# A Beginner's Guide to CASP: How to Interpret Critical Assessment of Protein Structure Prediction Results

CASP, the Critical Assessment of Protein Structure Prediction, is a biennial community experiment that evaluates methods for predicting protein three-dimensional structures from amino acid sequences. For researchers new to structural bioinformatics, CASP results provide the most objective benchmark available for comparing prediction methods, but interpreting those results requires understanding the experiment's format, scoring metrics, and limitations. This article explains what CASP measures, how to read its scoring systems, and how to use CASP outcomes to evaluate whether a prediction method suits your research needs.

## What CASP Measures and Why It Matters

CASP operates as a blind prediction experiment. Organizers collect protein sequences whose experimental structures have been determined but not yet publicly released. Participating groups submit predicted structures for these targets, and independent assessors compare the predictions against the experimentally determined structures once they become available. This blind design prevents groups from tailoring predictions to known answers, which makes CASP the most credible comparison of prediction methods in the field.

The experiment began in 1994 and has run every two years since. Each round introduces new targets, new participating methods, and refinements to the assessment protocols. For a beginner, the most important point is that CASP does not rank methods in a simple winner-takes-all fashion. Instead, it produces a complex dataset of scores across many targets, and understanding those scores is the key to interpreting the results.

CASP results matter for practical decisions in structural biology. If you need a reliable model of a protein for docking studies, mutagenesis planning, or functional annotation, the choice of prediction method directly affects the quality of your downstream analysis. CASP provides the evidence base for that choice, but only if you can read its output correctly.

## The CASP Experiment Format

### Target Selection and Blind Prediction

CASP organizers select target proteins from structures that have been solved experimentally but are not yet in public databases. The targets span a range of difficulty, from proteins with close homologs of known structure to proteins with no detectable structural relatives. This range is deliberate. Easy targets test whether methods can reliably use template information, while hard targets test whether methods can produce useful models without templates.

Participating groups receive the amino acid sequences and submit predictions within a defined time window. The time constraint is important because it simulates real-world conditions where researchers need answers quickly. Some groups submit multiple models per target, and the assessors evaluate each model independently.

### Assessment Categories

CASP evaluates predictions in several categories. The main categories include tertiary structure prediction, which covers the overall three-dimensional fold, and refinement, which tests whether methods can improve initial models toward experimental accuracy. Other categories have included model quality assessment, which asks methods to predict their own errors, and contact prediction, which asks methods to identify which residue pairs are close in space.

Each category has its own scoring protocol. A method that performs well in tertiary structure prediction may not perform well in refinement, and vice versa. When reading CASP results, you need to identify which category matches your intended use case.

### The Role of Human Groups and Servers

CASP distinguishes between human-expert groups and fully automated servers. Human groups may use any combination of software, manual inspection, and expert judgment. Servers must produce predictions without human intervention. This distinction matters for interpretation because server performance reflects the state of automated software, while human group performance reflects the best possible outcome with expert input.

For most laboratory researchers, server performance is the more relevant benchmark because it represents what you can achieve by running software yourself. Human group results may incorporate manual steps that are difficult to reproduce in a typical laboratory setting.

## Core Scoring Metrics in CASP

### GDT-TS: Global Distance Test Total Score

GDT-TS is the most widely cited metric in CASP results. It measures the fraction of residues in the predicted model that fall within certain distance cutoffs of their positions in the experimental structure. The standard cutoffs are 1, 2, 4, and 8 angstroms. The score is the average of the fractions at these four cutoffs, expressed as a percentage.

A GDT-TS score of 100 means the model perfectly matches the experimental structure at all cutoffs. A score of 70 or above generally indicates a model that captures the overall fold correctly. Scores below 50 suggest significant errors in the global topology. GDT-TS is useful because it rewards models that place most residues close to their correct positions, even if some regions are wrong.

When comparing methods, look at the distribution of GDT-TS scores across targets instead of a single average. A method that performs consistently well across many targets is more reliable than one that achieves a high score on a few targets but fails on others.

