Molecular simulations get stuck. A protein or peptide spends almost all of its simulated time vibrating inside one stable shape, and the transitions that matter, such as folding, switching between helix types, or opening a loop, happen too rarely to observe directly. The standard fix is to add an artificial energy "push" along a few hand-picked or learned summary coordinates, called collective variables (CVs), that are meant to track progress through the transition. Recent learned CVs are one-dimensional. They work well when there is a single route, but they cannot tell apart competing routes. Adding more Euclidean dimensions to separate the routes makes the push dilute, because the space that must be filled with bias grows as a power of the dimension.
We propose to embed molecular configurations in hyperbolic space, a curved geometry whose volume grows exponentially with radius and which can therefore hold trees with little distortion. The embedding is arranged so that the distance from the origin reads as "height in the energy landscape's tree of basins": stable states sit near the boundary, and the shared barrier regions sit near the center. The bias is then applied only along this single radial coordinate. That keeps the pushing as cheap as a 1D CV. The angular coordinates, which are never biased, record which channel the system escaped through. In two sentences: learn a hyperbolic map in which leaving any basin means moving toward the origin, and push only along that radius so that the physics chooses among channels. This gives a biasing coordinate that is one-dimensional and still able to distinguish channels, which neither 1D nor multi-dimensional Euclidean CVs provide.
Core Claim
The free-energy landscapes of peptides and atomic clusters are often organized as nested basins joined by saddles, which chemists draw as a tree. Poincaré Embeddings for Learning Hierarchical Representations and Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry showed that in such embeddings the norm encodes generality and the angle encodes branch identity. The Lorentz paper reports that a 10-dimensional Lorentz embedding of WordNet beats Euclidean models that needed up to 200 dimensions. We make a specific claim. If an encoder $f_\theta$ maps a configuration $x$ into the Lorentz model $\mathbb{H}^k$ so that the norm tracks basin depth, then the radial coordinate
$$r(x) = d_{\mathbb{H}}\big(o, f_\theta(x)\big) = \operatorname{arcosh}\big(-\langle o, f_\theta(x)\rangle\big),$$
where $o$ is the hyperboloid origin and $\langle\cdot,\cdot\rangle$ is the Minkowski inner product, is a channel-agnostic escape coordinate. Biasing $r$ alone should drive transitions out of every basin at the cost of a 1D bias. The unbiased angular part of $f_\theta(x)$ should cleanly separate reaction channels for analysis and reweighting.
The exponential volume of hyperbolic space is a hazard here, and it motivates the radial-only design. Depositing Gaussian bias in full hyperbolic coordinates would need exponentially many kernels near the boundary. We never do this; the geometry is used to organize the space, and the bias acts on one scalar.
Relation to Prior Work
The closest work is Learning Collective Variables from BioEmu with Time-Lagged Generation, BIOEMU-CV. It freezes a protein diffusion model and trains a small encoder whose 1D output must let the model generate the configuration a lag time later. Its authors name the 1D restriction as the reason it cannot handle branching mechanisms. We reuse its time-lagged training signal, change the output space to $\mathbb{H}^k$, and add a constraint on what the radius means. Two related lines of work supply alternatives. Path collective variables without paths builds CVs from fluctuations inside two chosen basins, and Enhanced Sampling of Protein Conformational Transitions via Dynamically Optimized Collective Variables adapts CVs on the fly along a minimum-energy route. Both assume the relevant pair of states, or the route, is known. Our radius requires no target basin.
The CV-free route-finders are the other point of reference. PIPS learns a Cartesian bias potential by stochastic optimal control toward a single fixed target structure. TPS-GFN learns a bias with a generative-flow-network objective to gain path diversity. PINN-MEP fits a continuous minimum-energy path and finds several routes on AIB9. All three need an endpoint. They produce paths, not a reusable coordinate for thermodynamic reweighting. Our method is closer in spirit to metadynamics, as described in Enhancing Important Fluctuations, and yields free energies directly.
