# How to Integrate Cryo-EM Maps with AlphaFold Models: A Step-by-Step Workflow for Hybrid Structure Determination

Structural biologists frequently face a practical problem: cryo-electron microscopy (cryo-EM) maps at intermediate resolution reveal the overall shape of a macromolecular assembly, but the density is too coarse to trace every amino acid side chain. AlphaFold predictions provide high-confidence atomic coordinates from sequence alone, yet those models do not automatically match the experimental map. This article gives a concrete workflow for combining cryo-EM maps with AlphaFold models using tools such as UCSF ChimeraX and Phenix, with specific attention to map preprocessing, rigid and flexible fitting, refinement, validation, and troubleshooting. The intended reader is a graduate student, postdoctoral researcher, or laboratory professional who has a cryo-EM map and an AlphaFold prediction and needs a reproducible path to a deposited atomic model.

## Scope and Reader Context

The workflow described here applies to single-particle cryo-EM maps in the approximate resolution range of 3.5 to 8 Å, where secondary structure elements are visible but side-chain density may be ambiguous. It also applies to subtomogram averaging maps from cryo-electron tomography in similar resolution ranges. The methods assume you have already performed particle picking, 2D classification, 3D reconstruction, and map sharpening in a package such as RELION or cryoSPARC. This article does not cover the initial reconstruction steps.

The core decision you face is how to place an AlphaFold model into your experimental map. The choice depends on map resolution, model confidence, conformational differences between the predicted state and the experimental state, and whether you are fitting a single chain or a multi-subunit complex. The workflow below gives you a decision sequence, quality checks at each stage, and criteria for escalating to manual intervention or alternative tools.

## At a Glance

The table below summarizes the main workflow stages, the tools commonly used at each stage, the key inputs and outputs, and the quality check that tells you whether to proceed.

| Workflow Stage | Primary Tools | Key Inputs | Output | Quality Check |
| --- | --- | --- | --- | --- |
| Map preprocessing | ChimeraX, Phenix | Raw reconstruction, mask, resolution estimate | Sharpened map, resampled map, map statistics | Map resolution and local resolution distribution |
| Model preparation | AlphaFold database or local prediction, ChimeraX | Target sequence, predicted model | Trimmed model, confidence-filtered regions | pLDDT scores and predicted aligned error |
| Initial placement | ChimeraX, Phenix, GMM-based fitting | Preprocessed map, prepared model | Rigid-body fitted model | Cross-correlation score and visual overlap |
| Flexible refinement | Phenix, ISOLDE, DeepTracer-Refine | Fitted model, map | Refined atomic model | Map-model FSC, Ramachandran statistics, clash score |
| Validation and deposition | Phenix, wwPDB validation | Refined model, map | Validation report, deposited model | MolProbity score, EMRinger score, map-model FSC |

## Core Principles of Hybrid Structure Determination

### Why AlphaFold Models Are Not Directly Usable

AlphaFold predictions are generated from sequence and evolutionary information without reference to your experimental map. The predicted structure may represent a conformational state that differs from the state captured in your cryo-EM experiment. Domain movements, loop rearrangements, ligand-induced changes, and oligomeric assembly interfaces can all differ between the prediction and the experimental density. The published nuclear pore complex studies illustrate this point: the researchers used AlphaFold to predict individual nucleoporin structures because no high-resolution experimental models of the Xenopus laevis proteins were available, then fitted those predictions into cryo-EM maps of the cytoplasmic ring at 6.9 and 6.7 Å resolution [8]. The predictions served as starting coordinates, not as final answers.

A second issue is that AlphaFold confidence varies along the sequence. Regions with high predicted local distance difference test (pLDDT) scores are generally reliable for backbone placement, while low-confidence regions, often loops and termini, may be poorly ordered in the prediction and may not match the map. You need to filter the model before fitting.

### What Cryo-EM Maps Provide at Intermediate Resolution

At 4 to 8 Å resolution, a cryo-EM map shows the path of the polypeptide backbone, the positions of alpha helices as rod-like densities, and beta sheets as flattened densities. Side-chain density is generally not interpretable at the lower end of this range. The map therefore constrains the overall fold and domain arrangement but does not give you atomic-level information. The published EMBuild work evaluated automatic model building on 47 single-particle maps at 4.0 to 8.0 Å and 16 subtomogram averaging maps at 3.7 to 9.3 Å, showing that this resolution range is the practical target for hybrid approaches [9].

The combination works because the two data sources are complementary. The map tells you where the protein is and how the domains are arranged in the complex. The AlphaFold model tells you the likely atomic details of each domain. The hybrid model is built by placing predicted domains into the map density and then refining the atomic coordinates against the map.

### The Role of Deep Learning in the Workflow

Recent methods add a deep learning step to improve the map before fitting or to improve the model after fitting. CryoFEM enhances cryo-EM map features before AlphaFold model fitting, and a benchmark on 36 maps showed that this approach improved Fourier shell correlations between maps and ground truth models [7]. In a subset of 17 datasets where initial AlphaFold predictions were less accurate, the workflow significantly improved model accuracy [7]. DeepTracer-Refine takes a different approach: it refines AlphaFold predicted structures by aligning them to a model built by the DeepTracer deep learning method, and evaluation on 39 multi-domain proteins improved average residue coverage from 78.2 to 90.0 percent and average local Distance Difference Test score from 0.67 to 0.71 [10]. These tools are options within the broader workflow, not replacements for the core fitting and refinement steps.

