Protein Structure Prediction Explained: From Amino Acid Sequence to AlphaFold

Quick answer: Protein structure prediction is the task of computing a protein’s three-dimensional shape from its one-dimensional amino acid sequence, because experimentally solving structure (X-ray crystallography, cryo-EM, NMR) is expensive and sometimes impossible. For decades the problem resisted physics-only and machine-learning attempts alike; AlphaFold2 (2020) crossed a qualitative threshold by predicting most structures to near-experimental accuracy, and the AlphaFold Protein Structure Database now ships precomputed models for over 200 million proteins. Using these predictions well means knowing what they are good for — and where they still fail.

The natural pairing with the sequence-alignment and BLAST posts on this blog: alignment finds relatives; structure prediction tells you what their shared fold looks like in space.

Why structure matters when the sequence already exists

A protein’s function is its shape: the pocket that binds a ligand, the groove that reads DNA, the interface that latches onto a partner. Sequence alone hides much of this, because two proteins can share barely 20% sequence identity while folding the same way. Experimental structures are the gold standard but are rate-limiting:

• X-ray crystallography needs crystals — some proteins, especially flexible membrane proteins, resist crystallization for years.
• Cryo-EM has transformed what is achievable for large complexes but still needs biochemical prep work per target.
• NMR covers smaller proteins and needs labeled sample.

The result is a persistent gap: sequences exist for hundreds of millions of proteins (from genome annotation), experimental structures for only a few hundred thousand. Structure prediction is the bridge across that gap.

The pre-AlphaFold landscape, in three strategies

For context, what people did before 2020 is worth skimming because the vocabulary is still everywhere:

• Homology modeling. If a relative with a solved structure exists, thread your sequence onto it and adjust loops and side chains. Tools: MODELLER, Swiss-Model. Accuracy tracks the sequence identity to the template — above roughly 30–40% identity, models are usually broadly correct.
• Fold recognition / threading. When no close relative is known, score your sequence against a library of known folds to detect remote similarity that sequence search missed (HHpred, Phyre2, RaptorX-era tools).
• Ab initio / physics-based folding. From first principles or fragment assembly (Rosetta), attempting to fold the chain without a template. Heroic, computationally expensive, and reliable only for small proteins.

The critical insight Alphafold built on: the information about what a fold is sits largely in evolutionary statistics — a Multiple Sequence Alignment of thousands of relatives, in which columns that co-vary reveal amino acids touching in 3D; such residue-pair couplings had long been shown good enough at reconstructing the fold topology from MSAs alone. AlphaFold’s leap was to combine MSA statistics with neural networks that learned how this geometry arises, plus a module that reasons a priori about chain geometry and symmetry.

How AlphaFold2 works, in one skim

A full treatment is a review paper. The working vocabulary a beginner needs:

• Input: a query sequence, and the MSA built from database searches (UniRef, BFD/MGnify). Optionally, known structures of homologous templates.
• Evoformer: the model’s repeated reasoning block, jointly processing MSA columns (what residues co-vary) and pair representations (which residues are near each other), exchanging information between the two computes.
• Structure module: predicts 3D coordinates from the pair representation, respecting backbone geometry, and iterating several times from coarse to fine.
• Confidence per residue: the model outputs pLDDT (0–100) per residue — how sure it is of each part of the chain. High pLDDT (>90) means near-experimental faith; low pLDDT (<50) usually means a genuinely disordered region, not an error in a fixed fold.
• AlphaFold-Multimer extends the same machinery to complexes of several chains, which matters because most biological function occurs in interactions.

Two properties of the whole thing matter in practice: predictions are fast enough to be web-based (AlphaFold Server returns a complex prediction in minutes), and a huge precomputed database means often you look up the structure instead of computing it.

