Metadiffusion: Inference-Time Meta-Energy Biasing of Biomolecular Diffusion Models 1. The authors introduce metadiffusion, a method that adds an inference-time meta-energy biasing layer on top of pretrained biomolecular diffusion models like Boltz-2, enabling diverse conformational ensemble generation without any retraining or fine-tuning. 2. The approach supports three complementary modes: optimisation (pushing collective variables to extrema), steering (targeting specific values), and exploration (maximizing inter-sample diversity through Gaussian repulsion potentials). 3. Metadiffusion generates conformational ensembles whose residue-level flexibility patterns closely match molecular dynamics simulations, achieving mean Pearson's R of 0.81 with ATLAS MD data—approaching the reproducibility ceiling between independent MD runs themselves. 4. The method enables controlled exploration of collective variables including radius of gyration, hinge angles, solvent-accessible surface area, and pairwise RMSD, allowing researchers to traverse alternative binding poses across proteins, nucleic acids, and ligands. 5. Through gradient-guided denoising, metadiffusion can steer ensembles to match experimental observables including SAXS pair distance distributions and NMR chemical shifts, as demonstrated with class V GTP aptamer and human calmodulin. 6. The technique operates orders of magnitude faster than microsecond-scale MD simulations while maintaining physical plausibility, with generated structures readily improvable through energy minimization to resolve occasional bond breaks and steric clashes. 7. As a model-agnostic framework, metadiffusion can theoretically extend to other diffusion-based structure predictors including AlphaFold3, OpenFold, and Chai-1, bridging the gap between single-structure prediction and ensemble-level experimental constraints. 💻Code: github.com/metadiffusion/met… 📜Paper: biorxiv.org/content/10.64898… #compbiol #structuralbiology #proteindynamics #diffusionmodels #machinelearning #alphafold #boltz2 #conformationalensemble #SAXS #NMR #metadynamics #biomolecularsimulation