Molecular dynamics simulations follow how every atom of a protein or RNA moves over time. They routinely get the proportions of a molecule's shapes wrong because the physical models behind them are imperfect. The standard repair is maximum entropy reweighting. It keeps the simulated snapshots, assigns each one a new probability so that the averaged prediction matches a lab measurement such as an X-ray scattering curve or an NMR distance, and changes the original probabilities as little as possible. This works well when the simulation is roughly right. When the simulation is badly biased, the repair fails in a characteristic way: it puts nearly all the probability on a handful of snapshots that happen to fit. The result matches the fitted data and disagrees with every independent measurement.
This note proposes a change to how the reweighting is allowed to act. Errors in physical models tend to mis-weight whole families of related shapes, such as "all compact states" or "the minor loop conformation". The correction should therefore be forced to move probability between whole branches of the molecule's shape hierarchy, not between isolated snapshots. To make "branch" precise and differentiable, we place the snapshots on a low-dimensional hyperbolic map, a curved space whose volume grows exponentially with radius. In such a space trees embed with little distortion. We then restrict the log-change in weights to a small family of functions that tilt probability toward chosen directions on that map. The question is whether this structured correction fits the data while keeping many more snapshots in play, and whether it predicts held-out measurements better than frame-by-frame reweighting.
Core Claim
Bayesian/maximum entropy reweighting (BME) chooses weights $w$ over a library of $M$ simulated frames with prior weights $w^0$. It minimizes $\chi^2(w) + \theta D_{\mathrm{KL}}(w \parallel w^0)$, where $\chi^2$ measures mismatch between weighted averages of computed observables and experiment, and $\theta$ sets how far the weights may drift from the prior. Every published variant treats the weights as an unstructured point on the simplex. That includes the scattering work of Larsen et al., the iterative hydration-aware iBME, the NMR study of the UUCG tetraloop, and the cryo-EM method cryoENsemble.
We embed each frame $i$ at a point $z_i$ in the Poincaré ball, the unit ball with hyperbolic distance. We then restrict the weights to
$$ w_i(c,\xi) \;=\; \frac{w_i^0}{Z}\exp\Big(\sum_{j=1}^{K} c_j\, B_{\xi_j}(z_i)\Big), \qquad B_{\xi}(z) = \log\frac{\lVert \xi - z\rVert^2}{1-\lVert z\rVert^2}, $$
and minimize the same BME objective over the coefficients $c_j$ and the boundary directions $\xi_j$, with $K$ small (one to five). Each $B_{\xi}$ is a Busemann function, the hyperbolic analogue of "distance toward a point at infinity" in direction $\xi$ on the ball's boundary, and $Z$ normalizes the weights. In tree terms, one term shifts weight toward an entire branch at every depth at once.
Two properties make this the right family rather than an arbitrary smoother. First, under any isometry of hyperbolic space a Busemann function changes only its direction and an additive constant, and the normalization $Z$ absorbs that constant. The fit therefore does not depend on which state sits at the origin. This sidesteps the single-root tension raised by Nickel & Kiela's Lorentz-model work for multi-basin landscapes such as open and closed adenylate kinase. Second, the family stays tiny however many frames there are, so it cannot concentrate weight on a few frames unless those frames form a branch.
The claim is conditional and testable. When the frame-to-frame distances of a trajectory are close to tree-like, measured by Gromov δ-hyperbolicity (a standard score of how close a distance matrix is to a tree's), branch-level reweighting should do three things at equal fit to the restrained data. It should retain a larger effective fraction of frames ($\Phi_{\mathrm{eff}}$) than frame-level BME. It should recover known ground-truth weights more accurately. It should predict held-out observables better. When the distances are not tree-like, we expect no advantage over a Euclidean version, and we will report that.
Research Program
Building the map. For each trajectory we compute pairwise frame distances in two ways. The structural version uses RMSD after alignment, or distance-matrix RMSD for disordered proteins. The kinetic version uses commute distances from a time-lagged transition graph, which separates states by barriers rather than by geometry. We fit 2–10-dimensional embeddings in the Lorentz model of Nickel & Kiela, which avoids the Poincaré ball's numerical problems near the boundary, by optimizing a distance-preserving stress loss with Riemannian SGD and the burn-in schedule from Poincaré Embeddings. We then map the points into the ball to evaluate the Busemann features. The method comes from these two papers. Neither embeds continuous physical trajectories, and neither uses the embedding to constrain a probability correction.
