AI Solves the Phase Problem in X-Ray Crystallography

This article explores how artificial intelligence (AI) structure-prediction models, such as AlphaFold and RoseTTAFold, have revolutionized macromolecular X-ray crystallography. It explains the fundamentals of the crystallographic phase problem, how the molecular replacement (MR) technique overcomes this hurdle, and how AI-generated protein models now serve as precise templates to rapidly determine experimental 3D macromolecular structures.

Macromolecular X-ray crystallography relies on shooting X-rays at crystallized proteins to determine their atomic structure. As the X-rays hit the crystal, they diffract and create a pattern of spots on a detector. To reconstruct a 3D map of the protein's electron density from this pattern, scientists need two pieces of information from the diffracted waves: their amplitudes and their phases. While detectors easily measure the amplitudes by recording the intensity of the spots, the phase information—which describes the timing and alignment of the waves—is completely lost during the measurement. This missing data is known as the crystallographic phase problem. Without phases, it is mathematically impossible to directly calculate the positions of the atoms.

Traditionally, the most common method to bypass the phase problem is Molecular Replacement (MR). MR uses a previously solved, structurally similar protein as a search model to estimate the missing phases of the unknown target protein. Software rotates and translates this search model within the unit cell of the new crystal until its theoretical diffraction pattern matches the experimentally measured amplitudes. Once the correct orientation and position are found, the phases calculated from the known search model are combined with the experimental amplitudes of the unknown protein to generate an initial electron density map. Historically, MR failed if a researcher was studying a completely novel protein with no similar structures available in the Protein Data Bank (PDB).

Structure-prediction AI models have eliminated this bottleneck by providing high-accuracy search models for virtually any protein sequence. Using deep learning algorithms trained on vast genomic and structural datasets, tools like AlphaFold and RoseTTAFold predict 3D atomic coordinates directly from a primary amino acid sequence. Instead of relying on the PDB to find a close evolutionary relative, crystallographers can now generate a custom, high-precision search model in silico. These AI predictions are frequently accurate enough to closely mimic the true fold of the target protein, making them ideal templates for MR.

Before an AI-generated model can be successfully used in MR, it requires specific preprocessing based on the AI's internal confidence metrics. Structure-prediction models output a per-residue confidence score, often called the pLDDT (predicted Local Distance Difference Test). Crystallographic software uses these scores to prepare the model in two ways:

• Pruning low-confidence regions: Flexible loops and unstructured tails typically have low confidence scores. Because these regions rarely match the static crystal conformation, they introduce structural noise and can cause the MR search to fail. Software automatically trims these inaccurate regions away, leaving only the rigid, high-confidence core.
• Converting scores to B-factors: In crystallography, B-factors indicate atomic mobility or uncertainty. The AI's confidence scores are mathematically converted into estimated B-factors, which tells the MR software how much weight to give each atom during the phase calculation.

Furthermore, large proteins with multiple domains often exhibit flexibility, meaning the relative orientation of the domains in the AI prediction might differ from the protein's state inside the crystal. To solve this, crystallographers split the AI model into smaller, independent rigid domains based on another AI metric called the Predicted Aligned Error (PAE). The MR software then places each domain into the crystal data separately, assembling the puzzle piece by piece.

Once the processed AI model is correctly placed into the experimental unit cell via MR, its calculated phases are merged with the measured X-ray amplitudes. This generates an initial electron density map. Because the AI model is only an approximation, this map will show differences between the prediction and the actual crystal data. Crystallographers then use automated refinement software to adjust the atomic coordinates to fit the experimental density, rebuild the previously trimmed flexible loops, and locate bound ligands or water molecules. By bridging the gap between sequence and structure, AI models have transformed molecular replacement from a conditional technique into a nearly universal solution for the phase problem, drastically reducing the need for expensive and time-consuming experimental phasing methods.