Structural biology studies the three-dimensional organisation of proteins, DNA, RNA and molecular complexes to explain how living systems function. It uses experiments such as X-ray crystallography, cryo-electron microscopy and NMR, alongside computational models, to connect molecular sequence, shape, interactions and biological evidence.
If a protein looks like a beautiful folded ribbon, how do we know the shape is correct, active or relevant inside a cell? This guide answers that question with 24 chapters, four worked evidence cases, an explanation of AlphaFold confidence and a connected library of molecular biology, bioinformatics and biotechnology articles.
Five reading routes
Build a structure-first model · Chapters 1–5 · Open route
Know what each experiment measures · Chapters 6–12 · Open route
Interpret models and predictions · Chapters 13–16 · Open route
Solve four structural evidence cases · Chapters 17–20 · Open route
Diagnose and connect the field · Chapters 21–24 · Open route
Chapter index · Specialist reading directory · Gold Standard master
Complete chapter index
Build a structure-first model
- 1. The shape of a biological molecule is a scientific claim
- 2. From sequence to shape to function—and back again
- 3. Proteins are chains with several levels of organisation
- 4. DNA, RNA and molecular machines also have structures
- 5. Forces, solvent and environment shape molecular states
Know what each experiment measures
- 6. Why structural biology cannot be reduced to one picture
- 7. Choose the method for the question
- 8. X-ray crystallography: an ordered sample, a diffraction pattern
- 9. Cryo-electron microscopy: particles, images and averaging
- 10. NMR: a set of constraints and often a dynamic perspective
- 11. Maps, coordinates and validation are different things
- 12. Resolution and confidence must be read locally
Interpret models and predictions
- 13. Predicted structures: useful models, not substitutes for every experiment
- 14. pLDDT and PAE answer different questions
- 15. Complexes, ligands and biological context
- 16. The Protein Data Bank is a research record, not a magic truth stamp
Solve four structural evidence cases
- 17. Worked example 1: do not confuse angstroms with nanometres
- 18. Worked example 2: single-site binding occupancy
- 19. Worked example 3: a prediction can be confident locally but uncertain globally
- 20. Worked example 4: two coordinate models can agree only after a fair comparison
Diagnose and connect the field
- 21. Common false moves and how to repair them
- 22. The connected biology and computation library
- 23. Questions readers most often ask
- 24. Teaching and research pathway: structure as a testable model
What counts as structural evidence?
| Evidence type | What it can support | What it does not establish alone |
|---|---|---|
| Diffraction or imaging data | Information supporting a structural model | Function under every cellular condition |
| NMR measurements | Constraints, interactions or dynamics in suitable samples | A uniquely fixed structure in all contexts |
| Computed coordinates | A model and testable hypotheses | Experimental confirmation or clinical efficacy |
| Functional assays | Observed behaviour under stated conditions | A full atomic mechanism without structural evidence |
1. The shape of a biological molecule is a scientific claim
Structural biology asks how the three-dimensional arrangements of proteins, nucleic acids and their complexes relate to what they do. It brings together molecular biology, chemistry, physics, mathematical modelling and experimental measurements. The attraction is clear: if we understand molecular architecture, we may better understand binding, catalysis, communication and disease.
Yet an attractive molecular rendering is not a photograph of a functioning cell. A structure is a model supported by experiments or computational inference. The strongest question is not simply ‘What shape did we draw?’ but ‘Under which conditions, and with what evidence, does this shape represent something biological?’
2. From sequence to shape to function—and back again
An amino-acid sequence constrains the possible forms of a protein, but a single sequence does not tell us every state, partner, modification or cellular condition. Likewise, seeing one three-dimensional structure does not prove its activity or role.
Think of four linked questions: which molecule was studied, which structural state was captured or predicted, what independent functional evidence exists, and where the observed state occurs in biology. A reader who keeps those four separate is far less likely to overclaim.
Structure and function are linked by a sequence of tests
Consider a protein whose shape suggests a pocket for a small molecule. This is a reasonable place to propose a binding experiment, not enough to claim that binding happens inside living cells. One can measure an interaction in a purified solution, then investigate activity in a cellular assay, and only later ask whether the effect is biologically useful under realistic conditions. Every experiment answers a different part of the original question.