## Preparing the Cryo-EM Map

### Resolution Assessment and Sharpening

Before fitting any model, you need a map that is optimally processed for interpretation. Start with your reconstruction and its reported global resolution from gold-standard FSC at 0.143. Examine the local resolution distribution, because a map with a global resolution of 5 Å may have regions at 4 Å and other regions at 7 Å. The fitting strategy should prioritize the highest-resolution regions for placing domains with the most conformational constraint.

Sharpening applies a negative B-factor to amplify high-resolution features. The goal is to make secondary structure elements more visible. Over-sharpening amplifies noise and creates artifacts that can mislead fitting. A common practice is to test several sharpening B-factors and visually inspect the map in ChimeraX. The map should show continuous density for helices without excessive noise peaks in solvent regions.

### Map Resampling and Masking

Cryo-EM maps are often reconstructed on a fine grid with a small voxel size, which increases file size and slows fitting calculations. Resampling the map to a coarser grid, typically 1 to 2 Å per voxel, speeds up rigid-body fitting without losing information needed for placement. Keep the original map for final refinement.

Apply a mask to remove noise in the solvent region. A soft mask around the density of interest improves cross-correlation calculations because the correlation is computed only over the region containing the particle. Phenix and ChimeraX both provide tools for mask creation and map editing. The mask should be generous enough to include all ordered density but tight enough to exclude most solvent noise.

### Map Normalization and Statistical Tools

Normalize the map so that the density values have a defined mean and standard deviation. Most fitting programs expect a normalized map. Phenix map tools and ChimeraX both provide normalization options. After normalization, check the map statistics, including the mean, standard deviation, and minimum and maximum values. Extreme outliers often indicate artifacts from sharpening or masking.

## Preparing the AlphaFold Model

### Obtaining the Prediction

You can obtain AlphaFold predictions from the AlphaFold Protein Structure Database or run your own prediction locally. The database version is convenient for well-annotated proteins from common organisms. For proteins from less common organisms or for engineered constructs, local prediction is necessary. The NCBI provides sequence resources and search systems that can help you identify the correct isoform and sequence for your target [1]. Verify that the sequence used for prediction exactly matches the construct used in your cryo-EM experiment, including any tags or mutations.

### Confidence Filtering

AlphaFold outputs a per-residue pLDDT score from 0 to 100. Residues with pLDDT above 90 are predicted with high confidence and are suitable for direct placement. Residues in the 70 to 90 range are generally reliable for backbone placement. Residues below 50 are low confidence and often correspond to disordered loops or termini that may not appear in the cryo-EM map.

Before fitting, trim the model to remove low-confidence regions that are unlikely to match the map. ChimeraX provides commands to select residues by B-factor, since AlphaFold stores pLDDT in the B-factor column of the PDB file. Remove or ignore regions with pLDDT below 50 for the initial fitting. You can revisit these regions later if the map shows density that the prediction missed.

### Domain Identification and Segmentation

For multi-domain proteins, identify the domain boundaries in the AlphaFold model. The predicted aligned error (PAE) matrix shows which residues are predicted to be in the same domain. Residues with low PAE relative to each other are likely in the same rigid domain. Domains are the units you will fit independently into the map. The published nuclear pore complex work used predicted complex structures to guide model fitting when individual subunit interactions were ambiguous [8]. This strategy of predicting complexes, beyond monomers, can resolve ambiguity in domain arrangement.

## Initial Placement of the Model into the Map

### Rigid-Body Fitting with ChimeraX

The first placement step is rigid-body fitting, where the model is moved as a single unit to maximize the correlation with the map. ChimeraX provides a fit-in-map tool that searches for the best placement. For a single domain or a compact monomer, this step often produces a good initial fit.

The procedure is as follows. Open the preprocessed map and the trimmed AlphaFold model in ChimeraX. Use the fitmap command with the model and map specified. The program will optimize the model position and orientation. Examine the resulting cross-correlation score and visually inspect the fit. Secondary structure elements in the model should align with corresponding density features in the map.

### Fitting Multi-Domain Proteins and Complexes

For multi-domain proteins, fit each domain separately. Start with the largest or highest-confidence domain. After fitting the first domain, fix its position and fit the next domain. This sequential approach prevents the search from being confused by multiple similar domains.

For multi-subunit complexes, the process is more involved. The published EMBuild method integrates AlphaFold structure prediction, FFT-based global fitting, domain-based semi-flexible refinement, and graph-based iterative assembly on a main-chain probability map predicted by a deep convolutional network [9]. This automated approach builds multi-chain complexes from intermediate-resolution maps. You can replicate the logic manually: predict each subunit, fit each into the map, then adjust the relative orientations of subunits to satisfy both the map density and the predicted interfaces.

