Guiding AlphaFold3 to generate experiment-consistent protein ensembles AlphaFold3 is a diffusion model, and like most generative models trained to predict a single most-probable output, it collapses to one dominant mode. For proteins that is a real problem, because proteins are not static. They breathe, flip loops, and sample multiple conformations, and the techniques used to measure them (NMR, X-ray crystallography, cryo-EM) report ensemble averages, not single snapshots. So you have a powerful prior that throws away exactly the heterogeneity the experiments are trying to capture. Advaith Maddipatla and coauthors fix this by reframing the whole task as posterior inference. They keep AlphaFold3 frozen as a sequence-conditioned prior and cast ensemble modeling as sampling from p(ensemble | sequence, data). Concretely, they modify the reverse diffusion steps to add a gradient-based guidance term derived from an experimental likelihood, so at each denoising step the sampler is nudged toward structures whose ensemble-averaged observables match the measurements. A single scaling hyperparameter interpolates between pure prior and strongly data-conditioned output. What makes it general is that any modality becomes a differentiable likelihood you can plug in: NOE distance restraints and order parameters from NMR, real-space electron density from crystallography, electrostatic potential maps from cryo-EM, and combinations of these. After sampling they run a short force-field relaxation and a matching-pursuit style pruning step to return a compact set of conformers. The results hold up. On ubiquitin, NOE-guided sampling produces far fewer distance-restraint violations than the deposited NMR ensemble, and adding an order-parameter term recovers backbone dynamics that match relaxation data, all in GPU minutes rather than the days that ensemble molecular dynamics needs. Across 91 NMR cases it beats the deposited PDB ensembles in about 77% of them. On crystallographic data it recovers alternate conformations hiding in electron density that the unguided model misses entirely. This is a practical inference tool, not a replacement for refinement. In drug discovery it can propose alternate loop conformations and occupancies during model building, surface cryptic states worth targeting, and speed up map fitting in cryo-EM workflows. The small compute footprint means it slots into existing structural pipelines as an interactive step, which matters for biotech and materials groups doing large-scale reanalysis of legacy structural data. Paper: Maddipatla et al., Nature Biotechnology (2026), CC BY 4.0 | nature.com/articles/s41587-0…