If the functional assay fails, several explanations remain: the proposed pocket may not bind, the protein may require a partner, the solution conditions may be unsuitable, or the molecule may not enter a cell. Good structural reasoning keeps these alternatives visible rather than treating a plausible pocket as a completed mechanism. The direction from sequence to structure to function should remain an evidence chain with checks at each arrow.
3. Proteins are chains with several levels of organisation
Primary structure describes the order of amino acids. Secondary structure describes local arrangements such as helices and sheets. Tertiary structure describes the folded form of a chain. Quaternary structure addresses how multiple chains associate in a complex.
These categories are useful labels rather than a promise that every protein folds into a stable, single rigid globule. Flexible regions, intrinsically disordered proteins and dynamic assemblies are biologically important too.
A helix is a local feature, not a whole protein
Students often recognise spiralling helices in a colourful protein cartoon and conclude they know its shape. In fact, the same set of helices can be packed into very different arrangements. The relative orientation of segments, connections between them, flexible loops and interactions with other chains give a protein its higher-order organisation. Two proteins can share a structural motif without performing the same task.
Imagine changing one amino acid in a buried core versus one in a long exposed tail. The structural consequence can be very different, depending on local chemistry and contacts. A mutation does not invariably unfold a protein, and an apparently intact fold does not prove normal function. Use biochemical measurements and an appropriate structural method to determine which property changed.
4. DNA, RNA and molecular machines also have structures
Double-helical DNA is a familiar beginning, but DNA can bend, package into chromatin, and participate in protein complexes. RNA forms stems, loops, folds and catalytic structures. Ribosomes and other assemblies combine proteins and nucleic acids into working molecular systems.
Research questions should identify the level of structure: an isolated chain, an interacting pair, a full complex or a component in a cellular setting. A structure solved outside its natural environment needs that limitation stated.
A molecular assembly is not merely the sum of its drawings
The ribosome is a useful reminder that biological structure can involve many components. Proteins and RNA together carry out an organised process that neither component’s isolated shape explains fully. The arrangement, timing of interaction and movement between states matter. Some structures represent a particular functional intermediate rather than one timeless configuration.
When viewing a large complex, read the chain assignments and the biological assembly record. A file may contain crystallographic neighbours or other copies that should not all be interpreted as a physiological complex. Ask which interactions were independently established, whether the sample was bound to a ligand and which alternative assemblies remain possible.
5. Forces, solvent and environment shape molecular states
Hydrogen bonds, electrostatics, hydrophobic effects, steric packing, covalent bonds and the surrounding solvent contribute to molecular behaviour. None should be treated as one universal explanation for every fold.
Temperature, pH, salt, ligands, membranes and molecular crowding can shift conformations and binding. Before comparing two structures, check whether the preparation and measurement conditions describe the same state.
What changes when the solvent changes?
A protein observed in one solution may behave differently when salt concentration, temperature or pH changes. Charged amino-acid side chains respond to their chemical environment, and nearby solvent contributes to molecular stability. Membrane proteins present an additional challenge because a lipid-like environment is part of their biological setting. Detergents or other preparations used for experiments can therefore affect what is observed.
An effective comparison describes sample conditions alongside the model. Suppose two studies report different loop positions. First check whether their ligands, solvent conditions and preparation methods match. A different position may indicate a real state shift, an environmental effect or an uncertain segment. Choosing among these explanations requires evidence, not a preference for the more attractive rendering.
6. Why structural biology cannot be reduced to one picture
Molecules move, and biology happens through time. A crystallographic structure, an averaged cryo-EM map or an NMR-derived structural ensemble can each report valuable information but need different interpretation.
Model uncertainty and real conformational flexibility are not the same thing. One can be uncertain because the data are weak; another can be variable because the molecule genuinely visits several states. Additional experiments help separate them.