### Using Gaussian Mixture Models for Low-Resolution Maps

When the map resolution is at the lower end of the useful range, around 8 Å or worse, direct cross-correlation fitting can be unreliable. Gaussian mixture model (GMM) fitting represents the map and the model as collections of Gaussian functions and fits the model by matching the Gaussian distributions. This approach is more robust to resolution differences and is implemented in several tools. The published nuclear pore complex work used integrative approaches combining cryo-EM reconstructions with structure prediction because the maps at 5 to 8 Å allowed identification of secondary structural elements but not unambiguous atomic placement [11]. GMM fitting is appropriate in this regime.

## Flexible Refinement Against the Map

### Real-Space Refinement with Phenix

After rigid-body placement, the model needs refinement to improve the fit to the map and to relieve any clashes introduced by the fitting. Phenix real-space refinement optimizes the atomic coordinates against the map while maintaining reasonable geometry. The program uses restraints on bond lengths, bond angles, and dihedral angles, and it can apply secondary structure restraints derived from the AlphaFold model.

The key parameters to set are the map weight, which balances the map-fitting term against the geometry restraints, and the secondary structure restraints. A high map weight forces the model to follow the density, which can distort geometry if the map is noisy. A low map weight preserves the AlphaFold geometry but may leave the model poorly fitted to the map. Start with default weights and adjust based on the validation statistics.

### Domain-Based Semi-Flexible Refinement

For multi-domain proteins, allow each domain to move as a rigid body during refinement while the linkers between domains are flexible. This approach, called domain-based semi-flexible refinement, is a component of the EMBuild workflow [9]. In Phenix, you can define rigid-body groups for each domain and allow the linkers to refine freely. This strategy accommodates conformational changes between the AlphaFold prediction and the experimental structure without distorting the well-predicted domain cores.

### Manual Adjustment with ISOLDE

When automated refinement stalls or produces poor geometry, manual adjustment in ISOLDE is the next step. ISOLDE runs inside ChimeraX and provides interactive, real-space refinement with molecular dynamics. You can drag atoms into density, and the program applies restraints to maintain reasonable geometry. This tool is valuable for fixing register errors, where the model is shifted by one or more residues along the sequence, and for building loops that were missing from the AlphaFold prediction.

Manual adjustment is labor-intensive and should be reserved for regions where automated methods fail. The published literature notes that AlphaFold predictions often require manual refinement, which is labor intensive and time consuming [10]. Budget your time accordingly and prioritize regions with clear density that the automated methods did not fit correctly.

### Deep Learning Refinement Options

DeepTracer-Refine offers an automated alternative to manual refinement. The method aligns AlphaFold predicted structures to a model built by DeepTracer, a deep learning model builder. Evaluation on 39 multi-domain proteins showed improved residue coverage and local Distance Difference Test scores compared to the starting AlphaFold models [10]. The same study compared DeepTracer-Refine with Phenix AlphaFold refinement and found that DeepTracer-Refine performed better when the initial AlphaFold model was less precise and surpassed Phenix in run-time performance [10]. This tool is worth trying when your initial AlphaFold model has low confidence in substantial regions.

## Validation and Quality Assessment

### Map-Model Fourier Shell Correlation

The primary metric for assessing how well the model explains the map is the map-model FSC. This curve compares the model-derived map with the experimental map as a function of spatial frequency. A model that fits well will show high correlation at low resolution and maintain correlation to the resolution limit of the map. Phenix calculates the map-model FSC as part of the validation output. The published CryoFEM work used Fourier Shell Correlations between maps and ground truth models as the benchmark metric, indicating that this is the standard measure of fit quality [7].

### Model Geometry Validation

A refined model must have acceptable geometry. The MolProbity score combines several geometry metrics into a single number, with lower scores indicating better geometry. Ramachandran statistics report the percentage of residues in favored and allowed regions of the backbone torsion angle distribution. The clash score reports the number of steric overlaps per thousand atoms. Phenix and the wwPDB validation server both provide these statistics.

The EMRinger score is specific to cryo-EM models and measures how well the model explains side-chain density. A higher EMRinger score indicates better side-chain placement. This metric is most useful when the map resolution is good enough to show side-chain density, typically better than 4 Å.

### Cross-Validation and Overfitting

Refinement can overfit the model to noise in the map. The standard protection is to refine against a map calculated from half of the data and validate against the other half. If your reconstruction workflow produced half-maps, use them for this purpose. The map-model FSC calculated against the half-map used in refinement will be optimistic. The FSC against the other half-map is a more honest estimate of fit quality.

### Local Resolution and Model Confidence

The model is not uniformly reliable. Regions of the map with lower local resolution will support less detailed model building. When you deposit the model, note the local resolution variation and consider whether the model in low-resolution regions should be interpreted with caution. The published nuclear pore complex work achieved reconstructions at 5 to 8 Å resolution that allowed identification of secondary structural elements, but the authors still needed additional approaches to define packing interactions among components [11]. This example shows that even a successful hybrid workflow leaves some regions underdetermined.

