Protein structure predictors such as Boltz-2 and AlphaFold3 usually return one conformation, even when the protein actually switches between several shapes. A recent trick turns these predictors into samplers. You blur a folded structure with noise, take a random walk guided by the model's learned "score," which points toward more probable structures, and then jump back to a clean structure. This recovers some motion. The walk, though, is undirected: it kicks every atomic coordinate with equal random noise, so it rarely crosses the large barriers that separate distinct functional states. Methods that do cross such barriers need a human to name a reaction coordinate in advance, meaning the one or two directions along which the change happens. We propose to let the model's own score field supply that direction. Using only extra score evaluations, we would find the direction in which the blurred probability landscape is softest (near a basin) or curves downward (near a transition state). We then push the sampler along that direction with a metadynamics-style bias, which fills in visited regions so the walk is driven elsewhere.
This technique comes from the theory of escaping saddle points in non-convex optimization, where a short run of gradient steps from a random perturbation acts as a power method that isolates negative-curvature directions without computing a Hessian. We move it into the application setting where getting stuck in local optima is the main failure. The payoff, if the idea works, is target-free and coordinate-free exploration of alternative protein states at the cost of a few hundred extra denoiser calls. It also yields a cheap, falsifiable diagnostic: does the softest direction of a predictor's blurred landscape point toward the known alternative state?
Core Claim
Our central hypothesis is that, at the blur levels already used for walk-jump sampling, the lowest-curvature eigenvectors of the model-implied energy $U_\sigma(y) = -\log p_\sigma(y)$ carry much of the displacement between a protein's known conformational states. Here $y$ is the noisy all-atom coordinate vector and $\sigma$ is the blur scale. If that holds, an automatically discovered one- or two-dimensional coordinate can replace the hand-chosen reaction coordinate that current enhanced samplers require. Biasing along it should then raise the rate at which multi-state proteins are fully sampled, beyond both isotropic walks and biasing along generic directions.
The curvature is reachable with first-order information only. Write $s_\sigma(y) = \nabla \log p_\sigma(y)$ for the score returned by the denoiser, so that $H(y) = -\nabla s_\sigma(y)$ is the Hessian of the energy. A Hessian-vector product then costs one extra score call:
$$H(y)\,v \approx -\frac{s_\sigma(y+\epsilon v) - s_\sigma(y)}{\epsilon},$$
where $v$ is a unit direction and $\epsilon$ a small step. Shifted power iteration, $v \leftarrow (cI - H)v$ followed by normalization with $c$ above the largest eigenvalue, converges to the lowest-curvature direction $v_{\min}$. This is the mechanism of NEON, carried out on a learned score rather than on a training loss.
What Prior Work Does and Leaves Open
The method comes from the saddle-escape literature. How to Escape Saddle Points Efficiently shows that occasional random perturbations let gradient descent leave strict saddles with only polylogarithmic dependence on dimension. Accelerated Gradient Descent Escapes Saddle Points Faster than Gradient Descent adds momentum and an explicit step that exploits negative curvature. First-order Stochastic Algorithms for Escaping From Saddle Points in Almost Linear Time shows that gradient steps started from a single perturbation amplify the negative-curvature component, so no Hessian is needed. We reuse this last idea directly.
Escaping Saddles with Stochastic Gradients plays a different role. It argues that intrinsic stochastic noise can already align with negative curvature. That raises a competing hypothesis we must test: a plain walk's own trajectory may already reveal the soft direction.
We adapt these tools in one important respect. They escape saddles, whereas a protein trapped in a basin sits at positive curvature. Near a basin we therefore use the softest mode rather than a negative one, and we use negative curvature only once the biased walk reaches a transition region.
The application regime is set by three papers. Boltz-Jump runs isotropic Langevin walks, which follow the score plus random noise, at a fixed blur of σ = 2 Å using Boltz-2's score, then jumps back in one denoising step. It reports faster sampling and better weak-contact recovery on ATLAS, but it has no directional mechanism and shows drift relative to molecular dynamics. ProMiSE motivates the work: it finds that current predictors sample all known states of intrinsically multi-state proteins only 8–24% of the time. It also traces the collapse to the structure module, even though the trunk's pairwise representations preserve multi-state information. COMPASS steers latents toward a known target and avoids the steric clashes caused by direct coordinate guidance, but it requires the destination structure. We require no target.