Static coordinates cannot tell every story about motion
A coordinate file lists positions for a represented model, but molecules fluctuate over many timescales. Atoms vibrate; side chains rotate; loops open and close; domains shift; and complexes form or dissociate. These motions may be relevant to function even if the deposited coordinates are highly precise. A structure may represent a dominant or selected state under the conditions used.
To study dynamics, researchers may combine structural observations with spectroscopy, kinetics, single-molecule techniques or molecular simulations that have been validated for the question. A simulation is not automatically a recording of what nature did. The model, force field, solvent treatment and sampling determine what the trajectory can support. This distinction between observed and calculated motion is essential.
7. Choose the method for the question
X-ray crystallography, cryogenic electron microscopy and nuclear magnetic resonance spectroscopy answer overlapping but different questions. Diffraction depends on suitable crystals; cryo-EM reconstructs information from imaged particles; NMR uses magnetic resonance signals and can reveal behaviour in appropriate samples.
No method guarantees an atomic-resolution answer for every target. Sample availability, heterogeneity, molecular size, flexibility and instrument access all influence the design of the experiment.
Make a method-selection matrix before using the instrument
Suppose the question is whether a protein’s small loop moves during ligand binding. A static structure may give a before-and-after comparison, but it may not reveal kinetics or every intermediate. If the question concerns a large heterogeneous complex, sample preparation and classification may dominate experimental feasibility. If the question concerns interactions in solution, a spectroscopic or biochemical approach may supply valuable complementary constraints.
Design the study backwards from the inference. Identify the proposed conclusion, the minimal observable evidence that would support it, the best available experimental method and the most serious alternative interpretation. Choosing a sophisticated technique first and inventing an interesting claim later is a common route to impressive but uninformative pictures.
8. X-ray crystallography: an ordered sample, a diffraction pattern
Crystallography uses diffraction measurements from a suitably ordered crystal and a mathematical reconstruction of electron density to support an atomic model. The process involves sample preparation, measurement, phasing, model building and refinement.
A reported numerical resolution is not the entire quality report. Local fit, geometry, disorder, occupancy and possible model bias matter. A model can contain precisely placed atoms in one region and uncertain loops elsewhere.
Read a structure-quality report, not only the headline resolution
A crystallographic model may have a reported overall resolution, refinement statistics and geometric validation. These summaries help readers inspect fit and plausibility, but no single metric replaces looking at the relevant atoms and electron density. A well-resolved core can coexist with a poorly determined loop or an ambiguous ligand position.
Before citing a catalytic mechanism, check the region surrounding the proposed active site. Are the relevant atoms supported by the map? Are occupancy and alternate conformations reported? Were hydrogen positions inferred rather than directly resolved? Do the assigned contacts make chemical sense? The purpose of validation is not to find a perfect-looking number; it is to locate where the model earns confidence.
9. Cryo-electron microscopy: particles, images and averaging
In single-particle cryo-EM, many particles embedded in vitrified samples contribute information used to reconstruct a three-dimensional map. Orientation estimates, heterogeneity, image processing and refinement each bring assumptions.
A beautifully sharpened map does not by itself prove that a flexible site is resolved. Inspect local resolution, independent half-map validation where applicable, fit to the density and conformational classes before interpreting fine details.
Independent halves test reconstruction rather than biology itself
In single-particle cryo-EM, information from many noisy particle images is combined in a reconstruction. Splitting the data into independent halves during processing can help evaluate whether fine-scale features are reproducible under the reconstruction procedure. It does not establish that a proposed conformation was the only state present in the sample.
Consider a flexible arm whose density is blurred while the central core is sharp. It may be inappropriate to use one global reported resolution to claim the arm’s side-chain contacts are certain. Inspect local evidence and classification, and be prepared to say that the arm is unresolved in this dataset. That is a scientifically useful result, not an embarrassment.
10. NMR: a set of constraints and often a dynamic perspective
NMR can supply information about atomic environments, distances, interactions and motions under suitable sample conditions. Structure determination uses experimentally derived constraints and model interpretation, often represented as an ensemble rather than one uniquely forced conformation.
A cluster of models is not a literal time-lapse video. It reflects the available measurements, restraints and sampling procedure. The exact question—fold, binding site or dynamics—should direct which NMR experiments are informative.