## Records and Measurements

### Documentation for Reproducibility

A hybrid structure determination is reproducible only if you document every step. Keep a laboratory notebook or electronic record that includes the following items. Record the AlphaFold database version or the local prediction software version and the exact sequence used. Record the map preprocessing steps, including the sharpening B-factor, the resampling voxel size, and the mask parameters. Record the fitting tool versions and the parameters used for each fitting run. Record the refinement parameters, including the map weight and the secondary structure restraints. Record the validation statistics at each stage.

The bioinformatics training resources from EMBL-EBI emphasize the importance of structured learning pathways for data-resource training and practical analysis education [2]. The same principle applies to your workflow documentation: a structured record of your analysis steps is a form of training for anyone who needs to reproduce or extend your work.

### Version Control for Scripts and Parameters

If you use command-line tools or scripts for any part of the workflow, store them in a version-controlled repository. The nf-core documentation describes community pipeline standards for reproducible workflow context [5]. While you may not use nf-core pipelines directly, the principle of versioned, documented analysis code applies. The Carpentries lessons provide foundational training in shell, Git, and programming that supports this practice [6]. A script that is not versioned is a script that cannot be reproduced.

### Quality Metrics Log

Create a table that records the key quality metrics at each stage of the workflow. Include the map resolution, the cross-correlation score after rigid-body fitting, the map-model FSC after refinement, the MolProbity score, the Ramachandran favored percentage, the clash score, and the EMRinger score. This log lets you track whether each step improved the model and provides the data you need for the deposition validation report.

## Common Failure Patterns and Troubleshooting

### The Model Does Not Fit the Map at All

If rigid-body fitting produces a placement with very low cross-correlation and no visual overlap between model and density, check the sequence match. A common error is using an AlphaFold model for a different isoform or a construct with a different domain composition than the cryo-EM sample. Verify the sequence identity between the predicted construct and the experimental construct. Also check whether the map is masked correctly. A map that includes substantial noise or a mask that excludes part of the particle will defeat the fitting search.

### The Model Fits but with Register Errors

A register error occurs when the model is shifted by one or more residues along the sequence, so that side chains and secondary structure elements are in the wrong positions. This failure is common in beta sheets and in regions with low sequence complexity. The map density for a beta sheet looks similar regardless of which strand is which, so the fitting algorithm can place the model in the wrong register. Fixing register errors requires careful inspection of side-chain density, which is only possible at resolutions better than about 4 Å. At lower resolution, you may need to accept the ambiguity and note it in the deposition.

### Domains Are Placed in the Wrong Order

For multi-domain proteins, the fitting algorithm may place domains in the correct positions but in the wrong order along the chain. This failure is more likely when the domains are similar in shape and size. The predicted aligned error matrix from AlphaFold can help resolve this problem, because it shows which residues are predicted to be close in space. If the PAE indicates that domain A is adjacent to domain B, the fitting should place them adjacent. The published nuclear pore complex work used predicted complex structures to guide model fitting when subunit interactions were ambiguous [8]. Predicting the complex, beyond the individual domains, provides the inter-domain constraints that resolve ordering ambiguity.

### The Refined Model Has Poor Geometry

If the refined model has a high clash score or poor Ramachandran statistics, the map weight in refinement was likely too high, forcing the model to follow noisy density. Reduce the map weight and refine again. Alternatively, the starting model may have had poor geometry in the low-confidence regions. Trim those regions before refinement and rebuild them manually if density supports them.

### The Map-Model FSC Is Worse Than Expected

If the map-model FSC is much worse than the map resolution suggests it should be, the model may be missing components. A common situation is a complex where some subunits are not predicted by AlphaFold or are not visible in the map. The published EMBuild work addressed this by building multi-chain complexes through iterative assembly [9]. Check whether your map contains density that is not explained by your model. If so, identify the missing component and add it to the model.

## Limitations and Interpretation Boundaries

### Resolution Limits of the Hybrid Approach

The hybrid workflow is most reliable when the map resolution is between 3.5 and 8 Å. At resolutions better than 3.5 Å, you can often build the model directly from the map without AlphaFold assistance, and the AlphaFold model may introduce bias. At resolutions worse than 8 Å, the map does not provide enough constraint to place domains reliably, and the resulting model is largely a prediction placed into a blob of density. The published EMBuild evaluation covered maps at 4.0 to 8.0 Å for single-particle and 3.7 to 9.3 Å for subtomogram averaging, defining the practical envelope for automated building [9].

### Conformational Heterogeneity

A cryo-EM map represents an average of many particles. If the sample is conformationally heterogeneous, the map may show blurred density in the mobile regions. The AlphaFold model represents a single conformation, likely the most stable state predicted from sequence. The hybrid model will not capture the conformational ensemble. The published CryoFEM work noted that model inaccuracy and conformational heterogeneity often necessitate labor-intensive manual model building and fitting into cryo-EM maps [7]. If your map shows clear heterogeneity, consider 3D classification to separate conformers before fitting.

