A Protein Model Is a Sequence, a Fold, and a Surface

Mon, 26 Oct 2026 00:00:00 +0000

A protein in a calculation is three objects, and they are not interchangeable. The sequence is which residues are present, including the isoform and the protonation. The fold is the ensemble of shapes that sequence occupies. The surface is the patch a partner actually touches. What enters the coordinate file, what the force field is allowed to do to it, and which observable the trajectory is asked to produce are three different questions.

Figure 1. A coordinate file is one of the two folds in the middle. An affinity, a clearance, or a pathway flux is a question about the patch on the right, usually with partners the file does not contain.

Twenty residues are free, and one is not

The standard genetic code builds proteins from twenty amino acids. Selenocysteine is inserted during translation where a UGA codon is recoded, not as a free choice at every position. Humans make it in a small set of selenoproteins, on the order of twenty-five. Pyrrolysine is not part of the human code. A rotamer library and a force field that offer selenocysteine at every site are describing a protein the cell does not make.

Length is a separate choice. Fifty residues is a peptide or a small domain. A chain of a few hundred residues is an ordinary globular protein. The reviewed human titin entry UniProt Q8WZ42 is 34,350 residues, and other isoforms differ, so one gene is not one chain to drop into a box [1]. What can actually be solvated is a domain, or a few domains. Titin is hundreds of those, a spring in muscle, not a single globular fold.

The sequence space of even one ordinary chain is already past any screen. For 375 positions and 20 residues at each position,

$$ 20^{375} \approx 7.7 \times 10^{487}. $$

That sits hundreds of orders of magnitude above the roughly $10^{80}$ atoms estimated for the observable universe. Almost every sequence is chemically writable and biologically unused. The gene is the filter. A docking screen or a generative model that samples protein-like strings without that filter is sampling chains no cell makes.

Figure 2. AUG, CAC, and GAC are ordinary residues. The dashed residue is the exception: UGA is selenocysteine only where that codon is recoded.

The cytoplasm those chains fold in is crowded. Diffusion there is not the diffusion of one protein in a large box of water. A periodic cell with one protein and a bath answers a different physical question.

About 20,000 protein-coding genes is the right order for the human set. It is not 20,000 structures. Alternative splicing and post-translational modification produce many chains from one gene. An AlphaFold model, or a PDB entry, is one sequence in one conformation. The is about whether that conformation is even a plausible fold, which comes before the question of whether it is the conformation that binds.

Charge is four terms in a force field

Amino acids differ in charge and polarity. The chain is synthesized as a line, and that line collapses into a fold. A force field does not carry this as one interaction called charge.

Figure 3. The bottom-right panel is the term a charge does not cover. Greasy side chains pack because water would rather not organize around them. A vacuum minimization will not reproduce that.

Partial charges are a model of the electrostatic box, assigned before the dynamics, and they depend on the protonation you chose. Histidine, aspartate, glutamate, and the termini are the usual decisions. The is about how much that assignment moves a docking score. It does not make the score a free energy. A quantum calculation on the ligand, as in the , reports the electrons of that molecule. It does not fold the protein and it does not prove the complex.

The hydrophobic effect is why an implicit-solvent or vacuum model can fold the wrong core. The chain is not clicking together by electrostatics alone. If the simulation has no water, or a solvent model that ignores that term, the compact structure it finds is answering a different physical question.

Allostery is a shift in how often each shape appears

A folded chain presents a surface. A partner binds, the local charge distribution changes, and the shape changes with it. A site that was hidden can open, the protein can join a larger assembly, or the new surface can pass a signal on. The error is to treat the protein as one shape that snaps to a second shape.

Figure 4. The bars are how often each shape is occupied. Induced fit and conformational selection are arguments about which population moves. One minimized structure cannot tell them apart.

One crystal structure, or one minimized frame, is a sample. The is the argument for watching the chain move, and for not reading a short trajectory as the ensemble. Protonation belongs in the same sentence as allostery. A change in the charge of a buried histidine can shift the same populations without any ligand at all.

A pathway is a queue of surfaces

A pathway is many proteins doing one job: taking up a nutrient, running glycolysis, answering a xenobiotic. Each step is a surface finding the next surface. There is no controller in the middle. Binding and rates are the whole mechanism.

The version a cytochrome P450 project actually computes is narrower. A ligand of the aryl hydrocarbon receptor binds the cytosolic receptor. The receptor changes partners and moves to the nucleus, binds DNA with ARNT, and one of the genes that comes on is CYP1A1. The enzyme can make the incoming molecule easier to clear. It can also make it more reactive. Which of those happens is a property of the substrate and of the reaction on that isoform, not a property of the gene having turned on. A model of the sensor, a model of the enzyme, and a model of clearance are three calculations. The is the same warning from the data side: a clearance number mixes mechanisms, and a $K_i$ on one isoform does not.

Figure 5. The drawings are schematic. Each one can be its own calculation. Turning the gene on does not decide whether CYP1A1 clears the molecule or makes it more reactive.

Protein-coding exons are on the order of one percent of the human genome. The rest includes introns, regulatory sequence, and a great deal of repetitive DNA. A simulation never sees that rest. It sees the chain you chose, in the protonation you chose, in the box you built.

The calculation ends at whichever object in Figure 1 the question actually named.

References