Technical Summary: CrystalBoltz

Problem Statement

Protein structure prediction models trained on public databases (e.g., AlphaFold, Boltz-2) provide powerful structural priors but are typically unconditioned on new experimental data. In X-ray crystallography, the primary challenge is the "phase problem": measured structure-factor amplitudes ($|F_o|$) lack phase information, meaning they do not uniquely determine an electron density map or atomic structure. Consequently, structure determination remains an inverse problem requiring candidate models to be both structurally plausible and consistent with measured diffraction data.

Current workflows often rely on substantial manual refinement by experts or iterative methods that struggle to incorporate experimental data directly into generative sampling. While recent methods attempt to steer predictors using experimental signals, they often operate on sequence inputs or latent representations rather than directly conditioning on observed structure-factor amplitudes. Furthermore, existing experimentally guided refinement methods can be computationally expensive, limiting their utility during active data collection sessions.

Methodology

CrystalBoltz is an end-to-end generative framework that formulates crystallographic structure determination as experiment-guided posterior sampling using a diffusion prior. The method operates in two distinct stages:

1. Experiment-Guided Diffusion Sampling

The core of CrystalBoltz is the integration of a sequence-conditioned diffusion prior (Boltz-2) with a differentiable crystallographic likelihood. The goal is to sample a structure $X$ from the posterior distribution $p(X|a, c, y)$, where $a$ is the protein sequence, $c$ represents crystal parameters (unit cell and space group), and $y$ are the experimental observables ($|F_o|$).

• Diffusion Posterior Sampling (DPS): The method augments the reverse stochastic differential equation (SDE) of the Boltz-2 prior with gradients from the experimental likelihood. Following the DPS framework, the intractable marginal likelihood is approximated using Tweedie's formula, utilizing the denoiser's one-step prediction ($\hat{X}_0$) as a point estimate of the clean structure.
• Differentiable Forward Model: To enable gradient-based guidance, CrystalBoltz employs SFCalculator, a tool that provides a differentiable calculation of bulk solvent scattering and structure factors. This allows for the direct computation of gradients with respect to atomic coordinates.
• Likelihood Guidance: The guidance signal is derived from two likelihood targets comparing calculated amplitudes ($|F_c|$) against observed amplitudes ($|F_o|$):
  1. A heteroscedastic least-squares loss (Gaussian error model).
  2. A Rice distribution likelihood, which accounts for the statistical properties of structure factors (complex normal for acentric reflections, real normal for centric reflections) and incorporates atomic coordinate errors via a $\sigma_A$ parameter.
• Scheduling: Guidance is not applied from the start of the diffusion process. Instead, sampling begins unguided to allow a coarse backbone fold to emerge. Experimental guidance is switched on at a specific timestep ($t_g$) once a plausible structure has formed, preventing the noisy early stages from being pulled toward spurious minima by the likelihood gradient.

2. Post-Guidance Refinement

The diffusion stage produces a global structure but may lack local precision (e.g., side-chain rotamers) and accurate B-factors. A second stage performs post-guidance refinement:

• The denoised structure is aligned to the crystal frame.
• Atomic coordinates and isotropic B-factors are jointly optimized against the experimental data.
• Unlike the diffusion stage, which relies on likelihood gradients propagated through the denoiser, this stage optimizes the crystallographic objective directly (either the correlation coefficient or the R-factor) using the forward model alone.
• B-factors are initialized from the prior's pLDDT scores and clamped to physical ranges ($[1, 80] \text{ \AA}^2$).

Key Contributions

1. Formulation as Posterior Sampling: The paper casts crystallographic structure determination as experiment-guided posterior sampling with a diffusion prior, directly linking experimental structure-factor amplitudes to atomic structures in an end-to-end fashion.
2. Native Likelihood Guidance: The authors derive crystallography-native likelihood guidance from differentiable structure-factor objectives (using SFCalculator), enabling the model to resolve large conformational changes that sequence-only priors might miss.
3. Hybrid Pipeline: The method combines guided sampling (for global conformational search) with a post-guidance refinement stage (for local coordinate and B-factor optimization), separating the tasks of fold determination and precise fitting.
4. Efficiency: The framework achieves state-of-the-art agreement metrics while significantly reducing runtime compared to existing experimentally guided methods.

Experimental Results

The authors evaluated CrystalBoltz on six protein targets from the PDB (8DWN, 4NTZ, 7O51, 7SEZ, 7VNX, 1L63) with resolutions ranging from 1.69 to 2.20 Å.

• Accuracy: CrystalBoltz consistently achieved lower Coordinate RMSD and lower R-factors ($R_{work}$ and $R_{free}$) compared to strong baselines, including Phenix (initialized from Boltz-2), unguided Boltz-2, and ROCKET (a recent data-guided method optimizing MSA embeddings).
  • On targets with significant conformational differences from the prior (e.g., 8DWN and 4NTZ), CrystalBoltz demonstrated the ability to correct large structural rearrangements, reducing RMSD from ~2.6–4.5 Å (baselines) to ~1.3 Å.
  • On targets where the prior was already close to the ground truth, CrystalBoltz still provided consistent improvements in R-factors.
• Runtime: CrystalBoltz achieved a 33.3× speedup over ROCKET on identical hardware (NVIDIA A6000). While ROCKET required ~376 minutes (involving iterative MSA optimization and multiple refinement phases), CrystalBoltz completed the task in approximately 11.3 minutes. This reduction makes experimentally guided refinement feasible during beamtime sessions rather than as an offline post-processing step.
• Ablation Studies:
  • Solvent Model: Including the differentiable bulk-solvent term in the forward model consistently improved the correlation coefficient (CC) by ~0.05–0.10, confirming the necessity of modeling solvent scattering.
  • B-factor Refinement: Jointly optimizing B-factors alongside coordinates improved RMSD by absorbing local disorder, allowing coordinates to relax more accurately toward the reference.
  • Scheduling: Delaying guidance until a coarse fold emerged (mid-way guidance) was critical; applying guidance from the first step degraded performance by pulling the trajectory toward poor minima before topology was established.

Significance and Claims

The paper claims that moving experimental conditioning into the generative sampling process is an effective path toward faster, data-consistent structure determination. By integrating learned structural priors with crystallography-native objectives in a single inference pipeline, CrystalBoltz bridges the gap between high-accuracy prediction and experimental validation.

The authors position the work as a proof-of-concept for end-to-end model building and refinement using differentiable likelihoods. They highlight that the method's ability to handle large conformational changes while maintaining minute-scale runtimes makes it competitive with classical refinement pipelines. The framework is presented as generalizable to other experimental modalities (e.g., cryo-EM) provided a differentiable forward model exists, though the current work focuses exclusively on X-ray crystallography. The authors acknowledge limitations, such as the dependency on reference structures for alignment (via molecular replacement) and the current restriction to single-chain proteins, noting these as natural directions for future engineering and algorithmic improvements.