### Model Bias

Using an AlphaFold model as a starting point introduces bias toward the predicted structure. The refinement may not move the model far from the prediction, even if the map supports a different conformation. This bias is most dangerous at intermediate resolution, where the map cannot strongly constrain the atomic details. The published DeepTracer-Refine work addressed this by aligning AlphaFold predictions to a deep learning model built directly from the map, providing an independent check on the AlphaFold model [10]. Consider running an independent model-building method and comparing the results.

### Missing Regions and Disorder

AlphaFold predictions often include regions that are not visible in the cryo-EM map, and cryo-EM maps often include density for regions that AlphaFold predicts as disordered. The published nuclear pore complex work identified additional components of the cytoplasmic ring subunit that were suggested by the EM maps but remained to be identified [11]. This example shows that the map can contain information beyond the prediction. Do not discard unexplained density without investigation.

## Safety and Regulatory Context

### Data Management and Reproducibility Standards

Structural biology is increasingly expected to follow reproducible research standards. The Galaxy Training Network provides accessible workflow training and analysis tutorials with a focus on reproducibility context [3]. The Bioconductor project provides official package, workflow, installation, and reproducible genomic-analysis documentation [4]. While these resources are not specific to cryo-EM, the principles of reproducible analysis apply. Store your raw data, intermediate files, and final models in organized directories. Document software versions. Record parameters. These practices protect you and your collaborators from errors and enable independent verification.

### Deposition Requirements

When you deposit a cryo-EM structure, the wwPDB requires both the map and the model. The validation report generated by the wwPDB includes the map-model FSC, the geometry statistics, and the fit of the model to the map. A model built by the hybrid workflow must meet the same validation standards as a model built entirely from the map. The NCBI provides official descriptions of database resources that support sequence and structure data management [1]. Familiarize yourself with the deposition requirements before you begin the workflow, so that you collect the necessary files and metrics along the way.

### Professional Escalation Criteria

Some problems require expertise beyond the standard workflow. Escalate to a structural biology core facility or a collaborator with cryo-EM modeling expertise when you encounter any of the following situations. The map resolution is better than 3.5 Å but you cannot build side chains into the density. The model has large regions with poor map-model FSC that do not improve after manual adjustment. The map shows density for components that you cannot identify. The conformational heterogeneity is severe enough that classification does not resolve it. The validation statistics do not meet the deposition standards after multiple refinement attempts. These situations indicate that the problem requires specialized experience or additional data collection.

## Decision Framework for Choosing Between Rigid-Body, Flexible, and Deep Learning Fitting Strategies

The existing workflow describes the individual fitting methods but does not give you a structured way to decide which method to use at each stage. This section provides a practical decision framework based on observable map and model characteristics, a record system for tracking fitting decisions, and a troubleshooting method for when the chosen strategy fails.

### Primary Decision Criteria

Before you open any fitting tool, assess three properties of your input data. These properties determine which fitting strategy will succeed.

**Map resolution and local variation.** The global resolution from gold-standard FSC at 0.143 tells you the overall quality, but the local resolution distribution matters more for fitting decisions. A map with a global resolution of 5 Å may have domains at 4 Å and linkers at 7 Å. The published EMBuild evaluation covered single-particle maps at 4.0 to 8.0 Å and subtomogram averaging maps at 3.7 to 9.3 Å, establishing this range as the practical envelope for automated hybrid building [9]. If your map has regions worse than 8 Å, those regions will not support reliable domain placement regardless of the fitting method.

**Model confidence distribution.** Examine the pLDDT distribution along the sequence, beyond the average. A model with an average pLDDT of 85 may have a 50-residue loop with pLDDT below 50 that will not match the map. The published DeepTracer-Refine work improved average residue coverage from 78.2 to 90.0 percent on 39 multi-domain proteins, showing that low-confidence regions can often be recovered with additional methods [10]. If substantial regions of your model have pLDDT below 70, plan for a deep learning refinement step instead of expecting standard flexible refinement to fix the problem.

**Conformational divergence between prediction and experiment.** The AlphaFold prediction represents a single conformational state derived from sequence and evolutionary information. Your cryo-EM map captures the state present in your sample. The published CryoFEM work noted that model inaccuracy and conformational heterogeneity often necessitate labor-intensive manual model building and fitting into cryo-EM maps [7]. If you suspect large domain movements between the predicted and experimental states, rigid-body fitting of individual domains will be necessary before any flexible refinement.

### Strategy Selection Matrix

Use the following matrix to select the initial fitting strategy. The matrix assumes you have already prepared the map and filtered the model as described in the main workflow.

| Map Resolution | Model Confidence | Conformational Divergence | Recommended Strategy |
| --- | --- | --- | --- |
| 3.5 to 5 Å | High pLDDT throughout | Low | Rigid-body fit whole model, then Phenix real-space refinement |
| 3.5 to 5 Å | High pLDDT throughout | High | Fit domains separately, then domain-based semi-flexible refinement |
| 3.5 to 5 Å | Mixed pLDDT | Low or high | Fit high-confidence domains, trim low-confidence regions, consider DeepTracer-Refine |
| 5 to 8 Å | High pLDDT throughout | Low | Rigid-body fit whole model, then flexible refinement with secondary structure restraints |
| 5 to 8 Å | High pLDDT throughout | High | Fit domains separately with GMM fitting, then domain-based semi-flexible refinement |
| 5 to 8 Å | Mixed pLDDT | Low or high | GMM fitting of high-confidence domains, deep learning refinement, manual adjustment for low-confidence regions |
| Worse than 8 Å | Any | Any | GMM fitting only, interpret model as low-resolution placement, consider additional data collection |

The published nuclear pore complex work used maps at 5 to 8 Å resolution that allowed identification of secondary structural elements but not unambiguous atomic placement [11]. The authors needed predicted complex structures to guide model fitting when individual subunit interactions were ambiguous [8]. This example shows that the strategy selection must account for the biological complexity of the assembly, beyond the map resolution.

