Structural Biology: Chapter 10 — From structure to prediction and design – Structural Biology Study Notes & Practical Guide

Module 10 • Chapter Study Notes Part of 📐 Structural Biology

Chapter 10: From structure to prediction and design

📌 Core Learning Outcomes

Apply the sequence-identity rule of thumb, read a PDB entry as an experiment, and judge pLDDT and the failure modes of AlphaFold and RoseTTAFold. State what structure-based design used HIV protease for, without a synthetic route.

🗺️ Curriculum Map 📐 High-Res Figures 📖 Terminology ⚡ Mind-Blowing Facts

10.1 Homology modelling #

10.1.1 Why sequence identity matters #

Structure changes more slowly than sequence, so genuine homologues share a fold and the closer relative is the better template. Homology modelling aligns the sequences, copies the conserved backbone, rebuilds insertions, and places side chains. The alignment decides the rest. Alignment algorithms are Bioinformatics; a shifted alignment is a confident model of the wrong residues. The rule of thumb: above about 30–40% identity a useful model is often possible; below that it is unsafe. The twilight zone, Rost’s term, is where similarity is hard to tell from chance. Above about 50% the core is often within about an ångström of the template. Between 30 and 50% the core may be useful and the details are not. A new domain, a biased composition, or a long insertion breaks a single cutoff. Identity belongs to the aligned core, not to a shared disordered tail.

10.1.2 What a model is not #

Loops, where insertions and binding sites sit, transfer badly. So do rotamers, ligands, waters, and hinge angles. A disease mutation may not move the backbone at all, so the picture looks wild-type while the cell degrades the protein. A homology model has no R-free, because nothing was measured. The test is external: honest molecular replacement, a mutation, a cross-link, a SAXS curve. A low energy from the program is not that test.

10.2 The Protein Data Bank #

The study of The Protein Data Bank provides fundamental principles governing cellular architecture, molecular interactions, and physiological regulation.

Glycine Ramachandran Data And Contours T8000
**Interpretation & Mechanism:** A Ramachandran plot of the phi,psi protein backbone dihedral angles for Glycine, from a dataset of 1.5 million residues in 8000 protein chains with resolution <2.0Å and backbone B-factors ≤30. The 116,789 individual-residue datapoints are color-coded by the number in each 0.1° bin. The inner contour encloses 98% of the data (the “favored” region, while the outer contour encloses 99.9% of the data, dividing “allowed” from “outlier” regions.

10.2.1 What an entry contains #

The Protein Data Bank, begun in 1971, archives experimental macromolecular models. NMR restraints sit in the BMRB and cryo-EM maps in the EMDB. There are well over two hundred thousand entries. An entry is one model under stated conditions. Before trusting a figure, read the experiment; the resolution, or the NMR ensemble, or the cryo-EM FSC and whether the map is deposited; R and R-free if it is a crystal; the sequence actually built, including mutations, tags, and residues missing from the coordinates; occupancy and B-factors; and the ligand’s density, not merely its name. Validation outliers clustered on the residue under discussion are a reason to stop. The biological assembly is an annotation of which contacts are oligomeric, and it is sometimes wrong. A crystal contact labelled as a dimer is still a crystal contact. Missing coordinates mean “not modelled”. “The PDB structure”, with no accession and no resolution, is not a scientific sentence.

10.3 Learned prediction #

10.3.1 AlphaFold and RoseTTAFold #

In CASP14 (2020), AlphaFold2 reached backbone accuracy of about 1 Å on many domains where earlier methods were often several ångströms out. Jumper and colleagues described it in Nature in 2021. The network uses the sequence and a multiple alignment: residues in contact tend to co-evolve. It outputs coordinates, a per-residue confidence pLDDT from 0 to 100, and a predicted aligned error for every pair of residues, which is the score for domain placement. RoseTTAFold, Baek and colleagues, Science 2021, is a three-track network of sequence, distance, and coordinates. It was less accurate than AlphaFold2 on CASP14 and far ahead of the previous generation. Both draw power from the alignment. A sequence with no homologues is a weaker case. Read pLDDT as the authors advise: above about 90, high local confidence; about 70–90, a generally good backbone; about 50–70, a low-confidence guess; below about 50, not a structure. Low scores often coincide with disorder. pLDDT is not an experimental B-factor. Two confident domains can still have a hopeless hinge, which only the aligned-error matrix shows.

10.3.2 What the confidence score does not cancel #

A high score is not an experiment. There is no map and no R-free. Disorder should not be drawn as the spaghetti a graphics program connects. Complexes were outside the original monomer model. Multimer variants exist and fail more often, especially on weak and on antibody complexes; a success is still a prediction. Ligands, metals, modifications, and membranes are not solved by a standard run of that original network. Later networks attempt some of them. A pose remains a hypothesis, and ligand binding is not automatically solved because a fold was predicted. One conformation is what comes back, often the state common in the training set, with no free-energy gap attached. A point mutation is often predicted as the wild-type fold, so the picture cannot be used to argue that the mutation does nothing. Shallow alignments, orphans, and designed sequences are weaker because coevolution is missing. A prediction that repeats a well-known fold family is not an independent observation of that fold. It can still be the right search model for a new crystal. These methods often answer “what is the domain fold?” well enough to plan an experiment. They move crystallography, NMR, and cryo-EM towards state, ligand, assembly, and dynamics.