On the entrapment side, Enhanced Diffusion Sampling supplies the biasing machinery we reuse: UmbrellaDiff and MetaDiff, which add biased score guidance, followed by reweighting. Its stated weakness is dependence on a predefined reaction coordinate, and that is exactly the gap we target.
Two further papers constrain the design. A Contour Stochastic Gradient Langevin Dynamics Algorithm flattens landscapes by partitioning energy values. A score model returns gradients, not normalized energies, so that route is unavailable without extra estimation. A Theoretical Analysis of Optimization by Gaussian Continuation warns that paths of minima can bifurcate as blur changes. For that reason we work at a fixed σ rather than tracking basins across scales.
Research Program
The first stage is a diagnostic that needs no new sampler. For each protein in the ProMiSE intrinsic multi-state bin and the COMPASS 98-pair apo–holo set, we would start from one experimental state and blur it at σ = 2 Å. We would then compute the few lowest-curvature eigenvectors of $H$ by finite-difference power or Lanczos iteration on Boltz-2's denoiser, keeping the trunk representation fixed so that each product costs one diffusion-module call. Before this, we project out the six rigid-body modes and align the structures.
The quantity of interest is the overlap $|v_{\min}^\top \Delta| / \|\Delta\|$, where $\Delta$ is the aligned displacement to the other known state. A random direction gives an overlap near $1/\sqrt{d}$. The informative comparisons are against the top principal components of a short unbiased Boltz-Jump walk and against elastic-network normal modes computed from the starting structure. This stage alone tells us whether the predictor's blurred landscape "knows" the transition. It also tests, in a new setting, the ProMiSE claim that information lost by the structure module survives upstream.
The second stage builds the sampler. It alternates three steps. First, a curvature estimate refreshes the soft directions every $K$ walk steps, warm-started from the previous estimate. Second, a MetaDiff-style history-dependent Gaussian bias is deposited along the projection onto those directions. Third, when the estimated lowest eigenvalue turns negative, meaning the walk is near a transition state, we inject a perturbation along $v_{\min}$ in the accelerated style, which commits the walk to the far basin instead of recrossing. Finally, the one-step jump returns clean structures. Where relative populations matter, we reweight using the recorded bias.
The third stage is analysis. We would examine how overlap and sampling success depend on σ, which connects the empirically chosen blur to how much structure the smoothed landscape retains. We would also examine where the approach fails, for instance ligand-induced transitions, where the ligand may be absent from the apo input.
Baselines
The naive reference is standard Boltz-2 diffusion from pure noise, with as many independent samples as the compute budget allows. ProMiSE reports that current unsteered models reach 8–24% multi-state success on the intrinsic bin, and BioEmu reaches 29%. The 29% is the strongest reported number on that metric and the bar we must clear. COMPASS reports that unsteered generation hits 0% of transitions larger than 15 Å on its benchmark.
The strongest direct competitor is Boltz-Jump itself, an isotropic walk at σ = 2 Å, since our method is a modification of its walk. On ATLAS it reports a weak-contact Jaccard index of 0.531, against 0.217 for standard Boltz-2. Any gain in state coverage from our method must not lose that recovery. We would rerun it ourselves on ProMiSE and the COMPASS set with the same number of denoiser calls as our method, curvature products included, and give it an equal tuning budget over σ, step size, and walk temperature.
MetaDiff with a hand-chosen generic coordinate, RMSD from the starting structure, is the representative of the enhanced-sampling line. BioEmu is the strongest pretrained ensemble generator in the file and is run as released. For ensemble fidelity on ATLAS, AlphaFLOW's reported median RMSF Pearson correlation of 0.85 is the reference; RMSF measures how much each residue fluctuates.
COMPASS with the true target is not a fair competitor, because it sees the answer. We report its 55.6% hit rate on transitions larger than 15 Å as an oracle ceiling, together with its clash score of 0.192 against 0.191 unsteered.
