Protein Structure Prediction

Protein structure prediction estimates 3D coordinates or structural constraints from sequence and optional context such as evolutionary information, templates, or learned representations.

A simplified sequence-to-structure view is:

$$\hat{X}=f_{\theta }(s_{1:L},c)$$

where $s_{1:L}$ is the residue sequence, $c$ is optional context, and $\hat{X}\in {\mathbb{R}}^{L\times 3}$ or atom-level coordinates are predicted.

Geometry losses often compare pairwise distances:

$${\mathcal{L}}_{\mathrm{dist}}=\sum _{i,j}{\left(\Vert {\hat{x}}_i-{\hat{x}}_j{\Vert }_2-\Vert x_i-x_j{\Vert }_2\right)}^2$$

Predictions can target different levels:

• contact map or distance map
• backbone coordinates
• all-atom coordinates
• confidence or error estimates
• complex or multimer structure
• pocket or binding-site geometry

The target choice changes evaluation. A good contact map does not imply all-atom pose quality, and a good backbone does not imply ligand-ready side-chain placement.

Prediction Levels

Target	Output	Evidence needed
contact map	binary residue-pair contact	precision/recall by sequence separation
distance map	pairwise distances or bins	distance error and calibration
backbone	$N$, $C_{\alpha }$, $C$, $O$ coordinates	RMSD/TM-like structural agreement and local geometry
all-atom	backbone plus side chains	rotamer quality, clashes, bond geometry
confidence	per-residue or pairwise uncertainty	calibration against actual error
complex	multimer or protein-ligand/protein-protein arrangement	interface quality, chain mapping, partner leakage checks

Coordinate Contract

Coordinate predictions should state what transforms are irrelevant:

$${\hat{X}}^{\prime }=R\hat{X}+t,X^{\prime }=RX+t$$

Global rotation and translation should not change the structural claim. If the model predicts vectors, forces, or coordinate updates, the output should transform equivariantly; if it predicts confidence or a scalar score, the output should be invariant.

Downstream Boundary

Downstream use	Extra requirement
docking	side-chain placement, pocket geometry, missing atoms, protonation states
binder or complex modeling	interface orientation, chain identity, template leakage control
function prediction	active-site residues, cofactors, conformational state
protein design	foldability and sequence-structure compatibility
representation learning	split-aware downstream evaluation, not only structure similarity

Checks

• Is the target backbone, all-atom structure, contact map, distance map, or confidence score?
• Does training or evaluation use templates that may leak test structures?
• Are missing residues, alternate conformations, chain breaks, and ligands handled?
• Is structure cleaning and residue indexing consistent between train and evaluation?
• Is downstream use docking, design, function prediction, or representation transfer?
• Is the confidence estimate calibrated to the reported error?
• Are monomer, multimer, and ligand-bound claims evaluated separately?

Related

• Protein representation
• Protein structure cleaning
• Residue indexing
• Contact map
• Equivariance