Diffusion-Model-Based Restart Selection for Adaptive Sampling of Adenylate Kinase: Structural Integrity Assessment and Comparison with Simple Baselines

Generative models are increasingly proposed for choosing where adaptive molecular dynamics should restart, but whether their proposals are physically sound starting points and whether they outperform simple rules has rarely been tested together. Using adenylate kinase, we first audited a workflow that reconstructs restart structures from a diffusion model's latent space: judged by per-domain native contacts calibrated on unbiased MD, most reconstructed structures were damaged from their first frame, and the rare states identified from them were partially unfolded LID domains. We then compared restart selection by a denoising diffusion probabilistic model with a latent-space Gaussian ablation, least-counts and two unbiased baselines, starting only from intact MD frames, with targets fixed and timestamped before any production run, equal simulation time and three replicates of every strategy and baseline. No strategy reached the primary target within the pre-specified 0.84 µs per series, so the binary comparison cannot distinguish them; adaptive restarts covered the reference distribution slightly better than the baselines, and in an exploratory closest-approach analysis the diffusion model did not come closer to the target than the simple strategies. The target differed from the data shown to the strategies only in a combination of domain states, which the diffusion model did not propose. Selecting starts by novelty in the latent space, with or without a learned model, tended to choose frames farther from the target than random picks, and the diffusion model's samples stayed close to the data. The value of a generative model lies in proposing structures beyond the data, yet that is the capability that produced the damaged restarts; confined to intact frames, its role reduces to a novelty rule on existing frames. We recommend domain-resolved integrity gates, pre-specified targets, simple baselines at equal simulation time and replicated runs for evaluating generative adaptive sampling.