What the output looks like, and how to read it honestly

A predicted structure is a PDB-format coordinate file plus a per-residue confidence map. The discipline of reading it:

• Check pLDDT first. Ribbon views color-coded by pLDDT (blue very high, orange/red low) show you what the model actually claims. High-confidence secondary-structure elements are usually right; low-confidence loops, termini, and domains may be arbitrary.
• Interpret low confidence as biology, not failure. Many low-confidence regions are intrinsically disordered — they have no single stable fold for the model to find. That is information, not noise.
• Predicted aligned error (PAE) tells you about relative positions. pLDDT grades local geometry; the PAE matrix says how good the model thinks the relative positions of two regions are — the number that decides whether you trust a model’s implication that domains A and B face each other.
• Structural quality is not functional proof. A confident model can still be missing ligands, cofactors, and partner proteins — so docking or mutagenesis predictions on a confident model should be described as hypothesis-generating, not confirmed.

Where AlphaFold surprisingly does not help

• Conformational change. AlphaFold mostly represents one state of a protein; proteins that switch between active and inactive states, or that fold differently on binding a partner, may need ensemble approaches or multiple predictions with different templates/MSAs.
• Ligands, cofactors, and metals. The structure module focuses on protein atoms; small molecules and metal coordination are weakly supported in the original AF2 and only partially in the successor model, so metal-site interpretation still requires ported expert judgment.
• Very long, disordered, or repeat-rich proteins. Tandem-repeat and unfolded stretches can dominate the model’s attention and erode confidence in the globally useful parts.
• Single-domain outliers with no homologs. Proteins with no MSA depth — orphan categories, fast-evolving parasite genes — sit in an MSA-thin regime, and here accuracy is strictly worse; this is the regime where older physics-based attempts were supposed to shine, and the MSA-poor regime is a live research frontier (a nice place to see single-chain approaches like ESMFold and language-model-based prediction).
• Alternative splicing and point mutations. The AlphaFold database models reference proteome sequences; a mutant or isoform is a new computation, with its accuracy not guaranteed and usually evaluated as a difference from reference rather than absolute validity.

What the prediction revolution changed in practice

• Annotating unknown genes. A confident predicted fold often reveals a family membership that sequence search BLAST missed, sharpening what you can infer about gene function from genome annotation pipelines.
• Designing wet-lab experiments by looking at where residues sit — which residues form a pocket, which are buried, which sit at a dimer interface — instead of guessing from a linear sequence.
• Interpreting variants: mapping a patient variant on a predicted structure plausibly positions it (buried core vs surface) and suggests mechanistic hypotheses, with the caveat that predicting the effect of any single amino-acid substitution still demands separate modeling.
• Cross-linking sequence-level analysis with 3D inspection — the everyday use is simply to look at the model, rotate it, and see the arrangement of helices and sheets that the sequence never disclosed.

The honest limitations worth internalizing

• Predictions are models, not measurements; a pLDDT of 95 expresses the model’s agreement with evolutionary statistics, not a cryo-EM map.
• Protein function in a cell depends on dynamics, partners, cellular localization, and post-translational modifications — none resolved by a static prediction.
• The database’s models are from reference proteomes; a species of interest outside those proteomes still requires its own prediction run.
• Established structural-biology techniques remain the way to establish precise molecular details, ligand positions, and conformational states; AlphaFold complements them, and increasingly guides them.

Key takeaways

• Structure prediction is a bridge across a 1,000-fold gap between known sequences and experimentally solved structures — and since 2020 the bridge is broadly reliable for well-characterized protein families.
• AlphaFold2’s ingredient is not magic but MSA-based evolutionary statistics plus a learned geometry model, and its per-residue pLDDT and pairwise PAE outputs tell you which parts to trust.
• Low confidence is often genuine disorder, not model failure — read it as biology.
• Structure prediction answers “what might this look like,” not “what does this protein do in a cell”; ligand binding, dynamics, mutations, and complexes still need targeted follow-up.
• The AlphaFold database turns most predictions into lookups rather than long runs, which changes the everyday workflow: check the database first, then predict.

Challenge

Pick a protein you care about. Look it up in the AlphaFold database and open the model in Mol* (in the browser, no install). Toggle the pLDDT coloring: which residues are blue, which are orange? Then find a functional annotation — an active-site residue or a domain — and check where it lands on the structure. You will quickly see where the model is trustworthy and where its claims are soft, which is exactly the reading skill this article argues for.

Need help reading predicted structures for a course or lab project? Ampersand Academy offers one-to-one mentoring in bioinformatics — part of Mahadhi, a development and digital marketing company.

FAQ