Multiscale Guidance of Protein Structure Prediction with Heterogeneous Cryo-EM Data

Conference: NeurIPS 2025 arXiv: 2506.04490 Code: GitHub Area: Medical Imaging Keywords: protein structure prediction, cryo-EM, diffusion model guidance, conformational diversity, Boltz-1

TL;DR

CryoBoltz leverages cryo-EM density maps to guide the sampling trajectory of a pretrained diffusion-based structure prediction model (Boltz-1) via a multiscale guidance mechanism (global → local), generating multi-conformational atomic models consistent with experimental data without any retraining.

Background & Motivation

The field of protein structure prediction faces two major challenges:

Single-conformation bias in structure prediction models: Diffusion models such as AlphaFold3 and Boltz-1 can generate structures, but their sampling distributions are heavily concentrated on a single conformation—for example, Boltz-1 samples only the outward-facing conformation of STP10, while AlphaFold3 samples only the inward-facing one. Existing MSA subsampling strategies moderately increase diversity but remain limited and cannot cover the conformational continuum.

Bottleneck from cryo-EM reconstruction to atomic models: Cryo-EM experiments can capture the conformational landscape of proteins, yet the resulting 3D density maps are not atomic models. Existing model-building methods (e.g., ModelAngelo) require high-resolution maps (<4 Å) and frequently fail on low-resolution (>4 Å) or heterogeneous complexes—for instance, ModelAngelo successfully modeled only 2.3%–40.3% of residues across four conformations of P-glycoprotein.

The core idea of this paper is to inject experimental information from cryo-EM density maps into the reverse sampling process of a pretrained diffusion model, thereby exploiting both the sequence and biophysical priors learned by the structure prediction model and the real conformational information captured by experimental data.

Method

Overall Architecture

CryoBoltz is built on the Diffusion Posterior Sampling (DPS) framework, treating the pretrained Boltz-1 diffusion model as an implicit prior \(p(\mathbf{x}|\mathbf{s})\) and the cryo-EM density map as observation \(\mathbf{y}\), guiding the reverse diffusion process to sample the posterior \(p(\mathbf{x}|\mathbf{y},\mathbf{s})\). The entire process consists of four stages over 200 diffusion steps:

Warm-up → Global Guidance → Local Guidance → Relaxation

Key Designs

1. Global Guidance: The density map \(\mathbf{y} \in \mathbb{R}^{w \times h \times d}\) is converted into a 3D point cloud \(\mathbf{Y} \in \mathbb{R}^{k \times 3}\) via weighted k-means clustering, where \(k = \lfloor N/(4r^3) \rfloor\) (\(N\) is the number of atoms, \(r\) is the voxel size). The guidance term is based on the Sinkhorn divergence (regularized Wasserstein distance):

\[\tilde{s}_\theta(\mathbf{y}, \mathbf{x}, \mathbf{s}, t) = -\nabla_\mathbf{x} \mathfrak{D}(\hat{\mathbf{x}}_\theta(\mathbf{x}, \mathbf{s}, t), \mathbf{Y})\]

This stage focuses solely on the global shape of the protein (e.g., which helices face inward or outward) without involving high-resolution details, thereby avoiding optimization difficulties arising from nonlinear likelihood functions in the early diffusion steps. Guidance strength is annealed via cosine scheduling from 0.25 to 0.05.

1. Local Guidance: Using the raw density map and a cryo-EM physical forward model (each non-hydrogen atom contributes a Gaussian scattering potential), the guidance term directly minimizes the L2 distance between the simulated and experimental density maps:

\[\tilde{s}_\theta(\mathbf{y}, \mathbf{x}, \mathbf{s}, t) = -\nabla_\mathbf{x} \|\mathbf{y} - \mathcal{B}(\Gamma(\hat{\mathbf{x}}_\theta(\mathbf{x}, \mathbf{s}, t), \mathbf{s}))\|^2\]

where \(\Gamma\) maps atomic coordinates to a density volume and \(\mathcal{B}\) simulates the blurring effect of B-factors. Guidance strength is fixed at \(\lambda=0.5\).

1. Multiscale Guidance Scheduling: For synthetic data, \(T_w=125, T_g=25, T_l=25, T_r=25\); for experimental data, \(T_w=100, T_g=50, T_l=25, T_r=25\) (more global guidance steps are used for experimental data to drive larger conformational changes). The warm-up and relaxation stages apply no guidance—warm-up establishes a reasonable initialization, while relaxation corrects steric clashes and other fine-grained details.

Loss & Training

CryoBoltz requires no training whatsoever and operates entirely as an inference-time guidance method. A key advantage is that it automatically benefits from continued improvements to the underlying base model. For each experiment, 25 structures × 3 replicates are sampled, and the structure with the lowest RMSD is selected as the final result. Computations are performed on a single A100 80 GB GPU.