Fitting. The objective is smooth in $c$ and in $\xi$, which lives on the sphere. We optimize with L-BFGS and use several restarts for the directions. $\theta$ is chosen exactly as in standard BME, by tracing $\chi^2$ against $\Phi_{\mathrm{eff}}$, and the same rule is applied to every method. For scattering data we keep iBME's default hydration parameters ($r_0 = 1.681$ Å, $\delta\rho = 3.34$ e/nm³) fixed across methods. This prevents differences in forward-model handling from contaminating the comparison. We will also report whether the structured family shrinks the degenerate valley in $(r_0, \delta\rho)$ that Pesce & Lindorff-Larsen had to search by grid.
Systems. We use four systems with published experiments and known failure modes.
α-synuclein. Ahmed et al. showed that an Amber TIP3P prior collapses under BME while a TIP4P-D prior refines cleanly against SAXS, with validation on hydrodynamic radius and paramagnetic relaxation enhancement (PRE). This motivates the proposal. It also constrains it: if the needed expanded states are missing from the poor prior, no reweighting can create them, so success here means degrading gracefully, not matching TIP4P-D.
TIA-1. We fit the three-domain protein to SAXS and hold out the contrast-variation SANS at 0%, 42% and 70% D₂O (Larsen et al.).
UUCG tetraloop. The force field puts State B near 40% and exact-NOE refinement brings it to about 10% (Bottaro et al.). We hold out J-couplings and residual dipolar couplings. The RDC residuals that flat refinement left unresolved are a natural test of whether branch structure finds sub-states.
Adenylate kinase. We use synthetic cryo-EM maps with a known open-to-closed ratio, following cryoENsemble.
Where the original trajectories are not released, we regenerate priors with the same force-field and water-model pairs.
Controlled bias. The cleanest test uses ground truth we construct ourselves. Starting from a well-behaved trajectory, such as TIP4P-D α-synuclein or adenylate kinase, we build a corrupted prior. We do this by multiplying the weights of selected regions by known factors, either one kinetic macrostate or a radius-of-gyration band that mimics compaction bias. We compute synthetic observables from the uncorrupted weights, add noise at the experimental error, and ask each method to recover the true weights. The error measure is KL divergence from the true weights, plus error on observables never used in fitting. Branch-shaped corruptions should favor our method. Corruptions that are deliberately not branch-shaped, such as reweighting by a random per-frame factor, are the fair negative control, since there the method should lose.
Baselines
The naive reference point is the unreweighted prior, with its $\chi^2$ on fitted and held-out data. It shows how much any reweighting buys.
The strongest existing method is frame-level BME with $\theta$ selected by the same $\chi^2$–$\Phi_{\mathrm{eff}}$ procedure, as used by Larsen et al. and Ahmed et al.. Its best reported operating point in the file is on TIA-1: reduced $\chi^2 = 1.0$ against SAXS with $\Phi_{\mathrm{eff}} = 83\%$, starting from $\chi^2 = 2.8$ before reweighting (Larsen et al.). Because that prior is already good, our method beats it there only if it matches $\chi^2 \approx 1.0$, keeps $\Phi_{\mathrm{eff}}$ at least as high, and improves SANS held-out $\chi^2$. On α-synuclein with TIP3P, the file reports a qualitative collapse but no $\Phi_{\mathrm{eff}}$ value, so we will measure it ourselves under identical forward models. Beating it means a substantially higher $\Phi_{\mathrm{eff}}$ at matched SAXS $\chi^2$, together with lower PRE and hydrodynamic-radius error.
For cryo-EM, the reference is cryoENsemble, which raised map–model correlation on the Trigger Factor complex from 0.68 to 0.73 (Włodarski et al.). On adenylate kinase it recovered the open/closed ratio with 10% noise at 10 Å; no numeric ratio error is reported in the file. We compare against it on adenylate kinase, where ground truth is known.
For the UUCG system, the target is the roughly 10% State B population from eNOE refinement in Bottaro et al.. Agreement with that population is the bar, and lower held-out RDC error is the win.
