populated low-FRET conformations and, after weighting, depended heavily on repeated use of relatively few structures; some individual structures contributed more than 7% to the final weighted collection. FARFAR2 and the MD ensemble sampled broader FRET ranges and retained greater diversity after weighting. This is a practically important point: agreement with FRET can be achieved by repeatedly reusing a small number of conformers if the starting pool lacks diversity, so FRET agreement alone does not guarantee a well-sampled structural ensemble.
5. Physical basis for interpreting FRETraj observables
The physical legitimacy of FRETraj-style forward modeling depends on the ordinary Förster regime. Nelson’s treatment of FRET as an open quantum system argues that the donor–acceptor pair is embedded in a fluctuating environment that rapidly destroys coherence between donor-excited/acceptor-ground and donor-ground/acceptor-excited states. In that limit, coherent oscillations are suppressed and the donor behaves as though FRET were an additional irreversible decay channel with first-order kinetics (Nelson, 2018).
In this description, donor loss by channels other than FRET occurs with rate 0, while FRET adds an extra contribution 1, so that the effective donor decay rate becomes
2
Within the electric-dipole approximation, the donor–acceptor coupling 3 is proportional to 4, which yields the familiar 5 dependence of the transfer rate. Nelson’s discussion is therefore a physical justification for interpreting donor lifetime shortening, transfer efficiency, and distance-dependent transfer rates with classical-looking rate expressions under weak coupling, fast dephasing, dipolar interaction, and suitable spectral overlap.
This justification also defines the interpretive boundary of FRETraj outputs. The 6 law is not presented as universally fundamental; it arises from weak coupling, dipole–dipole interaction, and rapid decoherence. If those assumptions fail—at short separations where exchange or multipole effects matter, in strongly coupled excitonic assemblies, or in environments requiring more complete electrodynamic treatment—then straightforward distance interpretation of FRET efficiencies may become biased. In that sense, Nelson’s paper is foundational background for FRETraj users rather than a software protocol.
6. Relation to adjacent methods and known limits
Two nearby strands of work are especially useful for locating FRETraj within the broader FRET-method landscape. The first concerns structured photonic environments. In the DNA-controlled emitter study with an epsilon-near-zero multilayer, the geometry of the donor–acceptor pair is set by a molecular beacon, but the glass-to-ENZ change is attributed to a modified photonic environment, specifically redistribution of the local density of optical states and strengthened near-field coupling. That paper explicitly notes that standard geometry-only FRET tools can model the DNA-defined states, but not the full glass-to-ENZ enhancement without electromagnetic-environment corrections such as Green tensors and substrate-modified decay channels (Aththanayeke et al., 24 Feb 2026). This implies a clear limitation for FRETraj-like approaches: they are naturally suited to geometry, dye placement, and ensemble averaging, but not to photonic-environment engineering unless additional electrodynamic modeling is added.
The second comparison is methodological. FRETtranslator addresses RNA smFRET by building a hidden Markov model whose hidden states are RNA secondary-structure minima, whose emission probabilities are derived from coarse-grained 3D sampling with Ernwin, and whose paths are decoded by Viterbi. The paper explicitly contrasts such methods with FRETraj-like approaches by noting that FRETtranslator does not explicitly model dye linker geometry, accessible volumes, anisotropic dye orientations, or orientation factor distributions; instead it uses approximate coarse-grained distances and ad hoc noise perturbation (Hecker et al., 2016). The contrast is instructive: FRETraj is more detailed at the dye-geometry and structure-to-observable layer, whereas FRETtranslator is centered on temporal pathway inference.
Taken together, these comparisons suggest that FRETraj occupies a specific methodological niche. It is strongest when the problem is to map candidate structures or trajectories to experimentally comparable FRET observables through explicit in silico labeling and dye-accessible volumes. It is less complete when the central issue is either nontrivial photonic environment modification or hidden-state kinetic decoding. The available evidence therefore supports a view of FRETraj as a forward model for structurally grounded FRET prediction and ensemble selection, with clear utility in integrative biomolecular modeling and equally clear dependence on the validity of the ordinary Förster approximation.