The hyperbolic machinery comes from three sources. We take Lorentz-model optimization from the Lorentz paper above, tangent-space layers from Hyperbolic Neural Networks, and the efficient hyperboloid layers of Fast and Geometrically Grounded Lorentz Neural Networks. The existing hyperbolic molecular models, Hyperbolic Neural Networks for Molecular Generation and Hyperbolic Molecular Representation Learning for Drug Repositioning, embed static 2D bond graphs and have no connection to dynamics.
Two limitations stated in the file constrain the design. Hyperbolic space distorts non-tree structure; the Lorentz paper says so explicitly. It also assumes a single root, whereas real landscapes can have several competing roots. We therefore learn the curvature, which lets the model fall back toward Euclidean geometry, and we test on systems both with and without branching.
We deliberately do not pursue two other hyperbolic ideas. Moving FEAT's interpolants onto hyperbolic geodesics was considered and rejected. Atoms live in $\mathbb{R}^{3N}$, and a geodesic in a learned latent space has no guaranteed physical meaning for the escorted Jarzynski work in FEAT. We also do not parameterize the PIPS control policy with hyperbolic layers. That change only alters the function class of a bias network, and we see no geometric argument that it would help.
Research Program
The encoder is an invariant network on pairwise distances, as in PIPS. Its output passes through an exponential map onto $\mathbb{H}^k$ with learnable curvature, with $k$ between 2 and 4. Training combines two terms. The first is the BIOEMU-CV time-lagged generation loss: a frozen BioEmu decoder, conditioned on the tangent-space image of $f_\theta(x_t)$, must generate $x_{t+\tau}$. This term makes the embedding keep only slow motions. The second is a radial ordering loss. Using BioEmu samples together with short unbiased simulations started inside basins, we estimate a local log-density for each configuration and penalize any case where denser (deeper) configurations sit at smaller radius than sparser (higher) ones. This loss pins "boundary means basin, center means barrier" without labeling any channel. Angular separation is left to emerge from the time-lagged loss. Diagnosing whether it does emerge is part of the study.
For sampling, the bias is well-tempered metadynamics or the OPES method on $r(x)$ alone. The bias force follows from the chain rule through $f_\theta$. Free energies are reweighted onto the channel labels, which we obtain by clustering the angles of reactive segments. A second-stage variant retrains the encoder on biased data so the radius improves once new basins are found. Our analyses measure three things: how well the recovered radial ordering matches barrier heights, whether angular clusters match independently identified channels, and how performance changes as learned curvature is forced to zero.
Baselines
The do-nothing reference is unbiased molecular dynamics with the same number of force evaluations. For ALDP, PIPS reports a 0% target hit percentage (THP) for 500 fs unbiased MD at 300 K, the fraction of short runs that reach the target state.
The strongest existing method on our question is BIOEMU-CV with metadynamics on its 1D output, run exactly as its authors describe. It reports an absolute $\Delta F$ error of 0.52 kJ/mol on Chignolin over 1 µs, against 1.71 for DeepTICA and 2.53 for TAE. It reports 93.8% THP on BBA and 100% THP for steered Chignolin folding. We retrain it on the same BioEmu samples and seed simulations, give it the same simulation length and the same hyperparameter search budget, and use identical metadynamics settings. On Chignolin and BBA, where one route dominates, beating it means matching its $\Delta F$ error and THP within seed variance while also resolving any secondary channels. On AIB9 and on an LJ cluster, beating it means lower reweighted free-energy error per channel and more distinct channels crossed per microsecond. For AIB9 we found no reported result on Lacuna for channel counts or free-energy error, so reference values will come from our own long unbiased runs.
The key ablation is the matched Euclidean twin. It uses the same encoder, losses, $k$, and radial ordering loss, but in $\mathbb{R}^k$, with the bias on the Euclidean norm. This isolates curvature from the radial design. A second ablation biases all $k$ hyperbolic coordinates instead of the radius, which tests our claim that biasing the full space is harmful. A third ablation runs the full model with curvature frozen near zero.
