Searching Over Learned Molecular Dynamics Rollouts to Find Rare Transitions

Machine-learned molecular dynamics models can now jump forward nanoseconds at a time instead of integrating femtosecond steps. They still sample open-loop: you launch a rollout, wait, and hope the protein folds or the ligand escapes its binding pocket. We propose to treat such a jump model the way recent spatial-reasoning work treats a video generator, as a world model to be searched at test time. Many short branches are launched from the current state. Each branch is scored on a small, rotation-invariant contact graph that records which residues touch and how far the ligand sits from its pocket. Branches that are physically valid and reach new regions are kept and the rest are pruned. The idea comes from MindJourney, which raised vision-language model (VLM) accuracy on spatial questions by running beam search over camera rollouts of a video diffusion model.

Our proposal departs from MindJourney on one point. MindJourney asks the language model to judge which rollout is helpful. For molecules, the spatial-reasoning literature gives strong reason to expect such a judge to fail. Language models confuse simple spatial relations, cannot take other viewpoints, and produce unphysical output when asked for 3D coordinates. We therefore replace the judge with a geometric verifier, and we test the language-model judge directly as a control arm. The claim is that guided branching over a learned dynamics model finds rare transitions, such as ligand unbinding and protein folding, at a lower compute cost than independent rollouts of the same model. Because branch weights are tracked during search, the paths it finds keep a known probability under the model.

Where the two directions meet, and what is already done

The molecular dynamics side supplies the generators and the open problem. DeepJump uses conditional flow matching (a generator that learns to transport noise into data along a learned velocity field) with equivariant networks. It jumps 1 to 100 ns and folds fast-folding proteins with up to about 1000x acceleration. BioMD forecasts coarse anchor frames and then interpolates between them. It reports complete ligand unbinding paths for 97.1% of DD-13M systems "within just ten attempts." FlashMD and BioDynaSpec attack drift through long strides and spectral generation. All of them sample open-loop, and the file names rare-event and out-of-distribution transitions as their shared weakness.

Energy-based Alignment improves the generator itself by reweighting training with Boltzmann factors. That is a training-time fix. Ours is an inference-time fix that leaves the generator untouched, so the two can be combined.

The closest bridging work treats language models as components of the dynamics model. EquiLLM feeds invariant distance summaries into a frozen LLM. MDLM tokenizes backbone dihedral angles. ConfRover runs a causal transformer over frames, behind an SE(3) diffusion head. None of them searches over rollouts.

The spatial-reasoning side contributes two things: a method and a constraint. The method is world-model search from MindJourney. The constraint is evidence that language models should not be the spatial judge. Models score near chance on controlled preposition tests (What's "up"). None exceeds chance on Level-2 perspective-taking (Vision Language Models See What You Want but not What You See). Models flip reference frames mid-answer (ViewSpatial-Bench). GPT-4o-mini fails at direct coordinate prediction (EquiLLM).

A third lesson comes from SpatialThinker, whose gains came from forcing an explicit intermediate scene graph. That supports scoring on an explicit relational graph rather than on raw coordinates or text.

Core Claim

At matched generator calls, branching search over a pretrained jump model raises the rate of reaching rare target states over independent rollouts. It should also shorten the time to first success. The search uses a verifier defined on an invariant contact graph, with weights tracked so that path probabilities stay correct. The gain should be largest exactly where independent sampling is weakest: the hardest unbinding systems and the slower folders.

A language-model scorer given the same information should do no better than the geometric verifier. It may do no better than random pruning. If instead it does help, that contradicts the deficiency literature in an informative way.

Research Program

We wrap an existing jump model $p_\theta(x_{t+\Delta} \mid x_t)$, where $x_t$ is the all-atom or Cα structure at time $t$ and $\Delta$ is the jump horizon. We use BioMD for protein–ligand unbinding and DeepJump for folding, so that each comparison is against the generator's own published regime.

At each search round, every surviving branch is extended $k$ times. Each child is mapped to a contact graph $G(x)$. Its nodes are residues, plus the ligand when present. Its edges mark pairs within a distance cutoff, and ligand–pocket distances are kept as edge features. The graph is invariant to rotation and translation by construction, which removes the frame-flipping failure seen in VLMs. Children are scored by

$$s(x) = -\alpha\, V(x) + \beta\, N\big(G(x) \mid \mathcal{B}\big) + \gamma\, P\big(G(x)\big).$$

Here $V$ counts physical violations: steric clashes, bond-length and bond-angle deviations, and Ramachandran outliers. $N$ is the novelty of the graph relative to the set $\mathcal{B}$ of graphs already visited, measured as distance to the nearest neighbour in contact-map space. $P$ is a task-agnostic progress term: the loss of native ligand–pocket contacts for unbinding, or the fraction of contacts predicted by a static structure predictor for folding. The weights $\alpha, \beta, \gamma$ are nonnegative.