### lDDT: Local Distance Difference Test

lDDT, or local Distance Difference Test, is a newer metric that evaluates local structural accuracy. It compares the distances between atoms in the predicted model with the corresponding distances in the experimental structure, focusing on local neighborhoods around each residue. lDDT is particularly useful for assessing models that will be used in applications requiring accurate local geometry, such as ligand binding site analysis.

lDDT scores range from 0 to 1, with higher values indicating better local accuracy. The metric is less sensitive to global misalignments than GDT-TS, which makes it complementary. A model can have a high GDT-TS but poor lDDT if the overall fold is correct but local details are wrong.

### TM-Score: Template Modeling Score

TM-score is another global metric that compares model and experimental structure. It normalizes the distance differences by the size of the protein, which makes it less biased toward smaller proteins than GDT-TS. TM-score values above 0.5 generally indicate the same fold, while values below 0.3 suggest unrelated folds.

CASP reports TM-score alongside GDT-TS in many assessments. The two metrics correlate strongly but not perfectly. TM-score is more sensitive to the overall topology, while GDT-TS is more sensitive to the fraction of correctly placed residues.

### RMSD: Root Mean Square Deviation

RMSD is the most familiar metric to structural biologists, but it is less central in CASP assessments. RMSD measures the average distance between corresponding atoms after optimal superposition of the model and experimental structure. The problem with RMSD is that a few poorly predicted regions can inflate the score even when most of the model is correct.

CASP assessors use RMSD as a supplementary metric instead of a primary ranking criterion. When you see RMSD values in CASP papers, interpret them in the context of the protein size and the fraction of residues included in the superposition.

## How to Read CASP Papers and Results Tables

### Identifying the Target Set

The first step in reading a CASP paper is to identify which targets were included in the analysis. CASP papers often report results for subsets of targets, such as only those with no detectable templates or only those with high-confidence experimental structures. The target set determines the generalizability of the conclusions.

Look for the number of targets in each category and the criteria used to classify target difficulty. A method that performs well on easy targets may not perform well on hard targets, and the paper should report both.

### Comparing Methods on Equal Footing

CASP results tables typically show scores for many methods across many targets. When comparing two methods, check whether they submitted predictions for the same set of targets. Methods that skip difficult targets can appear to perform better than they actually do.

Also check whether the comparison uses the same model number for each method. CASP allows multiple submissions per target, and the assessors may report the best model, the first model, or the average across models. The choice affects the ranking.

### Understanding Z-Scores

CASP often reports Z-scores, which normalize each method's performance relative to the distribution of all methods on each target. A Z-score of 1 means the method performed one standard deviation above the mean for that target. Z-scores are useful for comparing methods across targets with different difficulty levels, but they can be misleading if the underlying score distribution is not normal.

When reading Z-score tables, focus on the median Z-score across targets instead of the mean. The median is more robust to outliers and gives a better sense of typical performance.

### Reading the Prediction Accuracy Plots

Many CASP papers include plots showing prediction accuracy as a function of residue position along the sequence. These plots reveal which regions of the protein are predicted well and which are predicted poorly. For a protein of interest, these plots can indicate whether the method is likely to produce a useful model for your specific region of interest.

Look for patterns such as consistently poor prediction in loop regions or consistently accurate prediction in core secondary structure elements. These patterns help you anticipate where a method will succeed and where it will fail.

## Practical Workflow for Using CASP Results

### Step 1: Define Your Structural Question

Before consulting CASP results, define what you need from a protein structure model. Are you planning a docking study that requires accurate global topology? Are you designing mutations that require accurate local geometry around a binding site? Are you annotating function based on fold recognition? Each application has different accuracy requirements.

Write down your specific requirements in terms of the metrics that matter for your application. For example, a docking study might require a GDT-TS above 70 for the global fold, while a mutagenesis study might require high lDDT in the target region.

### Step 2: Identify Relevant CASP Categories

Match your structural question to the CASP assessment categories. If you need a full structure model, focus on tertiary structure prediction results. If you have an initial model that needs improvement, focus on refinement results. If you need to know which parts of a model are reliable, focus on model quality assessment results.

Each category has its own leaderboard and its own set of metrics. Reading the wrong category leads to incorrect conclusions about method performance.

