This chapter explains how a protein's 3D shape is read from a PDB entry, predicted from sequence by homology modelling and by deep learning, and compared by RMSD. An engineer needs it to judge whether a predicted structure can be trusted.
Updated 30 Oct 2026
What you will be able to do
You will read the atom lines and resolution of a PDB entry and outline how secondary structure is predicted. You will follow homology modelling and say why sequence identity sets how far to trust a model. You will read a deep-learning confidence score. You will compare two structures by calculating RMSD by hand.
Before you start
Chapter 3, Searching databases with BLAST (blastp, E-value, percent identity)
Amino acids, the peptide bond and hydrogen bonds (school biology)
Distance between two points in 3D (Pythagoras)
Mean, square root and percentages
From sequence to structure
A protein is a chain of amino acids, each called a residue. Residues share the backbone atoms N, CA (the alpha carbon), C and O, and differ in their side chains. Primary structure is the residue sequence; secondary, local alpha helices and beta strands held by hydrogen bonds between backbone C=O and N-H groups; tertiary, the 3D fold of one chain, with loops between its helices and strands; quaternary, how several chains pack together. For most globular proteins the sequence decides the fold, so the structure can in principle be predicted from it.
The PDB stores solved structures as text. An ATOM line gives serial number, atom name, residue name, chain, residue number, x, y and z in ångströms (1 Å = 0.1 nm), occupancy (the fraction of copies with the atom there), B-factor in Ų (how much the atom seems to move) and element. For X-ray and cryo-electron microscopy (cryo-EM) entries, the resolution is the finest detail the experiment could separate, in Å.
ATOM 2 CA ALA A 1 11.104 6.134 -6.504 1.00 20.00 C
Here: atom 2, the CA of alanine (ALA) residue 1 in chain A, at (11.104, 6.134, -6.504) Å; occupancy 1.00; B-factor 20.00 Ų; element C (carbon).
DISTANCE BETWEEN TWO ATOMS
d=(x2−x1)2+(y2−y1)2+(z2−z1)2
Holds for coordinates in one common frame, as in a single PDB entry. (x1,y1,z1) and (x2,y2,z2) are the two atoms' coordinates in Å; d is in Å. It is Pythagoras applied twice.
WORKED EXAMPLE
Made-up round coordinates in one PDB entry: atom N at (0, 0, 0) Å, atom CA at (1.2, 0.9, 0.0) Å. How far apart are they?
▸ Solution
Take the coordinate differences: Δx = 1.2, Δy = 0.9, Δz = 0.0 Å.
Square each, because the distance formula adds squares: 1.44, 0.81 and 0 Ų.
Add: 1.44 + 0.81 + 0 = 2.25 Ų.
Take the square root to return to a length: √2.25 = 1.5 Å.
Answer: 1.5 Å. A real N to CA bond is about 1.46 Å, so these rounded numbers are only an illustration.
Secondary structure can be guessed from sequence. Each residue gets one of three states: helix, strand or coil (everything else, such as turns and loops). Alanine and leucine sit in helices more often than average; other amino acids favour strands or turns. A simple method averages each residue's helix propensity, defined below, over a short window of neighbours. Modern methods instead feed a neural network a family of aligned related sequences.
HELIX PROPENSITY OF A RESIDUE
P(a)=fhelix(all)fhelix(a)
Counted from solved structures. fhelix(a) is the fraction of residues of amino acid a that lie in helices; fhelix(all), the fraction of all residues that do. P has no unit; P>1 favours helix. In the simplest version, a stretch whose window-averaged P exceeds 1 is labelled helix; real methods add rules for where helices end.
PREDICTION ACCURACY Q3
Q3=all residuesresidues given the right state×100
Measured on solved proteins left out when the method was built; each residue's right state comes from its solved structure. Q3 is a percentage. Propensity methods reach about 50 to 60%; network methods that use aligned families reach about 80% or more.
WATCH OUT FOR
Resolution is quoted as a distance, so a smaller number means finer detail: 1.5 Å shows more than 3.5 Å. Test: the entry with the smaller number in Å is the sharper one.
Homology modelling
Evolution changes sequences faster than folds, so two related proteins with similar sequences usually share a fold, even if only one has been solved. Homology modelling uses the solved protein (the template) as a scaffold for the unsolved one (the target). The higher their identity, the closer the target's backbone usually lies to the template's and the fewer alignment errors there are, so identity tells you how far to trust the model.
Find a template: run blastp (Chapter 3) of the target against the PDB's protein sequences. Pick a hit with a low E-value, high identity and good coverage (the share of the target the alignment spans).
Align target to template. Every later step rests on this alignment, so its errors become the model's errors.
Build: copy the template's backbone coordinates at aligned positions, model loops where the alignment has gaps, then place the side chains.
Refine: move atoms to remove clashes and bad bond geometry (energy minimisation).
Check: bond lengths, angles, backbone torsions (each residue's two backbone rotation angles, ϕ and ψ; only some pairs occur in real proteins) and a quality score that compares the model with solved structures. Rebuild what fails.
