⚠️
Read this chapter even if you're a beginner One of the most common misuses of AlphaFold is treating its predictions as ground truth without understanding where they break down. This chapter is as important as any technical explanation.
Limitation 1

Static snapshots: proteins are not statues

AlphaFold predicts one conformation of a protein — a single static 3-D structure. But real proteins in living cells are not static. They are dynamic, flexible molecules that constantly vibrate, breathe, and change shape.

This isn't a minor detail — protein dynamics are often central to function:

  • Allosteric regulation: Many proteins switch between active and inactive conformations when a regulatory molecule binds. AlphaFold gives you one conformation, not the conformational landscape.
  • Induced fit: Enzymes often change shape significantly when substrate binds. The drug-bound form of a protein can look quite different from the unbound form.
  • Domain motions: Large conformational changes — like the opening and closing of an ion channel pore — are not captured.
💡
Analogy: a photo vs. a video AlphaFold gives you a photograph of a protein. Real biology needs a video. A photo of a hummingbird's wings shows them in one position — but understanding how the bird flies requires seeing the full wing-beat cycle. Similarly, understanding enzyme catalysis, signalling, or transport often requires seeing the protein's conformational trajectory, not just one snapshot.

What to do instead: For dynamic questions, complement AlphaFold structures with molecular dynamics (MD) simulations (see Part 9), hydrogen-deuterium exchange mass spectrometry (HDX-MS), NMR relaxation experiments, or single-molecule FRET measurements.

Limitation 2

Intrinsically disordered regions: real disorder, not missing data

About 30–40% of the human proteome contains regions that are genuinely, functionally disordered — they don't adopt a stable 3-D structure under physiological conditions. These are called Intrinsically Disordered Regions (IDRs) or Intrinsically Disordered Proteins (IDPs).

IDRs are not "broken" or "incomplete" proteins. They are deliberately flexible, and that flexibility is often their function:

  • Flexible linkers: Allow two domains to move relative to each other.
  • Molecular recognition: IDRs often fold into a defined structure only when they bind a partner protein — a phenomenon called "coupled folding and binding."
  • Signalling hubs: Disordered regions often carry multiple short linear motifs (SLiMs) that can interact with many different partners — making them ideal for regulatory networks.
  • Phase separation: Many IDRs drive liquid-liquid phase separation — the formation of membrane-less compartments (like stress granules and the nucleolus) inside cells.
🚫
The critical mistake AlphaFold will predict a 3-D structure for IDRs, but with low pLDDT. The mistake is assuming that because AlphaFold shows a looping chain in a specific conformation, that conformation exists in the real protein. It doesn't — the real IDR samples many conformations in rapid exchange. The predicted coordinates are essentially a guess and should not be used for structural biology of those regions.

How to identify IDRs: Tools like IUPred3, PONDR, or ESpritz predict disorder propensity from sequence. Cross-reference with AlphaFold's pLDDT — regions that both tools flag as disordered are very likely genuinely disordered.

Limitation 3

Hallucination: confident but wrong

The most insidious limitation: AlphaFold can occasionally produce high-confidence predictions that are simply wrong — predicting a plausible-looking structure that doesn't reflect reality. This is sometimes called "hallucination" by analogy with large language models that confidently make up facts.

This happens for several reasons:

  • Sparse MSA: If your protein has few evolutionary relatives (a "shallow" MSA), AlphaFold has little co-evolutionary signal to work with. The model may still produce confident predictions based on general protein-folding patterns it has learned, but these are less reliable. Orphan proteins (those unique to one organism) are especially risky.
  • Novel topology: If your protein has a fold unlike anything in the training set, the model is extrapolating beyond its experience.
  • Long disordered regions interfering: Large low-complexity regions can sometimes confuse the model's representation, causing it to generate structured-looking but spurious predictions for adjacent regions.
⚠️
Hallucination is rare but real For mainstream globular proteins with good MSAs, hallucination is uncommon. It becomes more likely for: proteins with very few homologues, proteins with novel folds, membrane proteins, and proteins in unusual organisms. Always validate critical predictions experimentally.

Mitigation: Check the MSA depth AlphaFold used. Check pLDDT AND PAE — hallucinated structures sometimes have suspiciously uniform high confidence. Use multiple prediction methods (ESMFold, RoseTTAFold) and see if they agree. Disagreement between methods is a red flag.