### Step 3: Compare Methods Across Multiple Targets

Do not base your method choice on a single CASP round or a single target. Instead, look at performance across multiple rounds and multiple targets. Methods that consistently rank in the top tier across several CASP rounds are more likely to perform well on your protein.

Create a simple table for yourself that lists candidate methods, their median GDT-TS across recent CASP rounds, their performance on hard targets, and their server availability. This table becomes your decision tool.

### Step 4: Test the Method on Your Own Protein

CASP results provide a benchmark, but they do not guarantee performance on your specific protein. After selecting a method based on CASP evidence, run the method on your protein and evaluate the output using the same metrics you would use for any model.

If you have any experimental information about your protein, such as known secondary structure elements or distance constraints from crosslinking experiments, use that information to validate the model. CASP results tell you what a method can do in general, but only your own validation tells you what it did for your protein.

### Step 5: Document Your Method Choice

Record which method you used, which CASP evidence supported the choice, and how you validated the resulting model. This documentation is important for reproducibility and for justifying your methods in publications.

Include the CASP round, the method version, and the specific metrics you used for evaluation. This information allows others to understand the basis for your structural model.

## At a Glance: CASP Metrics and Their Uses

| Metric | What It Measures | Typical Use | Interpretation Guidance |
|--------|-----------------|-------------|-------------------------|
| GDT-TS | Global fold accuracy across distance cutoffs | Comparing overall model quality | Above 70 indicates correct fold, below 50 suggests topology errors |
| lDDT | Local distance accuracy around each residue | Evaluating models for local geometry applications | Higher values indicate better local accuracy, useful for binding site analysis |
| TM-score | Global topology similarity normalized by protein size | Comparing models of different sizes | Above 0.5 indicates same fold, below 0.3 suggests unrelated folds |
| RMSD | Average atomic distance after superposition | Familiar metric for structural comparison | Sensitive to outliers, interpret with fraction of residues included |
| Z-score | Performance relative to all methods on a target | Comparing methods across targets | Median Z-score is more robust than mean for ranking |

## Options and Tradeoffs in Prediction Methods

### Template-Based Methods

Template-based methods predict structure by identifying homologous proteins with known structures and using those structures as starting points. These methods perform well when a suitable template exists, which is the case for many proteins in well-studied families. The tradeoff is that template-based methods cannot produce accurate models for proteins without detectable homologs.

CASP results consistently show that template-based methods dominate the top rankings for easy targets. For hard targets with no templates, these methods fall back on ab initio approaches and their performance drops accordingly.

### Ab Initio and De Novo Methods

Ab initio methods predict structure from physical principles without using templates. These methods are essential for proteins with no detectable homologs, but they are computationally expensive and less accurate than template-based methods when templates exist.

Recent CASP rounds have shown dramatic improvements in ab initio methods, particularly with the introduction of deep learning approaches. These methods can now produce useful models for some proteins that previously had no structural information.

### Deep Learning Methods

Deep learning methods, particularly those based on attention mechanisms and evolutionary coupling analysis, have transformed the field. These methods learn patterns from large databases of known protein structures and sequences, and they can predict structures with accuracy that was unthinkable a decade ago.

The tradeoff with deep learning methods is interpretability. It is often difficult to understand why a method produces a particular prediction, which complicates troubleshooting when predictions fail. Additionally, these methods require substantial computational resources and may not be accessible to all laboratories.

### Hybrid Approaches

Many top-performing CASP groups use hybrid approaches that combine multiple methods. A typical pipeline might use a deep learning method to generate an initial model, a refinement method to improve local geometry, and a model quality assessment method to identify unreliable regions.

Hybrid approaches often achieve the best overall performance in CASP, but they are more complex to implement and require expertise in multiple tools. For a beginner, starting with a single well-documented method is usually more practical than building a complex pipeline.

## Observations and Measurements in CASP

### What the Scores Tell You About Method Behavior

CASP scores reveal systematic patterns in method behavior. For example, some methods consistently produce models with correct global topology but poor local geometry, while others produce models with accurate local details but occasional global errors. These patterns are visible in the distribution of GDT-TS and lDDT scores across targets.

