A diffusion model chooses where to restart MD, but simple baselines do just as well, and many reconstructed structures are damaged
- On adenylate kinase, the authors first audited a workflow that rebuilds restart structures from a diffusion model's latent space; by native-contact checks most were damaged from the first frame.
- They then compared diffusion-based restart selection with a latent Gaussian ablation, least-counts and unbiased baselines, using intact MD frames, pre-specified targets, equal simulation time and three replicates.
- No strategy reached the primary target within the budget; adaptive restarts covered the reference distribution slightly better than baselines, but the diffusion model got no closer to the target than the simple strategies.
- Choosing starts by latent-space novelty tended to select frames farther from the target than random picks.
💡 Points Based on the abstract
I like that the targets were fixed and timestamped before production runs; that is how an ML-for-MD claim should be tested, and it is rare. The practical message for anyone running enhanced sampling on a drug target is to try least-counts first and to check that a generative model's structures are physically intact before spending GPU time on them. One small protein limits how far this generalizes.