#30 · Structural bio & protein design

Generative models produce protein conformational ensembles guided by experimental data

The dominant thread is generative modeling of conformational distributions rather than single static structures, with experimental observables entering as constraints or benchmarks. A C-alpha flow-matching ensemble generator is dissected through ablation of architecture, sequence conditioning, alignment and training data, and benchmarked against SAXS. Empirical Bayes Flow Matching learns flow-based priors over atomic structures from noisy cryo-EM projections via stochastic EM. A differentiable spin-center diffusion model, coupled to Boltz-2, allows DEER distance distributions to guide generation of weighted protein ensembles. FoldScape, by contrast, conditions directly on sequence and a specified folding fraction to generate C-alpha ensembles, and the listed summary does not tie it to experimental data. The novelty lies in coupling generative conformational models to experimental observables rather than to static geometry alone. Not every item addresses the same question, but the shared structure holds across the set.