AlphaFold Protein Structure Prediction Tutorial

If you do protein research, you’ve probably heard the name AlphaFold countless times over the past few years. The AI model that wowed everyone at CASP14 in 2020 and was open-sourced in 2021 can actually predict protein structures with near-experimental accuracy?

But for many people, the experience often stops at 'knowing it’s amazing but not knowing how to use it.' Either they just throw a sequence online, get a 3D structure image, and that’s it; or they want to run it locally but get discouraged by environment setup and parameter tuning.

This AlphaFold protein structure prediction tutorial aims to help you get past that hurdle. We won’t go into vague theory or boast; we won’t pile up the obscure model architectures either. Instead, we cover how to use online tools, how to set it up locally, what new AI tricks there are, how to interpret results, common pitfalls, and even what structure prediction can and cannot do. Whether you’re a bioinformatics newbie or a researcher constantly dealing with structures, there’s something useful for you here.

I. First, let’s get clear: how accurate is AlphaFold really?

AlphaFold Protein Structure Prediction Accuracy

Before talking about how to use it, let's first answer the question everyone cares about: just how accurate are structures predicted by AlphaFold? Can you trust them?

The answer is: in the CASP14 global blind test, AlphaFold2 achieved a median GDT_TS of 92.4 (out of 100) across all targets, covering all difficulty categories. GDT_TS can be roughly understood as the proportion of residues within a certain correct position threshold; according to the CASP community, a GDT around 90 is considered close to experimental methods like X-ray crystallography or cryo-EM. Even for the hardest Free Modeling (FM) targets, the median reached 87.0. However, prediction accuracy heavily depends on the availability of rich homologous sequence information—the deeper the multiple sequence alignment (MSA) and the more homologs found, the higher the confidence usually is.

But that doesn't mean it's "all-powerful." There are several cases where AlphaFold's prediction quality drops noticeably:

· Intrinsically disordered regions (IDRs): regions without a fixed 3D structure; AlphaFold can't predict a definite conformation here and usually shows very low confidence.

· Orientation between domains in multi-domain proteins: each domain might be predicted accurately, but the relative orientation and positioning between domains often have lower confidence.

· Membrane proteins and protein complexes: helix bundles in membrane proteins can be predicted quite well, but they’re still more challenging than globular proteins. Also, note that the monomer version of AlphaFold2 doesn’t predict complexes—protein-protein complexes require AlphaFold-Multimer.

· Orphan proteins / proteins with very few homologs: AlphaFold relies heavily on MSA information, so fewer homologs mean lower prediction accuracy.

· Ligand binding and catalytic state details: AlphaFold predicts a single static conformation, generally close to an apo (unbound) state, but it might sometimes show a holo-like conformation (depending on homolog structures in the training data), so it doesn’t necessarily reflect conformational changes during substrate binding or catalysis.

So the conclusion is: AlphaFold is very powerful, but it’s not perfect. Understanding its limits before using it is far more important than blindly trusting the predictions. This is also the first thing this tutorial wants to make clear.

II. Getting Started: Online Version

If you only occasionally need to predict one or two sequences, or you're just starting out and want to get a feel for it, there's no need to immediately set up a local deployment. There are currently several commonly used online services where you can submit sequences and get prediction results directly.

Option 1: Using AlphaFold DB data

If you're looking for well-studied proteins or proteins with homologous sequences, you can first check AlphaFold DB (AFDB as it's commonly called)—many protein prediction results are already calculated, so you can just download and use them. The database launched in 2021, and after a major expansion in 2022, it now includes over 200 million predicted structures, covering almost all known protein sequences in UniProt (including model organisms, non-model organisms, and many pathogenic microbes). Almost all proteins with a UniProt ID can be found; only the newest sequences not yet included in UniProt might be missing.

Option 2: Using online prediction servers

If your protein isn’t in the database, you can submit a job to an online server. The process for commonly used AlphaFold-like online prediction services is roughly as follows:

1. Enter your protein sequence (FASTA format, note it should be an amino acid sequence, not a nucleic acid sequence)

2. Choose the model version and parameters (options differ by server)

3. Submit the job and wait for the computation to finish

4. Download the result files (structure PDB, confidence files, etc.)

Who are online tools suitable for? Researchers with a small number of sequences, who don’t want to mess with setting up environments, and only need preliminary results. The downsides: limited control, few customizable parameters, restrictions on sequence length and number, and data privacy considerations.

III. Advanced Plan: Local Deployment

If you need batch predictions, have custom parameter requirements, or your data can't be uploaded to the public internet, then local deployment is necessary. This part is a bit more challenging, but once it's set up, it becomes very convenient to use.