NMR constraints must be distinguished from pictures
An NMR structure is commonly presented as an ensemble of coordinate models. The observed signals and their interpretation impose constraints, while computational procedures find structures consistent with those constraints. Different conformations in the ensemble can indicate weakly constrained geometry or a genuine physical range, depending on the experiment and analysis.
A student comparing NMR and crystallography should not announce that one ‘shows motion’ and the other ‘shows the true shape’. Each works with a particular measurement and model. Ask what solvent conditions, restraints, temperature and timescale were studied. Functional dynamics may require additional experiments rather than visual inspection of several superimposed ribbon drawings.
11. Maps, coordinates and validation are different things
Experimental observations are processed into maps or constraints, and researchers propose atomic coordinates consistent with them. Every stage can add uncertainty. Validation examines geometry, clashes, stereochemistry, map fit and consistency with independent evidence.
A structure file should be read with its method and validation report, not treated as a universal answer sheet. A tiny clash near a proposed ligand can matter much more for a chemical mechanism than a tidy global appearance.
What validation can discover before a paper is published
A molecular model can exhibit steric clashes, unusual backbone conformations, implausible side-chain placements or a ligand whose chemistry does not match the reported geometry. Such flags are reasons to inspect the original data, not automatic proof of misconduct or total invalidity. Some unusual arrangements are genuinely functional and well supported.
Treat each flagged site as a specific question: which atoms conflict, what experimental observation supports them, and has an alternative conformation been considered? A summary score is most useful when it directs attention toward a claim that matters. The goal is an explainable model with known limits, not a cosmetic campaign to eliminate every warning.
12. Resolution and confidence must be read locally
An angstrom (Å) is 0.1 nanometres. Experimental resolution is often reported in angstroms, while predicted structures use different confidence metrics. A global number cannot describe all local regions of a complex.
Always ask whether the side chain, active site or inter-domain arrangement involved in the intended conclusion is actually supported. The answer may differ across parts of the same structure.
Local certainty matters more than distant perfection
Imagine a 500-residue model with a strong global quality score but a weakly supported ten-residue loop forming the proposed drug-binding pocket. The loop may be decisive for the biological claim even though it is a small part of the molecule. A high overall average cannot repair local uncertainty exactly where the inference depends on it.
Report the quality of the particular site involved. For comparisons across structures, note whether the same residues, chains and ligands are represented. Otherwise an apparently precise distance, angle or clash can be an artefact of choosing mismatched atoms or treating absent density as a confirmed position.
13. Predicted structures: useful models, not substitutes for every experiment
AI methods such as AlphaFold can predict many protein structures from sequence-based information and provide confidence estimates. Such models can help formulate hypotheses and target experiments, but a predicted structure is not an experimental observation of a particular bound, modified or cellular state.
Protein function, substrate specificity, binding strength, kinetics, post-translational modifications and alternative conformations generally need evidence beyond a coordinate file. The route forward is prediction plus critical interpretation plus targeted test.
Predictions are most powerful as experiment planners
A predicted fold may suggest which region to mutate, where a possible interface exists or which domain might be independently stable. Such suggestions can help prioritise research, especially when experimental structures are not yet available. But an algorithm may return one preferred coordinate configuration even when the biological system has several important states.
An effective next step states a discriminating experiment. If a predicted pocket is essential, ask what result would undermine that claim: failure to bind under controlled conditions, loss of a predicted contact without expected functional change, or a conflicting experimental structure. Scientific value grows when prediction makes a hypothesis easier to challenge, not harder.
14. pLDDT and PAE answer different questions
For AlphaFold2-style outputs, per-residue pLDDT estimates local structural confidence on a 0–100 scale. Predicted aligned error (PAE) concerns confidence in relative positional arrangement between regions or domains, with lower predicted error suggesting stronger confidence.
Two domains can each receive high local confidence while their relative orientation remains uncertain. It is especially risky to infer a hinge mechanism or binding site from that unverified placement. The EMBL-EBI guidance makes this distinction central to responsible interpretation.