When you observe a method with high GDT-TS but low lDDT, expect that the model will be useful for fold recognition but not for applications requiring accurate atomic positions. Conversely, a method with moderate GDT-TS but high lDDT may be more useful for local geometry applications even though the global fold is less accurate.

### Target Difficulty and Score Distributions

CASP targets are not equally difficult. Easy targets with clear templates produce high scores for most methods, which compresses the score distribution and makes it hard to distinguish methods. Hard targets produce low scores for most methods, which also compresses the distribution but at a different level.

The most informative comparisons come from medium-difficulty targets where methods show meaningful variation. When reading CASP results, pay attention to how methods perform on this middle range of targets.

### Consistency Across CASP Rounds

Method performance is not always consistent across CASP rounds. A method that ranks first in one round may rank fifth in the next round due to changes in target composition, improvements in competing methods, or updates to the method itself.

When evaluating a method for your work, look at its performance across at least two or three recent CASP rounds. Consistent top-tier performance is a stronger signal than a single outstanding result.

## Records and Documentation for Method Evaluation

### What to Record When Evaluating a Method

Keep a structured record of your method evaluation process. For each candidate method, record the CASP rounds you examined, the metrics you compared, the target sets used in the comparison, and the method version you tested.

Also record the computational requirements of each method, including runtime, memory usage, and whether the method requires specialized hardware such as GPUs. These practical factors matter for deciding whether a method is feasible in your environment.

### Building a Comparison Table

Create a comparison table that includes the following columns for each method: median GDT-TS across recent CASP rounds, median lDDT, performance on hard targets, server availability, computational requirements, and ease of use. This table becomes your reference for method selection.

Update the table after each CASP round. The field changes rapidly, and a method that was not competitive two years ago may now be the best option.

### Reproducibility Considerations

Document the exact version of each method you use, including any parameter settings. Prediction methods are updated frequently, and results can change between versions. For reproducible research, record the software version, the database version used for template searching, and any random seeds if applicable.

If you use a server-based method, record the date of submission and the server version if available. Server methods may change without notice, which complicates reproducibility.

## Common Failure Patterns in CASP Interpretation

### Overinterpreting Small Score Differences

A common mistake is treating small differences in GDT-TS or lDDT as meaningful when they fall within the noise of the assessment. CASP scores vary across targets, and a method that is one point higher on average may not be reliably better.

Look at the distribution of scores across targets instead of the average alone. If the distributions overlap substantially, the methods are likely equivalent in practice.

### Ignoring Target Difficulty

Comparing methods only on easy targets or only on hard targets gives a distorted picture. A method that excels on easy targets may fail on hard targets, and vice versa. Always examine performance across the full range of target difficulty.

### Confusing Correlation with Causation

CASP results show correlations between method features and performance, but they do not prove that a particular feature causes the performance. A method may perform well because of its training data, its architecture, or its post-processing steps, and the CASP results alone cannot distinguish these factors.

### Neglecting Practical Constraints

A method that performs best in CASP may be impractical for your laboratory due to computational requirements, licensing restrictions, or the need for specialized expertise. Always consider practical constraints alongside performance metrics.

### Using Results from a Single CASP Round

Single-round results are noisy and may not reflect long-term method quality. Methods improve, targets change, and the competitive landscape shifts. Use multiple rounds of evidence before making a method choice.

## Limitations of CASP Results

### The Gap Between Benchmark and Real-World Performance

CASP targets are selected for their experimental tractability and may not represent the full diversity of proteins that researchers study. Membrane proteins, intrinsically disordered proteins, and proteins with extensive post-translational modifications are underrepresented in CASP target sets.

Your protein may fall outside the range of CASP targets, which limits the applicability of CASP results to your specific case.

### The Time Constraint Effect

CASP imposes strict time limits on predictions, which may not reflect the time you can invest in a single protein. A method that performs poorly under CASP time constraints may perform better when given more time for careful analysis.

Conversely, a method that performs well under time pressure may not benefit from additional time. The time constraint effect varies by method and target.

### The Assessment Metric Limitations

Each CASP metric captures only part of model quality. GDT-TS rewards global topology but is insensitive to local errors. lDDT rewards local accuracy but may miss global misalignments. No single metric captures all aspects of model quality.

