Why structure matters and the folding problem

Structure dictates function

A protein's amino-acid sequence folds into a defined three-dimensional structure, and almost every biochemical function follows from that geometry rather than from the sequence directly. Catalysis depends on active-site residues that are far apart in sequence but positioned within a few angstroms of each other in space, such as the Ser-His-Asp catalytic triad of serine proteases. Molecular recognition depends on binding interfaces whose shape and electrostatics are complementary to a partner, and allostery propagates a conformational change from one site to a distal site through the folded scaffold. To see the raw scale of what has to fold, download hen egg-white lysozyme and count its residues by counting the one backbone alpha-carbon each amino acid contributes.

curl -s https://files.rcsb.org/download/1LYZ.pdb -o 1lyz.pdb
grep '^ATOM' 1lyz.pdb | awk '{print substr($0, 13, 4)}' | grep -c ' CA '

129

The folding problem

Anfinsen's thermodynamic hypothesis, drawn from the reversible refolding of ribonuclease A after denaturation, states that the native structure of a small globular protein is the conformation that minimizes the Gibbs free energy for a given environment, and that all the information needed to reach it is encoded in the sequence. Levinthal's paradox then observes that a protein cannot reach that minimum by random search: even a modest chain has an astronomical number of accessible conformations, so exhaustive sampling would take longer than the age of the universe, yet real proteins fold in microseconds to seconds. The resolution is that the energy landscape is not flat but funnel-shaped, biasing the ensemble downhill toward the native basin through partially ordered intermediates rather than a blind enumeration.

n_residues = 100
states_per_residue = 3          # coarse backbone rotamers per residue
sampling_rate = 1e13            # conformations sampled per second

conformations = states_per_residue ** n_residues
seconds = conformations / sampling_rate
years = seconds / (365.25 * 24 * 3600)

print(f"conformations: {conformations:.2e}")
print(f"exhaustive search: {years:.2e} years")

conformations: 5.15e+47
exhaustive search: 1.63e+27 years

The funnel is why ab initio prediction is hard for a computer even though physics solves it for the cell. Reproducing the funnel requires an energy function accurate to a fraction of a kcal/mol across an enormous conformational space, because folded proteins are only marginally stable, typically 5 to 15 kcal/mol below the unfolded state, and the decisive contributions are long-range contacts and solvent effects that are expensive to model and easy to get slightly wrong. Small errors in the potential move the predicted global minimum away from the true native structure.

Measuring progress with CASP

CASP, the biennial Critical Assessment of Structure Prediction experiment running since 1994, is the field's blind benchmark. Predictors receive only sequences whose experimental structures are solved but not yet released, submit models, and are scored against the withheld coordinates, which removes the risk of fitting to known answers. The headline backbone metric is GDT_TS, the average fraction of CA atoms superposable onto the native structure within 1, 2, 4, and 8 angstrom cutoffs, reported on a 0 to 100 scale. At CASP14 in 2020, AlphaFold2 reached a median GDT_TS near 92, close to experimental accuracy for many targets and a step change from earlier methods.