A two-domain reasoning case using PAE
Suppose two domains have high local confidence, but their relative placement has high predicted aligned error. A researcher might successfully design mutations within each domain based on plausible local geometry. The same evidence does not justify an exact rigid inter-domain distance or a definite mechanism that depends on a particular angle between domains.
The appropriate response could be to compare alternate predicted arrangements, inspect biological partners and seek experimental constraints across the interface. This is not a failure of prediction. It is a successful communication of uncertainty: local fold confidence and global arrangement confidence ask different questions, and the output must be read accordingly.
15. Complexes, ligands and biological context
A protein may bind another protein, DNA, RNA, an ion or a small molecule. Interfaces depend on compatible surfaces and chemistry, but a static close contact is not automatically evidence for functional binding or drug efficacy.
Determine the identity and state of each component, assay conditions, whether a ligand was observed experimentally or placed by a model, and whether binding or inhibition was separately measured.
Binding-site structure versus binding evidence
A ligand shown near a pocket in a visualisation may be experimentally observed, manually fitted, docked computationally or borrowed from another structural model. Those origins are not equivalent. Before claiming an interaction, check the experimental identity of the ligand, the measured density or other supporting data, and the sample’s chemical conditions.
Even confirmed binding does not establish that a candidate drug will work. Biological activity depends on context, selectivity, exposure, metabolism, toxicity and other properties far beyond a static contact map. Use structural evidence to guide investigation, but do not turn an atomic proximity into a medical recommendation. This distinction is especially important when science reaches a general audience.
16. The Protein Data Bank is a research record, not a magic truth stamp
Structures deposited in public archives can allow researchers to examine methods, samples, coordinates and validation. For a responsible comparison, record accession identifiers, structure versions, chain identifiers, ligands, the data-collection method and relevant quality indicators.
A structure may be revised or interpreted differently as new observations emerge. Proper citation and reproducible retrieval allow the next researcher to audit an important decision.
Keep a minimum provenance record for every model
For a published structure, note its archive accession and version, method, molecular identity, construct boundaries, chain identifiers, ligands and relevant validation summaries. For a predicted structure, record the software version or service, sequence, modelling conditions, confidence information and any templates or other major inputs reported.
Why this much bookkeeping? Two papers may say ‘the protein structure’ while discussing different fragments or different binding states. Without a provenance trail, a later reader can mistake a comparison between different constructs for a genuine conformational change. Clear identifiers and archived procedures turn a striking image into useful scientific infrastructure.
17. Worked example 1: do not confuse angstroms with nanometres
Invented distance conversion: a researcher measures a separation of 3.2 Å between two labelled atoms. Since 1 Å = 0.1 nm, 3.2 Å = 0.32 nm. Writing ‘3.2 nm’ would be a tenfold error.
Distance by itself does not prove a bond. The atom identities, protonation states, geometry, interactions, measurement uncertainty and chemical context determine what that separation can mean.
18. Worked example 2: single-site binding occupancy
Illustrative equilibrium model: assume a simple non-cooperative one-site interaction, with free ligand concentration [L] = 30 μM and dissociation constant Kd = 10 μM. Fractional occupancy is θ = [L]/(Kd + [L]) = 30/(10 + 30) = 0.75, or 75%.
The assumption of free, not nominal total, ligand concentration matters. This toy occupancy says nothing on its own about clinical dose, target selectivity, mechanism, residence time or therapeutic outcome.
What an occupancy calculation cannot tell us
The simple one-site occupancy formula assumes an equilibrium model in which one class of sites interacts with ligand without cooperation. Actual binding can involve several sites, conformational states, competing partners and depletion of free ligand. A total amount pipetted into a tube is not necessarily the free concentration available to bind.
For the illustrative 75% calculation in this chapter, teach students to write the underlying assumptions above the substitution. Then change the problem: What if there are two distinct binding sites, or the ligand is depleted? The original calculation remains a correct result for its model, but the model must change before it can describe the new case.
19. Worked example 3: a prediction can be confident locally but uncertain globally
Invented interpretation case: domain A has mean local confidence 92 and domain B has 90, yet the region-to-region PAE is high. The supported claim is that each domain may have a plausible local fold; the unsupported leap is that their mutual orientation is fixed.