When using CASP results, consider multiple metrics together instead of relying on a single score.

### The Rapid Evolution of Methods

The field of protein structure prediction is evolving rapidly, and CASP results become outdated quickly. A method that was state of the art in the most recent CASP round may be superseded within a year.

Check for updates to methods and for new benchmark results between CASP rounds. The CASP website and the websites of individual methods are the best sources for current information.

## Quality Controls and Validation Practices

### Independent Validation Beyond CASP

CASP provides a benchmark, but it should not be your only validation. For any model you use in your research, perform independent validation using whatever experimental data you have available.

Common validation approaches include checking the model against known secondary structure elements, comparing the model to homologous structures, and evaluating the model's stereochemical quality using tools that check bond lengths, bond angles, and Ramachandran plot outliers.

### Using Model Quality Assessment Tools

Model quality assessment tools predict which regions of a model are reliable and which are likely to be wrong. These tools are themselves evaluated in CASP, and the best-performing tools can help you interpret the reliability of your models.

When a model quality assessment tool identifies a region as unreliable, treat that region with caution in your downstream analysis. Consider whether you need additional experimental data to resolve the structure in that region.

### Cross-Validation with Experimental Data

If you have any experimental data about your protein, use it to cross-validate the model. Crosslinking data, mutagenesis data, and low-resolution electron microscopy data can all provide constraints that test the model's accuracy.

A model that is consistent with multiple independent experimental observations is more trustworthy than a model that is consistent with none.

## Safety and Ethical Context in Structural Biology

### Responsible Use of Structural Models

Structural models produced by prediction methods are computational hypotheses, not experimental facts. Using a model as if it were an experimentally determined structure can lead to incorrect conclusions and wasted experimental effort.

Always label predicted structures clearly in your records and publications. State the method used, the confidence of the prediction, and the limitations of the model.

### Data Management and Reproducibility

Structural biology research depends on reproducible workflows. Store your prediction inputs, method versions, and parameter settings in a way that allows others to reproduce your results.

The broader bioinformatics community emphasizes reproducible analysis through structured workflows and version control. Following these practices for your structure prediction work ensures that your results can be verified and built upon by others.

### Ethical Use of Prediction Methods

Prediction methods should be used to advance scientific understanding, not to misrepresent results. Do not present predicted structures as experimental structures, and do not use prediction methods to support conclusions that the data do not justify.

If you use a prediction method in your research, cite the method and acknowledge its limitations. This transparency allows readers to assess the reliability of your structural claims.

## Professional Escalation Criteria

### When to Seek Expert Help

If you are new to structural bioinformatics, you may encounter situations where CASP results are difficult to interpret or where your model quality is unexpectedly poor. In these situations, consider consulting a colleague with structural biology expertise.

Seek expert help when you need to make a high-stakes decision based on a structural model, such as designing experiments that depend on accurate atomic positions or making claims about protein function based on predicted structure.

### When to Consider Experimental Structure Determination

If your research depends critically on accurate structural information and prediction methods do not provide sufficient accuracy, consider experimental structure determination. Methods such as X-ray crystallography, cryo-electron microscopy, and NMR spectroscopy can provide experimentally determined structures, though they require substantial resources and expertise.

The decision to pursue experimental structure determination should be based on the accuracy requirements of your research question and the resources available to you.

### When to Update Your Method Choices

The field of protein structure prediction changes rapidly. If you have not evaluated new methods in more than a year, your method choices may be outdated.

Set a regular schedule for reviewing CASP results and testing new methods. The biennial CASP rounds provide a natural schedule for this review.

## Building a Practical Decision Framework for CASP-Based Method Selection

### Establishing Your Accuracy Thresholds Before Comparing Methods

The most common error new researchers make with CASP data is browsing results without predefined criteria. Before you open any CASP table, write down the minimum accuracy your application requires. This step converts CASP interpretation from an abstract exercise into a concrete selection process.

For global fold applications such as domain annotation or fold recognition, set a GDT-TS threshold. For local geometry applications such as binding site analysis or mutation effect prediction, set an lDDT threshold for the specific region of interest. For applications requiring both, define both thresholds and decide whether you need both satisfied simultaneously or whether one takes priority.

