Bioinformatics & Computational Biology: Chapter 9 — Protein bioinformatics

🔬 Key Illustrated Scientific Figure

Full-Screen Zoom

Click to enlarge on this page

9.1 Sequence properties before any fold #

9.1.1 Composition is a calculation #

Length, composition, a predicted mass, and a predicted isoelectric point follow from the sequence alone. An average residue is about 110 daltons, so 400 residues are about 44 kDa, before cleavage or modification. A calculated isoelectric point is the pH at which a model peptide has zero net charge. It is not the pH at which the protein precipitated, and a phosphate or a glycan moves the measured value.

A run of one residue may be a real feature, such as collagen or a glutamine tract, or a translation of repetitive DNA that is not a protein. Composition starts an argument. It is not a measurement of behaviour.

9.1.2 Hydrophobicity plots #

A hydrophobicity plot, classically Kyte–Doolittle, averages a scale in a sliding window. A transmembrane helix is about twenty residues, so a window near nineteen is the usual choice: long enough to ignore a soluble core, short enough to separate neighbouring helices. The bilayer itself is Cell Biology. The plot is a prediction, and it has a false-positive rate.

A sustained peak is a candidate, not a proof. Signal peptides are cleaved hydrophobic stretches at the N-terminus, and reading one as a helix invents a membrane anchor on a secreted protein. The positive-inside rule puts lysine and arginine preferentially on the cytoplasmic side. It is a hypothesis about orientation, with false calls in both directions, and a missed span flips everything downstream. The opposite pattern, low hydrophobicity and high net charge, is only a clue to disorder.

9.2 From sequence to a picture of the fold #

9.2.1 Secondary structure as a probability #

The states are helix, strand, and coil. Coil means neither of the other two. It does not mean disorder. Methods that used only one sequence reached about 60% three-state accuracy on the benchmarks of their time, because local sequence does not fix the state. Tertiary context does the rest.

A profile supplies some of that context: a site that stays hydrophobic, or that never accepts proline, is informative. With homologues, about 80% of residues are called correctly on standard soluble globular benchmarks. That figure belongs to those benchmarks. It falls for membrane proteins, disordered proteins, and orphans, and the errors gather at the ends of elements. Read the probability. A residue at 0.45 helix and 0.40 coil is not a helix, and a string of calls is not a coordinate model.

9.2.2 Homology modelling depends on a real template #

Homology modelling copies an experimental template: find it, align the target, copy backbone coordinates for the aligned residues, build the loops, and place side chains. The alignment is the model. A shift of one residue moves every side chain after it. Energy minimisation makes a wrong model more comfortable. It does not find the shift.

As a rule of thumb only, and only for a long globular alignment: above about 50% identity the conserved core is often useful and the loops are not; between about 30% and 50% the core can still help if every column is checked; below about 25–30%, the twilight zone, homology may be real and an automatic model is not evidence. A short alignment needs higher identity. Membrane proteins do not obey a globular rule. No named template means there is nothing of this kind to build.

9.2.3 Learned prediction, confidence, and failure #

AlphaFold, of which AlphaFold2 is the landmark of this class, names networks trained on experimental structures and on multiple alignments. They emit coordinates, a per-residue pLDDT from 0 to 100, and a predicted aligned error between residues. The usual reading of pLDDT is the developers’ convention, not a measurement on your protein: above about 90, local backbone and many side chains are offered as confident; about 70–90, the backbone more than the side chains; about 50–70, low; below about 50, do not read a stable fold. Low confidence often means disorder and sometimes means ignorance. The score does not say which.

Predicted aligned error blocks a standard misreading. Two domains can each score near 90 while the error between them is large: each domain is locally confident, and their packing is not. Disorder is not a broken protein; drawing through a low-confidence tail invents structure. Without homologues a major input is missing, and a compact cartoon can still appear. A confident fold with nothing in the alignment or the experimental record to justify it is a hallucination. A monomer model does not place a partner, a ligand, or a second conformation. Later methods that predict some complexes and ligands remain predictions, and they are weak where an interface has no evolutionary signal. A prediction has no map. Where it disagrees with an experiment, the map decides, which is Structural Biology. A test on a protein whose close homologue was in the training set measures memory. Report that identity.

9.2.4 Docking #

Docking searches orientations of two structures and scores shape, charge, and statistical preferences. The native pose need not win, and wrong poses score well. If either structure was itself predicted, the errors stack: two folds and one interface.

The legitimate product is a hypothesis about contacts, to be tested by mutation or by an experimental complex. The score is not a dissociation constant, not proof that the proteins meet in the cell, and not a drug lead.

9.3 Key terms, summary, and self-check #

Key terms. Hydrophobicity plot; Kyte–Doolittle; positive-inside rule; signal peptide; Q3; homology modelling; template; twilight zone; pLDDT; predicted aligned error; hallucination; docking.

Summary. Calculated mass and pI are model numbers. A hydrophobicity plot proposes transmembrane helices and confuses them with signal peptides unless told otherwise. Secondary-structure output is a probability, not a fold. Homology modelling copies a real template and stands or falls by the alignment; the identity bands are a rule of thumb, and below the twilight zone an automatic model is not evidence. AlphaFold-class models add pLDDT and predicted aligned error. Disorder, missing homologues, complexes, and the gap between prediction and experiment are the failures. Docking suggests an interface and does not demonstrate binding.

Self-check.

1. Why does a secreted soluble protein often show a hydrophobic peak at its N-terminus?

2. Helix probability 0.42 and coil 0.40. What should you draw?

3. State the modelling rule of thumb near 55% identity and near 20%. Why does minimisation not repair a one-residue shift?

4. Both domains have pLDDT 90 and the error between them is large. What may you believe?

5. Why is a docking score not a dissociation constant, even from two experimental structures?

—

—

—