The repair is to seek additional data on relative domain motion or contacts—such as experimental structural observations, cross-linking, functional assays or alternative conformational analyses—rather than reporting a single hinge angle as established.
A confidence number is not a probability of biological function
Predicted local confidence scores are useful for identifying regions whose atomic coordinates may be more or less trustworthy under a model. They should not be relabelled as the probability a protein is active, the chance two proteins bind or the likelihood a mutation causes disease. Those are different variables that require different methods and data.
A good explanation names the object of confidence. ‘The local backbone prediction is high-confidence’ is much narrower and more defensible than ‘we know what this protein does’. Readers also need to know whether the relevant biological state, modifications or interaction partners were represented in the model.
20. Worked example 4: two coordinate models can agree only after a fair comparison
Teaching case: consider two models of the same domain, compared after the same atom selections and an explicitly defined superposition. An RMSD of 0.8 Å is a summary of deviations for that matched set, not proof that every residue differs by 0.8 Å.
Exclude mismatched sequence positions only with justification, report alignment choices and inspect local regions. A low overall RMSD can hide a critical active-site difference; an apparently large value can arise from motion between otherwise similar domains.
Fair alignments start with matching atoms
RMSD calculations depend on the atom pairs chosen and how the structures are superimposed. If flexible tails are included in one analysis but excluded in another, the two RMSD values may answer different questions. Comparing only well-ordered core atoms can be appropriate for identifying conserved fold architecture, but that choice must be declared.
A change in overall RMSD cannot alone locate a mechanistic difference. Examine residue-by-residue deviations, local interactions and whether sequence alignment introduced gaps. Re-check the relevant functional region and whether the biological conditions match. A statistically neat summary is valuable when it supports interpretation rather than replacing it.
21. Common false moves and how to repair them
False move: a high-confidence prediction proves a protein binds a drug. Repair: separate fold confidence from ligand-binding evidence. False move: a 3 Å experimental structure is uniformly reliable everywhere. Repair: check local density and validation.
False move: one frozen structure tells the full dynamic mechanism. Repair: consider other conformations and kinetics. False move: visual similarity proves common function. Repair: examine sequence, evolutionary context and independent functional tests.
22. The connected biology and computation library
The Gold Standard ecosystem provides specialist explanations without asking this root to duplicate them. Use Molecular Biology for biological mechanisms, Bioinformatics for computational evidence, and Genomics for sequence context.
Then move toward Synthetic Biology and biotechnology applications, checking where a useful structural explanation ends and genuine experimental proof begins.
Specialist eduKate article directory
Biological foundations
Methods and evidence
Applications and systems
23. Questions readers most often ask
Is a predicted structure the same as an experimentally determined one? No. Methods, measurements and validation differ. Does a protein have only one shape? Often not; flexibility and multiple states can be essential.
Are low-confidence predicted regions necessarily wrong? Not always. They may be disordered, flexible or under-informed. Can a molecule look like a receptor without binding anything? Appearance does not establish a specific interaction. Follow the assay.
24. Teaching and research pathway: structure as a testable model
Give students a protein rendering, a method note and a validation summary. Ask them to identify the chain, ligand, evidence type, local quality and a claim the data do not support. Only after those checks should they attempt a mechanism.
An advanced reading exercise compares an experimental coordinate model with a prediction and writes three conclusions: well supported, plausible but provisional, and not established. The distinction is the transferable skill of structural biology.
Primary scientific resources and validation reading
- Protein structures and experimental methods — EMBL-EBI
- Structural prediction confidence — EMBL-EBI pLDDT
- Structural prediction confidence — EMBL-EBI PAE
- Limitations and accuracy of AlphaFold2 — EMBL-EBI
- Worldwide Protein Data Bank
- RCSB Protein Data Bank
Return to the Gold Standard master, or continue with Molecular Biology and Bioinformatics.
Find more connected learning routes in How X Works and The Gold Standard of Education. Good structural reasoning travels beyond proteins: name the evidence, state the model and test what changes when the context changes.