Two ablations isolate the key ingredient. The first is a Euclidean twin: the identical Busemann-free family, using linear tilts $c^\top z$ on a Euclidean embedding fit with the same stress loss, dimension and optimizer. It is also run at the larger dimension a Euclidean map needs to reach equal stress. The second is a shuffled-geometry control, which permutes $z_i$ across frames to test whether any low-parameter family would suffice.
All methods share the same frames, forward models, error estimates, $\theta$-selection rule and restart budget.
Why not just…?
Why not just tune θ in ordinary BME until diversity is acceptable? Raising $\theta$ preserves $\Phi_{\mathrm{eff}}$ only by giving up fit. In the TIP3P α-synuclein case, Ahmed et al. found that fitting SAXS required discarding most frames. We will plot both methods' full $\chi^2$–$\Phi_{\mathrm{eff}}$ curves. The claim holds only if our curve lies strictly above at matched fit. If it does not, the proposal fails.
Why not just cluster the trajectory into a dendrogram or Markov-state macrostates and reweight per node? This is the real threat, since it expresses the same idea of moving weight between branches without any curved geometry. We implement it directly, with one log-weight parameter per node at a chosen depth, regularized by the same KL term, and we sweep the cut level. Our expected advantage is that hard cuts force a choice of depth and misassign frames on boundaries between states, whereas Busemann tilts act at all depths smoothly. If tuned dendrogram reweighting matches us on the controlled-bias benchmark, we will report that the hierarchy matters and the hyperbolic geometry does not.
Why not just smooth weights over a Euclidean embedding (PCA or an autoencoder) like the representations used by Dynaformer? For tree-like data, Euclidean maps crowd distinct leaves together. On WordNet a 200-dimensional Euclidean embedding reached mean average precision 0.168 against 0.823 for a 5-dimensional Poincaré embedding (Nickel & Kiela). That evidence is linguistic, not molecular. The Euclidean twin ablation, stratified by each trajectory's measured δ-hyperbolicity, is what settles it here. Disordered proteins may not be tree-like, and on them we expect a tie.
Why not just fix the simulation, by retuning the force field or by restraining on the fly with metainference? These attack the root cause and remain the right answer when affordable. Larsen et al. themselves note that their 6% protein–water rescaling was specific to TIA-1 in MARTINI. Ahmed et al. compared post-hoc BME against metainference metadynamics, which requires many replicas and forward-model evaluations at every step. Post-hoc reweighting stays useful for large published trajectories and for expensive forward models such as cryo-EM maps. Our method is a better post-hoc tool, not a substitute for better physics.
Decisive Checks
The first check is the controlled-bias benchmark. Against the true weights, the hyperbolic family must beat frame-level BME, the Euclidean twin and the tuned dendrogram on branch-shaped corruptions, and it must lose on non-branch corruptions. A method that wins everywhere would suggest the gain comes from generic low capacity, not from the hierarchy.
The second is the δ-hyperbolicity gate. Across the four systems, any advantage over the Euclidean twin should grow as the trajectory's distances become more tree-like.
The third is held-out prediction on real data:
- SANS for TIA-1.
- PRE and hydrodynamic radius for TIP3P α-synuclein.
- J-couplings and RDCs for UUCG, with State B near 10%.
Each is compared at matched fitted $\chi^2$.
A negative result on the third check combined with a positive first check would still be informative. It would mean the real force-field errors in these systems are not branch-shaped, which is itself a finding about how simulation biases are organized.
References
- Combining molecular dynamics simulations with small-angle X-ray and neutron scattering data to study multi-domain proteins in solution
- Refining conformational ensembles of flexible proteins against small-angle x-ray scattering data
- Integrating NMR and simulations reveals motions in the UUCG tetraloop
- Bayesian reweighting of biomolecular structural ensembles using heterogeneous cryo-EM maps with the cryoENsemble method
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- Poincaré Embeddings for Learning Hierarchical Representations
- Refinement of α-Synuclein Ensembles Against SAXS Data: Comparison of Force Fields and Methods
- From Static to Dynamic Structures: Improving Binding Affinity Prediction with Graph‐Based Deep Learning