### Decision Points During the Fitting Process

The initial strategy selection is only the first decision. You need to evaluate the fitting result at each stage and decide whether to proceed, adjust parameters, or switch methods.

**Decision point 1: After rigid-body fitting.** Examine the cross-correlation score and the visual overlap between model and density. If the score is above 0.7 and secondary structure elements align with density features, proceed to flexible refinement. If the score is between 0.5 and 0.7, check whether individual domains are misplaced and refit them separately. If the score is below 0.5, the model may be in the wrong location entirely. Verify the sequence match and the map mask before retrying.

**Decision point 2: After flexible refinement.** Compare the validation statistics before and after refinement. The map-model FSC should improve, and the geometry statistics should remain acceptable. The published CryoFEM work used Fourier Shell Correlations between maps and ground truth models as the benchmark metric, confirming that this is the standard measure of fit quality [7]. If the map-model FSC improved but the MolProbity score worsened significantly, reduce the map weight and refine again. If the geometry is acceptable but the map-model FSC did not improve, the model may be trapped in a local minimum and need manual adjustment.

**Decision point 3: After manual adjustment or deep learning refinement.** Evaluate whether the additional effort produced a meaningful improvement. Compare the residue coverage and the local Distance Difference Test scores before and after the step. The published DeepTracer-Refine evaluation used these metrics to demonstrate improvement [10]. If the improvement is marginal, consider whether the time invested in further refinement is justified by the biological questions you need to answer.

### Record System for Fitting Decisions

Maintain a structured record of every fitting decision and its outcome. This record serves two purposes: it documents the workflow for reproducibility, and it provides a basis for troubleshooting when a strategy fails.

Create a fitting log with the following fields for each fitting attempt. Record the date, the tool and version, the input map and model files, the parameters used, the cross-correlation score, the map-model FSC, the geometry statistics, and the outcome classification. The outcome classification should be one of the following: success, partial success with noted limitations, or failure with suspected cause.

The bioinformatics training resources from EMBL-EBI emphasize the importance of structured learning pathways for data-resource training and practical analysis education [2]. A structured fitting log applies the same principle to your experimental workflow. The nf-core documentation describes community pipeline standards for reproducible workflow context [5]. While you may not use nf-core pipelines directly, the principle of versioned, documented analysis steps applies to your fitting log.

Store the fitting log alongside the version-controlled scripts and parameters described in the main workflow. The Carpentries lessons provide foundational training in shell, Git, and programming that supports this practice [6]. A fitting decision that is not recorded is a decision that cannot be reproduced or defended during peer review.

### Troubleshooting Method for Failed Fitting Strategies

When a fitting strategy fails, use the following systematic method to identify the cause and select the next action. Do not repeat the same strategy with different random seeds or minor parameter changes without first diagnosing the failure.

**Step 1: Classify the failure mode.** Determine whether the failure is a placement failure, a refinement failure, or a validation failure. A placement failure means the model did not end up in the correct position in the map. A refinement failure means the model is in the correct position but the refinement did not improve the fit or produced poor geometry. A validation failure means the model fits the map but does not meet deposition standards.

**Step 2: Check the input assumptions.** For a placement failure, verify the sequence match between the AlphaFold model and the experimental construct. Verify the map mask includes all ordered density and excludes solvent noise. Verify the map normalization is correct. The NCBI provides sequence resources and search systems that can help you confirm the correct isoform and sequence for your target [1].

**Step 3: Check the model preparation.** For a placement or refinement failure, examine the pLDDT distribution and the predicted aligned error matrix. Low-confidence regions may be dragging the fitting into incorrect positions. Trim these regions and retry. The published nuclear pore complex work used predicted complex structures to guide model fitting when individual subunit interactions were ambiguous [8]. If your model has ambiguous domain arrangements, predict the complex structure to resolve the ambiguity.

**Step 4: Check the map quality.** For a refinement failure, examine the local resolution in the region where the model does not fit. If the local resolution is worse than 8 Å, the map does not provide enough constraint for reliable atomic placement. The published EMBuild evaluation covered maps at 4.0 to 8.0 Å for single-particle and 3.7 to 9.3 Å for subtomogram averaging, defining the practical envelope for automated building [9]. If your map has regions outside this envelope, consider whether additional data collection or 3D classification is needed.