Limitation 4

No energetics: AlphaFold can't tell you stability or binding affinity

AlphaFold predicts where atoms will be. It does not predict:

  • Thermodynamic stability: How stable the protein is. Will it unfold at 37°C? Will mutation X make it more or less stable?
  • Binding affinity: How tightly a drug, another protein, DNA, or RNA binds. AlphaFold can predict the structure of a complex, but not the Kd or ΔG of binding.
  • Catalytic rates: How fast an enzyme catalyses its reaction.
  • Folding kinetics: How fast the protein folds, whether there are important intermediates, or what fraction folds correctly vs. misfolds in the cell.
🔬
What tools address these gaps? Thermodynamic stability: FoldX, Rosetta ddG, or experimental differential scanning fluorimetry (DSF). Binding affinity: Rosetta InterfaceAnalyzer, FEP+ (free energy perturbation), or experimental surface plasmon resonance (SPR) or isothermal titration calorimetry (ITC). Enzyme kinetics: Experimental Michaelis-Menten measurements, there's no robust computational shortcut.
Limitation 5

Rare folds and context-dependence

AlphaFold's training data (the PDB) is biased toward easily crystallisable, soluble, globular proteins. Certain protein classes are systematically harder:

Membrane proteins

Integral membrane proteins reside in lipid bilayers. AlphaFold often predicts reasonable backbone folds but side-chain orientations within the membrane can be unreliable, and the membrane environment matters enormously.

Repeat proteins

Proteins built from repeating structural units (ankyrin repeats, HEAT repeats, leucine-rich repeats) can confuse the MSA-based approach due to internal sequence similarity.

Novel folds

If a protein has a fold type not represented in the PDB training data, AlphaFold may force it into the nearest familiar fold. Very rare — most folds are already in the PDB — but not impossible.

Context-dependent structures

Some proteins fold differently depending on post-translational modifications (phosphorylation, glycosylation), binding partners, or cellular localisation. AlphaFold ignores all of these.

What AlphaFold doesn't know

The bigger picture: what's missing

AlphaFold predicts structure from sequence alone. It has no information about:

  • Post-translational modifications (PTMs): Phosphorylation, glycosylation, ubiquitination, acetylation, methylation, and dozens of other chemical modifications that alter protein behaviour after synthesis. These can dramatically change structure and function.
  • Cellular context: The crowded, pH-controlled, ionic-strength-specific environment inside a cell, the proximity of ribosomes, chaperones, and binding partners at the moment of folding.
  • Ligand-induced changes: How binding to a cofactor (like a metal ion or coenzyme) changes the structure — though AlphaFold 3 partially addresses this for some ligands.
  • Disease-associated mutations: AlphaFold predicts the wild-type sequence structure. For mutant proteins (relevant in cancer, genetic disease), you need to re-run the prediction with the mutated sequence and interpret carefully.
The positive view

When AlphaFold excels

To balance the limitations, here's where AlphaFold is genuinely transformative:

Globular, soluble proteins

With a good MSA, AlphaFold is often nearly as accurate as experimental structures. Ideal for initial structural characterisation.

Seeding further experiments

Even a rough structure is invaluable for planning crystallisation strategies, designing mutation experiments, or identifying binding pockets.

Comparing many related proteins

Screen entire protein families, identify conserved structural features, spot outliers — at a scale impossible with experimental methods.

Hypothesis generation

Even imperfect predictions can suggest hypotheses worth testing experimentally. The goal isn't perfection but a productive starting point.

Key points from this chapter

  • AlphaFold gives one static conformation — proteins are dynamic, and dynamics often matter for function.
  • Intrinsically disordered regions (IDRs) are real biology — they genuinely lack stable structure, and their low pLDDT reflects this. Don't trust their coordinates.
  • Hallucination (confident but wrong predictions) is rare but real, especially for proteins with sparse MSAs or novel folds. Cross-validate with other methods.
  • AlphaFold predicts structure, not thermodynamic stability, binding affinity, or catalytic rates. Use dedicated tools for these.
  • AlphaFold has no information about PTMs, cellular context, or ligand-induced conformational changes.
  • Despite limitations, AlphaFold excels at globular proteins, seeding experiments, and large-scale comparative analysis.