PIPS and TPS-GFN are included as CV-free references on ALDP and polyproline. We report their published ALDP numbers where available, for example PIPS at 63.5% THP within 500 fs. We note that their task, reaching one specified target, is narrower than ours.
Why not just…?
Why not just give BIOEMU-CV a 2- or 3-dimensional Euclidean output? This is the real threat. On a landscape with three or four channels, a 2D Euclidean CV may separate them perfectly well, and the capacity advantage the Lorentz paper found on WordNet trees with thousands of leaves may not matter at this scale. The matched Euclidean twin tests exactly this. It uses the same design with a Euclidean radius, and alongside it we run full 2D/3D Euclidean metadynamics. If the twin matches us on AIB9 and the LJ cluster, the contribution reduces to "bias a learned depth coordinate", and we will report it that way. Curvature should matter most when branches are nested, meaning basins within basins. So the LJ cluster, whose tree has many competing minima, is where we expect the gap to show, and it is the decisive system.
Why not just learn one 1D CV per pair of basins, with HLDA or path CVs? Path collective variables without paths needs both endpoint basins in advance, and the number of pairwise CVs grows with the square of the number of basins. Our radius needs no targets and can find basins nobody listed. We run pairwise HLDA on AIB9, which has known states, to measure the cost of this generality where the pairwise approach is at its best.
Why not just sample the diffusion model more, or use a CV-free path sampler? Enhanced Diffusion Sampling shows that unsteered generation cannot reach states of negligible equilibrium probability without enormous numbers of samples. That is the rare-state problem. COMPASS reports 0% hit rate for unsteered models on displacements larger than 15 Å. The CV-free path samplers need a fixed target, and they do not produce free energies without extra work. They answer a different question, so we compare against them only on ALDP and polyproline.
Proposed Experiments
The first check is a sanity test on ALDP, using the FEAT repository's data and alignment tools (jiajunhe98/FEAT). The radius should place the two basins at large norm and the barrier region near the center. On this system, whose tree is shallow, learned curvature should shrink toward zero; if it stays strongly negative, that is a warning sign of forced tree structure. Chignolin and BBA from the DESRES trajectories test whether we lose anything relative to BIOEMU-CV on largely single-route folding, using $\Delta F$ error and THP. AIB9 and an LJ cluster are the multi-channel tests. On these we measure per-channel free energies against long reference runs, count distinct channels crossed per microsecond, and check agreement between angular clusters and channels identified independently by committor or path analysis.
The decisive result has two parts. First, the hyperbolic radius must beat its Euclidean twin on the LJ cluster and AIB9 at equal $k$ and equal compute. Second, biasing the radius alone must outperform biasing all coordinates. If both hold, curvature is doing real work in making a 1D bias able to distinguish channels. If the Euclidean twin ties, the lasting finding is the radial depth coordinate itself, and hyperbolic geometry is not needed for it.
References
- Poincaré Embeddings for Learning Hierarchical Representations
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- Learning Collective Variables from BioEmu with Time-Lagged Generation
- Collective Variables from Local Fluctuations
- Enhanced Sampling of Protein Conformational Transitions via Dynamically Optimized Collective Variables
- Stochastic Optimal Control for Collective Variable Free Sampling of Molecular Transition Paths
- Collective Variable Free Transition Path Sampling with Generative Flow Network
- PINN-MEP: Continuous Neural Representations for Minimum Energy Path Discovery in Molecular Systems
- Enhancing Important Fluctuations: Rare Events and Metadynamics from a Conceptual Viewpoint
- Hyperbolic Neural Networks
- Fast and Geometrically Grounded Lorentz Neural Networks
- Hyperbolic Neural Networks for Molecular Generation
- Hyperbolic Molecular Representation Learning for Drug Repositioning
- FEAT: Free energy Estimators with Adaptive Transport
- Enhanced Diffusion Sampling: Efficient Rare Event Sampling and Free Energy Calculation with Diffusion Models
- COMPASS: Decoupled Latent Steering for Protein Conformational Transitions