The ablations isolate curvature. The first replaces $v_{\min}$ with a random direction of equal bias strength. The second replaces it with walk-trajectory principal components. The third replaces it with elastic-network modes. The fourth keeps the metadynamics bias but removes the negative-curvature kick. Beating the baselines means exceeding BioEmu's 29% on ProMiSE intrinsic success, and exceeding matched-compute Boltz-Jump in hit rate on transitions larger than 15 Å. It also means keeping the clash score within about 0.01 of the unsteered 0.191 and not falling below 0.531 on ATLAS weak-contact Jaccard.
Why not just…?
Why not just raise the blur or the walk temperature in Boltz-Jump? This is a real threat. More smoothing lowers barriers in every direction at once and needs no new machinery. Boltz-Jump already reports that its one-step jump loses fine-grained accuracy, measured as local distance difference test (lDDT), and that its samples drift beyond the molecular-dynamics ensemble, and larger σ should worsen both. We settle the question with a matched-compute sweep over σ and temperature. If a hotter isotropic walk reaches the same state coverage at equal clash rate and accuracy, the curvature component is unnecessary and we would say so.
Why not just use principal components of an unbiased walk as the reaction coordinate? This is the most serious alternative. If stochastic noise already aligns with soft or negative curvature, as the correlated-negative-curvature argument in Escaping Saddles with Stochastic Gradients suggests, a short walk's covariance would contain $v_{\min}$ for free. The stage-one overlap comparison tests this directly. If walk principal components match curvature eigenvectors in overlap and in downstream success, we would drop the power iteration and publish the simpler sampler.
Why not just use normal modes from an elastic network? Such modes are cheap and often capture domain hinges. They ignore everything the predictor learned, however, and cannot reflect the sequence-specific alternatives that ProMiSE finds preserved in trunk representations. Their overlap is measured side by side with ours. If they win, it would mean the learned score adds nothing beyond the structure's shape.
Why not just subsample MSAs or draw more diffusion samples? AlphaFold Meets Flow Matching for Generating Protein Ensembles reports that MSA subsampling is inconsistent across protein families and unavailable for single-sequence models. ProMiSE and COMPASS show that repeated unbiased sampling keeps landing in the dominant state. We still include both as matched-compute baselines rather than rely on those reports.
Decisive Experiments
Stage one is the gate. If the median overlap of $v_{\min}$ with the true inter-state displacement on the ProMiSE intrinsic bin is no better than walk principal components, the proposal reduces to PCA-based metadynamics and we stop there. If it is clearly better, stage two must show two things at matched denoiser-call budgets. First, curvature-directed biasing beats the random-direction and principal-component ablations in ProMiSE multi-state success and in the COMPASS hit rate for transitions larger than 15 Å. Second, it does so while keeping clash scores near the unsteered level and preserving ATLAS weak-contact recovery.
A split by transition class addresses the ProMiSE bottleneck claim. Success on intrinsic transitions combined with failure on ligand-induced ones would indicate that the missing information is chemical context, not landscape curvature. The final check is transfer: the same procedure run on a second score-based predictor, to confirm the effect is not specific to Boltz-2.
References
- How to Escape Saddle Points Efficiently
- Accelerated Gradient Descent Escapes Saddle Points Faster than Gradient Descent
- First-order Stochastic Algorithms for Escaping From Saddle Points in Almost Linear Time
- Escaping Saddles with Stochastic Gradients
- Boltz-Jump: Accelerated Sampling of the Conformational Landscape of Biomolecular Structure Prediction Models
- ProMiSE: Protein Multi-State Evaluation Benchmark in Biological Contexts
- COMPASS: Decoupled Latent Steering for Protein Conformational Transitions
- Enhanced Diffusion Sampling: Efficient Rare Event Sampling and Free Energy Calculation with Diffusion Models
- A Contour Stochastic Gradient Langevin Dynamics Algorithm for Simulations of Multi-modal Distributions
- A Theoretical Analysis of Optimization by Gaussian Continuation
- AlphaFold Meets Flow Matching for Generating Protein Ensembles