**Step 5: Select the next strategy based on the diagnosis.** If the failure is due to low model confidence, use a deep learning refinement tool such as DeepTracer-Refine. The published comparison showed that DeepTracer-Refine performed better than Phenix AlphaFold refinement when the initial AlphaFold model was less precise and surpassed Phenix in run-time performance [10]. If the failure is due to conformational divergence, fit domains separately and use domain-based semi-flexible refinement. If the failure is due to map quality, consider GMM fitting or additional data collection.

**Step 6: Escalate when appropriate.** If two different strategies fail for the same region, escalate to a structural biology core facility or a collaborator with cryo-EM modeling expertise. The published literature notes that AlphaFold predictions often require manual refinement, which is labor intensive and time consuming [10]. Some problems require specialized experience or additional data collection that you cannot perform in your current setup.

### Comparison of Fitting Strategies by Failure Mode

The following comparison helps you select the backup strategy when your initial choice fails.

**Rigid-body fitting failures.** Rigid-body fitting fails when the model has multiple domains that move relative to each other between the predicted and experimental states. The failure mode is a low cross-correlation score with the model placed in a plausible but incorrect position. The backup strategy is to split the model into domains and fit each separately. The published EMBuild workflow uses FFT-based global fitting before domain-based semi-flexible refinement, confirming that global fitting alone is insufficient for multi-domain assemblies [9].

**Flexible refinement failures.** Flexible refinement fails when the map weight is too high, causing geometry distortion, or too low, leaving the model poorly fitted. The failure mode is either poor geometry statistics or a map-model FSC that does not improve. The backup strategy is to adjust the map weight systematically, testing a range of values and recording the validation statistics for each. If no map weight produces acceptable results, the starting model may be too far from the experimental structure, and manual adjustment or deep learning refinement is needed.

**Deep learning refinement failures.** Deep learning refinement fails when the deep learning model builder cannot interpret the map, often because the map resolution is too low or the map has artifacts from over-sharpening. The failure mode is a refined model that does not improve on the starting AlphaFold model. The backup strategy is to return to the map preprocessing stage and generate a new map with different sharpening parameters or a different mask.

### Practical Implementation Steps

Implement the decision framework in your laboratory with the following steps.

First, create a template fitting log that includes all the fields described above. Distribute the template to all members of your group who perform hybrid structure determination. The Galaxy Training Network provides accessible workflow training and analysis tutorials with a focus on reproducibility context [3]. Use these resources to train group members on the documentation standards.

Second, establish a regular review process where fitting logs are examined for patterns. If multiple projects fail at the same stage, the problem is likely systematic, such as a consistent issue with map preprocessing or model preparation. The Bioconductor project provides official package, workflow, installation, and reproducible genomic-analysis documentation [4]. While not specific to cryo-EM, the principle of reviewing analysis workflows for systematic errors applies.

Third, document the decision framework itself in your laboratory protocols. The framework should be a living document that evolves as you gain experience with different map resolutions, model confidence distributions, and biological systems. Record which strategies work for which input characteristics, and update the strategy selection matrix accordingly.

Fourth, use the framework to guide your time allocation. The published literature notes that manual refinement is labor intensive and time consuming [10]. The decision framework helps you avoid spending hours on manual adjustment when a deep learning refinement tool would produce a better result in less time. Conversely, the framework helps you recognize when manual adjustment is necessary and budget your time accordingly.

## Frequently Asked Questions

### What resolution range is suitable for the hybrid cryo-EM and AlphaFold workflow?

The workflow is most reliable for maps in the 3.5 to 8 Å resolution range. At this resolution, secondary structure elements are visible and provide constraints for domain placement, but side-chain density is generally not sufficient for de novo model building. The published EMBuild evaluation covered single-particle maps at 4.0 to 8.0 Å and subtomogram averaging maps at 3.7 to 9.3 Å, demonstrating that this range is the practical target for automated hybrid building [9]. At better than 3.5 Å, direct model building from the map is often possible. At worse than 8 Å, the map provides limited constraint and the model is largely a prediction placed into density.

### How do I decide whether to use rigid-body fitting or flexible refinement first?

Always start with rigid-body fitting to place the model or individual domains into the map. Rigid-body fitting establishes the correct position and orientation of each domain. Flexible refinement then adjusts the atomic coordinates to improve the fit and relieve clashes. Skipping rigid-body fitting and going directly to flexible refinement risks trapping the model in a local minimum where the overall placement is wrong but the local geometry is reasonable. The published EMBuild workflow uses FFT-based global fitting before domain-based semi-flexible refinement, confirming this order of operations [9].

### What does the pLDDT score tell me about which parts of the AlphaFold model to trust?

The pLDDT score, stored in the B-factor column of the AlphaFold PDB file, indicates per-residue confidence. Scores above 90 indicate high confidence and are reliable for direct placement. Scores in the 70 to 90 range are generally reliable for backbone placement. Scores below 50 indicate low confidence, often in disordered loops and termini, and these regions may not match the cryo-EM map. Trim low-confidence regions before fitting and revisit them later if the map shows density. The published DeepTracer-Refine work improved average residue coverage from 78.2 to 90.0 percent by refining AlphaFold predictions, showing that low-coverage regions can often be recovered with additional methods [10].