Key Experimental Results

Main Results (Synthetic Data)

System	Metric	CryoBoltz	Boltz-1	Boltz-1+MSA Subsampling	AlphaFold3
STP10 (inward)	All-atom RMSD (Å)	1.057	3.815	3.768	1.263
STP10 (outward)	All-atom RMSD (Å)	0.888	2.656	2.542	4.478
CH67 antibody	Local RMSD (Å)	1.269	3.120	3.270	3.191
CH67 antibody	TM score	0.994	0.972	0.969	0.971

Main Results (Experimental Data)

System	Resolution (Å)	CryoBoltz RMSD	Boltz-1 RMSD	AF3 RMSD	ModelAngelo Coverage
P-gp (apo)	4.3	1.382	6.994	3.827	40.3%
P-gp (inward)	4.4	1.348	5.630	2.692	18.3%
P-gp (occluded)	4.1	1.727	2.929	3.440	2.3%
Pma1 (inhibited)	3.52	1.999	6.140	8.017	72.8%
CYP (closed)	4.4	2.004	8.784	3.585	18.9%
CYP (open)	6.5	4.167	8.532	6.490	0.0%

Ablation Study

Configuration	STP10 Inward RMSD	STP10 Outward RMSD	Notes
Full CryoBoltz	1.057	0.888	Global + local guidance
Local guidance only	3.860	2.722	Lacks global guidance; fails to drive large conformational changes
Global guidance only	1.287	1.164	Lacks local guidance; insufficient high-resolution detail

Key Findings

• CryoBoltz outperforms both Boltz-1 and AlphaFold3 on all 10 experimental density maps.
• Unguided Boltz-1/AF3 typically samples only one conformation; CryoBoltz successfully recovers two to four distinct conformations.
• ModelAngelo fails severely at low-resolution maps (>4 Å), achieving 0% coverage at 6.5 Å.
• Global guidance drives large-scale conformational changes, while local guidance refines high-resolution details; both components are indispensable.

Highlights & Insights

• Zero training overhead: A purely inference-time method that can be plug-and-play integrated into any diffusion-based structure prediction model.
• Necessity of multiscale guidance: The forward mapping from density maps to atomic models is highly nonlinear; directly optimizing the likelihood leads to multimodality issues, while the point-cloud intermediate representation elegantly resolves optimization instability in the early guidance stages.
• High practical value: cryo-EM model building typically requires hours of manual refinement; CryoBoltz accomplishes this within minutes.

Limitations & Future Work

• Optimization stability is affected by the multimodality of the likelihood \(p(\mathbf{y}|\mathbf{x})\), necessitating sampling of multiple structures and selecting the best, which increases computational cost.
• The warm-up stage relies on reasonable initialization by the base model; initialization fails for some complex systems (e.g., the DSL1/SNARE complex).
• The step allocation across guidance stages is set heuristically, and no automatic selection strategy has been established.
• In multi-density-map scenarios, a shared deformation model as an alternative to independent optimization remains unexplored.

Related Work & Insights

• DPS / ScoreALD: Diffusion model posterior sampling frameworks; CryoBoltz is directly built upon DPS.
• ROCKET: Uses AF2 as a regularizer but optimizes in AF2's latent space; CryoBoltz optimizes directly in atomic space.
• ModelAngelo: The current state-of-the-art de novo model-building method, but constrained by high-resolution requirements.
• Insight: Inference-time guidance of diffusion models represents a general paradigm for integrating experimental data with learned priors.

Rating

• Novelty: ⭐⭐⭐⭐⭐ First application of the DPS guidance framework to protein structure prediction integrated with cryo-EM data.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Synthetic and real data, six biological systems, multiple baselines, comprehensive ablations, and statistical tests.
• Writing Quality: ⭐⭐⭐⭐⭐ Clear problem formulation, rigorous methodological derivation, and high-quality figures.
• Value: ⭐⭐⭐⭐⭐ Addresses an important bottleneck in structural biology, with open-source code and strong practical utility.

Related Papers

• [NeurIPS 2025] Quantifying the Role of OpenFold Components in Protein Structure Prediction
• [NeurIPS 2025] Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs
• [ICCV 2025] CryoFastAR: Fast Cryo-EM Ab initio Reconstruction Made Easy
• [CVPR 2026] cryoSENSE: Compressive Sensing Enables High-throughput Microscopy with Sparse and Generative Priors on the Protein Cryo-EM Image Manifold
• [NeurIPS 2025] Steering Generative Models with Experimental Data for Protein Fitness Optimization