Record these thresholds in a lab notebook or electronic document. When you later compare methods, you will evaluate each candidate against your stated requirements instead of against an abstract notion of best performance. This approach prevents the common failure of choosing a method because it ranks highly overall when a lower-ranked method actually meets your specific accuracy needs.

### Creating a Weighted Scoring Matrix for Method Comparison

A weighted scoring matrix transforms qualitative CASP observations into a quantitative decision tool. Start by listing the criteria that matter for your application. Typical criteria include median GDT-TS across recent rounds, median lDDT, performance on hard targets, server availability, computational cost, ease of use, and documentation quality.

Assign each criterion a weight that reflects its importance to your specific project. For example, if you work in a laboratory without GPU access, computational cost might receive a weight of 30 percent while median GDT-TS receives 25 percent. If you have access to high-performance computing, reverse those weights.

For each candidate method, assign a score from 1 to 5 for each criterion based on the CASP evidence you have gathered. Multiply each score by the criterion weight and sum the results. The method with the highest total is your top candidate, but the matrix also shows you why it ranks highest and which tradeoffs you are accepting.

Update this matrix after each CASP round and whenever a method releases a significant update. The matrix becomes a living document that tracks your reasoning and makes your method selection transparent and reproducible.

### Building a Target Difficulty Stratification Table

CASP results become more informative when you stratify targets by difficulty. Create a table with rows for easy, medium, and hard targets and columns for each candidate method. For each cell, record the median GDT-TS or lDDT for that method on that target subset.

This stratification reveals patterns that aggregate scores hide. A method may show excellent median performance because it excels on easy targets while failing on hard ones. Another method may show moderate overall performance but consistent results across all difficulty levels. For most research applications, the consistent method is more reliable because your protein may not resemble the easy targets in the CASP set.

To build this table, you need access to the target classification information in CASP papers or on the CASP website. The organizers classify targets by the availability of templates and by the difficulty of the prediction task. Use these classifications to group targets instead of creating your own subjective categories.

### Implementing a Two-Stage Validation Protocol

A two-stage validation protocol separates method selection from model validation. In the first stage, use CASP results to narrow your candidate methods to two or three options. In the second stage, run those methods on your protein and evaluate the outputs using your predefined thresholds.

The first stage should take hours, not days. You are looking for methods that meet your accuracy thresholds, are practically accessible, and have shown consistent performance across multiple CASP rounds. Do not overanalyze small differences between top methods at this stage because those differences may not persist on your protein.

The second stage requires running the actual software or servers. For each candidate method, record the runtime, the memory usage, and the output format. Evaluate the resulting models using the same metrics you would use for any structural model. If you have experimental data such as known secondary structure elements or crosslinking constraints, check the model against those data.

This two-stage approach prevents the common failure of selecting a method based solely on CASP rankings and then discovering that the method is impractical or produces poor results on your specific protein.

### Recording Method Version and Parameter Information

CASP results apply to specific method versions. A method that performed well in CASP14 may have changed substantially by the time you read this article. When you record CASP evidence for a method, note the CASP round and the method version that participated.

When you run a method yourself, record the exact version, the parameter settings, the database versions used for template searching, and the date of the run. This documentation serves two purposes. First, it allows others to reproduce your results. Second, it allows you to diagnose problems if the method behaves differently than expected.

For server-based methods, record the submission date and any version information provided by the server. Server methods may update without notice, which means your results may not be reproducible if you do not document when you ran the prediction.

### Establishing a Regular Review Schedule

The protein structure prediction field changes rapidly. A method that was state of the art at the last CASP round may be superseded within months. Establish a regular review schedule that aligns with the biennial CASP cycle.

At minimum, review your method choices after each CASP round. Check whether the methods you use participated, how they performed relative to previous rounds, and whether new methods have entered the top tier. Also check the websites of the methods you use for version updates and new benchmark results between CASP rounds.

During your review, ask whether your accuracy thresholds are still appropriate for your application. As methods improve, you may be able to raise your thresholds and achieve better models. Conversely, if your application has changed, you may need to adjust which metrics matter most.