### How do I fit a multi-subunit complex when I have AlphaFold models for each subunit?

Fit each subunit separately into the map using rigid-body fitting, starting with the largest or highest-confidence subunit. After placing the first subunit, fix its position and fit the next. Use the predicted interfaces from AlphaFold to guide the relative orientations of subunits. If individual subunit predictions leave interface ambiguity, predict the complex structure with AlphaFold and use that prediction to guide fitting. The published nuclear pore complex work used predicted complex structures to guide model fitting when individual subunit interactions were ambiguous [8]. The EMBuild method automates this process using graph-based iterative assembly [9].

### What is the difference between map-model FSC and the gold-standard FSC used for resolution estimation?

The gold-standard FSC is calculated between two half-maps reconstructed from independent halves of the data and is used to estimate the map resolution. The map-model FSC is calculated between a map simulated from your atomic model and the experimental map. The map-model FSC tells you how well your model explains the experimental density as a function of spatial frequency. A good model will show high correlation across the resolution range of the map. The published CryoFEM work used Fourier Shell Correlations between maps and ground truth models as the benchmark metric for model quality [7].

### When should I use a deep learning refinement tool like DeepTracer-Refine instead of Phenix real-space refinement?

Use DeepTracer-Refine when your initial AlphaFold model has substantial low-confidence regions or when Phenix refinement does not improve the model sufficiently. The published comparison showed that DeepTracer-Refine performed better than Phenix AlphaFold refinement when the initial AlphaFold model was less precise and surpassed Phenix in run-time performance [10]. The method improved average residue coverage and local Distance Difference Test scores on 39 multi-domain proteins [10]. If your model has high confidence throughout and Phenix refinement produces good validation statistics, the deep learning refinement step may not be necessary.

### How do I know if my model is overfitted to the map?

The standard protection against overfitting is to refine against one half-map and validate against the other half-map. If your reconstruction workflow produced half-maps, use them for this purpose. The map-model FSC calculated against the half-map used in refinement will be optimistic. The FSC against the other half-map is a more honest estimate. If the model fits the refinement half-map much better than the validation half-map, overfitting has occurred. Reduce the map weight in refinement and consider simplifying the model by removing poorly supported regions.

### What should I do if the map contains density that my AlphaFold model does not explain?

Investigate the unexplained density before discarding it. The density may correspond to a component that was not included in your prediction, such as an additional subunit, a ligand, or a post-translational modification. The published nuclear pore complex work identified additional components of the cytoplasmic ring subunit that were suggested by the EM maps but remained to be identified [11]. Use the map to guide identification of the missing component. If you cannot identify it, note the unexplained density in your deposition and consider whether additional biochemical characterization is needed.

## Related Bioinformatics Guides

- [Multi-Omics Integration: A Practical Workflow for Combining Proteomics, Metabolomics, and Epigenomics Data](/knowledge/bioinformatics/multi-omics-integration-a-practical-workflow-for-combining-proteomics-metabolomics-and-epigenomi)
- [Metabolomics Data Analysis in R: A Practical Workflow](/knowledge/bioinformatics/metabolomics-data-analysis-in-r-a-practical-workflow)
- [Structural and Evolutionary Dynamics of Coronavirus Spike Protein: Integrating Cryo-EM, Molecular Dynamics, and Phylogenetic Surveillance](/knowledge/bioinformatics/structural-evolutionary-dynamics-coronavirus-spike-protein)
- [Whole Slide Image Analysis: A Practical Workflow for Pathologists](/knowledge/bioinformatics/whole-slide-image-analysis-a-practical-workflow-for-pathologists)
- [Protein Language Models in Bioinformatics: A Practical Guide to Selection and Application](/knowledge/bioinformatics/protein-language-models-in-bioinformatics-a-practical-guide-to-selection-and-application)

## References and Further Reading

- [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information.
- [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute.
- [Bioconductor](https://bioconductor.org/). Bioconductor Project.
- [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project.
- [nf-core Documentation](https://nf-co.re/docs). nf-core.
- [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries.
- [Integrating AlphaFold and deep learning for atomistic interpretation of cryo-EM maps.](https://pubmed.ncbi.nlm.nih.gov/37982712). Briefings in bioinformatics, 2023.
- [Structure of cytoplasmic ring of nuclear pore complex by integrative cryo-EM and AlphaFold.](https://pubmed.ncbi.nlm.nih.gov/35679401). Science (New York, N.Y.), 2022.
- [Model building of protein complexes from intermediate-resolution cryo-EM maps with deep learning-guided automatic assembly.](https://pubmed.ncbi.nlm.nih.gov/35831370). Nature communications, 2022.
- [Enhancing cryo-EM structure prediction with DeepTracer and AlphaFold2 integration.](https://pubmed.ncbi.nlm.nih.gov/38609330). Briefings in bioinformatics, 2024.
- [Structure of the cytoplasmic ring of the Xenopus laevis nuclear pore complex.](https://pubmed.ncbi.nlm.nih.gov/35679404). Science (New York, N.Y.), 2022.

> This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.