Aligned positions are the columns where both sequences have a residue; gap columns are left out. Some tools divide by the full alignment length instead, so state your denominator. The result is a percentage. Over a short alignment it means little: a database search easily finds a high identity over 10 positions by chance.
Identity to template
What to expect (rough; the bands blend)
above about 50%
Backbone close to the template; side chains mostly right
about 30% to 50%
Fold right; loops, side chains and the alignment need checking
below about 30%
Alignment errors dominate; the model may be wrong
WATCH OUT FOR
A very low E-value shows the two proteins are related, not that they are close. Test: read the identity and the coverage too; a template that wins on both is the clear pick.
Deep learning and comparing structures
Deep-learning predictors such as AlphaFold need no template. A neural network trained on solved structures takes the target sequence and a multiple sequence alignment: related sequences from a database search, stacked so that matching positions form columns. Positions that mutate together tend to touch in 3D, and the network turns that pattern into coordinates. Each residue gets a confidence score, pLDDT, from 0 to 100: an estimate of the lDDT score below.
LDDT, THE SCORE PLDDT PREDICTS
lDDT=4100(f0.5+f1+f2+f4)
Per residue, against the true structure, with no superposition. Pair each atom of the residue with each atom of other residues within 15 Å of it in the true structure. ft is the fraction of pairs whose model distance differs from the true one by less than t Å. lDDT runs 0 to 100. pLDDT estimates the CA-only version, without the true structure.
pLDDT
How to read it
above 90
Very high: backbone and side chains likely right
70 to 90
Confident: backbone likely right
50 to 70
Low: treat with care
below 50
Very low: often a disordered stretch; do not interpret
pLDDT is per residue: a well-predicted domain (a compact, separately folding part of the chain) can sit beside a poorly predicted linker, and pLDDT says nothing about how domains sit relative to each other. To compare two structures with matched atoms, first superpose them: rotate and shift one rigidly so that the sum of squared distances between matched atoms is as small as possible. Programs find this best fit exactly; below, the superposed distances are given.
ROOT-MEAN-SQUARE DEVIATION (RMSD)
RMSD=N1i=1∑Ndi2
Holds after superposition, over N matched atoms (usually the CA atoms of aligned residues). di is the distance in Å between matched atom i in the two structures. RMSD is in Å; 0 means the matched atoms coincide.
WORKED EXAMPLE
After superposition, four matched CA atoms lie 0, 1, 2 and 2 Å apart (a real comparison uses hundreds of atoms). Find the RMSD and compare it with the plain mean distance.
▸ Solution
Square each distance, as the formula asks; squaring makes large distances count more: 0, 1, 4, 4 Ų.
Add them: 0 + 1 + 4 + 4 = 9 Ų.
Divide by N = 4 to get the mean square: 9 / 4 = 2.25 Ų.
Take the square root to return to Å: √2.25 = 1.5 Å.
The plain mean is (0 + 1 + 2 + 2) / 4 = 1.25 Å. RMSD is larger because squaring gives the 2 Å atoms more weight.
Answer: RMSD = 1.5 Å, against a mean distance of 1.25 Å.
IN ONE LINE
RMSD weights the worst-placed atoms most, so one floppy loop can make two similar folds look far apart.
Check yourself
CHECK YOURSELF 1
Two PDB entries of the same protein have resolutions of 1.5 Å and 3.0 Å. Which statement is correct?
(A)
The 3.0 Å entry shows finer detail.
(B)
The 1.5 Å entry shows finer detail.
(C)
Both show the same detail.
(D)
Resolution applies only to the B-factor.
▸ Answer and solution
Answer: (B)
Resolution is the smallest detail the experiment can separate.
A smaller value separates finer detail, so 1.5 Å is sharper than 3.0 Å.
CHECK YOURSELF 2
Two atoms lie at (0, 0, 0) Å and (3, 4, 12) Å. Find the distance between them in Å.
▸ Answer and solution
Answer:13 (accepted range ±0.01)
d² = 3² + 4² + 12² = 9 + 16 + 144 = 169.
d = √169 = 13 Å.
CHECK YOURSELF 3
A deep-learning model gives pLDDT = 45 for a stretch of 30 residues. What is the best reading?
(A)
The stretch is predicted with very high accuracy.
(B)
The stretch is a confident helix.
(C)
The prediction there is unreliable; it may be a disordered stretch.
(D)
The stretch has 45% identity to a template.
▸ Answer and solution
Answer: (C)
pLDDT below 50 is very low confidence.
Such stretches are often disordered, so their coordinates should not be interpreted.
pLDDT is a confidence score, not a percent identity.
CHECK YOURSELF 4One or more options are correct
Which of these are correct steps or rules in homology modelling?
(A)
Align the target sequence to the template sequence.
(B)
Choose the template with the highest E-value.
(C)
Copy backbone coordinates for aligned positions from the template.
(D)
Check the model for bad bond geometry and clashes.
▸ Answer and solution
Answer: (A), (C), (D)
A, C and D are steps of the method.
B is wrong: a lower E-value means a more significant hit, so the best template has a low E-value, not a high one.