### Troubleshooting When CASP Results Do Not Match Your Experience

When a method that performs well in CASP produces poor results on your protein, work through a systematic troubleshooting process before abandoning the method. First, verify that you used the correct method version and parameter settings. Second, check whether your protein falls into a category that is underrepresented in CASP targets, such as membrane proteins or proteins with extensive disordered regions.

Third, compare your protein to the CASP targets on which the method performed well. If your protein is substantially larger, more hydrophobic, or has a different domain architecture than the targets where the method excelled, the performance difference may be explained by target composition instead of method failure.

Fourth, check whether your evaluation metrics match the metrics used in CASP. If you are evaluating your model with a different metric or a different implementation of the same metric, the scores may not be directly comparable.

If the method still performs poorly after this troubleshooting, consider whether your accuracy thresholds are realistic for your protein class. Some proteins are genuinely harder to predict than others, and no method will meet high thresholds for every target.

### Documenting Your Decision Rationale

When you select a method for your research, document the reasoning behind your choice. Record the CASP rounds you examined, the metrics you compared, the thresholds you set, and the practical constraints you considered. This documentation serves multiple purposes.

For your own research, the documentation helps you revisit your decision when new methods become available or when your application changes. For publications, the documentation provides transparency about why you chose a particular method. For reproducibility, the documentation allows others to understand the basis for your structural models.

A simple template for this documentation includes the following fields: application type, accuracy thresholds, candidate methods considered, CASP evidence examined, practical constraints, final method choice, and date of decision. Store this template in a shared laboratory document so that collaborators can see the basis for method choices.

### Comparing CASP Performance with Independent Benchmarks

CASP is the most rigorous benchmark in the field, but it is not the only source of evidence. Some methods publish additional benchmarks on specific protein classes or application types. These independent benchmarks can supplement CASP results, particularly when your protein falls outside the typical CASP target distribution.

When using independent benchmarks, check whether they used a blind prediction format or whether the test structures were available during method development. Benchmarks where the test structures were known during development may overestimate real-world performance. Also check whether the benchmark used the same metrics as CASP or different metrics that may not be directly comparable.

Use independent benchmarks to confirm CASP findings instead of to replace them. If an independent benchmark contradicts CASP results, investigate the source of the discrepancy before drawing conclusions about method quality.

### Escalating to Expert Consultation

If your method selection involves high-stakes decisions, such as choosing a method for a drug discovery project or for a publication that makes strong structural claims, consider consulting a structural bioinformatics expert. Experts can help you interpret CASP results in the context of your specific protein and application.

Seek expert consultation when your troubleshooting does not resolve discrepancies between CASP results and your experience, when your protein falls into a class that is poorly represented in CASP targets, or when you need to justify your method choice to reviewers or funders. An expert can also help you identify methods that are not in CASP but may be appropriate for your application.

The CASP website maintains lists of participating groups and their affiliations. Many of these groups are responsive to inquiries about their methods and can provide guidance on appropriate use cases. Professional societies in structural biology and bioinformatics also maintain directories of experts who may be available for consultation.

## Frequently Asked Questions

### What does a GDT-TS score of 70 mean for my model?

A GDT-TS score of 70 indicates that, on average, 70 percent of residues in your model fall within the distance cutoffs used in the calculation. In practice, this score generally means the model captures the correct overall fold, though local regions may contain errors. For applications that require accurate global topology, such as fold recognition or domain annotation, a GDT-TS above 70 is often sufficient. For applications that require accurate atomic positions, such as docking or drug design, you need additional validation of local accuracy using metrics like lDDT.

### How is CASP different from other structure prediction benchmarks?

CASP uses a blind prediction format where the experimental structures are not available to participants at the time of prediction. This design prevents methods from being tailored to known answers and provides a more realistic assessment of prediction performance. Other benchmarks may use structures that are already in public databases, which allows methods to benefit from information that would not be available in a real prediction scenario. The blind format is the key feature that makes CASP results credible for comparing methods.

