Inference-time optimization for experiment-grounded protein ensemble generation

Protein function relies on dynamic conformational ensembles, yet models like AlphaFold3 (AF3) often fail to produce ensembles that match experimental data. Recent experiment-guided approaches address this by steering the reverse diffusion process, but remain limited by fixed sampling horizons and sensitivity to initialization, often yielding thermodynamically implausible results. We introduce a general inference-time optimization framework that instead optimizes latent representations to maximize ensemble likelihood, removing dependence on diffusion length, reducing initialization bias, and enabling flexible incorporation of external constraints. Further, we present sampling schemes for generating Boltzmann-weighted ensembles by combining AF3 structural priors with force-field energies. Across X-ray and NMR benchmarks, our method improves diversity, physical energy, and agreement with experimental data, often surpassing deposited PDB structures. Finally, ipTM maximization shows that perturbing AF3 embeddings can artificially inflate model confidence, exposing a vulnerability in current design metrics and suggesting ways to reduce false discoveries in binder engineering.