Before deployment, make sure your hardware is adequate:

AlphaFold has significant computational resource demands, especially for GPUs. Suggested reference configuration:

· GPU: At least one NVIDIA GPU with 16GB or more of VRAM (A100, A6000, RTX 3090/4090, etc. are all fine). The VRAM mainly determines the maximum sequence length that can be predicted. But speed isn't entirely dependent on the GPU—steps like MSA sequence searches are mostly done on the CPU, and overall speed is influenced by multiple factors. There is indeed a performance difference between GPU cores (for example, A100 is significantly faster than V100), but this is just one factor affecting overall speed.

· CPU and RAM: The official minimum reference is roughly an 8-core CPU and 32GB RAM; to run smoothly and support parallel batch processing, a CPU with 32 cores or more and at least 128GB RAM is preferable.

· Storage: The reduced database is about 550GB, the full version (including BFD) is about 2.2TB; including intermediate files and output results, it’s recommended to have at least 3TB of disk space for the full version.

· System: Mainstream Linux distributions (Ubuntu is the easiest).

Basic deployment steps:

1. Install NVIDIA drivers and CUDA: This is the basic environment for GPU computing. If you use the official Docker image or JAX GPU wheel, CUDA/cuDNN comes packaged, saving you the hassle of manual version matching.

2. Download the AlphaFold code: Get it from the official code repository.

3. Install Python dependencies: It’s recommended to use conda or Docker to isolate the environment and avoid dependency conflicts.

4. Download the genetic and structure template databases: This is the basis for AlphaFold to search for homologous information.

5. Run a test example: Start by running the official test sequences to make sure the whole process works.

6. Submit batch tasks: Once you have the batch script ready, you can start large-scale predictions.

Key points to pay attention to for local deployment:

· Make sure the database is fully and correctly downloaded: Missing databases or incorrect versions are among the most common reasons for abnormal prediction results.

· Sequence length has limits: Because of GPU memory constraints, longer sequences require more memory. AlphaFold v2.3 supports predicting whole proteins with up to around 2700 residues; longer sequences need more GPU memory. If you need to split predictions by domain, you can only model each domain separately based on domain boundaries, but be aware that information about the relative orientation between domains will be lost.

· Don’t just chase the latest version: Different versions make trade-offs in speed, accuracy, and resource requirements, so the best one is the one that suits you.

IV . AI upgrade trick: Unlocking AlphaFold on the MatwingsVenus™ (XiaoWu™) platform

If online tools feel too limiting and local deployment feels too troublesome, there’s a third option—using an all-in-one AI protein research platform to get stronger analysis capabilities without setting up the environment yourself.

The MatwingsVenus™ (XiaoWu™) AI platform already has built-in AlphaFold structure prediction. Users don’t need to set up a complex GPU environment or download hundreds of GBs of databases—they can just log in and submit sequences to predict structures. The platform comes with a database of billions of protein sequences, massive amounts of labeled protein data, over 200 protein design tools, and dozens of expert-tuned skills. After structure prediction, users can continue with downstream analysis directly on the platform without switching between multiple tools.

MatwingsVenus™

Compared to pure structure prediction tools, the real value of MatwingsVenus™ (Xiaowu™) lies in the "complete workflow after prediction." Many people run AlphaFold and stare at a structure image without knowing what to do next. The MatwingsVenus™ (Xiaowu™) platform, on the other hand, doesn’t stop at structure prediction—it integrates the full chain from structure quality assessment, functional site mining, stability and physical property analysis, to intelligent design for directed evolution, making the predicted structures actually "usable" instead of just sitting on your hard drive taking up space.

If your research is more challenging, the platform also offers custom services—AlphaFold is embedded into a complete AI protein R&D workflow. Prediction is just the first step; the true value of AI structure prediction lies in the complete chain of functional interpretation, performance optimization, and experimental validation.

V. How to interpret the results?

After running a prediction, you’ll end up with a bunch of files. The real question is: how do you know if the results are reliable? This is just as much of a skill as running the prediction itself.

1. Check pLDDT first: confidence for each residue

pLDDT (predicted Local Distance Difference Test) is the confidence score AlphaFold gives for each amino acid position, ranging from 0 to 100:

· > 90: very high confidence, the structure model is extremely reliable, usually close to experimental accuracy

· 70–90: confident, the backbone is basically reliable, side-chain details might be off

· 50–70: low confidence, should be interpreted cautiously

· < 50: very low confidence, the region might be intrinsically disordered or the prediction could be unreliable

A simple rule of thumb: if most regions have pLDDT above 70, the overall prediction is fairly trustworthy. If large regions have pLDDT below 50, you shouldn’t take those areas too seriously.