### Should I use the top-ranked CASP method for my protein?

The top-ranked CASP method is a reasonable starting point, but it is not automatically the best choice for your specific protein. Consider the target difficulty of your protein, the computational resources available to you, and the specific accuracy requirements of your application. A method that ranks second or third in CASP may be more practical for your situation if it is easier to use or requires fewer computational resources. Test the method on your protein and validate the output before committing to it.

### How often are CASP results updated?

CASP runs every two years, and each round produces a new set of results. Between rounds, individual methods may release updated versions with improved performance. The CASP website and the websites of individual methods are the best sources for the most current information. If you are making a method choice, check whether the method has been updated since the last CASP round and whether the update has been benchmarked.

### Can I use CASP results to evaluate a method that was not in CASP?

CASP results only evaluate methods that participated in the experiment. If a method did not participate, you cannot directly compare it to CASP participants using CASP data. You can, however, run the method on CASP targets yourself and compare its performance to the published CASP results. This approach requires access to the CASP target sequences and the experimental structures, which are available after the assessment is complete.

### What is the difference between GDT-TS and lDDT?

GDT-TS measures global fold accuracy by calculating the fraction of residues within distance cutoffs of their experimental positions. lDDT measures local accuracy by comparing distances between atoms in local neighborhoods. GDT-TS is more sensitive to the overall topology of the model, while lDDT is more sensitive to local geometric accuracy. A model can score well on one metric and poorly on the other, so you should consider both when evaluating a prediction.

### How do I know if my model is good enough for docking studies?

The accuracy required for docking studies depends on the specific docking application. For protein-protein docking, a model with correct global topology and accurate surface geometry is essential. For small molecule docking, accurate local geometry around the binding site is critical. Use lDDT to evaluate local accuracy in the regions that matter for your docking study, and use model quality assessment tools to identify unreliable regions. If the model is unreliable in the regions you need, consider experimental structure determination or additional modeling.

### What should I do if my predicted structure does not match my experimental data?

If your predicted structure conflicts with experimental data, first check whether the experimental data are correctly interpreted. Then check whether the prediction method was appropriate for your protein. If the conflict persists, consider whether the protein undergoes conformational changes that the prediction method did not capture. In some cases, the protein may adopt multiple conformations, and the predicted structure may represent one conformation while your experimental data reflect another. Consult with a structural biology expert if the conflict affects your research conclusions.

## Related Bioinformatics Guides

- [How To Use Alphafold To Predict Structure: Structural Analysis and Computational Methodologies in Bioinformatics](/knowledge/bioinformatics/how-to-use-alphafold-to-predict-structure)
- [Lipidomic Analysis: A Beginner's Guide to Workflows and Data Interpretation](/knowledge/bioinformatics/lipidomic-analysis-a-beginner-s-guide-to-workflows-and-data-interpretation)
- [How to Interpret Gene Set Enrichment Analysis Results](/knowledge/bioinformatics/how-to-interpret-gene-set-enrichment-analysis-results)
- [Conformational Sampling Algorithms in Protein Structure Prediction](/knowledge/bioinformatics/conformational-sampling-algorithms-in-protein-structure-prediction)
- [AlphaFold and Beyond: Deep Learning for Protein Structure Prediction in Veterinary Virology](/knowledge/bioinformatics/alphafold-deep-learning-protein-structure-prediction-veterinary-virology)

## 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.
- [Language barriers between nurses and patients: A scoping review.](https://pubmed.ncbi.nlm.nih.gov/32994104). Patient education and counseling, 2021.
- [Interpreter services and effect on healthcare - a systematic review of the impact of different types of interpreters on patient outcome.](https://pubmed.ncbi.nlm.nih.gov/36816444). Journal of migration and health, 2023.
- [Women's psychological experiences of physiological childbirth: a meta-synthesis.](https://pubmed.ncbi.nlm.nih.gov/30341110). BMJ open, 2018.
- [Suicide care from the nursing perspective: A meta-synthesis of qualitative studies.](https://pubmed.ncbi.nlm.nih.gov/33591582). Journal of advanced nursing, 2021.
- [A comparative analysis of three online appraisal instruments' ability to assess validity in qualitative research.](https://pubmed.ncbi.nlm.nih.gov/20671302). Qualitative health research, 2010.

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