ROCKET: MSA-bias in Crystallographic Refinement

• ROCKET is a computational method that refines macromolecular X-ray crystallography models by integrating AlphaFold2-informed priors with experimental diffraction data.
• It optimizes multiple sequence alignment inputs to improve R-factors and coordinate RMSDs, positioning it as a hybrid alternative to traditional methods like Phenix.refine.
• Benchmarking suggests ROCKET serves as a valuable baseline for integrating deep learning in crystallography, despite slower runtimes compared to diffusion-guided methods.

A ROCKET is a recent computational method for macromolecular X-ray crystallographic refinement that leverages AlphaFold2 and MSA-bias optimization to guide atomic model refinement using measured diffraction data. Developed in the context of protein structure determination, ROCKET represents an advance over traditional model rebuilding and refinement workflows by directly integrating state-of-the-art sequence-conditioned priors and large-scale structural prediction models into crystallographic pipelines. Its approach and performance must be considered relative to existing methods such as Phenix.refine, REFMAC5, and more recent diffusion-based experiment-guided platforms such as CrystalBoltz (Kim et al., 15 May 2026).

1. Background in X-ray Crystallography Inverse Problems

Macromolecular X-ray crystallography recovers atomic-resolution protein structures from measured magnitudes of reciprocal-space structure factors, $\|\mathbf{F}_o(\mathbf{h})\|$, without access to the associated phases, resulting in the canonical “phase problem.” The objective is to reconcile candidate atomic structures $\mathbf{X}$ with observed amplitudes through an ill-posed inverse problem where structural plausibility (determined by chemical and physical constraints and sequence priors) and experimental consistency (described by likelihoods comparing $\|\mathbf{F}_c\|$ and $\|\mathbf{F}_o\|$) must both be optimized. Classical refinement begins from a close-enough model—often obtained by molecular replacement—and iteratively updates coordinates and B-factors using tools such as REFMAC5 and Phenix.refine, frequently incorporating substantial manual expert intervention.

2. Incorporating Sequence-Conditioned Priors in Refinement

Recent advances in protein structure prediction—AlphaFold2/3, Boltz-2, RF3, Protenix—provide highly informative, sequence-conditioned priors $p(\mathbf{X}|\mathbf{a})$ over atomic configurations. ROCKET introduces a strategy to couple these priors with crystallographic experimental data by tuning AlphaFold2 model parameters and MSA weights to bias the generative prediction in the direction that improves agreement with measured reflection amplitudes. This process is often referred to as MSA-bias optimization, where the evolutionary information in multiple sequence alignments (MSA) is utilized for enhanced model construction.

3. Crystallographic Refinement Frameworks: ROCKET and Comparators

The primary computational challenge in conditioning structure prediction on measured amplitude data $\|\mathbf{F}_o\|$ stems from (i) the nonlinearity and complexity of the forward transform $\mathbf{X}\mapsto\|\mathbf{F}_c\|$ (impacted by factors such as bulk solvent, Debye–Waller B-factors, and symmetry averaging), and (ii) the nonconvex likelihood landscape due to the absence of phase information.

ROCKET, as evaluated alongside baselines such as unguided Boltz-2, Phenix.refine, and the diffusion/posterior-guided CrystalBoltz framework, integrates experimental amplitudes into AlphaFold2 model predictions by optimizing MSA inputs to favor better R-factors and coordinate RMSDs, thus leveraging both the generative power of machine learning models and experimental data consistency. While the details of ROCKET’s internal algorithmic steps are proprietary, it operationalizes the paradigm in which sequence-based priors are modulated on-the-fly in direct response to diffraction data (Kim et al., 15 May 2026).

Method	Structure Prior	Experimental Guidance	Refinement Strategy
ROCKET	AlphaFold2 (MSA-bias)	Optimizes wrt $\|\mathbf{F}_o\|$	MSA input optimization
Phenix.refine	Initial model/manual input	Iterative likelihood fitting	Gradual coordinate/B-factor
CrystalBoltz	Boltz-2 diffusion	Data-guided SDE sampling	Posterior-guided, gradient
Boltz-2	Boltz-2, unguided	None	Prior-only

4. Benchmarking and Performance Evaluation

Empirical benchmarks were conducted on six single-chain PDB targets (resolution 1.69–2.20 Å, lengths 164–306 residues). Metrics include global RMSD, $C_\alpha$ RMSD, R_work, and R_free. ROCKET is explicitly compared with Boltz-2 (unguided), Phenix.refine, and CrystalBoltz. Notably, CrystalBoltz outperforms ROCKET and the other baselines in both accuracy and computational efficiency—with superior global RMSD (down to approximately 0.65 Å), $C_\alpha$ RMSD (approximately 0.22 Å), R_work (approximately 0.26–0.35), and R_free (approximately 0.28–0.37) on most targets (Kim et al., 15 May 2026).

A salient finding is the dramatic reduction in runtime afforded by CrystalBoltz over ROCKET: 11.3 minutes total (CrystalBoltz) versus 376 minutes (ROCKET), representing a 33.3× speedup. This suggests that ROCKET, while integrating machine learning–derived structure prediction into refinement, is significantly less computationally efficient than modern diffusion-based approaches on the evaluated benchmarks.

5. Methodological Comparison: Diffusion-Guided Versus MSA-Bias Approaches

ROCKET’s optimization of MSA input space for AlphaFold2 enables model updating through direct evolutionary prior modulation, which is distinct from the explicit sampling of structure-conditioned posteriors with experiment-guided diffusion models (as in CrystalBoltz). The latter injects crystallographic data gradients directly into the generative diffusion process, likely contributing to both enhanced model-data consistency and large conformational shift recovery in challenging cases (e.g., RMSD improvements on 8DWN: 2.65 Å → 1.32 Å; 4NTZ: 4.54 Å → 1.30 Å) (Kim et al., 15 May 2026).

A plausible implication is that posterior-guided diffusion models offer a principled Bayesian inference formulation with tractable gradient-based refinement and direct likelihood integration, whereas MSA-bias methods, as in ROCKET, provide a more heuristic route combining strong machine learning priors with indirect experimental guidance.

6. Impact and Prospects in Crystallographic Model Building

ROCKET marks a significant step in bridging the gap between machine learning–powered priors and traditional crystallographic workflows. Its ability to guide AlphaFold2 structures with experimental amplitude data broadens the scope of hybrid refinement methodologies. However, empirical evaluation against contemporary frameworks such as CrystalBoltz indicates substantial advantages of guided diffusion models in terms of both structural fidelity and computational speed.

Consequently, the current research landscape positions ROCKET as a baseline comparator for future methods that unify deep learning, experiment-guided inference, and scalable refinement, with attention to both methodological rigor and practical throughput (Kim et al., 15 May 2026).