10.4 Structure-based drug design #

10.4.1 A site is a constraint, not a medicine #

Structure-based design uses the shape and the chemistry of a site to suggest the next molecule, then solves the complex and measures affinity in solution. A docked pose is not a dissociation constant. Docking scores inherit every error in the coordinates and usually treat the protein as nearly rigid. HIV protease is the historical case where the cycle contributed to medicines. The enzyme is a homodimer. Each chain contributes Asp 25 to one active site, and a flap from each chain closes over the substrate. Unlike a serine protease it forms no acyl-enzyme: two aspartates activate water, and inhibitors mimic the tetrahedral species so they bind and are not cleaved. Structures from 1989 onward showed the flaps on the compound and which group lay in which subsite. Early inhibitors, saquinavir among them, came from that geometry. No synthetic route and no dose are given here. Resistance continues the lesson. Pocket and flap mutations can spare the viral substrates and weaken the inhibitor. A site is not a fixed lock. Most targets are less kind: the site moves, the surface is flat, or the conformer in hand is the wrong one. Absorption and toxicity are not in the coordinate file. Design chooses the next experiment. It does not guarantee a medicine.

10.5 Key terms, summary, and self-check #

Key terms. Homology modelling; twilight zone; Protein Data Bank; biological assembly; pLDDT; predicted aligned error; AlphaFold; RoseTTAFold; structure-based drug design.

Summary. Above about 30–40% identity a homology model of the core is often useful; below that it is unsafe, usually because the alignment is wrong. A model has no R-free. A PDB entry records method, resolution, missing residues, and validation, and its biological assembly can be wrong. AlphaFold and RoseTTAFold predict folds from sequence and alignments; pLDDT below about 50 is not a structure. Disorder, complexes, other conformations, ligands, mutations, and shallow alignments remain failure modes. HIV protease shows design at its strongest, a dimeric aspartic site and transition-state analogues, and resistance shows that the site can change.

Self-check.

1. State the identity rule of thumb, and why models below it usually fail.

2. Name five things to read in a PDB entry before believing a ligand picture.

3. What should be done with pLDDT below about 50?

4. Give four failure modes that a high confidence score does not cancel.

5. What did HIV protease structures contribute, and why was resistance not a surprise of principle?

—

# Closing: how the chapters fit together

A structure is a conformation preferred, often, by only tens of kilojoules per mole, and observed by a method that averages, restrains, or classifies. Chapter 1 names the variables and the symmetries. Chapter 2 says why one conformation wins, and why sickle haemoglobin, amyloid, and prions are three different second conformations. Chapters 3 to 5 ask, for three experiments, what was measured and what the specimen could not show. Chapter 6 stops a low-resolution method from being narrated as atomic. Chapter 7 ties a coordinate change to a chemical step. Chapter 8 puts back the bilayer and uses one filter to show that selectivity is chemistry. Chapter 9 treats nucleic acids the same way, then uses ribozymes to show that catalysis belongs to a fold. Chapter 10 asks for the same scepticism of a model no detector measured. Kinetic constants and the proton gradient are Biochemistry. The instrument is Biophysical Techniques. The alignment algorithm is Bioinformatics. The factor cycle of decoding is Molecular Biology.

Revision checklist #

1. Mark the α and β basins of a Ramachandran plot and say why glycine and proline are special.

2. State Anfinsen’s hypothesis, assign the hydrophobic effect, hydrogen bonds, and packing their roles, and name one exception to unaided folding.

3. Separate sickle-cell polymerisation, cross-β amyloid, and a prion in three sentences.

4. Write Bragg’s law, say what the phase problem removes, and name one experimental and one model-based route back to phases.

5. Say what 3 Å and 1.5 Å each justify, and why R-free exists.

6. Choose among crystallography, NMR, cryo-EM, and one Chapter 6 method for a stated question, and give the limit.

7. Walk from the proximal histidine to a Hill coefficient of about 2.8 and the 2,3-BPG site, and from the triad to the pocket that separates trypsin from elastase.

8. Say why the peptidyl-transferase centre is RNA, and what three β subunits in one F1 crystal showed.

9. Explain potassium selectivity without calling the filter a sieve, name the two membrane folds, and state one groove difference between A-form and B-form.

10. Apply the 30–40% identity rule, say what pLDDT below about 50 means, and say what HIV protease structures did not guarantee.

Selected sources #

These notes are an original teaching summary, not a transcript of any textbook.

Brändén C, Tooze J. Introduction to Protein Structure. 2nd edition. Garland.
Lesk AM. Introduction to Protein Architecture. Oxford University Press.
Rhodes G. Crystallography Made Crystal Clear. 3rd edition. Academic Press.
Voet D, Voet JG. Biochemistry. Wiley.
Kühlbrandt W. The resolution revolution. Science 343:1443–1444 (2014).
Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature 596:583–589 (2021).

—

# Part C. Visual Figures & Scientific Diagrams Gallery

Comprehensive high-resolution visual reference repository for Structural Biology from the B10logy Visual Atlas.