Robust Inference-Time Steering of Protein Diffusion Models via Embedding Optimization
Minhuan Li, Jiequn Han, Pilar Cossio, Luhuan Wu · 2026
Read Robust Inference-Time Steering of Protein Diffusion Models via Embedding Optimization by Minhuan Li, Jiequn Han, Pilar Cossio, Luhuan Wu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-05285.txt.
Description
Protein diffusion models generate plausible molecular shapes, but experimenters need shapes that also match their measurements. Minhuan Li and 3 co-authors present EmbedOpt in this 2026 machine learning paper, a way to steer generation at run time. The usual approach perturbs atomic coordinates with gradients from the experimental likelihood, which gets unstable when the target shape sits in a region the model considers unlikely.
EmbedOpt optimizes a different knob: the model's conditional embedding, which carries rich coevolutionary information. Updating the embedding moves the prior itself toward the experimental constraints instead of fighting it at the coordinate level. On sparse distance constraints the method matched coordinate-based posterior sampling, and on fitting cryo-electron microscopy maps, including real noisy ones, it did better. It also shrugged off hyperparameter changes across two orders of magnitude and reached comparable quality with fewer diffusion steps.
Catalog record
| CREATOR | Minhuan Li, Jiequn Han, Pilar Cossio, Luhuan Wu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-05285