The top-$B$ children survive. When a branch is split, its weight is divided among its copies; when branches are pruned, the survivors absorb the pruned branches' weight. This bookkeeping keeps the final ensemble of paths an unbiased sample under $p_\theta$. As a result, path statistics remain estimable after the search has steered sampling. This separates the method from greedy best-of-N selection, which loses that property.

The analysis question is which score term carries the gain, and whether the progress term $P$ is needed at all. A version without $P$ uses no target information and is the cleanest result.

Baselines

The do-nothing reference is the jump model run as independent rollouts at the same total number of generator calls. For unbinding, this is BioMD's own protocol, for which the file reports 97.1% unbinding success on DD-13M within ten attempts (BioMD). Because 97.1% is near ceiling, we will not claim a win on that headline. A win means one of three things at a budget equal to ten BioMD attempts: higher success on the residual roughly 3% of hard systems, the same success at a smaller budget (a fraction of an attempt's calls), or lower clash and bond-error scores along successful paths.

For folding, the reference is DeepJump's independent rollouts on the 12 fast-folding proteins. The file reports only that 1 ns and 10 ns jumps fold proteins better than 100 ns jumps. We found no reported numerical folding-success rate on Lacuna for DeepJump, so we will rerun it ourselves with identical seeds and horizons.

The strongest existing competitor is the same generator after Energy-based Alignment, which reports 0.90 per-target RMSF correlation on ATLAS (Aligning Protein Conformation Ensemble Generation with Physical Feedback). It is the strongest because it fixes the same physical-validity failures that our verifier filters for. To be fair, we report both the aligned generator alone and the aligned generator with search, all at matched generator calls. Every arm has the same tuning budget for its hyperparameters: beam width, branching factor, and score weights for search, or the number of rollouts for the baselines.

Classical molecular dynamics supplies the compute anchor: about 160 GPU days for 100 µs of WW domain simulation (Str2Str). It also supplies the reference paths and free-energy landscapes.

Four ablations isolate the key ingredient. The first is search with random pruning, which keeps the branching but drops the verifier. The second keeps only the validity term. The third keeps validity and novelty but drops the progress term. The fourth, the MindJourney-style arm, replaces $s$ with an LLM scorer. That scorer receives EquiLLM-style invariant text summaries of each child's contact graph and ranks the children for helpfulness.

Why not just…?

Why not just draw more independent samples? At a large enough budget, independent rollouts will eventually find any transition the model can produce. BioMD's ten-attempt result shows that for most unbinding systems this already works. Search only earns its place if it helps where success per rollout is small: the hard residual systems and the slow folders. The random-pruning ablation separates the effect of branching from the effect of scoring. If matched-budget independent sampling equals search on the hard subsets, the idea fails, and we will report it as such.

Why not just train a better generator with physical feedback, as Energy-based Alignment does? This is a real threat for the validity metrics. An aligned generator may produce so few clashes that the $V$ term becomes redundant. It does not, however, address open-loop exploration of rare states, which is the term $N$. Running search on top of the aligned model settles which effect remains. If search adds nothing to the aligned model on time-to-first-transition, then the opportunity lies in training rather than inference.

Why not let a frontier language or vision-language model judge the rollouts, as MindJourney does? The evidence argues against it. GPT-4o scores 34.98% on ViewSpatial-Bench, against roughly 26% chance. No model beats chance on PerspectBench. GPT-4-class models get 39% of numerical physical questions right on MaScQA (Are LLMs Ready for Real-World Materials Discovery?). We include this arm anyway, because the prediction is falsifiable. If the LLM scorer beats random pruning, then language priors about molecular plausibility carry usable signal. That would bear on the deficiency debate.

Why not just use a pure equivariant network as the scorer? A learned equivariant critic is a reasonable variant, and EquiLLM's ESTAG baseline is one candidate. We use a hand-built contact-graph score first because it requires no training and cannot overfit to the generator's errors. A learned critic is a follow-up arm, judged against the same budget.

Proposed Experiments

The first experiment runs on DD-13M with BioMD as the generator. We stratify systems by BioMD's own per-attempt success rate. For each arm, we report unbinding success and generator calls to first success as functions of budget. We also report clash and bond-geometry scores and agreement with the reference unbinding paths.

The second experiment uses DeepJump on the 12 fast-folding proteins. Its metrics are folding success, time to first native-like structure, and agreement of the weighted path ensemble's free-energy landscape with the long reference trajectories. Str2Str's BPTI kinetic clusters serve as an additional state-coverage check.

A sanity experiment on ATLAS confirms that searching with only the validity and novelty terms does not distort equilibrium flexibility. RMSF correlation should stay near the generator's baseline once branch weights are applied.

The decisive check comes from comparing the full verifier, random pruning, and the LLM scorer on the hardest stratum, all at a budget of ten attempts' worth of generator calls. If the full verifier clearly beats random pruning there, and the target-free version without $P$ keeps most of that gain, the claim holds. If not, test-time search is not a useful lever for rare events in learned dynamics models.