Predicting alternative protein conformations using cryo-EM guided AlphaFold3
Abstract
The function of a protein is determined through its three-dimensional structures. The long-standing challenge of predicting this structure from sequence data has recently seen great progress driven by diffusion models like AlphaFold 3 [Abramson et al. Nature, 2024] and Boltz [Wohlwend et al. bioRxiv, 2024]. However, these models often output static, dominant conformations, failing to reflect the conformational ensemble of proteins. Integrating experimental data, such as cryo-EM maps, into the diffusion models has been proposed as a solution to capture conformational heterogeneity [Raghu et al. NeurIPS, 2026; Li et al. Arxiv, 2026]. However, current guidance methodologies are not always successful and often fail on more challenging targets. To address this gap, we propose enhancing cryo-EM guided diffusion through three mechanisms: conditioning the Pairformer representations directly through spectral guidance, applying a resolution annealing strategy during diffusion guidance, and an updated diffusion guidance scheme.
