Research map / Probabilistic and scientific modeling / Molecules and proteins

Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data

NeurIPS 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Guides the sampling trajectory of a pretrained diffusion structure predictor (Boltz-1) with multiscale global-to-local cryo-EM density guidance to recover multiple conformational states from heterogeneous maps, without retraining.

Problem

Structure prediction models generate ensembles peaked around a single conformation and fail to capture conformational diversity of dynamic biomolecular complexes, while cryo-EM density maps of heterogeneous states lack atomic models and existing model-building tools are incomplete at low resolution.

Approach

Apply DPS-style likelihood guidance to a pretrained AlphaFold3-like diffusion model in atomic space using a two-stage multiscale schedule: a global guidance term matching a weighted k-means point cloud of the map via Sinkhorn divergence, followed by local guidance using a simulated-map likelihood with the original density map, bracketed by unguided warm-up and relaxation steps.

Key results

• On synthetic STP10 inward state, CryoBoltz achieves all-atom RMSD 1.057 Å versus 3.815 Å for Boltz-1 and 1.263 Å for AlphaFold3.
• On four P-glycoprotein states, CryoBoltz reaches all-atom RMSD 1.31–1.73 Å versus 2.93–6.99 Å for Boltz-1, while ModelAngelo completes only 2.3–40.3% of residues.
• On CYP102A1 closed state, guidance improves all-atom RMSD from 8.78 Å (Boltz-1) to 2.00 Å.

Measured effects reported in the paper

Surprising findings

• Unguided AlphaFold3 predicts only the inward STP10 state and Boltz-1 only the outward state, showing single-state bias in both models.
• MSA subsampling of Boltz-1 gives only modest gains and does not sample the alternative STP10 state.

Limitations

• Optimization is unstable due to multimodality of the likelihood; mitigated only by sampling many structures, at higher memory and time cost.
• Relies on the unguided base model providing a good warm-up initialization, which failed on some systems (DSL1/SNARE complex, full IgG antibody).
• Heuristic choices such as stage durations and guidance schedules, and the global point-cloud term, are not physics-derived.

Experimental setup

Datasets: STP10 synthetic map from PDB 7AAQ/7AAR CH67 Fab synthetic map from PDB 4HKX P-glycoprotein EMD-40226/40259/40258/40227 Pma1 EMD-64135/64136 CYP102A1 EMD-27534/27536 YbbAP EMD-51292/51291

Benchmarks: STP10 CH67 antibody P-glycoprotein Pma1 CYP102A1 YbbAP

Metrics: all-atom RMSD C-alpha RMSD TM score local RMSD RSCC

Baselines: Boltz-1 AlphaFold3 Boltz-1 + MSA subsampling ModelAngelo

Builds on: Boltz-1 diffusion posterior sampling (DPS) AlphaFold3 ChimeraX molmap GeomLoss Sinkhorn divergence

Closest papers by meaning

• CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM (NeurIPS 2024 · similarity 0.92)
• CryoFM: A Flow-based Foundation Model for Cryo-EM Densities (ICLR 2025 · similarity 0.92)
• cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.92)
• Mixture of neural fields for heterogeneous reconstruction in cryo-EM (NeurIPS 2024 · similarity 0.91)
• Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery (NeurIPS 2025 · similarity 0.90)
• CryoDomain: Sequence-free Protein Domain Identification from Low-resolution Cryo-EM Density Maps (AAAI 2025 · similarity 0.90)