2. Then check PAE: are the relative positions of domains reliable?

PAE (Predicted Aligned Error) predicts the confidence of the relative positions between two residues, mainly useful to check if the orientation between domains in multi-domain proteins is accurate.

How to read a PAE map? In a PAE map, blue/green (low values) indicates high confidence in the relative positions of residues, while red (high values) indicates low confidence. If the PAE map shows clear blocks, it means the structure within each block (usually a domain) is very reliable; if there are red areas between blocks, the relative orientation between domains isn’t highly confident.

3. Make a comprehensive judgment based on biological knowledge

Besides looking at the confidence scores, ask yourself a few questions:

· Does this structure resemble known homologous structures?

· Are the residue positions in the active site/binding pocket reasonable?

· Do key interactions like disulfide bonds and salt bridges meet expectations?

· Does the predicted fold type match the known structures of this protein family?

The numbers are just a reference; in the end, you still have to rely on biological plausibility.

VI. Eight common pitfalls in AlphaFold predictions

8 Common Pitfalls to Avoid

Many people run into pitfalls with AlphaFold, not because the tool is bad, but because they overlook some details. Here are 8 common problems—how many have you encountered?

Pitfall 1: Using DNA sequences directly for prediction

AlphaFold predicts protein structures, so the input must be an amino acid sequence. Submitting DNA or CDS sequences directly will either throw errors or produce meaningless results—you need to first translate the CDS into a protein sequence using a tool like EMBOSS transeq.

Pitfall 2: Believing the structure just because the 3D image looks nice

Just because the structure looks pretty doesn’t mean the prediction is accurate. Always check the pLDDT and PAE first to figure out which regions are reliable and which are not.

Pitfall 3: Treating low-confidence regions as real

Especially loops, termini, and disordered regions—pLDDT can be very low here, and these regions are highly uncertain. Don’t discuss fine interactions as if the predicted structure is experimental data.

Pitfall 4: Assuming "the predicted structure is the only conformation"

Proteins are dynamic in solution and can have multiple conformations. AlphaFold outputs a single static model (by default, 5 ranked models, each a point estimate), usually close to the apo conformation—it doesn’t mean this protein only has one structure, nor does it represent a dynamic average.

Pitfall 5: Ignoring ligands and post-translational modifications

By default, AlphaFold doesn’t model ligands, metal ions, or post-translational modifications (heme and a few covalently bound cofactors may occasionally appear). If you need complexed or modified structures, you have to combine experimental data or additional modeling.

Pitfall 6: Thinking longer sequences are more accurate

Actually, the opposite is true: longer sequences (especially multi-domain proteins) may have lower overall global confidence, though the local confidence within individual domains may still be high. The main issue is the relative orientation between domains (reflected in the PAE).

Pitfall 7: Predict it and forget it, no validation

AlphaFold is a prediction tool, not a gold standard. Important conclusions must be verified with experiments or known structures; you can’t take predicted results as fact.

Pitfall 8: Using AlphaFold for any protein

For highly disordered proteins, membrane protein complexes, or completely novel folds, AlphaFold’s predictions may be limited—you can’t just force it.

VII. Final Summary: Five Things to Remember About AlphaFold

1. First ask, 'Is it accurate?' then ask, 'How do I use it?' Understanding AlphaFold’s limitations is key to using it correctly.

2. Use the online version whenever possible, and only go for local deployment if needed. Not everyone needs to run it locally—just choose whatever suits your needs.

3. Upgrading your AI game isn’t just about prediction—you need to know how to use it. With an integrated platform like MatwingsVenus™ (XiaoWu™), you can directly link structure prediction to functional analysis and directed evolution for real value.

4. pLDDT and PAE are your best friends. Check the confidence of your predictions first, don’t just go straight to looking at the structures.

5. Structure is the starting point, not the end point. The goal of predicting structures is to study function, not to treat the method as the purpose.

There are many ways to learn AlphaFold protein structure prediction, but the most fundamental rule is: tools exist to serve humans. Understanding what it can do, making the most of its strengths, and knowing its limitations is the real way to empower your research. Platforms like MatwingsVenus™ (XiaoWu™) are putting the 'next step' in structure prediction into the hands of every researcher—from predicting structures, to understanding functions, to improving performance. AI is completely reshaping how protein research is done.