For a detailed description of our approach and analysis, check out our preprint!. Predicting conformational dynamics is a key challenge for structural biology. CryoBoltz is our attempt to unlock the full potential of structure prediction models using experimental data. 🚀 7/7.

We also demonstrate our method on real cryo-EM data of P-glycoprotein. CryoBoltz models 4 conformational states of this transporter given experimental maps, whose resolutions are too low (> 4Å) for model building algorithms. 6/7

We demonstrate CryoBoltz on antibodies, whose CDR loops are challenging targets for structure prediction models due to a lack of evolutionary signal. We show that guidance with a synthetic map enables accurate modeling of the highly variable CDR H3 loop. 5/7

CryoBoltz can sample both the outward and inward state of the STP10 transporter given a corresponding synthetic density map, whereas unguided Boltz-1 and AlphaFold3 only sample the outward or inward state, respectively. 4/7

To fit both the low-resolution shape and high-resolution details of a cryo-EM map, we propose a guidance mechanism that combines global, point cloud-based constraints with local, density-based constraints, enabling sampling of conformations supported by the experimental data. 3/7

Protein structure prediction models can now map sequence to structure with high accuracy, but are often biased towards a single low-energy conformational state. Can we guide them to sample multiple functionally relevant states using cryo-EM measurements? 2/7.

Excited to present CryoBoltz ❄️⚡, a multiscale guidance approach for steering AlphaFold3/Boltz-1 to sample structures that are consistent with experimental cryo-EM density maps. 🧵1/7. Joint work with @axlevy0 @GordonWetzstein & @ZhongingAlong!

RT @ZhongingAlong: 🚨New paper at #NeurIPS2024! We present Hydra, a mixture of neural fields 🐉🐉🐉 for ab initio cryo-EM reconstruction of com….